Datasets:

tukey

/

sampled-arxiv-ocr001_validation

like 0

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "## B. Variance Distribution Estimation Based on Marginal Max- imum Likelihood\n\nLet us consider the estimation of $\\alpha$ and $\\beta$, given $N$ samples of EMG signals $X = \\{x_n\\}^N_{n=1}$. The parameters where observations are most likely to occur can be estimated by maximizing the marginal likelihood $P(X) = \\prod^{N}_{n=1} P(x_n)$. Hayashi et al. maximized this using the steepest descent method [0], but the convergence stability of this approach is sensitive to the initial values and step size. This is because direct optimization of marginal likelihood is generally prob- lematic [0]. Therefore, we conduct this optimization based on the expectation maximization (EM) algorithm [0], which is an effective approach for models having latent variables. In this algorithm, marginal likelihood is maximized indirectly by maximizing the expectation of the complete-data log likeli- hood, as this is equivalent to maximizing the lower bound of the marginal likelihood [0]. The EM algorithm is iterated via application of an expectation step (E-step) and a maximization step (M-step).\n\nEquation (3) is transformed via the introduction of the new parameterization $\\nu = 2\\alpha$ and $s = \\beta/\\alpha$, with the latent variable redefined as $\\tau_n = \\sigma^2_n \\alpha/\\beta$. Marginal distribution can then be expressed as\n\n$$\n\\begin{aligned}\n\t\tP(x_n) &=& \\int {\\mathcal N}(x_n|s\\tau_n) {\\rm IG}(\\tau_n;\\nu/2,\\nu/2) {\\rm d}\\tau_n \\\\\n\t\t&=& \\frac{\\Gamma \\left(\\frac{\\nu+1}{2}\\right)s^{-\\frac{1}{2}}}\n\t\t{\\sqrt{\\pi\\nu}\\Gamma(\\frac{\\nu}{2}) (1 + \\frac{x_n^2}{\\nu s})^{\\frac{\\nu+1}{2}}}.\n\n\\end{aligned}\n$$\n\nThis expression is equivalent to Student’s $t$ -distribution with a mean of zero, in which $\\nu$ is the degree of freedom and $s$ is the scale parameter of the $t$ -distribution [0]. Here, the posterior distribution of the latent variable is given by\n\n$$\n\\begin{aligned}\n\tP(\\tau_n|x_n) &\\propto & {\\mathcal N}(x_n|s\\tau_n) {\\rm IG}(\\tau_n;\\nu/2,\\nu/2) \\\\\n\t&\\propto & {\\rm IG}\\left(\\tau_n;\\frac{\\nu+1}{2},\\frac{\\nu+x_n^2/s}{2}\\right).\n\t\n\\end{aligned}\n$$\n\nThe parameters $[\\nu, s]$ are estimated as outlined below based on the EM algorithm.\n\n- (i) Initialize $\\nu$ and $s$ with the selection of arbitrary starting values. (ii) E-step. Calculate the expectation of the complete-data log likelihood. This expectation, denoted as $Q(\\nu,s)$, is given by\n\n$$\n\\begin{aligned}\n\t\t\tQ(\\nu,s) &=& \\mathbb{E} \\left[\\log \\left(\\prod\\right)_{n=1}^{N} {\\mathcal N}(x_n|s\\tau_n)\n\t\t\t\t\t\t {\\rm IG}(\\tau_n;\\nu/2,\\nu/2)\\right] \\\\\n\t\t\t\t\t &=& \\sum_{n=1}^{N} \\biggl\\{-\\frac{1}{2} \\log \\left((\\right)2\\pi) -\\frac{1}{2} \n\t\t\t\t\t\t \\log \\left(s\\right) - \\frac{x_n^2}{2s} \\omega_n + \\frac{\\nu}{2} \\log \\left(\\frac\\right){\\nu}{2} \\\\\n\t\t\t\t\t&\\quad& - \\log \\left(\\Gamma\\right) \\left(\\frac{\\nu}{2} \\right) - \\frac{\\nu}{2} \\omega_n \n\t\t\t\t\t\t -\\left(\\frac{\\nu}{2}+\\frac{3}{2} \\right) \\lambda_n \\biggr\\} \n\t \\end{aligned}\n$$\n\n- where $\\omega_n$ and $\\lambda_n$ are derived as follows by calculating the posterior distribution of the latent variable from (6).\n\n$$\n\\begin{aligned}\n\t\t\t\\omega_n &:=& \\mathbb{E}\\,[\\tau^{-1}_n|x_n] = \\frac{\\nu+1}{\\nu + x_n^2/s}, \\\\\n\t\t\t\\lambda_n &:=& \\mathbb{E}\\,[\\log \\left(\\tau\\right)_n|x_n] \\\\\n\t\t\t\t\t\t&=& -\\log \\left(\\omega\\right)_n + \\log \\left(\\left\\right)(\\frac{\\nu + 1}{2} \\right) - \\psi \\left(\\frac{\\nu + 1}{2} \\right)\n\t\t\\end{aligned}\n$$\n\n- where $\\psi(\\cdot)$ is a digamma function. M-step. Update the parameters by maximizing $Q(\\nu, s)$. By setting the derivative of $Q(\\nu, s)$ with respect to $s$ equal to zero, the new scale parameter is obtained as\n\n$$\ns^{\\textrm{new}} = \\frac{1}{N} \\sum_{n=1}^{N} \\omega_n x_n^2.\n$$\n\n- There is no closed-form expression for the degree of freedom. Hence, it is updated by finding $\\nu$, where $Q(\\nu, s)$ is maximized numerically based on a line search.\n\n$$\n\\nu^{\\textrm{new}} = \\mathop{\\rm arg\\ max}\\limits_{\\nu}\\, Q(\\nu,s^{\\textrm{new}}).\n$$\n\n- Evaluate the log-marginal likelihood\n\n$$\n\\log \\left(P\\right)(X) = \\log \\left(\\prod\\right)^{N}_{n=1} P(x_n)\n$$\n\n- and determine whether the convergence of the estimation is reached when the relative change in the log-marginal likelihood between iterations falls below the predetermined threshold $\\varepsilon$. If the convergence criterion is not satisfied, return to step (ii).\n\nAs each cycle of the EM algorithm will increase the log- marginal likelihood monotonically [0], the convergence of pa- rameter estimation is guaranteed. The estimated $t$ -distribution parameters are finally transformed to variance distribution parameters using $\\alpha = \\nu/2$ and $\\beta = \\nu s/2$. Using this procedure, the variance distribution can be estimated from measured EMG signals.\n\n## III. E XPERIMENTS\n\n## A. Simulation\n\nTo evaluate the accuracy of the estimated variance distribu- tion, we performed a simulation experiment using artificially generated EMG signals. First, a discrete series $\\{\\sigma^2_t; t=1, \\cdots, T\\}$ was produced from random numbers following an inverse gamma distribution ${\\rm IG}(\\alpha_0, \\beta_0)$, and a random number $\\{x_t\\}$ following a Gaussian distribution with a mean of zero and a variance of $\\sigma_t$ was then generated for each value of $t$. The $\\{x_t\\}$ values were regarded as a time series of an EMG signal measured at a sampling frequency of $F_s$ Hz. The accuracy of the distribution estimation was verified by comparing the true values $\\alpha_0$ and $\\beta_0$ with the estimated values $\\alpha$ and $\\beta$. As an index of the estimation accuracy, the absolute percentage error was defined as $|{\\rm true\\ value} - {\\rm estimated\\ value}|/({\\rm true\\ value}) \\times 100$.\n\nIn the estimation of $\\alpha$ and $\\beta$, the last $L$ values of the signals $\\{x_t\\}$ were used. The window length $L$ took values of $100$, $50$, $10$, $5$, $2$, and $1$ s. The average absolute percentage errors", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "FIG. 2. (a) Transmissivity measurement performed with an FTIR on a µ m-thin film exfoliated from a bulk specimen of BiTeI. The inset shows the band structure of BiTeI obtained by first-principles calculations [12]. The transition $\\beta$ (red dashed line at 0.37 eV) represents the highest energy for which a transition between the Rashba-split bands is possible and $\\alpha$ (green) the lowest. The transition $\\gamma$ (blue dashed line at 0.52 eV) represents the lowest energy for which electrons can be excited from the valence band to a free state above the Fermi energy $E_{\\textnormal{F}}$ (dashed black line). In the narrow window between $\\beta$ and $\\gamma$ lies a range of increased transmission. (b) Spectral absorption coefficient of BiTeI digitized from [13]. The triangular markers indicate the values of the wavelength of the pump beam employed in our experiment.\n\nsmaller than in our sample [13]. However, this difference affects only the exact value of the energy corresponding to $\\gamma$. The values of the wavelength chosen for the pump beam in our experiments are marked in Fig. 2 (b). For the interband transitions (0.52 µ m to 2.08 µ m) $\\alpha_\\textnormal{abs}$ is of the order of $10^5$ cm -1. These considerations about the absorption coefficient of BiTeI are relevant to our objective of impulsive excitation of coherent lattice dynamics. In particular, this process can be successfully induced in two different light-matter interaction regimes. If dissipation is negligible (i.e. negligible absorption of the pump beam), impulsive stimulated Raman scattering (ISRS) of elementary excitations can be realised [10,16]. On the other hand, the absorption of light by electrons can also lead to an impulsive force triggering Raman- active phonons. In this case a displacive excitation of coherent phonons (DECP) is achieved [10]. We note that the ISRS scenario seems favourable in view of our goal, as it allows to exclusively excite the quasiparticles of interest, without parasitic electronic dissipation. As these two mechanisms were established in dielectrics and semiconductors with a much more simple band structure than BiTeI, a spectrally broad tunability of the pump-photon energy is key for our purposes. In particular, it can be expected that the interband transitions in BiTeI may allow a DECP generation [10,17] of the A 1g lattice mode. However, the role of the incoherent electronic background must be quantitatively assessed by an experimental procedure, to realise the conditions under which such background does not hinder the coherent lattice dynamics [18]. The situation for the intraband transitions is less obvious, as it has not yet been explored. The intraband absorption coefficient is an order of magnitude smaller than in the case of interband processes (33 times lower at 7.5 µ m than at $0.52$ µ m). Importantly, it is not straightforward to predict how the vastly different natures of intraband and interband pro- cesses affect the photoinduced coherent lattice dynamics. Fig. 3 (a) shows the dynamics of the normalised reflection coefficient $\\Delta R/R_0$ photoinduced by laser pulses with a wavelength of 0.52 µ m. The pulse duration is 198 fs and the fluence is set to $\\Phi=0.77$ mJ/cm 2. The signal consists of two components, (i) coherent oscilla- tions with a frequency of 2.7 THz superimposed to (ii) an exponentially decaying background with an amplitude of $\\approx 15\\cdot10^{-3}$. We ascribe the incoherent component of the observed signal to the relaxation dynamics and thermalisation processes of the photoinduced electrons reported in the literature [7]. As we are interested in the coherent response of the lattice, we subtract the incoherent background (details of data analysis in Supplementary Materials) and analyse the oscillatory pattern depicted in Fig. 3 (b). The peak-to-peak amplitude of the oscillations amounts to $\\approx 1\\cdot 10^{-3}$ and decays in approximately 7 ps. A Fourier transform of the time trace in Fig. 3 (b) reveals the spectrum of the oscillations displayed in Fig. 3 (c), which we fit with a Lorentzian function (see Supplementary Materials). As a result we estimate the center frequency to be $f_0=2.76$ THz and the lifetime to be $\\tau=1.82$ ps. These values are consistent with the particular A 1g mode, in which the atomic displacement has the correct symmetry to modulate $E_z$, as previously determined by spontaneous Raman scattering [9]. We investigate the dependence of the electronic background and the oscillation amplitude on the pump wavelength. We perform measurements of $\\Delta R/R_0$ employing the pump wavelengths indicated in Fig. 2 (b). For each pump wavelength we analyze the dependence of electronic-background and oscillation amplitude, $A_\\textnormal{e}$ and $A_\\textnormal{osci}$, on the excitation fluence $\\Phi$. The results are reported in Fig. 4 (a) and (b), in which intraband and interband processes are color coded. We calculate the values of the absorbed fluence for every pump wave-", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "# Combining guaranteed and spot markets in display advertising: Selling guaranteed page views with stochastic demand ∗\n\nBowei Chen † Adam Smith Business School University of Glasgow School of Computing Science University of Glasgow Trinity Business School Trinity College Dublin School of Computer Science University of Lincoln\n\nWhile page views are often sold instantly through real-time auctions when users visit websites, they can also be sold in advance via guaranteed contracts. In this paper, we present a dynamic programming model to study how an online publisher should op- timally allocate and price page views between guar- anteed and spot markets. The problem is challeng- ing because the allocation and pricing of guaran- teed contracts affect how advertisers split their pur- chases between the two markets, and the terminal value of the model is endogenously determined by the updated dual force of supply and demand in auctions. We take the advertisers’ purchasing be- haviour into consideration, i.e., risk aversion and stochastic demand arrivals, and present a scalable and efficient algorithm for the optimal solution. The model is also empirically validated with a commer- cial dataset. The experimental results show that sell- ing page views via both channels can increase the publisher’s expected total revenue, and the optimal pricing and allocation strategies are robust to differ- ent market and advertiser types.\n\n## Introduction\n\nDisplay advertising is one of the most popular forms of online marketing. It uses the Internet and the World Wide Web as an advertising medium, and when users visit websites, the pro- motional messages (i.e., the ads), appear on the pages. They usually come in terms of rectangular images or photos placed on a web page either above, below or on the sides of the page’s main content and are linked to other web pages. On- line publishers make profits by selling the page views, namely the impressions, through two channels: (i) selling them in ad- vance via contracts; or (ii) auctioning them off in real time\n\nEuropean Journal of Oper- ational Research\n\nwhen users visit the web pages. The former is called guar- anteed contracts (or reservation contracts) while the latter is called real-time bidding (RTB). Over the past decades, RTB has become the widely used sales model for display adver- tising, in which advertisers come to a common marketplace, i.e., ad exchange, to compete for impressions from their tar- geted users (Muthukrishnan, 2009; Mansour et al., 2012). It is real-time, impression-level and auction-based, thus has achieved a significant level of automation, integration and user- targeting (Yuanetal., 2014b; Sunetal., 2016).\n\nAlthough RTB is more widely used, guaranteed contracts in fact have a longer history. In 1994, wire.com signed four- teen contracts with companies, such as AT & T, Club Med and Coor’z Zima, being recognised as the start of display advertis- ing (DoubleClick, 2005). A guaranteed contract is an agree- ment and it is usually negotiated privately between a publisher and an advertiser for bulk sales. Only a small portion of im- pressions on the market is sold through guaranteed contracts but they bring in much more revenue than RTB (eMarketer, 2013). In order to meet the demand for automation due to the huge number of site visits, standardised guaranteed con- tracts have been recently discussed. This is known as pro- grammatic guarantee (PG). In essence, PG is a sales system that sells future impressions via standardised guaranteed con- tracts in addition to RTB (OpenX, 2013). Examples include Google DoubleClick’s Programmatic Guaranteed, AOL’s Pro- grammatic Upfront and Rubicon Project’s Reserved Premium Media Buys. Recent studies have investigated PG from differ- ent perspectives and we provide an extensive review in Sec- tion 2. These studies aim to answer the following two main questions: (i) how many future impressions should be allo- cated to guaranteed contracts? and (ii) how to price the guar- anteed contracts?\n\nIn this paper, we use a revenue maximisation model to study how to sell impressions using RTB and PG. More specifically, some impressions are sold in advance via standardised guar- anteed contracts before the delivery day while the rest of the impressions are auctioned off in RTB using the second-price auctions (Ben-Zwietal., 2015). We focus on one ad slot and study how the estimated total impressions in a future period, i.e., the ad delivery day, should be allocated and priced algo- rithmically between guaranteed and spot markets. For exam-\n\n∗ B.Chen, J.Huang, Y.Huang, S.Kollias, and S.Yue. Combining guaranteed and spot markets in display advertising: selling guaran- teed page views with stochastic demand., 280(3):1144–1159, 2020. [PublicationLink] † Corresponding author", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "## INTRODUCTION\n\nDynamical chiral symmetry breaking is an important feature of Quantum Chromodynamics(QCD). With the rapid development in heavy ion experiment and observation in astronomy, restoration of the chiral symmetry and the deconfinement phase transition are expected to occur in ultra-relativistic heavy-ion collisions or in the interior of neutron stars [1-4]. Many famous research institutions, such as FAIR GSI, NICA JINR, and J-PARK, have carried out experiments to study the properties of high-density matter which could help to clarify the phase structure of quark matter.\n\nAs a gauge theory at short distances, perturbative Quantum Chromodynamics is a remarkably successful and rich theory of the strong interactions. However, many physical phenomena in a long distance have to be dealt with non- perturbative method, especially in handling the problems related to low-energy physics, the non-perturbation model is of particularly great use. Among many effective models pion plays a crucial role with the fact that it occupies a special place in nuclear and particle physics. In the standard picture, pion is explained as the (pseudo-) Nambu- Goldstone boson. It arises as a consequence of the dynamical breakdown of the chiral symmetry [5,6]. In the chiral limit, where the masses of the two lightest quarks are turned off, the pion mass is zero. When the chiral symmetry is broken spontaneously, the quark condensate represents the leading order parameter. The Nambu-Goldstone bosons can interact if they carry momentum. Weinberg’s low energy theorems state that pion-pion scattering lengths are related to the pion mass with $a_0\\sim7M_\\pi^2/(32\\pi f_\\pi^2)$, $a_2\\sim-M_\\pi^2/(16\\pi f_\\pi^2)$. In the chiral limit, the pion-pion S-wave scattering lengths vanish, and these quantities can be used as a sensitive probe of the chiral symmetry breaking.\n\nThe pion-pion scattering, as one of the most fundamental hadronic processes of QCD at the mesonic level, provides a direct link between the theoretical formalisms of chiral symmetry and experiments. Most calculations so far have been done in the infinite volume limit[7-13]. Works on the chiral phase transition are stressed on the influence of temperature, baryon densities and other outside parameters such as magnetic field [14-17]. But the quark-gluon plasma(QGP) system produced by heavy-ion collisions experiments always has a finite size. The volume of homo- geneity before freeze out for Au–Au and Pb–Pb collisions is between approximately $50 \\sim250$ fm $^3$. It is estimated that the volume of the smallest QGP system could be as low as (2 fm) $^3$ [18], motivated by the estimated plasma size presumably formed in high-energy nucleus-nucleus collision at RHIC [1]. Studies show that the chiral behavior depends on the volume size of quark matter and the chiral symmetry breaking is closely related to finite volume effects in QCD [19-22]. Besides, the finite size effect on the dissociation and diffusion of chiral partners, on the phase structure of the NJL model in D=3 Euclidean dimensions and on the viscosity, bulk viscosity, electrical conductivity have been studied [23-25].\n\nSince pion meson plays an important role in low-energy physics, it is worthwhile to clarify its behaviors as extrap- olation to high temperature and small volume size.\n\nThe chiral phase transition of a finite system depends on the choice of the boundary conditions [26,27]. For quark fields, the time direction is constrained by anti-period requirement, but the choice of boundary condition in spatial direction is at large. Typical boundary conditions are anti-period boundary condition (APBC) and period boundary condition(PBC) which are used to weaken the impact from physical boundary.\n\nIn this paper, the scheme of proper time regularization is adopted [28]. The virtue of using proper time regularization here is that it enables us to perform the sum over the thermal Matsubara frequencies analytically. Other than most previous works which use cutoff scheme, we first use the proper time regularization to study the pion-pion scattering length in a finite volume size. We expect to extract the signals of chiral phase transition from pion-pion interactions in the hot medium while considering the finite volume effect.\n\nThe paper is organized as follows. In section II, the volume and temperature dependence of effective quark mass is deduced from Nambu – Jona-Lasinio (NJL) model under the proper time regularization. In section III, the equations for the meson mass and the pion decay constant at finite volume are deduced. We also give numerical results and analyze them in detail. The scattering length formula and its corresponding numerical results are presented in section IV, closely followed by a short conclusion in section V.\n\n## NJL MODEL AT FINITE VOLUME SIZE\n\nWe use a two-flavor NJL Lagrangian model, which is motivated by the symmetries of QCD, to describe the coupling", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "[email protected]\n\[email protected]\n\[email protected]\n\[email protected]\n\[email protected]\n\n# Measurement of the energy relaxation time of quantum states in quantum annealing with a D-Wave machine\n\nTakashi Imoto\n\nYuki Susa and\n\nRyoji Miyazaki\n\nTadashi Kadowaki\n\nYuichiro Matsuzaki\n\nQuantum annealing has been demonstrated with superconducting qubits. Such a quantum an- nealer has been used to solve combinational optimization problems and is also useful as a quantum simulator to investigate the properties of the quantum many-body systems. However, the coher- ence properties of actual devices provided by D-Wave Quantum Inc. are not sufficiently explored. Here, we propose and demonstrate a method to measure the coherence time of the excited state in quantum annealing with the D-Wave device. More specifically, we investigate the energy relaxation time of the first excited states of a fully connected Ising model with a transverse field. We find that the energy relaxation time of the excited states of the model is orders of magnitude longer than that of the excited state of a single qubit, and we qualitatively explain this phenomenon by using a theoretical model. The reported technique provides new possibilities to explore the decoherence properties of quantum many-body systems with the D-Wave machine.\n\n## INTRODUCTION\n\nThere have been significant developments in quantum annealing (QA). QA is useful to solve combinatorial op- timization problems [1–3]. D-Wave Quantum inc. has developed a quantum annealing machine consisting of thousands of qubits [4]. It is possible to map the com- binatorial optimization problem into the search for the ground state of the Ising Hamiltonian [5, 6]. Efficient clustering and machine learning using QA have been re- ported [7–14]. Moreover, QA can be used for topological data analysis (TDA) [15].\n\nFurthermore, QA is used not only for solving prob- lems but also as a quantum simulator of quantum many- body systems in and out of equilibrium [16–19]. There is a great interest in whether the D-Wave machines ex- ploit quantum properties, and recent experiments with a D-Wave machine have confirmed good agreement with theoretical predictions based on Schrodinger dynamics [20, 21]. Focusing on this point, an attempt has been reported to use longitudinal magnetic fields to probe the dynamics of a qubit and distinguish the effects of noise in the D-Wave machine [22]. However, the coherence prop- erties of actual devices provided by D-Wave are not suffi- ciently explored, and further studies are still needed. In particular, since an eigenstate of the Hamiltonian pro- vided by the D-Wave machine can be entangled, the co- herence properties of entangled states need to be clarified.\n\nIn this paper, we propose and demonstrate a method to measure the energy relaxation time (called $T_1$) of the excited states during QA by using the D-Wave machine. We adopt a fully connected Ising model of four qubits for the problem Hamiltonian of QA. We show that the energy relaxation time of the first excited states of the", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "## Preliminary considerations\n\nThe most straightforward way to present both positive and nan search results is that of working in the context of a given BSM theory; an approach of this type is, by construction, a top-down one. While statistically clean, top-down procedures have two main drawbacks. Firstly, they often have to be repeated, even if the datasets are unchanged, whenever a different theoretical model is chosen. Secondly, the details of how BSM theories are treated in such procedures are under the control of the experimental collaborations, which, among other things, renders it difficult for theorists to assess how tweaking different aspects of the models might improve, worsen, or otherwise affect the search results.\n\nFor these reasons, it is interesting to consider the opposite viewpoint, namely that of a bottom-up approach in which data are manipulated, and the search results presented using the fewest possible number of theoretical assumptions. This is the goal of the present paper. More specifically, a model-independent functional form for describing the lineshape of a resonance and its interference with the background is employed, and the search results presented as allowed or forbidden regions in the space of parameters relevant to such a form. The idea is that if experimental results are given in this way, any theoretical model can be quickly checked to be compatible or incompatible with the data by means of a simple computation whose results are expressed in terms of the same parameters.\n\nIn order to simplify the approach we are proposing, a number of assumptions need to be made. In particular:\n\n- 1. We consider one resonance at a time; if several resonances are present, they must be sufficiently well separated for the procedure to work independently for each of them.\n\n- 2. We work with the invariant mass of the resonance, which can be reconstructed by means of the four-momenta of the decay products.\n\n- 3. A single partonic process is responsible for the signal – background interference pat- tern.\n\nThere is at least one implication of item # 3 that requires an immediate explanation. The overarching understanding is that we presently have a solid confidence in the SM, as well as in the correctness of the theoretical tools that are used to simulate both SM and BSM physics processes with a good control on the systematics. Thus, in the entirely realistic possibility that the background to the search proceeds through more than one partonic channel ($H\\rightarrow\\gamma\\gamma$ being a chief example of this situation), the channels that are not interfering must be subtracted from the data prior to the fitting procedure that we shall describe below. This operation will contribute to the overall systematics of the procedure we are proposing.\n\n## Model-independent template functional form\n\nGiven the assumptions listed in section 2, let us denote by $m$ and $\\Gamma$ the mass and width, respectively, of the resonance whose characteristics we seek to determine. We write the amplitude for the partonic process that features the signal – background interference as follows:\n\n$$\n\\begin{aligned}\n\\bar{A}_h(q^2)=\\frac{S_h(q^2)}{q^2-m^2+im\\Gamma}+\\frac{B_h(q^2)}{m^2}\\,,\n\n\\end{aligned}\n$$\n\n", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "# Discovering Gateway Ports in Maritime Using Temporal Graph Neural Network Port Classification\n\n\\begin{tabular}{cc}\nDogan Altan\\upstairs{\\mathbf{S},*}, Mohammad Etemad\\upstairs{\\mathbf{S}}, Dusica Marijan\\upstairs{\\mathbf{S}}, Tetyana Kholodna\\upstairs{\\affiltwo}\n[0.25ex]\n{\\small \\upstairs{\\mathbf{S}} Simula Research Laboratory, Norway} {\\small \\upstairs{\\affiltwo} Navtor AS, Norway} \\end{tabular}\n\nVessel navigation is influenced by various factors, such as dynamic environmental factors that change over time or static features such as vessel type or depth of the ocean. These dynamic and static navigational factors impose limitations on vessels, such as long waiting times in regions outside the actual ports, and we call these waiting regions gateway ports. Identifying gateway ports and their associated features such as congestion and available utilities can enhance vessel navigation by planning on fuel optimization or saving time in cargo operation. In this paper, we propose a novel temporal graph neural network (TGNN) based port classification method to enable vessels to discover gateway ports efficiently, thus optimizing their operations. The proposed method processes vessel trajectory data to build dynamic graphs capturing spatio-temporal dependencies between a set of static and dynamic navigational features in the data, and it is evaluated in terms of port classification accuracy on a real-world data set collected from ten vessels operating in Halifax, NS, Canada. The experimental results indicate that our TGNN-based port classification method provides an f-score of 95% in classifying ports.\n\n20pt 20pt Keywords: Keywords: maritime situational awareness, port classification, port con- gestion, vessel trajectories, AIS data, actual ports, gateway ports, temporal graph neural networks, spatio-temporal, maritime traffic, port area\n\n## 1. Introduction\n\nAccording to the United Nations Conference on Trade and Development (UNCTAD) report in 2018, around 80% of global trade by volume is transferred by vessels and operated cargo activities at ports worldwide [1]. Therefore, it is crucial to ensure the efficiency of navigation operations to minimize port congestion, reduce vessels’ waiting time and optimize travel for reduced vessel fuel consumption. Vessels often need to wait in specific regions outside the actual ports because of port congestion, unfavorable weather conditions, or particular properties of the vessels or the ocean. We call these waiting regions gateway ports. To illustrate this type of regions, Figure 1 depicts a port area of Singapore that includes actual and gateway ports. Colored polygons represent the extracted ports, and the extraction is achieved by taking into account the frequency of visits; the darker the color, the higher the number of visits. As can be seen from the figure, gateway ports are mostly clustered outside of the actual ports on the right side of the figure. They can be formed by cargo vessels exchanging items or waiting for better weather conditions. Being able to distinguish between actual and gateway ports allows expert systems to extract the properties of these areas and result in vessels to optimize trajectories for fuel-saving [2] and optimize the estimated time of arrival (ETA) [3, 4]. Furthermore, these types of optimizations contribute to the efficiency of port operations, such as loading/unloading [5] made by port authorities.\n\nTo address the problem of automatically discovering gateways ports, we need to con- sider spatial relationships between gateway ports since they are generally clustered in close proximity to the actual ports. This type of information is typically obtained by analysing trajectory data collected from multiple vessels over time. Moreover, we need to consider the", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Figure3: MDS plot obtained using Euclidean distance on (Welch’s) spectral density estimates on airflow signals of OSA patients. The clusters were obtained using fuzzy clustering with 4 clusters and a fuzzy parameter value of $m=1.5$.\n\nwith $ahi\\geq 2$ and may be considered the mildest case. It appears that unlike cluster 1, these patients primarily suffer from sleep issues which do not affect the other aspects of their life to the same extent.\n\nClusters 2 and 3 show results that are intermediate to Clusters 1 and 4, both reporting high sleep efficiencies. Cluster 3 stands out with a high response to emotional stress survey questions, despite having a low percentage of wake sleep stages. Children in this cluster may benefit from being referred to a psychiatrist in order to study the connection between their emotional symptoms and their sleep quality. Patients in cluster 2 on the other hand have relatively low emotional and school functioning issue scores,but report high physical suffering scores. These patients may benefit from a dentist referral - for example, there are multiple patients with very high CFI in this groups. Overall, patients in these two clusters may need specific treatments to address their individual issues.\n\n## 7. TDA analysis of airflow signals\n\nIn this section, we present an analysis of OSA patients based on the topo- logical features of their airflow signals. To conduct TDA on these signals, we first transformed them into periodicity scores over time frames i.e. we divided each time series into blocks or frames of fixed time lengths and calculated the periodicity for each block. This was done using two steps: (1) transforming time series into point cloud data using a sliding window and (2) calculating periodicity over time frames of the point cloud data through 1D persistence.\n\nWe followed this up with a three-fold analysis of the periodicity score data by: 1) clustering using zero dimensional persistence (complete-linkage hierar- chical clustering), 2) applying ANOVA to evaluate periodicity sores over time", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "4.3.\n\n## Corollaries on return probability and drift\n\n. By the general relation between isoperime- try and decay of return probability (see e.g. [39,Theorem14.3]), a lower bound on F ø lner function implies an upper bound on the return probability $\\mu^{{n}}(e)$. In particular, if there is some $\\delta>0$ such that the F lner function of $G$ satisfies ${\\rm F{\\o}l}_G(n)\\succcurlyeq \\exp(n^{2+\\delta})$ over a sequence of sufficiently long sub-intervals, then there exists a time subsequence $(n_i)$ such that\n\n$$\n\\mu^{(2n_i)}(e)\\preccurlyeq\\exp\\left(-n_i^{\\frac{2+\\delta}{4+\\delta}}\\right).\n$$\n\nRecall the drift function of a random walk with step distribution $\\mu$ is defined as\n\n$$\nL_{\\mu}(n)=\\sum_{x\\in G}d_S(e,x)\\mu^{(n)}(x).\n$$\n\nRemark4.3. Although in general it is not sharp, one can use the fact that return probability $\\mu^{(2n)}$ attains its maximum at identity $e$ to give the following bounds:\n\n$$\n\\sum_{g\\in B(e,r)}\\mu^{(2n)}(g)\\leqslant\\sum_{g\\in B(e,r)}\\mu^{(2n)}(e)=v_{G,S}(r)\\mu^{(2n)}(e).\n$$\n\nSince the entropy satisfies $H_{\\mu}(2n)\\geqslant -\\log\\mu^{(2n)}(e)$, we have\n\n$$\nL_{\\mu}(2n)\\geqslant cH_{\\mu}(2n)\\geqslant -c\\log\\mu^{(2n)}(e).\n$$\n\nIt follows that in this case if there is some $\\delta>0$ such that the F ø lner function of $G$ satisfies ${\\rm F{\\o}l}_G(n)\\succcurlyeq \\exp(n^{2+\\delta})$ over a sequence of sufficiently long sub-intervals, the $\\mu$ - random walk is neither cautious nor diffusive.\n\nWe now show that groups considered in the previous section with appropriate chosen parameters provide examples with oscillating drift function as stated in Corollary 1.6 in the Introduction.\n\nProofofCorollary1.6. Given $\\beta$, let $\\alpha$ be an exponent such that $\\frac{\\alpha}{2+\\alpha}>\\beta$. Take a rapidly growing sequence $(m_i)$, and set\n\n$$\n\\tau(n)=n^{\\alpha}\\mbox{for}n\\in[m_{2j},m_{2j+1}]\\mbox{and}\\tau(n)=m_{2j+1}^{\\alpha}\\mbox{for}n\\in[m_{2j+1},m_{2j+2}).\n$$\n\nBy Theorem 1.5, there is a nilpotent-by-cyclic group $G$ such that\n\n$$\n\\frac{1}{C}(n+\\tau(n))\\leqslant\\log{\\rm F{\\o}l}_{G}(n)\\leqslant Cn+C\\tau(n).\n$$\n\nRecall that in the proof of Theorem 1.5, the group $G$ is defined as the smallest group that projects onto a sequence $\\bar{G}_i$ marked with generating set $\\{b_0,t\\}$, where by construction the Cayley graph of $\\bar{G}_i$ coincide with the Cayley graph of the lamplighter ${\\mathbb Z}\\wr ({\\mathbb Z}/2{\\mathbb Z})$ in the ball of radius $2^{k_i}$ around identity. The sequence $(k_i)$ is determined from the prescribed function $\\tau$ as in (4.1).\n\nFrom the definition of the function $\\tau$ and the sequence $(k_i)$ as in (4.1), we have that there exists a subsequence $(p_j)$ such that\n\n$$\nk_{p_j+1}/k_{p_j}\\geqslant m_{2j+2}/m_{2j+1}.\n$$\n\nIndeed, this is because the function $\\tau$ is constant on the interval $(m_{2j+1},m_{2j+2})$. Mark $\\bar{G}_0={\\mathbb Z}\\wr({\\mathbb Z}/2{\\mathbb Z})$ with the generating set $\\{b_0,t\\}$ and write\n\n$$\nG_j=\\Gamma/\\cap_{i=0}^{j}\\mathrm{ker}\\left(\\pi_{i}:\\Gamma\\rightarrow\\bar{G}_{i}\\right).\n$$\n\n", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "\n\n$$\n\\begin{split}\n\\sum_{1\\leqslant i,j\\leqslant n^{1-\\gamma}} K_{ij}\\binom{M}{k}^{-2}\\geqslant \\mathopen{}\\mathclose\\bgroup\\originalleft(2n^{2-\\gamma_k}-n\\aftergroup\\egroup\\originalright)\\binom{M}{k}^{-2} \\\\\n%\\geq \\frac{(2n^{2-\\gamma_k}-n^{1+\\gamma_k})k!}{e\\cdot n^{k\\gamma_k}}\n\\geqslant k!^2\\mathopen{}\\mathclose\\bgroup\\originalleft(2n^{2-(2k+1)\\gamma_k}-n^{(1-2k\\gamma_k)}\\aftergroup\\egroup\\originalright)\n\\geqslant k!^2n^{2-(2k+1)\\gamma_k}.\n\\end{split}\n$$\n\nLet us define an auxiliary bipartite graph where the vertices are the selected $k$ -tuples of points, and there is an edge between two if there was a $k$ -branch selected on the $2k$ points. This graph has $n^{1-\\gamma_k}$ vertices in each vertex class and at least ${k!}^2n^{2-(2k+1)\\gamma_k}$ edges. By the K ˝o v ´a ri-S ´o s-Tur ´a n theorem [0] if\n\n$$\n(1-\\gamma_k)\\mathopen{}\\mathclose\\bgroup\\originalleft(2-\\frac{1}{k}\\aftergroup\\egroup\\originalright)\\leqslant 2-(2k+1)\\gamma_k \\quad \\text{i.e. when}\\quad \\gamma_k\\leqslant \\frac{1}{2k^2-k+1,}\n$$\n\nthen the graph contains a complete bipartite graph $K_{k,k}$ which proves the existence of a $k$ -system among the triples.\n\nWith the same $\\gamma_k$ exponent, by increasing the constant multiplier, $\\mu_k,$ a larger bipartite graph, $K_{k,t}$ will appear in the bipartite graph. Using a larger $t$, one can guarantee a $k$ -system where all $b\\in B$ vertices are distinct, so the system contains exactly $3k^2$ points. It is a particular $(3k^2,k^3)$ configuration, which is an unavoidable collinear triple configuration, even in sparse systems.\n\n## When one set of points is collinear\n\nIn this section, we assume that set $C$ consists of $n$ collinear points, the convex hulls of $A$ and $B$ are not intersected by the line of $C$, and no line determined by the points of $A$ intersects the convex hull of $B$. Let us suppose that that there are $\\delta n^2$ collinear $(a,b,c)$ triples, where $a\\in A, b\\in B,$ and $c\\in C.$ In the case when $A$ is also a set of collinear points, Elekes proved that a positive fraction of the points of $B$ should also be collinear [0]. Elekes and Szab ´o proved the stronger statement that if the points of $A\\cup C$ are on a conic, then a positive fraction of the points of $B$ is collinear (Theorem 5.2 in [0]). Our main result in this section is to show that if $C$ is collinear, then — under some additional conditions — many points of $A$ and $B$ are on a conic. Let’s recall another conjecture of Elekes, similar to the earlier ones:\n\nConjecture For every $C > 0$ there is an $n_0 = n_0 (C)$ with the following property. If $A\\subset \\mathbb{R}^2,$\n\n$|A|\\geqslant n_0$ and\n\nthe number of directions determined by $A$ is at most $C|A|$ then $A$ contains six points of a (possibly degenerate) conic. (As usual, a pair of lines is considered a degenerate conic.)\n\nIn his classical paper Sums vs Products [0] Elekes writes the following about his conjecture:\n\nWe will show a stronger, density version of this later statement.\n\nTheorem 14 For any $D>0$ and $c>0$, there is a constant, $\\delta>0,$ and a threshold $n_0=n_0(c,D)$, so the following holds. Let $A$ be a set of $n$ points in the plane in convex position (the vertex set of a convex polygon) such that at least $cn^2$ point pairs determine at most $Dn$ distinct directions. If $|A|\\geqslant n_0$, then at least $\\delta n$ points of $A$ are on a conic.\n\nTheorem 14 is a direct corollary of Lemma 18 below. Before proving Elekes’ conjecture, we recall a definition from projective geometry.", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Figure2. Effect of the filament radius ($R$) and the magnetic field ($B$) on the GEOS filament instability with $\\kappa=0.1$ (upper left panel), $\\kappa=0.2$ (upper right panel), $\\kappa=0.5$ (lower left panel) and $\\kappa=1$ (lower right panel). In each panel, the horizontal axis shows the radius of filaments, the vertical axis shows the logarithm of magnetic field strength and the colour bar represents $\\omega_{\\rm fast}$. Units are given in Table 1. The darker shaded areas are more unstable.\n\nthe effect of filament radius and the magnetic field on $k_{\\rm critic}$. Here, the overall picture is very similar to the GEOSs, spe- cially one with $\\kappa = 0.1$. The only difference is that the very thin MPEOS filaments in the absence or presence of very weak magnetic fields, have $k_{\\rm critic}$ almost twice larger than their GEOS counterparts.\n\nFig. 6 represents the dispersion relations for the PEOS. As already mentioned in $\\Sigma$, we set $n=-1.5, -2$ and $-3$ as the polytropic indices and use $R$ and $B$ values same as in the MPEOS. We also compute the dispersion relation for $n=-4$ in order to examine the effect of smaller $n$ on the in- stability. In this figure, panels in each row indicate filaments with a specific polytropic index, while those in each column show specific radii. The filaments radii $R>10$, do not show noticeable changes in their dispersion relations, so we select only $R=0.25, 0.5, 0.75, 1, 2, 5$ and $10$ in order to see their effect on the filaments instability. Nevertheless, to see the impact of these parameters on the shape of dispersion rela- tions more clearly, we draw the plots only for $R = 1, 2, 5$ and $10$ and $B = 0, 0.2, 0.5$ and $5$. Here, once more the magnetic field strength $B>5$ causes the dispersion relation to nearly overlap similar to the GEOS and the MPEOS (Nagasawa 1987; Gehmanetal.1996b, H18). In all the panels of Fig. 6 it is obvious that increasing the magnetic field leads to the more stability by decreasing $\\omega_{\\rm fast}$ and for $R\\lesssim5$, reduction of $k_{\\rm critic}$ value. H18 showed that the infinite filaments with the PEOS are prone to the radial instability and a strongly enough magnetic field could suppress the instability. Focus- ing on the $R=5$ and $R=10$ panels, we see that this is also the case for the pressure confined filaments with PEOSs and relatively large radii. It should be noted that the smaller ra- dius is radially unstable just for $n=-1.5$ while the larger one is unstable for all the polytropic indices. The occurrence of RI could be probably due to this fact that the effective sound speed decreases when the density of the filament increases. For $R<5$ there is no sign of RI in the dispersion relations.\n\nIn the following, the effect of $R$, $B$ and $n$ on the instabil- ity of filaments with the PEOS is studied by leveraging the result of four large surveys in more details. Fig. 7 illustrates the outcome for $\\omega_{\\rm fast}$. The upper left stable region is a com- mon clear feature of all the panels which states that similar to the previous GEOS and MPEOS, the thinner filaments in", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Since the sequence $(w_n)_{n\\geqslant 1}$ is bounded in $H^1_0(D)$, up to a subsequence, there exists a function $z\\in H^1_0(D)$ such that $w_{\\varphi(n)}$ converges to $z$ weakly in $H^1_0(D)$ and strongly in $L^2(D)$. Now, since\n\n$$\nL\\, w_n=f_n+V\\cdot \\nabla w_n+cw_n:=g_n\\quad\\mbox{in}\\quad\\Omega_n,\n$$\n\nand $w_n\\in H^1_0(\\Omega_n)$, $w_n=R^L_{\\Omega_n}(g_n)$. But $g_{\\varphi(n)}\\rightharpoonup g:=f+V\\cdot \\nabla z+cz$ weakly in $L^2(D)$, so that, by assumption $(2)$, $w_{\\varphi(n)}\\rightarrow R^{L}_{\\Omega}(g)$ strongly in $L^2(D)$. Then the convergence of $w_{\\varphi(n)}$ to $z$ yields that $z=R^{L}_{\\Omega}(g)$, is a solution of\n\n$$\nL\\,z=f+V\\cdot \\nabla z+cz\\quad\\mbox{in}\\quad\\Omega,\\qquad z\\in H^1_0(\\Omega),\n$$\n\nor, in other words, $z=R^{-\\Delta}_{\\Omega}(f)=w$. Thus, provides a contradiction, therefore showing that $w_n\\rightarrow w$ strongly in $L^2(D)$, which means that holds.\n\n## 3. The principal eigenvalue on quasi-open sets\n\nFor a bounded domain $\\mathcal{O}\\subset \\mathbb{R}^d$ and $V\\in L^{\\infty}(\\Omega,\\mathbb{R}^d)$, the principal eigenvalue $\\lambda_1(\\Omega,V)$, of the (non self-adjoint) elliptic operator $L = -\\Delta + V \\cdot \\nabla$ on $\\mathcal{O}$ with Dirichlet boundary condition on $\\partial\\Omega$, was defined in [4] by\n\n$$\n\\lambda_1(\\mathcal{O},V)=\\sup\\big\\{\\lambda\\in\\mathbb{R}\\ :\\ \\exists \\varphi\\in W^{2,d}(\\Omega)\\quad\\text{such that}\\quad \\varphi>0\\quad\\text{and}\\quad -L\\varphi+\\lambda\\varphi\\leqslant0\\quad \\text{in}\\quad \\Omega\\big\\},\n$$\n\nwhere it was proved that $\\lambda_1(\\Omega,V)\\in\\mathbb{R}$ has the following properties:\n\n$(i)$ There\n\nis a positive eigenfunction $u:\\Omega\\rightarrow\\mathbb{R}$ such that $u \\in W^{2,p}_{loc}(\\mathcal{O})$, for all $p\\in [1,+\\infty)$, and\n\n$$\nLu = \\lambda_1(\\mathcal{O},V)u\\quad\\text{in}\\quad \\Omega,\\qquad u\\in H^1_0(\\Omega),\\qquad \\int_\\Omega u^2\\,dx=1,\n$$\n\n(see [4, Theorem 2.1]).\n\n$(ii)$ $\\lambda_1(\\mathcal{O},V) < \\text{Re}\\,(\\lambda)$ for\n\nevery eigenvalue $\\lambda \\neq \\lambda_1(\\mathcal{O},V)$ of $L$ in $\\mathcal{O}$ (see).\n\n$(iii)$ The\n\nfunctional $\\Omega\\mapsto\\lambda_1(\\Omega,V)$ is decreasing with respect to the domain inclusion.\n\nIn the sequel we extend the definition of $\\lambda_1(\\Omega,V)$ to quasi-open sets. We first recall that the definition can be extended to an arbitrary open set $\\mathcal{O}\\subset D$ by\n\n$$\n\\lambda_1(\\mathcal{O},V) = \\inf \\lambda_1(\\mathcal{O}_n,V),\n$$\n\nwhere the infimum is taken over all the connected component $\\mathcal{O}_n$ of $\\mathcal{O}$. Now, in view of property $(iii)$ above, for any quasi-open set $\\mathcal{O} \\subset D$, we define\n\n$$\n\\lambda_1(\\mathcal{O},V) := \\sup \\big\\{\\lambda_1(\\tilde{\\mathcal{O}},V) \\ : \\ \\tilde{\\mathcal{O}} \\text{open,} \\ \\mathcal{O} \\subset \\tilde{\\mathcal{O}} \\subset D \\big\\}.\n$$\n\nRemark 3.1. Notice that, these two definitions coincide for open sets.\n\nRemark 3.2. The functional $\\mathcal{O}\\mapsto \\lambda_1(\\mathcal{O} ,V)$, defined on the family of quasi-open sets, is still non-increasing with respect to the set inclusion, that is $\\lambda_1(\\Omega_2,V)\\leqslant \\lambda_1(\\Omega_1,V)$, whenever $\\Omega _1\\subset\\Omega_2$.\n\nWe will show that $\\lambda_1(\\mathcal{O},V)$ is finite and is an eigenvalue of $L$ in $\\Omega$ satisfying the minimality property $(ii)$. Recall that, for a quasi-open set of finite Lebesgue measure $\\Omega\\subset D$, we say that $\\lambda\\in \\mathbb{C}$ is an eigenvalue of the operator $L=-\\Delta+V\\cdot \\nabla$ in $\\Omega$ if there is an eigenfunction $u:\\mathbb{R}^d\\rightarrow \\mathbb{C}$, (weak) solution to the problem\n\n$$\n-\\Delta u+V\\cdot\\nabla u=\\lambda u\\quad\\text{in}\\quad\\Omega,\\qquad u\\in H^1_0(\\Omega;\\mathbb{C}),\\qquad \\int_\\Omega |u|^2\\,dx=1.\n$$\n\nLet now $c>0$ be the constant from Subsection 2.2 and $L' = L + c$. Note that $\\lambda\\in \\mathbb{C}$ is an eigenvalue of $L$ in $\\mathcal{O}$, if and only if, $\\lambda+c$ is an eigenvalue of $L'$ in $\\mathcal{O}$. By the argument from Subsection 2.2, we have that the bilinear form associated to the operator $L'$ is coercive and so, $R_{\\mathcal{O}}^{L'}$ is a compact operator on $L^2(D)$. In particular, the spectrum is a discrete set of eigenvalues with no accumulation points except zero and $\\lambda\\in\\mathbb{C}$ is an eigenvalue of $L$ in the sense of (3.2) if and only if $(\\lambda+c)^{-1}$ is an eigenvalue of $R_\\Omega^{L'}$.", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "## ELECTRONS, PHONONS, AND ELECTRON-PHONON COUPLING\n\nWe consider a model for electrons on the graphene lat- tice, and allow for lattice site vibrations. For the elec- trons, we use a nearest neighbour tight binding model [4] describing the $\\pi$ -bands, as explained in further detail in Appendix A. Other bands are disregarded, since only the $\\pi$ -bands are close to the Fermi surface for realistically achievable doping levels in graphene. For the phonons, we use a force constant model with nearest and next-to- nearest neighbour couplings as introduced in Refs. 63and 64 and elaborated in Appendices B and C. These mod- els give a realistic band structure and realistic phonon spectra.\n\nThe electron-phonon coupling model is derived by as- suming the electrons to follow the lattice site ions adia- batically, and by Taylor expanding the overlap integral $t_{ij}$ in the hopping Hamiltonian\n\n$$\nH = - \\sum_{\\langle i, j \\rangle, \\sigma} (t_{ij} c_{i\\sigma}^\\dagger c_{j\\sigma} + \\mathrm{h.c.})\n$$\n\nto linear order in the deviations. Here, $c_{i\\sigma}^\\dagger$ and $c_{i\\sigma}$ are creation and annihilation operators for an electron at site $i$ with spin $\\sigma \\in \\{\\uparrow, \\downarrow\\}$. Considering only the nearest neigh- bour hoppings, we obtain\n\n$$\nt_{i+\\pmb{\\delta}_A, i} = t_1 + (\\mathbf{u}_{i+\\pmb{\\delta}_A} - \\mathbf{u}_{i}) \\cdot \\nabla_{\\pmb{\\delta}} t_1(\\pmb{\\delta}),\n$$\n\nwhere $t_1$ is the nearest neighbour hopping amplitude, $\\mathbf{u}_i$ is the ionic displacement of lattice site $i$ from its equilib- rium position, and the overlap integral $t_1(\\pmb{\\delta})$ is regarded as a function of the relative position $\\pmb{\\delta}$ of the two lattice sites $i$ and $i+\\pmb{\\delta}_A$, where $\\pmb{\\delta}_A$ is the equilibrium nearest neighbour vector from the $A$ to the $B$ sublattice. Due to the mirror-symmetry about the line connecting the lat- tice sites $i$ and $i+\\pmb{\\delta}_A$, the electron-phonon coupling can be written as\n\n$$\nH_\\mathrm{el-ph} = \\frac{\\gamma t_1}{d^2} \n\\sum_{i\\in A, \\pmb{\\delta}_A, \\sigma} \n\\pmb{\\delta}_A \\cdot (\\mathbf{u}_{i+\\pmb{\\delta}_A} -\\mathbf{u}_i) (c_{i+\\pmb{\\delta}_A,\\sigma}^\\dagger c_{i,\\sigma} + \\mathrm{h.c.}),\n$$\n\nwhere $\\gamma = - {\\mathrm{d} \\ln t_1}/{\\mathrm{d} \\ln d}$ is a dimensionless number of order 1, and $d$ is the equilibrium nearest neighbour distance, which we use as our unit of length. In quasimomentum-space, this gives the electron-phonon coupling\n\n$$\n\\begin{aligned}\nH_\\mathrm{el-ph} &= \n\\sum_{\\mathbf{k},\\mathbf{q}} \\sum_{\\eta\\eta'} \\sum_{\\nu,\\sigma} \ng_{\\mathbf{k},\\mathbf{k}+\\mathbf{q}}^{\\eta\\eta',\\nu}\n(a_{\\mathbf{q}\\nu} + a_{-\\mathbf{q},\\nu}^\\dagger) c_{\\eta'\\sigma}^\\dagger(\\mathbf{k}+\\mathbf{q}) \nc_{\\eta\\sigma}(\\mathbf{k}) ,\n\\end{aligned}\n$$\n\nwhere $a_{-\\mathbf{q},\\nu}^\\dagger$ and $a_{\\mathbf{q}\\nu}$ are creation and annihilation oper- ators for in-plane phonons labelled by $\\nu \\in \\{0,1,2,3\\}$, and $\\eta, \\eta' = \\pm$ denote electron bands. To linear order in the lattice site deviations, the out-of-plane phonon modes do not couple to the electrons due to the assumed $z\\rightarrow -z$ mirror symmetry of the system [65]. The coupling matrix element $g^{\\eta\\eta',\\nu}_{\\mathbf{k},\\mathbf{k}+\\mathbf{q}}$ is given by\n\n$$\n\\begin{aligned}\ng_{\\mathbf{k},\\mathbf{k}+\\mathbf{q}}^{\\eta\\eta',\\nu} = &\n\\frac{g_0}{\\sqrt{N_A}} \\sqrt{\\frac{\\omega_\\Gamma}{\\omega_{\\mathbf{q}\\nu}}} \\sum_{\\pmb{\\delta}_A} \\mathopen{}\\mathclose\\bgroup\\originalleft(\\frac{\\pmb{\\delta}_A}{d} \\aftergroup\\egroup\\originalright) \n \\mathopen{}\\mathclose\\bgroup\\originalleft[e^{i\\mathbf{q} \\cdot \\pmb{\\delta}_A} \\mathbf{e}_\\nu^B(\\mathbf{q}) - \\mathbf{e}_\\nu^A(\\mathbf{q}) \\aftergroup\\egroup\\originalright] \\\\\n &\\times \\Big[e^{i \\mathbf{k}\\cdot \\pmb{\\delta}_A} F_{A\\eta'}^*(\\mathbf{k}+\\mathbf{q}) F_{B\\eta}(\\mathbf{k}) \n \\\\ & + e^{-i (\\mathbf{k}+\\mathbf{q})\\cdot \\pmb{\\delta}_A} F_{B\\eta'}^*(\\mathbf{k}+\\mathbf{q}) F_{A\\eta}(\\mathbf{k}) \\Big],\n\\end{aligned}\n$$\n\nwhere $F_{D\\eta}(\\mathbf{k})$ is the sublattice amplitude of electron band $\\eta$ at quasimomentum $\\mathbf{k}$ and follows from the di- agonalization of the free electron model, as elaborated in Appendix A. Similarly, $\\mathbf{e}_\\nu^D(\\mathbf{q})$ is the phonon polar- ization vector at sublattice $D\\in \\{A,B\\}$ for the phonon mode $(\\mathbf{q},\\nu)$, and follows from diagonalization of the in- plane phonon Hamiltonian (see Appendix B for details). The phonon mode frequencies are denoted by $\\omega_{\\mathbf{q}\\nu}$, $N_A$ is the number of lattice sites on the A sublattice, and the energy scale $g_0$ is given by\n\n$$\ng_0=\\sqrt{\\mathopen{}\\mathclose\\bgroup\\originalleft(\\frac{\\hbar^2}{2Md^2} \\aftergroup\\egroup\\originalright) \\frac{1}{\\hbar\\omega_\\Gamma}} \\gamma t_1,\n$$\n\nwhere $M$ is the carbon atom mass, and $\\omega_\\Gamma$ is a phonon energy scale given by the optical phonon frequency at the $\\Gamma$ -point $\\mathbf{q} = (0,0)$.\n\nTo quantify the strength of the electron-phonon cou- pling, one may introduce the dimensionless electron- phonon coupling strength parameter [47, 66]\n\n$$\n\\lambda_{\\mathbf{k}\\eta} = \\sum_{\\mathbf{q} \\nu} \\frac{2}{\\hbar \\omega_{\\mathbf{q}\\nu}} |g^{\\eta \\eta,\\nu}_{\\mathbf{k},\\mathbf{k}+\\mathbf{q}}|^2 \\delta(\\varepsilon_{\\mathbf{k}+\\mathbf{q}, \\eta}-\\varepsilon_{\\mathbf{k}\\eta}),\n$$\n\nwhere $\\varepsilon_{\\mathbf{k}\\eta}$ is the electron single particle energy. We have neglected interband scattering processes since the $\\pi$ -band only overlaps with the lower lying $\\sigma$ -bands at unrealistic doping levels [48, 67].\n\nAveraging $\\lambda_{\\mathbf{k}\\eta}$ over the Fermi surface corresponding to the energy of the incoming momentum often provides a simple and useful tool for understanding the depen- dence of the critical temperature of a superconductor on other system parameters through the BCS formula $k_B T_c \\approx \\hbar \\omega_D \\exp (-1/\\lambda)$, where $\\lambda$ is the Fermi surface av- erage of $\\lambda_{\\mathbf{k}\\eta}$.\n\n## COULOMB INTERACTION", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "\\begin{algorithm}[t]\n\\SetAlgoLined\n\\KwResult{$\\pi_\\theta$}\n{\\bf Initialization}: $\\mathcal{S}_{{}} = \\lbrace (z_{0,i} , e_i)\\rbrace_{i=1,...,N}$ where $(z_{0,i}, e_i) \\sim \\mathrm{Uniform}[\\mathcal{X}_0 \\times \\mathcal{E}]$\\;\n$k \\leftarrow 0$\\;\n{\\bf Model training}: Optimize $\\theta$ using $N_\\mathrm{full}$ training epochs on \\eqref{eq:loss_training} over $\\mathcal{S}_{{}}$\\;\n \\While{$k < K$}{\n {\\bf Adversarial search}: Find $\\mathcal{S}_{{\\mathrm{adv}}} = \\lbrace (z_{0,j}, e_j) \\mid \\xi_{x_{0,j}, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}} \\not\\models \\phi \\rbrace_{i=1,...,N_\\mathrm{adv}}$ with $\\theta$ fixed (e.g., acceptance-rejection sampling)\\;\n {\\bf Update initial states}: Re-sample from $\\mathrm{Uniform}[\\mathcal{X}_0 \\times \\mathcal{E}]$ to get $\\mathcal{S}_{{0}} = \\lbrace (z_{0,i}, e_i) \\rbrace_{i=1,...,N}$\\;\n $\\mathcal{S}_{{}} \\leftarrow \\mathcal{S}_{{0}} \\cup \\mathcal{S}_{{\\mathrm{adv}}}$\\;\n {\\bf Model training}: Optimize $\\theta$ using $N_\\mathrm{mini}$ training epochs on \\eqref{eq:loss_training} over $\\mathcal{S}_{{}}$\\;\n $k \\leftarrow k + 1$\\;\n}\n \\caption{STL controller synthesis (offline)}\n\\end{algorithm} 1:\n\nSTL controller synthesis (offline)\n\n\\begin{algorithm}[t]\n\\SetAlgoLined\n\\KwResult{$\\pi_\\theta$}\n{\\bf Initialization}: $\\mathcal{S}_{{}} = \\lbrace (z_{0,i} , e_i)\\rbrace_{i=1,...,N}$ where $(z_{0,i}, e_i) \\sim \\mathrm{Uniform}[\\mathcal{X}_0 \\times \\mathcal{E}]$\\;\n$k \\leftarrow 0$\\;\n{\\bf Model training}: Optimize $\\theta$ using $N_\\mathrm{full}$ training epochs on \\eqref{eq:loss_training} over $\\mathcal{S}_{{}}$\\;\n \\While{$k < K$}{\n {\\bf Adversarial search}: Find $\\mathcal{S}_{{\\mathrm{adv}}} = \\lbrace (z_{0,j}, e_j) \\mid \\xi_{x_{0,j}, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}} \\not\\models \\phi \\rbrace_{i=1,...,N_\\mathrm{adv}}$ with $\\theta$ fixed (e.g., acceptance-rejection sampling)\\;\n {\\bf Update initial states}: Re-sample from $\\mathrm{Uniform}[\\mathcal{X}_0 \\times \\mathcal{E}]$ to get $\\mathcal{S}_{{0}} = \\lbrace (z_{0,i}, e_i) \\rbrace_{i=1,...,N}$\\;\n $\\mathcal{S}_{{}} \\leftarrow \\mathcal{S}_{{0}} \\cup \\mathcal{S}_{{\\mathrm{adv}}}$\\;\n {\\bf Model training}: Optimize $\\theta$ using $N_\\mathrm{mini}$ training epochs on \\eqref{eq:loss_training} over $\\mathcal{S}_{{}}$\\;\n $k \\leftarrow k + 1$\\;\n}\n \\caption{STL controller synthesis (offline)}\n\\end{algorithm} π θ 1 Initialization: S = {(z 0,i,e i)} i =1,...,N where (z 0,i,e i) ∼ Uniform[X 0 ×E]; 2 k ← 0; 3 Model training: Optimize θ using N full training epochs on (2 \\begin{algorithm}[t]\n\\SetAlgoLined\n\\KwResult{$\\pi_\\theta$}\n{\\bf Initialization}: $\\mathcal{S}_{{}} = \\lbrace (z_{0,i} , e_i)\\rbrace_{i=1,...,N}$ where $(z_{0,i}, e_i) \\sim \\mathrm{Uniform}[\\mathcal{X}_0 \\times \\mathcal{E}]$\\;\n$k \\leftarrow 0$\\;\n{\\bf Model training}: Optimize $\\theta$ using $N_\\mathrm{full}$ training epochs on \\eqref{eq:loss_training} over $\\mathcal{S}_{{}}$\\;\n \\While{$k < K$}{\n {\\bf Adversarial search}: Find $\\mathcal{S}_{{\\mathrm{adv}}} = \\lbrace (z_{0,j}, e_j) \\mid \\xi_{x_{0,j}, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}} \\not\\models \\phi \\rbrace_{i=1,...,N_\\mathrm{adv}}$ with $\\theta$ fixed (e.g., acceptance-rejection sampling)\\;\n {\\bf Update initial states}: Re-sample from $\\mathrm{Uniform}[\\mathcal{X}_0 \\times \\mathcal{E}]$ to get $\\mathcal{S}_{{0}} = \\lbrace (z_{0,i}, e_i) \\rbrace_{i=1,...,N}$\\;\n $\\mathcal{S}_{{}} \\leftarrow \\mathcal{S}_{{0}} \\cup \\mathcal{S}_{{\\mathrm{adv}}}$\\;\n {\\bf Model training}: Optimize $\\theta$ using $N_\\mathrm{mini}$ training epochs on \\eqref{eq:loss_training} over $\\mathcal{S}_{{}}$\\;\n $k \\leftarrow k + 1$\\;\n}\n \\caption{STL controller synthesis (offline)}\n\\end{algorithm} over S; 4 while k <K do 5 Adversarial search: Find S adv = {(z 0,j,e j) | ξ π θ, P D x 0,j,T ̸| = ϕ} i =1,...,N adv with θ fixed (e.g., acceptance-rejection sampling); 6 Update initial states: Re-sample from Uniform[X 0 ×E] to get S 0 = {(z 0,i,e i)} i =1,...,N; 7 S ←S 0 ∪S adv; 8 Model training: Optimize θ using N mini training epochs on (2 \\begin{algorithm}[t]\n\\SetAlgoLined\n\\KwResult{$\\pi_\\theta$}\n{\\bf Initialization}: $\\mathcal{S}_{{}} = \\lbrace (z_{0,i} , e_i)\\rbrace_{i=1,...,N}$ where $(z_{0,i}, e_i) \\sim \\mathrm{Uniform}[\\mathcal{X}_0 \\times \\mathcal{E}]$\\;\n$k \\leftarrow 0$\\;\n{\\bf Model training}: Optimize $\\theta$ using $N_\\mathrm{full}$ training epochs on \\eqref{eq:loss_training} over $\\mathcal{S}_{{}}$\\;\n \\While{$k < K$}{\n {\\bf Adversarial search}: Find $\\mathcal{S}_{{\\mathrm{adv}}} = \\lbrace (z_{0,j}, e_j) \\mid \\xi_{x_{0,j}, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}} \\not\\models \\phi \\rbrace_{i=1,...,N_\\mathrm{adv}}$ with $\\theta$ fixed (e.g., acceptance-rejection sampling)\\;\n {\\bf Update initial states}: Re-sample from $\\mathrm{Uniform}[\\mathcal{X}_0 \\times \\mathcal{E}]$ to get $\\mathcal{S}_{{0}} = \\lbrace (z_{0,i}, e_i) \\rbrace_{i=1,...,N}$\\;\n $\\mathcal{S}_{{}} \\leftarrow \\mathcal{S}_{{0}} \\cup \\mathcal{S}_{{\\mathrm{adv}}}$\\;\n {\\bf Model training}: Optimize $\\theta$ using $N_\\mathrm{mini}$ training epochs on \\eqref{eq:loss_training} over $\\mathcal{S}_{{}}$\\;\n $k \\leftarrow k + 1$\\;\n}\n \\caption{STL controller synthesis (offline)}\n\\end{algorithm} over S; 9 k ← k +1; 10 end\n\nplex specifications and systems. However, our approach is semi -supervised since the demonstrations are used for reg- ularization instead of the main training objective. Therefore we do not require a significant amount of demonstrations compared to fully supervised approaches (e.g., [10]), and favorably so if data collection is expensive or demonstrations are scarce.\n\nWe apply stochastic gradient descent on $\\theta$ to minimize the following loss objective,\n\n$$\n\\begin{aligned}\n \\mathcal{L}_\\mathrm{train}(\\theta;\\mathcal{S}, \\Xi^\\mathrm{exp}) = \\mathcal{L}_\\mathrm{STL}(\\theta; \\mathcal{S}) + \\gamma \\mathcal{L}_\\mathrm{imit}(\\theta; \\Xi^\\mathrm{exp}), \\\\\n \\mathcal{L}_\\mathrm{STL}(\\theta; \\mathcal{S}) = \\frac{1}{|\\mathcal{S}_{{}}|}\\sum_{(z_{0}, e) \\in \\mathcal{S}}\\mathrm{LeakyReLU}(-\\rho_\\phi(\\xi_{z_{0}, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}})), \\\\\n \\mathcal{L}_\\mathrm{imit}(\\theta; \\Xi^\\mathrm{exp}) = \\frac{1}{N_\\mathrm{exp}} \\sum_{i = 1}^{N_\\mathrm{exp}}\\frac{1}{T_i} \\Delta(\\xi_{z_{0, i}^\\mathrm{exp},T_i}^\\mathrm{exp},\\xi_{z_{0, i}^\\mathrm{exp},T_i}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}}; e_i^\\mathrm{exp}) \n\\end{aligned}\n$$\n\n \n\n$$\n\\begin{aligned}\n \\mathcal{L}_\\mathrm{train}(\\theta;\\mathcal{S}, \\Xi^\\mathrm{exp}) = \\mathcal{L}_\\mathrm{STL}(\\theta; \\mathcal{S}) + \\gamma \\mathcal{L}_\\mathrm{imit}(\\theta; \\Xi^\\mathrm{exp}), \\\\\n \\mathcal{L}_\\mathrm{STL}(\\theta; \\mathcal{S}) = \\frac{1}{|\\mathcal{S}_{{}}|}\\sum_{(z_{0}, e) \\in \\mathcal{S}}\\mathrm{LeakyReLU}(-\\rho_\\phi(\\xi_{z_{0}, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}})), \\\\\n \\mathcal{L}_\\mathrm{imit}(\\theta; \\Xi^\\mathrm{exp}) = \\frac{1}{N_\\mathrm{exp}} \\sum_{i = 1}^{N_\\mathrm{exp}}\\frac{1}{T_i} \\Delta(\\xi_{z_{0, i}^\\mathrm{exp},T_i}^\\mathrm{exp},\\xi_{z_{0, i}^\\mathrm{exp},T_i}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}}; e_i^\\mathrm{exp}) \n\\end{aligned}\n$$\n\n \n\n$$\n\\begin{aligned}\n \\mathcal{L}_\\mathrm{train}(\\theta;\\mathcal{S}, \\Xi^\\mathrm{exp}) = \\mathcal{L}_\\mathrm{STL}(\\theta; \\mathcal{S}) + \\gamma \\mathcal{L}_\\mathrm{imit}(\\theta; \\Xi^\\mathrm{exp}), \\\\\n \\mathcal{L}_\\mathrm{STL}(\\theta; \\mathcal{S}) = \\frac{1}{|\\mathcal{S}_{{}}|}\\sum_{(z_{0}, e) \\in \\mathcal{S}}\\mathrm{LeakyReLU}(-\\rho_\\phi(\\xi_{z_{0}, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}})), \\\\\n \\mathcal{L}_\\mathrm{imit}(\\theta; \\Xi^\\mathrm{exp}) = \\frac{1}{N_\\mathrm{exp}} \\sum_{i = 1}^{N_\\mathrm{exp}}\\frac{1}{T_i} \\Delta(\\xi_{z_{0, i}^\\mathrm{exp},T_i}^\\mathrm{exp},\\xi_{z_{0, i}^\\mathrm{exp},T_i}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}}; e_i^\\mathrm{exp}) \n\\end{aligned}\n$$\n\nwhere $\\mathrm{LeakyReLU}(x)=\\max(0.01x,x)$ and $\\Delta(\\xi_a, \\xi_b; e) = \\mathrm{MSE}(z_{a}, z_{b}) + \\gamma_\\mathrm{imit} \\mathrm{MSE}(u_{a}, u_{b})$ is a weighted sum of the mean-square-error in state and controls between trajecto- ries $\\xi_a$ and $\\xi_b$ under environment $e$. Note that (3) differs from maximizing robustness as done in [7], [8], [9]. The LeakyReLU function focuses primarily on minimizing the amount of violation and focuses less on increasing the amount of satisfaction (by a factor of 0.01). The purpose of (4) is to help regularize the training process since (3) is non- linear and non-convex. Simultaneously, (4) helps guide the exploration towards regions where the controller produces trajectories consistent with how humans would behave and therefore avoid superfluous or unreasonable trajectories (e.g., taking unnecessary detours but still satisfy $\\phi$).\n\n## 2) Adversarial search (Line 5 in Algorithm 1)\n\nThe training step is optimized using samples of initial states and environments. As such, there could still exist initial states and environments that lead to negative robustness. The goal of the adversarial search is to find a set of initial states and environments $\\mathcal{S}^\\mathrm{adv} = \\lbrace (z_{0,i}^\\mathrm{adv}, e_i^\\mathrm{adv})\\rbrace_{i=1,...,N_\\mathrm{adv}}$ such that the resulting trajectories violate $\\phi$. A number of methods can be used to search for adversarial samples, such as batched (projected) gradient descent on $z_{0}$ to minimize robustness, cross-entropy method, or simulated annealing [29]. Since we strive to find any samples that produce\n\n\\begin{algorithm}[t]\n\\SetAlgoLined\n\\KwResult{$\\tau_T$}\n\\textbf{Initialization:}STL specification $\\phi$, initial state $z_{0}$, environment $e$, time horizon $T$, neural network weights $\\theta_0$, step size $\\eta$, and maximum number of gradient steps $N_\\mathrm{gd}$\\;\n$\\theta \\leftarrow [\\tilde{\\theta}, \\hat{\\theta}]$ where $\\tilde{\\theta}$ are neural network parameters of $g_\\mathrm{MLP}^{n_h \\rightarrow m}$, and $\\hat{\\theta}$ are the remaining neural network parameters (held fixed)\\;\n$\\tau_t \\leftarrow z_{0}$\\;\n\n \\For{$t = 0:T$}{\n\n \\If{$\\mathbb{E}[\\rho(\\xi_{\\tau_t, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}}, \\phi) < 0$}{\n $j \\leftarrow 0$\\;\n \\While{$(\\mathbb{E}[\\rho(\\xi_{\\tau_t, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}}, \\phi) < 0) \\wedge (j < N_\\mathrm{gd})$}{\n $\\tilde{\\theta} \\leftarrow \\tilde{\\theta} + \\eta\\nabla_{\\tilde{\\theta}} \\mathbb{E}_{d\\sim \\mathcal{P}_\\mathcal{D}}[\\rho(\\xi_{\\tau_t, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}}, \\phi)$\\;\n $\\theta \\leftarrow [\\tilde{\\theta}, \\hat{\\theta}]$\\;\n $j \\leftarrow j + 1$\\;\n }\n }\n $u_{t} = \\pi_\\theta(\\tau_t; e)$\\;\n $z_{t+1} = f (z_{t}, u_{t}, d_{t})$ where $d_{t} \\sim \\mathcal{P}_\\mathcal{D}$\\;\n $\\tau_{t} \\leftarrow z_{0:t+1}$\\;\n}\n \\caption{Deploying $\\pi_\\theta$ (online)}\n\\end{algorithm} 2:\n\nDeploying $\\pi_\\theta$ (online)\n\n\\begin{algorithm}[t]\n\\SetAlgoLined\n\\KwResult{$\\tau_T$}\n\\textbf{Initialization:}STL specification $\\phi$, initial state $z_{0}$, environment $e$, time horizon $T$, neural network weights $\\theta_0$, step size $\\eta$, and maximum number of gradient steps $N_\\mathrm{gd}$\\;\n$\\theta \\leftarrow [\\tilde{\\theta}, \\hat{\\theta}]$ where $\\tilde{\\theta}$ are neural network parameters of $g_\\mathrm{MLP}^{n_h \\rightarrow m}$, and $\\hat{\\theta}$ are the remaining neural network parameters (held fixed)\\;\n$\\tau_t \\leftarrow z_{0}$\\;\n\n \\For{$t = 0:T$}{\n\n \\If{$\\mathbb{E}[\\rho(\\xi_{\\tau_t, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}}, \\phi) < 0$}{\n $j \\leftarrow 0$\\;\n \\While{$(\\mathbb{E}[\\rho(\\xi_{\\tau_t, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}}, \\phi) < 0) \\wedge (j < N_\\mathrm{gd})$}{\n $\\tilde{\\theta} \\leftarrow \\tilde{\\theta} + \\eta\\nabla_{\\tilde{\\theta}} \\mathbb{E}_{d\\sim \\mathcal{P}_\\mathcal{D}}[\\rho(\\xi_{\\tau_t, T}^{\\pi_\\theta, \\mathcal{P}_\\mathcal{D}}, \\phi)$\\;\n $\\theta \\leftarrow [\\tilde{\\theta}, \\hat{\\theta}]$\\;\n $j \\leftarrow j + 1$\\;\n }\n }\n $u_{t} = \\pi_\\theta(\\tau_t; e)$\\;\n $z_{t+1} = f (z_{t}, u_{t}, d_{t})$ where $d_{t} \\sim \\mathcal{P}_\\mathcal{D}$\\;\n $\\tau_{t} \\leftarrow z_{0:t+1}$\\;\n}\n \\caption{Deploying $\\pi_\\theta$ (online)}\n\\end{algorithm} τ T 1 Initialization: STL specification ϕ, initial state z 0, environment e, time horizon T, neural network weights θ 0, step size η, and maximum number of gradient steps N gd; 2 θ ← [˜ θ, ˆ θ] where ˜ θ are neural network parameters of g n h → m MLP, and ˆ θ are the remaining neural network parameters (held fixed); 3 τ t ← z 0; 4 for t =0: T do 5 if E [ρ (ξ π θ, P D τ t,T,ϕ) < 0 then 6 j ← 0; 7 while (E [ρ (ξ π θ, P D τ t,T,ϕ) < 0) ∧ (j <N gd) do 8 ˜ θ ← ˜ θ + η ∇ ˜ θ E d ∼P D [ρ (ξ π θ, P D τ t,T,ϕ); 9 θ ← [˜ θ, ˆ θ]; 10 j ← j +1; 11 end 12 end 13 u t = π θ (τ t; e); 14 z t +1 = f (z t,u t,d t) where d t ∼P D; 15 τ t ← z 0: t +1; 16 end\n\na violating trajectory, we opt for a simpler approach of acceptance-rejection sampling whereby we uniformly sample from $\\mathcal{X}_0 \\times \\mathcal{E}$ and reject any samples that produce satisfying trajectories. We continue sampling until $N_\\mathrm{adv}$ adversarial samples are found or a termination criterion is met. Using the adversarial samples, and newly sampled initial states and environments, we can continue training the model (lines 6 – 8 in Algorithm 1).\n\n## D. Deploying the Controller Online\n\nUnfortunately, after running Algorithm 1, there are no guar- antees that the resulting controller will produce satisfying trajectories for all initial states and environments, especially under the presence of disturbances to the system. There is a lot of effort towards neural network verification, though ver- ifying RNNs especially with temporal logic considerations remains challenging.\n\nTo address the limitations of the LSTM controller, Algo- rithm 2 proposes updating some of the controller parameters online whenever the controller is expected to produce a violating trajectory. That is, we perform gradient steps on $\\tilde{\\theta}$, the parameters of $g_\\mathrm{MLP}^{n_h \\rightarrow m}$, when needed. At each time step, we perform a Monte Carlo estimate of the expected robustness value (line 5). If the expected robustness is negative, then we perform at most $N_\\mathrm{gd}$ gradient descent steps on $\\tilde{\\theta}$ to increase the robustness value (lines 7 – 10). If desired, a different risk metric (e.g., Value at Risk) could be used instead. Then the resulting control is passed into the system and a step is taken forward in time (lines 13 – 15).\n\n## VI. E XPERIMENTS\n\nWe investigate the offline and online performance of our proposed controller applied to a nonlinear system.\n\n## A. Case-Study Set-Up\n\nWe investigate a car-like robot where the discrete-time dy- namics are given by applying a zero-order hold on controls", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Let $\\xi : \\mathfrak{g}\\rightarrow\\bigwedge^2\\mathfrak{g}$ be a cocyle on $\\mathfrak{g}$. The coefficients of $\\xi(s_k)$, $\\xi(z_k)$, $\\xi(d_{2k-1})$, $\\xi(d_{2k})$ are denoted respectively $a_{ij}^k,a_{ij}^{((k))},a_{ij}^{(2k-1)},a_{ij}^{(2k)}\\ldots$ and so on, similarly for a bivector $r\\in\\bigwedge^2\\mathfrak{g}$, namely\n\n\\begin{align*} \\xi(s_k)=&\\sum_{1\\leqslant i<j\\leqslant k_0}a_{ij}^k\\,s_i\\wedge s_j+\\sum_{\\substack{1\\leqslant i\\leqslant k_0\\\n 1\\leqslant j\\leqslant\\ell_0}}b_{ij}^k\\,s_i\\wedge z_j+\\sum_{\\substack{1\\leqslant i\\leqslant k_0\\\n 1\\leqslant j\\leqslant m}}c_{ij}^k\\,s_i\\wedge d_{2j-1}+\\sum_{\\substack{1\\leqslant i\\leqslant k_0\\\n 1\\leqslant j\\leqslant m}}e_{ij}^k\\,s_i\\wedge d_{2j}\\\\\n&+\\sum_{1\\leqslant i<j\\leqslant\\ell_0}f_{ij}^k\\,z_i\\wedge z_j+\\sum_{\\substack{1\\leqslant i\\leqslant\\ell_0\\\n 1\\leqslant j\\leqslant m}}g_{ij}^k\\,z_i\\wedge d_{2j-1}+\\sum_{\\substack{1\\leqslant i\\leqslant\\ell_0\\\n 1\\leqslant j\\leqslant m}}h_{ij}^k\\,z_i\\wedge d_{2j}\\\\\n&+\\sum_{1\\leqslant i<j\\leqslant m}m_{ij}^k\\,d_{2i-1}\\wedge d_{2j-1}+\\sum_{1\\leqslant i,j\\leqslant m}n_{ij}^k\\,d_{2i-1}\\wedge d_{2j}+\\sum_{1\\leqslant i<j\\leqslant m}p_{ij}^k\\,d_{2i}\\wedge d_{2j}. \\end{align*}\n\n## 3.2. The set of invariants\n\nLet $\\mathfrak{g}$ a flat Lie algebra. The invariants of $\\bigwedge^2\\mathfrak{g}$ are defined as follows\n\n\\begin{equation*} \\left(\\bigwedge^2\\mathfrak{g}\\right)^{\\mathfrak{g}}=\\left\\{r\\in\\bigwedge^2\\mathfrak{g},\\ \\operatorname{ad}_xr=0\\ \\text{for all}\\ x\\in\\mathfrak{g}\\right\\}. \\end{equation*}\n\nLemma 3.1. For a flat Lie algebra $\\mathfrak{g}$ we have\n\n\\begin{equation*} \\left(\\bigwedge^2\\mathfrak{g}\\right)^{\\mathfrak{g}}=\\left\\{\\begin{array}{cc} \\left(\\bigwedge^2\\mathfrak{z}\\right)\\oplus\\left(\\bigwedge^2P_i\\right) & \\text{if}\\ \\mathfrak{g}\\ \\text{is nondegenerate},\\\\\n\\left(\\bigwedge^2\\mathfrak{z}\\right)\\oplus\\left(\\bigwedge^2P_i\\right)\\oplus\\left(\\underset{i<j}{\\oplus}A_{ij}\\oplus B_{ij}\\right) & \\text{if}\\ \\mathfrak{g}\\ \\text{is degenerate}, \\end{array}\\right. \\end{equation*}\n\nwhere $P_i=\\spann\\{d_{2i-1},d_{2i}\\}$ and the last sum stands for the couples $i<j$ such that $L_j=\\varepsilon L_i$ with $A_{ij}=\\mathbb{R}(d_{2i-1}\\wedge d_{2j-1}+\\varepsilon d_{2i}\\wedge d_{2j})$ and $B_{ij}=\\mathbb{R}(d_{2i-1}\\wedge d_{2j}+\\varepsilon d_{2j-1}\\wedge d_{2i})$.\n\nProof. Following our notations, for $r\\in\\mathfrak{g}$ we have $\\operatorname{ad}_{s_\\ell}r=0$ for all $\\ell=1,\\ldots,k_0$ if and only if\n\n\\begin{equation*}c_{ij}=e_{ij}=g_{ij}=h_{ij}=0\\end{equation*}\n\nand\n\n\\begin{equation*} \\begin{pmatrix} -\\lambda_{\\ell j}&\\phantom{-}\\lambda_{\\ell i}\\\\\n\\phantom{-}\\lambda_{\\ell i}&-\\lambda_{\\ell j} \\end{pmatrix}\\begin{pmatrix} n_{ij}\\\\\nn_{ji} \\end{pmatrix}= \\begin{pmatrix} 0\\\\\n0 \\end{pmatrix}=\\begin{pmatrix} \\phantom{-}\\lambda_{\\ell j}&-\\lambda_{\\ell i}\\\\\n-\\lambda_{\\ell i}&\\phantom{-}\\lambda_{\\ell j} \\end{pmatrix} \\begin{pmatrix} m_{ij}\\\\\np_{ij} \\end{pmatrix}. \\end{equation*}\n\nIf $\\mathfrak{g}$ is nondegenerate, there is $\\ell\\in\\{1,\\ldots,k_0\\}$ such that $\\lambda_{\\ell i}^2-\\lambda_{\\ell j}^2\\neq0$, then\n\n\\begin{equation*}m_{ij}=p_{ij}=n_{ij}=n_{ji}=0.\\end{equation*}\n\nIf $\\mathfrak{g}$ is degenerate and for all $\\ell\\in\\{1,\\ldots,k_0\\}$\n\n$\\lambda_{\\ell i}^2-\\lambda_{\\ell j}^2=0$ then\n\n$\\lambda_{\\ell j}=\\varepsilon_{ij}\\lambda_{\\ell i}$, ($\\varepsilon_{ij}=\\pm1$) and\n\n\\begin{equation*}p_{ij}=\\varepsilon_{ij}m_{ij},\\ n_{ji}=\\varepsilon_{ij}n_{ij}.\\end{equation*}\n\nLet $\\{s_1^*,\\ldots,s_{k_0}^*,z_1^*,\\ldots,z_{\\ell_0}^*,d_1^*,\\ldots,d_{2m}^*\\}$ be the dual basis of the dual vector space $\\mathfrak{g}^*$. We have\n\n\\begin{equation*}s_i^*\\wedge d_{2\\ell}^*\\left(\\operatorname{ad}_{d_{2\\ell-1}}r\\right)=0=z_i^*\\wedge d_{2\\ell}^*\\left(\\operatorname{ad}_{d_{2\\ell-1}}r\\right),\\end{equation*}\n\n", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Fig.5. Convergence curves of JPLTD on ORL, NGs, BBCSport, and RGB-D with missing rate $p$ = 0.7.\n\nA BLATION S TUDY R ESULTS ON F IVE DIFFERENT DATASETS.\n\n\\begin{tabular}{|p{0.1cm}|p{1.5cm}|p{0.87cm} p{0.87cm} p{1.0cm} | p{0.87cm} p{0.87cm} p{1.0cm} |p{0.87cm}p{0.87cm} p{1.0cm}|p{0.87cm}p{0.87cm} p{1.0cm}|}\n%\n\\hline\n& Missing rates&\\multicolumn{3}{c}{0.1}\\vline & \\multicolumn{3}{c}{0.3}\\vline & \\multicolumn{3}{c}{0.5}\\vline& \\multicolumn{3}{c}{0.7}\\vline \\hline\n& Metrics & ACC~(\\%) & NMI~(\\%) & ARI~(\\%) & ACC~(\\%) & NMI~(\\%) & ARI~(\\%) & ACC~(\\%) & NMI~(\\%) & ARI~(\\%) & ACC~(\\%) & NMI~(\\%) & ARI~(\\%) %\t\t\\midrule\n\\hline\n\\multirow{4}{*}{\\rotatebox{90}{\\textbf{ORL}}}\n&JPLTD-N\t&96.76±1.31\t&99.34±0.27\t&96.78±1.29\t&96.26±1.30\t&99.04±0.29\t&96.00±1.27\t&94.90±2.10\t&98.38±0.46\t&93.99±1.85\t&93.88±1.79\t&97.88±0.64\t&92.39±2.16&JPLTD-B\t&95.91±1.40\t&98.89±0.41\t&95.46±1.40\t&95.28±2.34\t&98.79±0.57\t&94.99±2.33\t&92.29±1.98\t&97.59±0.59\t&91.06±2.03\t&87.81±2.74\t&94.66±1.21\t&84.02±3.26&JPLTD-O &93.65±1.95\t&98.57±0.46\t&93.50±1.83\t&92.64±2.35\t&98.00±0.56\t&91.91±2.29\t&90.15±2.45\t&96.84±0.58\t&88.50±2.24\t&87.34±2.76\t&94.40±1.20\t&83.51±3.32&\\textbf{JPLTD}\t&\\textbf{97.65±1.72}&\\textbf{99.47±0.38}&\\textbf{97.50±1.73}&\\textbf{97.51±1.50} &\\textbf{99.43±0.32}&\\textbf{97.46±1.51}&\\textbf{96.36±1.33}&\\textbf{98.68±0.32}&\\textbf{95.29±1.36}&\\textbf{95.66±1.72}&\\textbf{98.62±0.45}&\\textbf{95.03±1.76}\\hline\n\\multirow{4}{*}{\\rotatebox{90}{\\textbf{NGs}}}\n&JPLTD-N\t&99.60±0.00\t&98.61±0.00\t&99.00±0.00\t&99.60±0.00\t&98.61±0.00\t&99.00±0.00\t&99.60±0.00\t&98.61±0.00\t&99.00±0.00\t&99.20±0.00\t&97.39±0.00\t&98.00±0.00&JPLTD-B\t&98.60±0.00\t&95.30±0.00\t&96.52±0.00\t&98.60±0.00\t&95.30±0.00\t&96.52±0.00\t&98.00±0.00\t&93.62±0.00\t&95.02±0.00\t&95.00±0.00\t&85.75±0.00\t&87.88±0.00&JPLTD-O\t&98.00±0.00 &93.80±0.00\t&95.06±0.00\t&98.00±0.00\t&94.38±0.00\t&95.04±0.00\t&97.20±0.00\t&91.26±0.00\t&93.10±0.00\t&93.80±0.00\t&81.98±0.00\t&85.14±0.00&\\textbf{JPLTD}\t&\\textbf{100.0±0.00}\t&\\textbf{100.0±0.00}\t&\\textbf{100.0±0.00}\t&\\textbf{99.80±0.00}\t&\\textbf{99.30±0.00}\t&\\textbf{99.50±0.00}\t&\\textbf{99.60±0.00}\t&\\textbf{98.61±0.00}\t&\\textbf{98.99±0.00}\t&\\textbf{99.60±0.00}\t&\\textbf{98.61±0.00}\t&\\textbf{98.99±0.00}\\hline\n\\multirow{4}{*}{\\rotatebox{90}{\\textbf{MSRC-v1}}}\n&JPLTD-N\t&95.12±0.21\t&89.77±0.34\t&88.80±0.51\t&93.67±0.22\t&87.76±0.51\t&85.94±0.55\t&92.38±0.29\t&84.31±0.18\t&82.58±0.29\t&87.14±0.32\t&79.89±0.30\t&74.81±0.64&JPLTD-B\t&93.83±0.11\t&88.56±0.13\t&86.07±0.22\t&91.88±0.11\t&85.83±0.19\t&82.46±0.20\t&87.62±0.48\t&83.72±0.61\t&77.81±0.83\t&83.76±0.17\t&75.66±0.39\t&68.77±0.40&JPLTD-O\t&86.19±0.29\t&79.64±0.24\t&74.63±0.14\t&82.69±0.32\t&75.10±0.51\t&67.27±0.62\t&76.45±0.89\t&71.27±0.46\t&63.76±0.23\t&70.76±0.47\t&64.09±0.56\t&52.96±0.59&\\textbf{JPLTD}\t&\\textbf{100.0±0.00} &\\textbf{100.0±0.00}\t&\\textbf{100.0±0.00}\t&\\textbf{99.52±0.00}\t&\\textbf{98.92±0.00}\t&\\textbf{98.88±0.00}\t&\\textbf{99.28±0.24}\t&\\textbf{98.38±0.54}\t&\\textbf{98.31±0.57}\t&\\textbf{99.27±0.24}\t&\\textbf{98.36±0.54}\t&\\textbf{98.30±0.56}\\hline\n\\multirow{4}{*}{\\rotatebox{90}{\\textbf{BBCSport}}}\n&JPLTD-N\t&98.90±0.00\t&96.28±0.00\t&96.86±0.00\t&98.90±0.00\t&96.20±0.00\t&97.07±0.00\t&98.34±0.00\t&94.52±0.00\t&95.60±0.00\t&97.24±0.00\t&91.71±0.00\t&93.63±0.00&JPLTD-B\t&97.98±0.00\t&93.28±0.00\t&94.65±0.00\t&97.61±0.00\t&92.93±0.00\t&95.23±0.00\t&96.14±0.00\t&93.12±0.00\t&95.23±0.00\t&95.22±0.00\t&90.26±0.00\t&92.35±0.00&JPLTD-O &96.51±0.00\t&92.70±0.00\t&94.80±0.00\t&96.14±0.00\t&87.63±0.00\t&89.62±0.00\t&95.58±0.00\t&87.59±0.00\t&89.84±0.00\t&93.58±0.00\t&88.07±0.00\t&89.31±0.00&\\textbf{JPLTD}\t&\\textbf{99.91±0.09} &\\textbf{99.67±0.34}\t&\\textbf{99.83±0.17}\t&\\textbf{99.63±0.19}\t&\\textbf{98.69±0.70}\t&\\textbf{99.32±0.45}\t&\\textbf{99.45±0.19}\t&\\textbf{98.13±0.57}\t&\\textbf{98.76±0.55}\t&\\textbf{98.90±0.37}\t&\\textbf{96.50±0.92}\t&\\textbf{97.30±0.96}\\hline\n\\multirow{4}{*}{\\rotatebox{90}{\\textbf{RGB-D}}}\n&JPLTD-N\t&63.19±0.12\t&60.54±0.17\t&47.35±0.15\t&60.98±0.14\t&58.65±0.13\t&44.96±0.15\t&58.44±0.23\t&52.42±0.16\t&40.08±0.25 &54.45±0.73\t&51.52±0.25\t&36.70±0.66&JPLTD-B\t&61.02±0.43\t&60.09±0.35\t&45.38±0.36\t&57.95±0.13\t&54.66±0.14\t&41.00±0.19\t&53.25±0.59\t&50.44±0.31\t&36.10±0.30\t&50.22±0.26\t&41.83±0.37\t&32.19±0.39&JPLTD-O &58.57±0.12\t&60.58±0.16\t&44.98±0.16\t&56.39±0.40 &53.13±0.26\t&40.67±0.17\t&52.38±0.62\t&49.80±0.43\t&36.95±0.27\t&49.41±0.32\t&41.83±0.32\t&31.97±0.40&\\textbf{JPLTD}\t&\\textbf{64.73±0.04}\t&\\textbf{62.26±0.10}\t&\\textbf{50.10±0.10}\t&\\textbf{62.52±0.08}\t&\\textbf{59.65±0.13}\t&\\textbf{46.16±0.12}\t&\\textbf{62.29±0.06}\t&\\textbf{59.25±0.08}\t&\\textbf{45.45±0.08}\t&\\textbf{61.70±0.33}\t&\\textbf{55.23±0.18}\t&\\textbf{43.72±0.19}%\t\t\\bottomrule\n\\hline\n\\end{tabular}\n\nFig.6. An example to explain the recovered graphs of JPLTD-B and JPLTD on ORL dataset with missing rate $p$ = 0.5. Figure (a) is the recovered graph of JPLTD-B. Figure (b) is the learned sparse noise graph of JPLTD, and (c) is the learned intrinsic graph of JPLTD.\n\nvalues in each iteration on ORL, NGs, BBCSport, and RGB- D datasets with missing rate $p$ = 0.7, where the error value under each experiment is normalized by dividing the maximum value. We can observe that all the residuals can converge to 0 in less than 80 iterations, indicating that our algorithm has good convergence properties.\n\n## F. Ablation Study\n\nJPLTD projects the original data into a lower-dimensional space for graph learning, and it decomposes the initial tensor into an intrinsic tensor and a sparse tensor to better depict the similarities of incomplete data. Herein, we conduct an ablation study to investigate the effectiveness of projection learning and tensor decomposition in JPLTD. We derive three variants based on the original JPLTD model, i.e., JPLTD-N, JPLTD-B, and JPLTD-O. JPLTD-N gives up the dimension reduction of the original data, its loss function is:\n\n$$\n\\begin{aligned}\n&\\min\\limits_{\\Lambda}\\ \\sum_{v=1}^{m}\\|\\textbf{E}^{v}\\|_{2,1} + \\lambda \\|\\mathcal{B}\\|_{\\circledast}+ \\theta\\|\\mathcal{P}\\|_{1}\\\\\n&s.t.\\ \\textbf{E}^{v} = \\textbf{X}_{o}^{v} - \\textbf{X}_{o}^{v}\\textbf{A}^{v}\\textbf{G}^{v}\\textbf{A}^{v^{T}},\\ \\mathcal{G} = \\mathcal{B}+\\mathcal{P},\\ \\\\\n &\\mathcal{G} = \\Psi(\\textbf{G}^{1},\\ \\textbf{G}^{2},\\cdots,\\ \\textbf{G}^{m}),\n\\end{aligned}\n$$\n\nJPLTD-B drops the sparse tensor term and directly utilizes the original tensor $\\mathcal{G}$ for clustering, its loss function can be described as:\n\n$$\n\\begin{aligned}\n&\\min\\limits_{\\Lambda}\\ \\sum_{v=1}^{m}\\|\\textbf{E}^{v}\\|_{2,1} + \\lambda \\|\\mathcal{G}\\|_{\\circledast}\\\\\ns.t.\\ &\\textbf{E}^{v} = \\textbf{W}^{v}\\textbf{X}_{o}^{v} - \\textbf{W}^{v}\\textbf{X}_{o}^{v}\\textbf{A}^{v}\\textbf{G}^{v}\\textbf{A}^{v^{T}},\\ \\textbf{W}^{v}\\textbf{W}^{v^{T}} = \\textbf{I}_{k},\\\\ \n&\\mathcal{G} = \\Psi(\\textbf{G}^{1},\\ \\textbf{G}^{2},\\cdots,\\ \\textbf{G}^{m}),\n\\end{aligned}\n$$\n\nJPLTD-O considers neither the dimension reduction of the", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "private, where\n\n\\begin{gather*} % \\varepsilon_A = \\varepsilon_{1:T} \\sqrt{2T \\log \\frac{N}{\\delta^\\prime}} + \\mathcal{O}\\left(\\frac{1}{N^2} \\right), \\\\\n % \\delta_A = \\frac{\\delta'}{N}, \\end{gather*}\n\nwhere\n\n\\begin{gather*} % \\varepsilon_{1:T} = \\frac{2 L^\\mathrm{ERM}_{1:T}}{N b} I_{1:T}, \\end{gather*}\n\nand $L^\\mathrm{ERM}_{1:T} := \\left(\\frac{1}{T} \\sum_{t=1}^T \\left(L^\\mathrm{ERM}_t\\right)^4\\right)^{\\frac{1}{4}}$, $\\delta'$ is a positive real, $I_{1:T}$ is the robustified intensity for the whole training procedure (see Definition 4.2), and $b$ is the Laplacian parameter.\n\nRemark 5.1. Theorems 5.1 and 5.2 suggest a negative correlation between adversarial robustness and privacy preservation.\n\nRemark5.2. Theorem 5.1 approximate the differential privacy of a learning algorithm based on the adver- sarial robustness of its every iteration ($I_1, \\ldots, I_t$), while Theorem 5.2 approximate the differential privacy based on the adversarial robustness of the whole algorithm ($I_{1:T}$).\n\nRemark5.3. The approximation of differential privacy given by Theorem is $(\\mathcal O (\\sqrt{\\log N}/N),\\mathcal O (1/N))$.\n\nSimilarly, we can approximate the differential privacy of ERM in the following corollary.\n\nCorollary 5.1. Suppose one employs SGD for ERM and the whole training procedure has $T$ iterations. Then, the ERM is $(\\varepsilon, \\delta)$ -differentially private, where\n\n\\begin{gather*} \\varepsilon = \\varepsilon^\\mathrm{ERM}_{1:T} \\sqrt{2 T \\log \\frac{N}{\\delta'}} + \\mathcal{O}\\left(\\frac{1}{N^2} \\right),\\\\\n % \\delta = \\frac{\\delta'}{N}, \\end{gather*}\n\nin which,\n\n\\begin{equation*} \\varepsilon^\\mathrm{ERM}_{1:T} = \\frac{2 L^\\mathrm{ERM}_{1:T}}{N b}, \\end{equation*}\n\nand $L^\\mathrm{ERM}_{1:T} := \\left(\\frac{1}{T} \\sum_{t=1}^T \\left(L^\\mathrm{ERM}_t\\right)^4\\right)^{\\frac{1}{4}}$, $\\delta^\\prime$ is a positive real.\n\nComparing the results for adversarial training and ERM, we have\n\n\\begin{equation*} \\varepsilon_{1:T} = I_{1:T} \\cdot \\varepsilon^{\\text{ERM}}_{1:T} + \\mathcal{O}(1/N^2). \\end{equation*}\n\nTheorem 5.2 and Corollary 5.1 show that both factors $\\varepsilon$ and $\\delta$ have positive correlations with the robustified intensity, which suggests a trade-off between privacy preservation and adversarial robustness.\n\n## Proof skeleton\n\nThis section presents the proof skeleton. Detailed proofs for Theorems 5.1 and 5.2 are given in Sections 7.2 and 7.3, respectively.\n\nIt is usually hard to calculate the differential privacy of iterative algorithm directly. However, the differential", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "and global probability density, thus considering both local and global features to accurately detect outliers and cluster anomalies.\n\nNext, we present the experimental results of the EDROD method and its competing counterparts on the ten-dimensional synthetic 300 SAMPLE and 700 SAMPLE datasets. We have to mention here that the value of kernel width in the Gaussian kernel function for kernel density estimation can affect the anomaly-detection performance of the EDROD method. In our studies, by using the grid search method, the optimal value of kernel width on the ten-dimensional synthetic datasets studied here is around 0.36. Therefore, we set the kernel width $h=0.36$ in the EDROD method for studying the synthetic datasets here.\n\nWe first compare the performance of six KNN-based meth- ods, which include DCROD, EDROD, KNN, LOF, COF, and ABOD. We vary the value of $K$ from 4 to 140. Fig. 8 shows the relationship between average AUC values and the parameter $K$ for all competing algorithms on these two synthetic datasets. It should be noted that each average AUC value is obtained over the results of all 10 independent dataset instances. It can be observed from both Fig. 8 (a) and Fig. 8 (b) that, as the value of $K$ changes, variations of resulting AUC values present different characteristics for all competing algo- rithms. KNN and LOF algorithms exhibit the most significant changes as the value of $K$ varies. Specifically, when $K$ is set to 140, the AUC value drops to below 0.7 (on the 300 SAMPLE dataset) and around 0.8 (on the 700 SAMPLE dataset) when using KNN and LOF algorithms. Thus, the effectiveness of both KNN and LOF algorithms is highly dependent on the value of $K$. With a small sample size in the dataset, increasing the value of $K$ gradually leads to the estimation of both normal and anomalous samples together during the process of local kernel density estimation. As a result, the detection accuracy decreases, leading to a noticeable decline in AUC values. Besides, after comparing Fig. 8 (a) with Fig. 8 (b), we have also observed that the size of the dataset also has an impact on the accuracy of anomaly detection for LOF, KNN, and DCROD methods. It is evident that the AUC values of LOF, KNN, and DCROD algorithms on the 300 SAMPLE dataset are lower than those on the 700 SAMPLE dataset. This result indicates that the LOF, KNN, and DCROD algorithms are also sensitive to the change in dataset sizes. In contrast, EDROD exhibits high stability in terms of AUC values when varying the value of $K$. As $K$ increases from 4 to 140, the AUC values only fluctuate within an order of magnitude of $10^{-4}$. Moreover, EDROD consistently maintains higher AUC values than all other competing algorithms on both 300 SAMPLE and 700 SAMPLE datasets. The AUC values remain to be close to 1 when varying the value of $K$ on both 300 SAMPLE and 700 SAMPLE datasets, which indicates that the EDROD algorithms are also robust to the change of dataset sizes.\n\nOn the synthetic 300 SAMPLE and 700 SAMPLE datasets, we have also conducted anomaly detection using various non-KNN-based algorithms, which include OCSVM, COPOD, KDE, PCA, IForest, ECOD, and LUNAR. To compare the per- formance of both KNN-based and non-KNN-based methods, we demonstrate the AUC values of all competing algorithms in\n\nAverage AUC values (over 10 different dataset instances) versus the parameter $K$ on the synthetic 300 SAMPLE (a) and 700 SAMPLE (b) respectively.\n\nBoxplot of all KNN-based and non-KNN-based methods over all 10 dataset realizations for 300 SAMPLE (a) and 700 SAMPLE (b) datasets respectively.\n\nTABLE V (on the 300 SAMPLE dataset) and TABLE VI (on the 700 SAMPLE dataset). It should be noted that each AUC value of the KNN-based algorithm in TABLE V and VI is the average over the results with the parameter $K$ ranging from 4 to 140. It can be observed from TABLE V and TABLE VI that the EDROD method can yield the highest average AUC values over all 10 dataset instances.\n\nThen, based on TABLE V and TABLE VI, we draw the boxplot of AUC values over the 10 datasets for all compet- ing algorithms, which is demonstrated in Fig. 9 (a) (on the 300 SAMPLE dataset) and Fig. 9 (b) (on the 700 SAMPLE dataset) respectively. It can be found that the EDROD method can achieve the highest median AUC values. The algorithm that comes closest to the performance of the EDROD al- gorithm on the 300 SAMPLE dataset is LUNAR. However, the LUNAR algorithm lags behind EDROD by 1.3% on the 300 SAMPLE in terms of median AUC values. In addition, the algorithm that comes closest to the performance of the EDROD algorithm on the 700 SAMPLE dataset is ABOD, which lags behind EDROD by 1.5% in terms of median AUC values. Besides, we have also found that the EDROD algorithm exhibits the smallest range of fluctuations compared to other algorithms, which means that the EDROD algorithm has the most stable performance when faced with datasets of different sizes. To sum up, for the synthetic 300 SAMPLE and 700 SAMPLE datasets, the EDROD method can achieve a higher AUC value and more stable performance of anomaly detection when compared to other competing algorithms.\n\nBased on the experiments conducted on synthetic datasets, the following conclusions can be drawn. The proposed EDROD method shows lower sensitivity to the choice of parameter $K$ compared to other KNN-based algorithms, in- dicating enhanced robustness to the change of parameter", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "The inclusion of $\\dot{H}$ into the dark energy density is identified by a constant term mediated by the $\\gamma$ parameter in the critical solution. In terms of the reduced variables, we can write Eq. (33) as follows\n\n$$\n\\bar{P}=\\frac{\\bar{T}}{\\bar{v}}\\frac{T_{c}}{v_{c}P_{c}}(1+12\\pi \\gamma)+\\left(\\frac{2^{2n}6 \\pi \\beta}{v_{c}^{2(n-1)}}\\frac{1}{\\bar{v}^{2(n-1)}} - 48\\pi \\gamma - 1\\right)\\frac{1}{4\\pi \\bar{v}^{2}}\\frac{1}{P_{c}v^{2}_{c}}.\n$$\n\nFor comparative purposes between (32) and (23), we focus only on the behavior of the reduced Gibbs free energy in the case $n=2$. In Fig. (4), we show such behavior; as can be seen, the value of the concomitant parameter to $\\dot{H}$, namely $\\gamma$, modifies the size of the loop that characterizes the first-order phase transition; as its value increases, we observe that the loop’s size decreases. Then, in this scenario, the thermodynamic instability can be attenuated, which means that the isotherms obtained in the $P-v$ diagram are smoothed. Similar to the energy density (23) results, the outcomes emerging from (32) are independent of the parameter $\\beta$.\n\nFIG. 4: Reduced Gibbs free energy against reduced pressure for energy density (23) (dotted line) and (32) (solid and dashed lines).\n\n## CONCLUSIONS", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Proof. he projection $S_{W}=\\pi^{\\operatorname*{Mp}}(\\widehat{S}_{W,m})$ is a free symplectic matrix, in block-matrix form\n\n$$\nS_{W}=%\n\\begin{pmatrix}\nA & B\\\\\nC & D\n\\end{pmatrix}\n\\text{\\ , \\}\\det B\\neq0.\n$$\n\nA straightforward calculation yields the factorization\n\n$$\nS_{W}-I=%\n\\begin{pmatrix}\n0 & B\\\\\nI & D-I\n\\end{pmatrix}%\n\\begin{pmatrix}\nC-(D-I)B^{-1}(A-I) & 0\\\\\nB^{-1}(A-I) & I\n\\end{pmatrix}\n.\n$$\n\nSince $S_{W}\\in\\operatorname*{Sp}(n)$ we have $C-DB^{-1}A=-(B^{T})^{-1}$ and hence\n\n$$\nC-(D-I)B^{-1}(A-I)=B^{-1}A+DB^{-1}-(B^{T})^{-1}=W_{xx}%\n$$\n\nso that\n\n$$\nS_{W}-I=%\n\\begin{pmatrix}\n0 & B\\\\\nI & D-I\n\\end{pmatrix}%\n\\begin{pmatrix}\nW_{xx} & 0\\\\\nB^{-1}(A-I) & I\n\\end{pmatrix}\n.\n$$\n\nIt follows that\n\n$$\n\\det(S_{W}-I)=(-1)^{n}\\det B\\det W_{xx}%\n$$\n\nand hence\n\n$$\n\\arg\\det(S_{W}-I)=n\\pi+\\arg\\det B+\\arg\\det W_{xx}\\text{\\}\\operatorname{mod}%\n2\\pi.\n$$\n\nNoticing that $\\arg\\det W_{xx}=\\pi\\operatorname{Inert}W_{xx}$ and that this is\n\n$$\n\\arg\\det(S_{W}-I)=n\\pi+\\arg\\det B+\\pi\\operatorname{Inert}W_{xx}\\text{\n\\}\\operatorname{mod}2\\pi.\n$$\n\nIn view of formula (21) and (43) we have $\\arg\\det(B) =\nm\\pi$ (see Appendix A) and hence\n\n$$\n\\arg\\det(S_{W}-I)=(n+m-\\operatorname{Inert}W_{xx})\\pi\\text{\\}%\n\\operatorname{mod}2\\pi\n$$\n\nthat is\n\n$$\n\\arg\\det(S_{W}-I)=(n+\\nu(\\widehat{S}_{W,m}))\\pi\\text{\\}\\operatorname{mod}%\n2\\pi\n$$\n\nwhich yields (44).", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Figure12. Measured nebular Heii 4686 / H $\\beta$ values (a proxy for the ratio of $\\mathrm{He^+}$ -ionizing to H-ionizing photons) versus gas-phase metallicity. Both our galaxies (blue circles, backed by red squares if nebular Civ was detected) and the Wolf-Rayet galaxies presented by López-Sánchez&Esteban (2010) are presented. The right axis indicates the corresponding $\\mathrm{Q(He^+)/Q(H)}$ assuming Case B recombination for both species. Both observational samples show a sharp transition with metallicity, with substantially larger ratios and harder spectra below $12+\\log\\mathrm{O/H}\\sim 8$.\n\nused stellar population synthesis models. Shirazi & Brinchmann (2012) found that both Starburst 99 (Leithereretal.1999, 2010) and BPASS (Eldridge & Stanway 2009) can reproduce nebular Heii /H $\\beta$ ratios as high as 1-10% for instantaneous bursts of star formation, but only for a short period of time approximately 3 Myr after the burst; and only above $Z/Z_\\odot > 1/5$. This modeling sug- gests that strong nebular Heii should only occur when Wolf-Rayet (WR) stars are active, and that large WR populations should power strong Heii /H $\\beta$\n\n$\\gtrsim 1\\%$.\n\nThis is in contrast to the observations, where nebular Heii is commonly seen in systems without WR features at low metallicity (Figure 5; see also Brinchmann et al. 2008b; Shi- razi & Brinchmann 2012). At the relatively high-metallicity end ($12+\\log\\mathrm{O/H}>8.1$) where strong WR features are seen, our high- resolution view of Heii reveals that nebular Heii is consistently weaker than predicted (Heii $\\lambda 4686$ /H $\\beta< 1\\%$). As discussed in Sec. 4, fits to the UV spectra presented here indicate that the promi- nent nebular Heii $\\lambda 1640$ emission visible in SB 111, 2, and 182 cannot be fit with the models of Gutkin et al. (2016) without si- multaneously invoking $M_{up}=300 M_\\odot$ and metallicity substantially lower than the optically-derived gas-phase metallicity. If we are to accurately interpret high-ionization emission in distant galaxies, we must understand the origin of these discrepancies.\n\n## 5.3 Sources of Ionizing Radiation\n\nAccurate predictions for extreme ultraviolet (EUV) ionizing flux from star-forming galaxies are critical for modeling and interpreting the high-ionization UV nebular lines that this flux powers. As dis- cussed above, current stellar population synthesis models do not re- produce the strong metallicity dependence observed in Heii /H $\\beta$ near $Z/Z_\\odot \\sim 1/4$ (Fig. 12). Our data reveals a sharp transition in this ra- tio consistent with an order of magnitude increase in $\\mathrm{Q(He^+)/Q(HI)}$ as gas-phase metallicity is decreased from $Z_\\odot$ to $Z_\\odot/5$ (Sec. 5.2). Identifying the physical origin of this transition will help direct ad- justments or additions to these models.\n\nOur deep moderate-resolution HST /COS and ESI data allow us to put direct constraints on the massive stellar populations and other ionizing sources potentially present in these systems. By virtue of their selection using Heii diagnostics (see Sec. 2.1 and Shirazi & Brinchmann2012) and as inferred from sensitive UV line diagnostics (Fig. 11), an AGN contribution to photoionization is highly unlikely in these systems. This leaves three primary ionizing sources which could contribute to the high-ionization emission lines, which we discuss in-turn: stars, fast radiative shocks, and X-ray binaries.\n\n## 5.3.1 Stars\n\nThe EUV flux output from young stellar populations which powers high-ionization nebular emission lines is very uncertain and inti- mately linked to stellar winds. Both OB stars and their hydrogen- stripped relatives Wolf-Rayet (WR) stars drive winds through metal line absorption in the EUV, which diminish in strength as stellar metallicity is decreased (e.g. Castoretal.1975; Kudritzkietal.1987; 2006) The amount of EUV flux which passes through these winds to power nebular emission is decreased substantially by high-density winds and difficult to calibrate directly, especially beyond the $\\mathrm{He^+}$ ionizing edge (e.g. Smith et al. 2002; Crowther & Hadfield 2006; Smith2014). As metallicity decreases and winds weaken, WR stars will be produced less efficiently by wind-driven mass loss. However, mass transfer in binaries can efficiently remove the outer layers of a donor star, producing a hot stripped star without the need for high wind mass loss (e.g. Maeder&Meynet1994; Eldridgeetal.2008; Gotbergetal.2017; Smithetal.2017). In addition, the mass gainer in a binary system may be sufficiently spun-up by angular momentum transport to become fully-mixed during its hydrogen burning phase, leading to quasi-chemically homogeneous evolution and similarly high effective temperatures (Maeder1987; Yoon&Langer2005; El- dridge&Stanway2012; deMinketal.2013). If sufficient metals are available, these stars will reveal themselves in broad wind emission lines such as Heii like canonical WR stars; but at low metallicities, these sources of hard ionizing radiation will be effectively invisible except for their impact on nebular emission. The efficiency of mass transfer and the effects of rotation are both likely enhanced at lower metallicities, where weaker stellar winds allow massive stars to retain their mass longer (e.g. Maeder&Meynet2000; Eldridge&Stanway 2009; Szécsietal.2015). Our moderate-resolution HST /COS and ESI spectra allow us to measure stellar wind features directly, and thus ascertain whether the wind properties of the massive stars present correlate with the nebular emission we detect.\n\nThe UV spectra reveal a significant weakening in stellar wind lines with declining gas-phase metallicity (Fig. 4). The five systems above $12+\\log\\mathrm{O/H}\\gtrsim 8.2$ have P-Cygni absorption deeper than $-4$ Å; whereas the feature is uniformly weaker than this for the lower- metallicity systems. Quantitatively, the trend is consistent with other local galaxies. The empirical relation derived by Crowther et al. (predicts a gas-phase oxygen abundance of $12+\\log\\mathrm{O/H}=8.2$ for a P-Cygni absorption depth of $-4$ Å, in good agreement with our direct- $T_e$ gas-phase metallicities. The largely monotonic decrease in the strength of this feature with decreasing gas-phase oxygen abun- dance is strong evidence that the stellar metallicity and thus stellar wind strengths track the gas-phase metallicity, as expected. This con- firms that the systems with the strongest observed high-ionization UV emission harbor populations of very metal-poor stars in addition to hot metal-poor gas.\n\nThe five relatively high-metallicity systems show surprisingly strong WR star signatures. The stacked G160M spectra of the sys- tems above $12+\\log\\mathrm{O/H}>8.1$ reveals very strong broad $\\lambda 1640$", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "behavior of all physical quantities is expressed via the only nonperturbative parameter – the string tension, $\\sigma_E=\\sigma_H= \\sigma$. The resulting field correlators $D^E(x),D^H(x)$ are the so-called gluelump Green’s functions, which define all confining QCD dynamics. These gluelumps have been calculated in good agreement between the FCM [0] and the lattice data [0], while $D^E(x),D^H(X)$ were also studied in detail at $T>0$ on the lattice [0]. It was found that the situation is drastically changing at $T>0$, where both $\\sigma_E(T)$ and $\\sigma_H(T)= \\sigma_s(T)$ have a different behavior. Moreover $\\sigma_E(T)$ disappears above $T=T_c$, while in contrast to that $\\sigma_s(T)$ grows almost linearly at large $T$, as it was found on the lattice [0, 0, 0, 0] and supported by the studies in the framework of the FCM [0, 0, 0]. This dependence contains the coefficient, denoted as $c_{\\sigma}$ [0, 0],\n\n$$\n\\sqrt{\\sigma_s(T)}= c_{\\sigma} g^2(T) T,\n$$\n\ndefined numerically in the lattice calculations [0, 0] in the case $N_c=3, N_f=0$ as\n\n$$\nc_{\\sigma}= 0.566\\pm 0.013.\n$$\n\nFor the $SU(2)$ theory a similar lattice calculation [0, 0] has given a smaller value,\n\n$$\nc_{\\sigma}^2= (0.136 + 0.011),~~ c_{\\sigma}= 0.369\\pm 0.015.~~\n$$\n\nOn the theoretical side the linear temperature growth of $\\sigma_s(T)$ was derived in the framework of FCM [0, 0, 0], where the basic notion is the colormag- netic (CM) field correlator $D^H(z)$, which can be expressed via the gluelump Green’s function $G(x.y|z)$ and in this case the (two-gluon) gluelump is the string triangle, created by two gluons and the adjoint spectator point. The exact theory of these gluelumps was given in [0, 0, 0] and will be discussed in the next section in the case of the CM correlators. Note that the gluelump Green’s functions define the field correlators in the self-consistent way, with- out fitting parameters. In the case of the CM field the resulting $\\sigma_s(T)$ has exactly the same form as in (1) [0] with $c_{\\sigma}$, expressed via the gluelump Green’s function integral $G(0,0|z)$,\n\n$$\nc_{\\sigma}^2= \\frac{N_c^2 -1}{4} \\int d^2 z G(0,0|z).\n$$\n\nHowever, in the FCM the numerical values of $c_{\\sigma}$ were not calculated before", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Figure 3: The proposed architecture and method for power transient stability analysis (TSA).\n\nevents as quickly as possible through TFIs, predicting the stability of the system and detecting any anomalies. Moreover, TFIs can also aid in condition monitoring and predictive maintenance of power systems. By continuously monitoring the system signals and analyzing the TFIs, we can detect any gradual changes in the frequency content, which may indicate the presence of faults or degradation in the system compo- nents. This information can be used to schedule maintenance activities and prevent equipment failure, leading to reduced downtime and cost savings.\n\nFig. 3 presents a framework of TFIs and 2D-CNNs to monitor the transient stability before and after electric vehicle fast-charging stations (EVFCS) are loaded into the smart power grid [8]. The phasor measurement units (PMUs) placed in the WSCC 127-bus system, shown in Fig. 4, provides a vast collection of synthetic data to train our ML/DL models. Certain types of electrical quantities in the sequential PMU data are utilized to forecast the system stability. In this case, the voltage magnitude $V_{M}$ is chosen as the input feature to train conventional ML models such as XGboost, RF, DT, MLP.\n\nOn the other hand, to train the CNN, a 3-channel input, TFIs of voltage magnitude $V_{M}$, voltage angle $V_{A}$, voltage frequency $V_{F}$, is created with a time window of 0.1s through RP, GADF, and DWT transformations. Then, the TFIs are fed into 2D-CNNs. During the feature learning stage, a series of operations are performed, including convolution (filtering), activation, and pooling. Trainable kernels are used to convolve the multi-channel input, and the resulting output is passed through an activation function. The data is then subjected to a pooling operation, which reduces its resolution and makes it more resistant to small variations in previously learned features. Finally, at the end of this stage, the feature maps are flattened for further processing. After flattening the processed data into one-dimensional representations, it is fed into a fully connected feed-forward network for the final classification stage. In the last layer of the network, a softmax function is utilized as the activation function to produce the final prediction result. This function is highly appropriate for binary classification problems, including the prediction of power system transient stability, as it is capable of transforming a vector of real values into probabilities that add up to 1.\n\nFigure 4: The WSCC 127-bus system’s single line diagram showcasing the placement of both the EVFCS and PMUs\n\nThe PMU data captured in the WSCC 127-bus system were used to evaluate the performance of different classifiers, based on three metrics. While ACC represents the overall accuracy, the TUR and TSR represent the proportion of correctly predicted unstable instances, and the proportion of correctly predicted stable instances, respectively. As shown in Table I, the DWT and CNN combination provides the highest accuracy, and reduces the computational cost of training and inference as the DWT can reduce the size of the input data by decomposing the images into frequency sub-bands, which can be processed more efficiently by the CNN.\n\nTABLE I Accuracy results before/after integration of EVFCS.", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "BYRON HEERSINK\n\nTheorem Let $\\mathcal D\\subseteq\\{\\boldsymbol{x}\\in[0,1]^{n+1}:0<x_1,\\ldots,x_n\\leq x_{n+1}\\}$ have boundary of Lebesgue measure zero, and $f:\\overline{\\mathcal D}\\times\\Omega_\\Delta\\rightarrow\\mathbb R$ be bounded and continuous. Then we have\n\n$$\n\\begin{aligned}\n&\\lim_{T\\rightarrow\\infty}\\frac{1}{T^{n+1}}\\sum_{\\boldsymbol{a}\\in\\mathbf A\\cap T\\mathcal D}f\\left(\\frac{\\boldsymbol{a}}{T},\\Delta h\\left(\\frac{\\boldsymbol{a}'}{a_{n+1}}\\right)a(T)\\right)\\notag\\\\\n&\\qquad\\qquad\\qquad=\\frac{1}{\\zeta(n+1)}\\int_{\\mathcal D\\times\\Delta\\backslash\\mathbf A^*H}f(\\boldsymbol{y},Ma(y_{n+1}))\\,d\\lambda(\\boldsymbol{y})\\frac{d\\mu_H(M)}{[\\Gamma:\\Delta]}.\n\\end{aligned}\n$$\n\nProof. First, let $g:[0,1]\\times{\\mathbb T_\\Delta^n}\\times\\Delta\\backslash\\Gamma H_a\\rightarrow\\mathbb R$ be continuous with compact support, and for every pair of constants $b,c\\in[0,1]$ with $b<c$, define $\\mathcal{L}_{g,b.c},\\mathcal{U}_{g,b,c}:{\\mathbb T_\\Delta^n}\\times\\Delta\\backslash G\\rightarrow\\mathbb R$ by\n\n$$\n\\mathcal{L}_{g,b,c}(\\boldsymbol{x},M)=\\underset{_}{\\text{inf}}{u\\in[b,c]}g(u,\\boldsymbol{x},M)\\quad\\text{and}\\quad\\mathcal{U}_{g,b,c}(\\boldsymbol{x},M)=\\sup_{u\\in[b,c]}g(u,\\boldsymbol{x},M).\n$$\n\nThen for $T>0$, define $\\mathcal F_\\Delta^{b,c}(T)=\\{\\boldsymbol{p}/q\\in{\\mathcal F_\\Delta}(T):bT\\leq q\\leq cT\\}$. By the proof of Theorem 5(a) (in particular, see (17)), we have\n\n\\begin{align*} &\\limsup_{T\\rightarrow\\infty}\\frac{1}{\\#{\\mathcal F_\\Delta}(T)}\\sum_{\\boldsymbol{p}/q\\in\\mathcal F_\\Delta^{b,c}(T)}g\\left(\\frac{q}{T},\\frac{\\boldsymbol{p}}{q},\\Delta h\\left(\\frac{\\boldsymbol{p}}{q}\\right)a(T)\\right)\\\\\n&\\qquad\\qquad\\qquad\\leq\\lim_{T\\rightarrow\\infty}\\frac{1}{\\#{\\mathcal F_\\Delta}(T)}\\sum_{\\boldsymbol{r}\\in\\mathcal F_\\Delta^{b,c}(T)}\\mathcal{U}_{g,b,c}(\\boldsymbol{r},\\Delta h(\\boldsymbol{r})a(T))\\\\\n&\\qquad\\qquad\\qquad=\\int_b^c\\int_{{\\mathbb T_\\Delta^n}\\times\\Delta\\backslash\\Gamma H}\\mathcal{U}_{g,b,c}(\\boldsymbol{x},Ma(y))(n+1)y^n\\,d\\lambda_\\Delta(\\boldsymbol{x})\\,\\frac{d\\mu_H(M)}{[\\Gamma:\\Delta]}\\,d\\lambda(y). \\end{align*}\n\nSo for any finite partition $0=b_0<b_1<\\cdots<b_m=1$ of $[0,1]$,\n\n\\begin{align*} &\\limsup_{T\\rightarrow\\infty}\\frac{1}{\\#{\\mathcal F_\\Delta}(T)}\\sum_{\\boldsymbol{p}/q\\in\\mathcal F_\\Delta(T)}g\\left(\\frac{q}{T},\\frac{\\boldsymbol{p}}{q},\\Delta h\\left(\\frac{\\boldsymbol{p}}{q}\\right)a(T)\\right)\\\\\n&\\qquad\\qquad\\qquad\\leq\\sum_{j=0}^{m-1}\\int_{b_j}^{b_{j+1}}\\int_{{\\mathbb T_\\Delta^n}\\times\\Delta\\backslash\\Gamma H}\\mathcal{U}_{g,b,c}(\\boldsymbol{x},Ma(y))(n+1)y^n\\,d\\lambda_\\Delta(\\boldsymbol{x})\\,\\frac{d\\mu_H(M)}{[\\Gamma:\\Delta]}\\,d\\lambda(y). \\end{align*}\n\nBy the uniform continuity of $g$, we can take the infimum over all finite partitions $\\{b_j\\}_j$ of $[0,1]$ to obtain\n\n\\begin{align*} &\\limsup_{T\\rightarrow\\infty}\\frac{1}{\\#{\\mathcal F_\\Delta}(T)}\\sum_{\\boldsymbol{p}/q\\in\\mathcal F_\\Delta(T)}g\\left(\\frac{q}{T},\\frac{\\boldsymbol{p}}{q},\\Delta h\\left(\\frac{\\boldsymbol{p}}{q}\\right)a(T)\\right)\\\\\n&\\qquad\\qquad\\qquad\\leq\\int_0^1\\int_{{\\mathbb T_\\Delta^n}\\times\\Delta\\backslash\\Gamma H}g(y,\\boldsymbol{x},Ma(y))(n+1)y^n\\,d\\lambda_\\Delta(\\boldsymbol{x})\\,\\frac{d\\mu_H(M)}{[\\Gamma:\\Delta]}\\,d\\lambda(y). \\end{align*}\n\nWe can similarly use the functions $\\mathcal{L}_{g,b,c}$ to get the reverse inequality while replacing $\\limsup$ with $\\liminf$; and hence\n\n$$\n\\begin{aligned}\n&\\lim_{T\\rightarrow\\infty}\\frac{1}{\\#{\\mathcal F_\\Delta}(T)}\\sum_{\\boldsymbol{p}/q\\in\\mathcal F_\\Delta(T)}g\\left(\\frac{q}{T},\\frac{\\boldsymbol{p}}{q},\\Delta h\\left(\\frac{\\boldsymbol{p}}{q}\\right)a(T)\\right)\\notag\\\\\n&\\qquad\\qquad\\qquad=\\int_0^1\\int_{{\\mathbb T_\\Delta^n}\\times\\Delta\\backslash\\Gamma H}g(y,\\boldsymbol{x},Ma(y))(n+1)y^n\\,d\\lambda_\\Delta(\\boldsymbol{x})\\,\\frac{d\\mu_H(M)}{[\\Gamma:\\Delta]}\\,d\\lambda(y).\n\\end{aligned}\n$$\n\n", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "where $\\mathcal{B}\\subset\\mathcal{X}$ and $\\lambda_{i}$ are a training batch and the weights corresponding to each sign, respectively. Note that $\\mathcal{L}_{\\mathcal{B}}^{sign}$ enforce the network to reconstruct the desired surface and refuse to generate the redundant surfaces simultaneously. Moreover, the MSE loss forces further attention on dis- tinguishing the inside and outside points near the recon- structed surface because their $\\mathop{\\rm th}\\nolimits{\\mathcal{S}(\\textbf{p}_i)}$ are near zero while their ground truths are $-1$ and $+1$, respectively.\n\nLoss normal. To improve the reconstruction, we use the normal vectors as second-order information. Therefore, we define a MSE loss between the ground-truth normal vectors of the sample points $\\textbf{p}_{on}$ on the surface of a target mesh model $\\textbf{n}_{\\text{g}}$ and their normal vectors $\\textbf{n}_{\\text{r}}$ obtained by Eq. 8:\n\n$$\n\\begin{split}\n \\mathcal{L}^{n}_{\\mathcal{B}}=\\mathbb{E}_{\\textbf{p}_{on}\\sim P}\\lVert \\textbf{n}_{r}(\\textbf{p}_{on}) - \\textbf{n}_{g}(\\textbf{p}_{on})\\rVert^2,\n\\end{split}\n$$\n\nwhere the normal vectors on the union surface can be di- rectly determined for any point on the surface as Eq. 9:\n\n$$\n\\begin{split}\n \\textbf{n}_r=&\\frac{\\nabla \\mathcal{S}(x,y,z)}{||\\nabla \\mathcal{S}(x,y,z)||}=\\frac{(\\frac{\\partial \\mathcal{S}}{\\partial x},\\frac{\\partial \\mathcal{S}}{\\partial x},\\frac{\\partial \\mathcal{S}}{\\partial z})}{\\sqrt{(\\frac{\\partial \\mathcal{S}}{\\partial x})^2+(\\frac{\\partial \\mathcal{S}}{\\partial y})^2+(\\frac{\\partial \\mathcal{S}}{\\partial z})^2}},\\\\\n &\\text{subject to:}\n \\quad \\nabla \\mathcal{S}(x,y,z) = \\nabla p_{m^{\\ast}}(x,y,z), \\\\\n &\\hspace{15mm} \\quad m^{\\ast}= \\underset{_}{\\operatorname{arg}\\,\\operatorname{min}}\\;m(p_m(x,y,z)).\n\\end{split}\n$$\n\nNote that $m^{\\ast}$ is the index of the closest primitive (i.e., the primitive with the smallest value $p_{m^{\\ast}}(x,y,z)$) to the point.\n\nFinally, the total loss $\\mathcal{L}^{\\text{total}}_\\mathcal{B}$ is the weighting average of all defined losses with the corresponding weights as Eq. 10:\n\n$$\n\\mathcal{L}^{\\text{total}}_\\mathcal{B} = \\mathcal{L}^{\\mathrm{sign}}_{\\mathcal{B}}+\\lambda_n\\mathcal{L}^{n}_{\\mathcal{B}}.\n$$\n\n## Implementation details\n\nWe consider the bounding box $\\mathcal{C}=[-1,1]^3$ and fit the given input 3D shapes into it by keeping its aspect ratio. Then we extract $1M$ points from $[-1.1,1.1]^3\\in\\mathbb{R}^3$ sur- rounding the 3D shapes, and at each iteration, we randomly select $1\\%$ of them as jointly inside points $\\textbf{p}_{in}$ and outside points $\\textbf{p}_{out}$ for all shapes in the batch $\\mathcal{B}$. Moreover, we pick $10k$ points $\\textbf{p}_{on}$ on the surface of each 3D shape and randomly select $20\\%$ of them at each iteration.\n\nTo capture the information of the input observation $o\\in\\mathcal{X}$ we employ the pretrained ResNet-18 [18] as the en- coder. Then three sets of independent fully connected (FC) layers $(4096,4096,4096,55\\times M),(1024,512,256,3\\times M),(256,M)$ decodes the encoded features to obtain the parameters of each symmetric matrix $B$, scalar $R$ and cen- ter $c$ for $M=100$ primitives in Eq. 2, respectively. All FC layers except the last layers are empowered by the $ReLU$ non-linear activation function. We also apply three batch normalization layers after the first three FC layers of $B$ de- coder to accelerate training and boost the performance.\n\nTo ensure $H$ is a PD matrix, we parameterize it as Eq. 11:\n\n$$\n\\begin{split}\n H=BB^T+\\alpha I\\succ 0\n\\end{split}\n$$\n\nwhere $\\alpha=0.0001$ is a small scalar factor, and $B$ and $I$ are $10\\times 10$ symmetric and identity matrices, respectively. Moreover, we apply the $sigmoid$ function on the output of the $R$ decoder to generate a value in $(0,1)\\subset\\mathbb{R}^3$ as the parameter $R$ of Eq. 5 to control the size of all primitives and guarantee that they are not larger than the size of the bounding box $\\mathcal{C}$. Furthermore, we apply $\\mathop{\\rm th}\\nolimits$ on the output of the $c$ decoder to generate the centers inside the bounding box $\\mathcal{C}$. Therefore, each primitive is parameterized with 59 parameters in total including three parameter for the center $c=(c_1,c_2,c_3)$, one parameter for the $R$, and 55 parameters for the matrix $B$.\n\nWe set the parameters $\\lambda_{on}$, $\\lambda_{in}$, $\\lambda_{out}$, and $\\lambda_{n}$ as 2, 1, 10, and 1, respectively. We train our encoder-decoder architec- ture with Adam optimizer with the initial learning rate 1e-4, weight decay 1e-7, and batch size of 64. We implement our model in Python3.7 using PyTorch via CUDA instruction.\n\n## 4. Experiments\n\nIn this section, we provide information about the eval- uation setups and show qualitative and quantitative results of our method compared to state-of-the-art methods on sin- gle RGB image 3D shape reconstruction. We also perform various ablation studies to analyze our method better. More experiments are available in the supplementary material.\n\n## 4.1. Dataset and Metrics\n\nWe evaluate our approach on the subset of the ShapeNet dataset [6] with the same image renderings and train- ing/testing split provided by Choy et al. [9]. We also em- ploy mesh-fusion [36] to generate watertight meshes from the 3D CAD models. We then use Houdini [35] to extract the inside/outside/on points and the normal vectors. For evaluation, we use the volumetric IoU, Chamfer [11], and F-Score [23] metrics. Volumetric IoU is used to measure the overlapped volume between the ground truth meshes and the reconstructed surfaces. Chamfer is the mean of the accuracy and the completeness score. The mean distance of points on the reconstructed surface to their nearest neigh- bors on the ground truth mesh is defined as the accuracy metric. The completeness metric is defined in the opposite direction of the accuracy metric. F-score is the harmonic mean of precision which shows the percentage of correctly reconstructed surfaces. To compute IoU, we sample $100k$ points from the bounding box. To evaluate Chamfer and F- score, we first transfer the reconstructed surfaces to meshes,", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Repeat the procedure for each of the four intervals composing $F(I)$ in place of $I$, and iterate. Concerning how our sets $K$ look globally, we consider two examples of Koch-type snowflakes $S_\\alpha^1$ and $S_\\alpha^2$ which are locally the same, but $S^1_\\alpha$ is purely fractal, and $S^2_\\alpha$, on a large scale, resembles a line. To get $S_\\alpha^2$, we decompose $\\mathbb{R} = \\cup_{k \\in \\mathbb{Z}}{[k - 1/2, k + 1/2]}$, and for all $k \\neq 0$ repeat the procedure we applied to the unit interval $[-1/2, 1/2]$.\n\n$S^2_\\alpha$\n\nTo get $S_\\alpha^1$, first denote by $l$ the length of any segment the transformation $F$ described above produces from the unit interval $I$. It is easy to compute that $l = \\frac{1}{2(1 + \\cos{\\alpha})}$. By $H_l$ denote the scaling transform (homothety) with the coefficient $1/l$ and center $-1/2$ (the left end of the unit interval $I$), and denote by $S_\\alpha |_I$ the fractal set we built from $I = [-1/2, 1/2]$. Define the part of the set $S^1_\\alpha$ to the right of $\\{x = -1/2\\}$ to be $\\cup_{k \\ge 1}{H_l(S_\\alpha |_I)}$. To get the part of the set to the left of the axis, we reflect the part to the right with respect to it: $S^1_\\alpha = \\cup_{k \\ge 1}{H_l(S_\\alpha |_I)} \\cup \\left(\\overline{\\left(\\cup_{k \\ge 1}{H_l(S_\\alpha |_I)} + 1/2\\right)} - 1/2\\right)$. Here is how $S^1_\\alpha$ looks like to the right of $-1/2$.\n\n$S^1_\\alpha$\n\nIn the theorem below we will call $S_\\alpha$ both $S_\\alpha^1$ and $S_\\alpha^2$ for simplicity. A classical com- putation, essentially the same as for the standard Koch snowflake, shows that the Haus- dorff dimension $d$ of $S_\\alpha$ (both $S_\\alpha^1$ and $S_\\alpha^2$) is $\\frac{\\ln{(4)}}{\\ln{(2(1 + \\cos{\\alpha}))}}$. We also choose the domains $\\Omega_1 = \\left(\\mathbb{R}^2 \\setminus S_\\alpha^1\\right)^{+}$ and $\\Omega_2 = \\left(\\mathbb{R}^2 \\setminus S_\\alpha^2\\right)^{+}$, and denote them both $\\Omega$ in what follows.", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "families of the associated subgraphs, as shown in Table 2. We also use the fol- lowing observations (for $p \\ll m^{-1/2} \\ll 1$) to cancel some of the redundant terms in the expressions.\n\n4-path: $m^2p^7 \\ll m^2 p^6$ and $m^3 p^8 \\ll m^3 p^7$\n\n4-cycle: $m^2 p^6 \\ll mp^4$\n\n3-pan: $m^3 p^7 \\ll m^2 p^5$\n\nDiamond: $m^2 p^6 \\ll mp^4$ and $m^4 p^9 \\ll m^3 p^7$\n\nButterfly: $m^5 p^{12} \\ll m^5 p^{11}$, $m^4 p^{11} \\ll m^3 p^8 \\ll m^2 p^6$, $m^3 p^{10} \\le m^3 p^8 \\ll m^2 p^6$.\n\n\\begin{tabular}{lccll}\n\\rightarrowprule\n$R$ & $|V(R)|$ & $|E(R)|$ & Appr.\\ density ($p \\ll m^{-1/2} \\ll 1$) & Appr.\\ density ($p \\sim \\mu m^{-1}$) \\midrule\n1-star & 2 & 1 & $mp^2$ & $\\mu^2 m^{-1}$ [1.5ex]\n2-star & 3 & 2 & $mp^3 + m^2 p^4$ & $(1+\\mu)\\mu^3 m^{-2}$ 3-cycle & 3 & 3 & $mp^3 + m^3 p^6$ & $\\mu^3 m^{-2}$ [1.5ex]\n3-star & 4 & 3 & $mp^4 + 3m^2 p^5 + m^3 p^6$ & $(1+3\\mu+\\mu^2)\\mu^4 m^{-3}$ 3-path & 4 & 3 & $mp^4 + 2m^2 p^5 + m^3 p^6$ & $(1+2\\mu+\\mu^2)\\mu^4 m^{-3}$ 4-cycle & 4 & 4 & $mp^4 + 4m^3 p^7 + m^4 p^8$ & $\\mu^4 m^{-3}$ 3-pan & 4 & 4 & $mp^4 + m^2 p^5 + m^4 p^8$ & $(1+\\mu)\\mu^4 m^{-3}$ Diamond & 4 & 5 & $mp^4 + 2 m^3 p^7 + m^5 p^{10}$ & $\\mu^4 m^{-3}$[1.5ex]\n4-star & 5 & 4 & $mp^5 + 7m^2 p^6 + 6 m^3 p^7 + m^4 p^8$ & $(1+7\\mu+6\\mu^2+\\mu^3)\\mu^5 m^{-4}$ 4-path & 5 & 4 & $mp^5 + 3 m^2 p^6 + 3 m^3 p^7 + m^4 p^8$ & $(1+3\\mu+3\\mu^2+\\mu^3)\\mu^5 m^{-4}$ Chair & 5 & 4 & $mp^5 + 4 m^2 p^6 + 4 m^3 p^7 + m^4 p^8$ & $(1+4\\mu+4\\mu^2+\\mu^3)\\mu^5 m^{-4}$ Butterfly & 5 & 6 & $mp^5 + m^2 p^6 + 2 m^4 p^9 + 4 m^5 p^{11} + m^6 p^{12}$ & $(1+\\mu)\\mu^5 m^{-4}$ \\bottomrule\n\\end{tabular}\n\nTable 1: Approximate densities of some subgraphs.\n\n## Proofs of Theorems 3.1, 3.2, and 3.3\n\nProof (of Theorem 3.1). Denote $\\hat\\lambda = \\hat\\lambda(G^{(n_0)})$ and $\\hat N = N_{K_2}(G^{(n_0)})$. Then the variance of $\\hat\\lambda$ is given by\n\n$$\n\\operatorname{Var}(\\hat\\lambda)\n \\ = \\ 4 \\frac{n^2}{n_0^4} \\operatorname{Var}(\\hat N).\n$$\n\nBy writing\n\n$$\n\\hat N = \\sum_{e \\in \\binom{[n_0]}{2}} 1(G \\supset e)\n \\qquad\\text{and}\\qquad \n \\hat N^2 = \\sum_{e \\in \\binom{[n_0]}{2}} \\sum_{e' \\in \\binom{[n_0]}{2}} 1(G \\supset e) 1(G \\supset e'),\n$$\n\nwe find that $\n \\mathbb{E} \\hat N\n = \\binom{n_0}{2} \\mathbb{P}(G \\supset K_2)\n$ and\n\n$$\n\\mathbb{E} \\hat N^2\n \\ = \\ \\binom{n_0}{2} \\mathbb{P}(G \\supset K_2) + 2(n_0-2) \\binom{n_0}{2} \\mathbb{P}(G \\supset S_2) + \\binom{n_0}{2} \\binom{n_0-2}{2} \\mathbb{P}(G \\supset K_2)^2.\n$$\n\n", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "GIL SOLANES AND JUAN ANDR ´ ES TRILLO\n\nspace and $N(A)$ is compact, we also have that $N(A)$ is a closed subset of $N'(A)$. Since the number of connected components of both $N(A),N'(A)$ clearly equals the number of connected components of $A$, we necessarily have $N(A)=N'(A)$.\n\n4.3.\n\n## Derivative operators in $\\ensuremath{\\mathbb R}\\xspace{m}$ and $\\CC{n}$\n\n. Given $\\lambda \\in \\mathbb{R}$ let $\\derivada{\\lambda,\\mathbb{R}}\\colon \\mathcal{V}_{\\lambda,\\mathbb{R}}^m\\rightarrow\\mathcal{V}_{\\lambda,\\mathbb{R}}^m$ be the restriction of $\\derivada{\\ensuremath{\\mathbb R}\\xspace{m}}$ to $\\mathcal{V}_{\\lambda,\\mathbb{R}}^m$, and let $\\derivada{\\lambda,\\mathbb{C}}$ be the restriction of $\\derivada{\\CC{m}}$ to $\\mathcal{V}_{\\lambda,\\mathbb{C}}^n$.\n\n$$\n\\begin{aligned}\n&\\derivada{\\lambda,\\mathbb{R}} \\sigma_i^{\\lambda} = (m-i)\\sigma_{i-1}^{\\lambda}-\\lambda (i+1)\\sigma_{i+1}^{\\lambda},& 0 \\leqslant i \\leqslant m-2,\\\\\n&\\derivada{\\lambda,\\mathbb{R}} \\sigma_{m-1}^{\\lambda} = \\sigma_{m-2}^{\\lambda},\\\\\n&\\derivada{\\lambda,\\mathbb{R}} \\sigma_{m}^{\\lambda} = \\sigma_{m-1}^{\\lambda},\n\\end{aligned}\n$$\n\nwhere it is understood that $\\sigma_{-1}^\\lambda=0$.\n\nLet us emphasize that (40) would make formal sense but does not hold for $i=m-1$.\n\nProof. By [21, Lemma 3.1], putting $\\kappa_m=0$, we have\n\n$$\nd\\kappa_{i} = \\alpha \\wedge\\left((m-i) \\kappa_{i-1}-\\lambda(i+1) \\kappa_{i+1}\\right), \\quad 0 \\leqslant i \\leqslant m-1.\n$$\n\nContracting with $T$ yields\n\n$$\ni_T d\\kappa_i = (m-i)\\kappa_{i-1} - \\lambda (i+1)\\kappa_{i+1}, \\quad [(\\,\\, \\mathrm{mod} \\,\\, \\alpha],d\\alpha).\n$$\n\nBy Proposition 4.2 the result follows.\n\nLemma 4.9. The following equalities hold modulo $\\alpha,d\\alpha$\n\n- i) For $k > 2q$\n\n$$\n\\begin{split}\n\t\t\\dfrac{\\omega_{2n-k}}{\\omega_{2n-k+1}}i_{T} d\\beta_{k,q}\\equiv &\\, 2 (n-k+q+1)\\gamma_{k-1,q}\\\\ &\n\t\t+{(k-2q+1)}\\beta_{k-1,q-1} \\\\\n\t\t&-\\frac{\\lambda}{2\\pi} {(k-2q+1)(2n-k+1)}\\beta_{k+1,q}.\n \\end{split}\n$$\n\n- ii) For $n > k-q$\n\n$$\n\\begin{split}\n\t\t\\dfrac{\\omega_{2n-2q}}{\\omega_{2n-2q+1}}i_T d\\gamma_{2q,q} \\equiv&\\, \\beta_{2q-1,q-1} \\\\\n\t\t&-\\frac{\\lambda}{2\\pi} \\frac{(q+2)(2n-2q+1)}{n-q}\\beta_{2q+1,q} \\\\\n\t\t&-\\frac{\\lambda}{2\\pi} \\frac{(n-q-1)(2n-2q+1)}{n-q}\\gamma_{2q+1,q}.\n \\end{split}\n$$\n\nProof. This is a straightforward computation using [1, Lemma3.3,Lemma3.6].\n\nProposition For $k > 2q$\n\n$$\n\\begin{aligned}\n\\dfrac{\\omega_{2n-k}}{\\omega_{2n-k+1}}\\derivada{\\lambda,\\mathbb{C}}\\mu_{k, q}^\\lambda&=(k-2q+1) \\mu_{k-1,q-1}^\\lambda + 2 (n-k+q+1)\\mu_{k-1,q}^\\lambda \\\\\n&- \\dfrac{\\lambda}{2 \\pi} (2 n-k+1)\\left((k-2 q+1) \\mu _{k+1,q}^\\lambda +2 (q+1) \\mu _{k+1,q+1}^\\lambda \\right)\\notag\n\\end{aligned}\n$$\n\n", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Figure8: Research Design as proposed by [31]\n\nThe approach used by the Bitcoin Energy Consumption Index has been widely debated both in academia ([11]) and non-academic ([9]) literature. Our review suggests that the foundational approach used by Digiconomist as outlined in [24] is of questionable scientific rigor. We specifically question the choices made regarding the life span of the hardware and the subsequent calculation of the 60% ratio of electricity cost in miners’ revenue. Unlike the CBECI model, the Digiconomist model does not provide a clear change log of the model evaluation.\n\nWe suggest that this work should be considered with caution. As indicated in the earlier section, the follow-up work [42] of Digiconomist suffers from some issues as well, most noticeably the assumptions made by the author regarding the equal distribution of hardware sales.\n\n## Code of Practices\n\nAs alluded to in the introduction section, one of the main contributions of this review is the preparation of novel research guidelines that could assist in improving the rigor of research in blockchain energy sciences. In the previous section, we highlighted some of the common issues in terms of the scientific quality of the reviewed studies.\n\nThe primary recommendation is to avoid those pitfalls however we acknowledge that this might be difficult to accomplish due to the lack of reliable datasets and difficulty in acquisition of newer data. In case it is difficult to avoid any of the above-mentioned pitfalls, we give some specific recommendations on how the researcher could present their research to avoid giving inaccurate predictions, measurements, or estimates.\n\nIn this section, we propose novel codes of practices segmented into three broad categories: basic research design, quantitative energy modeling, and other research recommendations.\n\n## Basic Research Design\n\nWe have three broad recommendations when designing and executing a blockchain energy study, these guidelines are intended for both academic and non-academic studies.\n\n## Explicit Research Methodology\n\nAny study measuring, estimating, or predicting the energy or environmental footprint of a crypto-asset should be designed to be easily reproduced. To this end, our first recommendation is to explicitly document the methodology employed in the study. We recommend that the research be documented using the 7-step model proposed by [31] as illustrated in Figure 8. We provide a brief overview of these 7 steps in the following text however it is recommended that the reader refers to the detailed description as provided by [31].\n\n- 1. Clear research problem: The research problem addressed by the project must be sufficiently documented in a clear manner. This assists the reader understand the project’s intended goal without any ambiguity. 2. Review the literature: As discussed in the results section, due to the lack of coherence in this newly developing space, a lot of research development is done in small silos resulting in likely slow progress. Thus we recommend that newer models review existing literature thoroughly. To that end, our survey provides the reader with an extensive list of relevant academic and non-academic literature. We also document progress made by non-academic source 30. 3. Development of a hypothesis: Ideally after doing an extensive literature review, the researcher should have a starting point within the literature, this could be an existing model that the researcher intends to build upon or a research methodology that the researcher intends to adopt. Before the execution of their research methodology, it is important to clearly state the working hypothesis or hypotheses. The hypothesis should be as concise as possible to allow testing through the research at the end.\n\n30 This document is public, we hope researchers and practitioners would refine and update it as needed, it can be accessed at: https://bit.ly/3csZLg9", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "- a) For $\\kappa>0$ we let $M_{\\kappa}$ denote the metric space given by the two-dimensional sphere $\\mathbb{S}^{2}$ with its intrinsic metric, scaled by a factor of $1/\\sqrt{\\kappa}$. b) We define $M_{0}$ as the Euclidean space $\\mathbb{R}^{2}$. c) For $\\kappa<0$ we write $M_{\\kappa}$ for the hyperbolic space $\\mathbb{H}^{2}$ (see [4, Definition 2.10]) with the metric scaled by a factor of $1/\\sqrt{-\\kappa}$.\n\n- We write $d_{\\kappa}$ for the metric on $M_{\\kappa}$. 2. Let $\\kappa\\in \\mathbb{R}$ and $(X,\\rho_{X})$ be a metric space. Given three points $x_{1},x_{2},x_{3}\\in X$ and metric segments of the form $[x_{1},x_{2}],[x_{2},x_{3}],[x_{3},x_{1}]\\subseteq X$, we call the union $[x_{1},x_{2}]\\cup[x_{2},x_{3}]\\cup[x_{3},x_{1}]$ a geodesic triangle with vertices $x_{1},x_{2},x_{3}$. A geodesic triangle with vertices $\\overline{x}_{1},\\overline{x}_{2}, \\overline{x}_{3}$ in $M_{\\kappa}$ is said to be a comparison triangle for a geodesic triangle with vertices $x_{1},x_{2},x_{3}$ in $X$ if $d_{\\kappa}(\\overline{x}_{i},\\overline{x}_{j})=\\rho_{X}(x_{i},x_{j})$. A point $\\overline{x}\\in [\\overline{x}_{i},\\overline{x}_{j}]$ is called a comparison point for $x\\in[x_{i},x_{j}]$ if $d_{\\kappa}(\\overline{x},\\overline{x}_{k})=\\rho_{X}(x,x_{k})$ for $k=i,j$. 3. Let $(X,\\rho_{X})$ be a metric space. If $\\kappa\\leq 0$, then $(X,\\rho_{X})$ is called a $\\mathrm{CAT}(\\kappa)$ -space if it is geodesic and every geodesic triangle $T$ in $X$ has a comparison triangle $\\overline{T}$ in $M_{\\kappa}$ such that\n\n$$\n\\rho_{X}(x,y)\\leq d_{\\kappa}(\\overline{x},\\overline{y})\n$$\n\n- whenever $\\overline{x},\\overline{y}\\in\\overline{T}$ are comparison points for $x,y\\in T$. If $\\kappa>0$, then we define a constant $D_{\\kappa}=\\mathrm{diam} M_{\\kappa}=\\frac{\\pi}{\\sqrt{\\kappa}}$ and we say that $(X,\\rho_{X})$ is a $\\mathrm{CAT}(\\kappa)$ space if for every pair of points $x,y\\in X$ with $\\rho_X(x,y)< D_\\kappa$ there is a metric segment joining $x$ and $y$ and (1) is satisfied for all geodesic triangles $T\\subseteq X$ with perimeter smaller than $2D_\\kappa$, that is, $\\rho_X(x,y)+\\rho_X(y,z)+\\rho_X(z,x)< 2D_{\\kappa}$, where $x,y,z$ denote the vertices of $T$.\n\nThus, $\\mathrm{CAT}(\\kappa)$ spaces can be thought of as metric spaces for which every sufficiently small geodesic triangle is ‘thinner’ in all directions than a corresponding triangle in the model space $M_{\\kappa}$. The classes of $\\mathrm{CAT}(\\kappa)$ spaces are increasing in the sense that whenever $X$ is a $\\mathrm{CAT}(\\kappa)$ space, it is also a $\\operatorname{CAT}(\\kappa')$ space for all $\\kappa'\\geq \\kappa$; see [4, Theorem 1.12].\n\nFinally, we recall the definition of $\\sigma$ -porous sets.\n\nDefinition 2.2. Given a metric space $(M,d)$, a subset $A\\subset M$ is called porous at a point $x\\in A$ if there exist $\\varepsilon_0>0$ and $\\alpha>0$ such that for all $\\varepsilon\\in(0,\\varepsilon_0)$, there exists a point $y\\in B(x,\\varepsilon)$ such that $B(y,\\alpha\\varepsilon)\\cap A = \\emptyset$. The set $A$ is called porous if it is porous at all its points and $A$ is called $\\sigma$ -porous if it is the countable union of porous sets.\n\nNote that the definition of porous sets differs slightly from the one used in [8]. For $\\sigma$ -porous sets both definitions agree. A detailed discussion of these differences and why for $\\sigma$ -porosity they do not matter can be found in [2, p. 93].\n\n## 3 The typical nonexpansive self-mapping on a small enough set is a Rakotch contraction", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "- Despite major changes in the course structure, the lab course improved my knowledge.\n\n- The online physics lab helped me to better under- stand the physics concepts behind the experiments.\n\n- Online discussions about the labs with the experi- ment supervisor or with peers reinforced my physics knowledge.\n\nExperimental skills:\n\n- The limited hands-on activities had negative im- pact on my learning\n\n- The online lab course in this semester had a similar effect on my experimental skills as the standard lab course.\n\n- Not having gathered the measurement data myself complicates data analysis and discussion\n\n- Missing the hands-on experience, I am concerned that I cannot handle experimental equipment in the next lab/in the future.\n\n- I feel that I gained less experimental skills due to the modified course format.\n\nProblem-solving sessions / recitations General\n\n- Did you attend online problem-solving sessions dur- ing the semester?\n\n- To which type of course did the sessions belong to? E.g. Introductory physics lecture with a relatively large audience, special course in physics with few students or other?\n\n- Were exercise sheets provided? Were exercise sheets compulsory?\n\n- What fraction of the exercise sheets did you com- plete?\n\n- Did you work alone or with peers on the tasks?\n\n- If you solved the homework problems in a study group, how did you organize the study groups?\n\n- Did you meet online to discuss the solutions to the tasks?\n\n- If online discussions of the problem sheets existed, what fraction did you attend?\n\nStructure\n\n- How was the problem-solving course structured? Click all that apply and/or describe the structure yourself.\n\n- – The students’ solutions were submitted, cor- rected by the tutor and discussed in an online meeting – The solutions were reconstructed during a live online session in real time – Exercise sheets were solved online live and dis- cussed in groups – The solutions to the exercise sheets were handed out to the students as text or as a video (worked-out solutions) – Forums were used to discuss exercise sheets without time constraints.\n\n- Which format (one of the above) would you con- sider best in terms of fostering your learning achievement? Think about what format you would like to see in the future.\n\nPerceived effectiveness of the problem-solving sessions\n\n- The online problem-solving sessions in this semester had a similar effect on understanding physical concepts as the regular exercises.\n\n- The online format led to less discussion about the individual solution steps.\n\n- In regular recitations a better exchange between students and teachers is achieved.\n\n- When presenting solutions online, it is easier for me to understand the individual steps and approaches.\n\n- During the online sessions I felt well supervised con- cerning my questions\n\n- Lacking handwriting opportunities in online ses- sions makes discussions about the tasks less effec- tive\n\n- Working on exercise sheets in a study group was more difficult than in normal semesters.", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Therefore,\n\n\\begin{equation*} \\left\\lfloor \\frac{(k'-1)N}{2^{K'}} \\right\\rfloor<i\\leqslant \\left\\lfloor \\frac{k'N}{2^{K'}} \\right\\rfloor. \\end{equation*}\n\nThis means $i\\in B_{\\lfloor(k-1)2^{K'-K}\\rfloor +1}^{(K')}$.\n\nThe proof of (ii) proceeds as follows. Let $k'\\in [2^{K'}]$ and $i\\in B_{k'}^{(K')}$, meaning that\n\n\\begin{equation*} \\left\\lfloor \\frac{(k'-1)N}{2^{K'}} \\right\\rfloor<i\\leqslant \\left\\lfloor \\frac{k'N}{2^{K'}} \\right\\rfloor. \\end{equation*}\n\nThis can be rewritten as\n\n\\begin{equation*} \\left\\lfloor \\frac{2^{K-K'}(k'-1)N}{2^{K}} \\right\\rfloor<i\\leqslant \\left\\lfloor \\frac{2^{K-K'}k'N}{2^{K}} \\right\\rfloor, \\end{equation*}\n\nAs $(B_k^{(K)})_{k\\in [2^K]}$ is a partition of $[N]$, there exists a unique $k\\in [2^K]$ such that $i\\in B_k^{(K)}$, i.e.\n\n\\begin{equation*} \\left\\lfloor \\frac{(k-1)N}{2^K} \\right\\rfloor< i\\leqslant \\left\\lfloor \\frac{kN}{2^K} \\right\\rfloor. \\end{equation*}\n\nCombining the two previous displays implies that $(k'-1)2^{K-K'}<k\\leqslant k'2^{K-K'}$, meaning that $i\\in B_k^{(K)}$. Moreover, applying (i) to $k$ gives that $B_k^{(K)}$ is a subset of $B_{k'}^{(K')}$, hence the result.\n\n## Proof of Lemma 4.3\n\nUsing (i) from Lemma 4.2, we have:\n\n\\begin{align*} B_{k_1}^{K_1}&\\subset B_{k_1'}^{(3)},\\quad \\text{with}\\quad k_1'=\\lfloor(k_1-1)2^{3-K_1}\\rfloor +1,\\\\\nB_{k_2}^{K_2}&\\subset B_{k_2'}^{(3)},\\quad \\text{with} \\quad k_2'=\\lfloor(k_2-1)2^{3-K_2}\\rfloor +1. \\end{align*}\n\nIn addition, $K_0$ is by definition larger than $3$ (see (6)) and (ii) from Lemma 4.3 then gives:\n\n\\begin{align*} (B_{k}^{(K_0)})_{k\\in \\left\\rrbracket (k_1'-1)2^{K_0-3},\\ k_1'2^{K_0-3} \\right\\rrbracket}&\\quad \\text{is a partition of}\\quad B_{k_1'}^{(3)},\\\\\n(B_{k}^{(K_0)})_{k\\in \\left\\rrbracket (k_2'-1)2^{K_0-3},\\ k_2'2^{K_0-3} \\right\\rrbracket}&\\quad \\text{is a partition of}\\quad B_{k_2'}^{(3)}. \\end{align*}\n\nThen, using the latter result and starting from the definition of $\\mathcal{K}_0(K_1,k_1,K_2,k_2)$, we can write\n\n\\begin{align*} \\mathcal{K}_0(K_1,k_1,K_2,k_2)&=\\left\\{k\\in \\left[2^{K_0} \\right]\\,\\middle|\\,B_k^{(K_0)}\\cap (B_{k_1}^{(K_1)}\\cup B_{k_2}^{(K_2)})=\\varnothing \\right\\}\\\\\n&\\supset \\left\\{k\\in \\left[2^{K_0} \\right]\\,\\middle|\\,B_k^{(K_0)}\\cap (B_{k_1'}^{(3)}\\cup B_{k_2'}^{(3)})=\\varnothing \\right\\}\\\\\n&= \\left[2^{K_0} \\right]\\setminus \\left(\\left\\rrbracket (k_1'-1)2^{K_0-3},\\ k_1'2^{K_0-3} \\right\\rrbracket \\cup \\left\\rrbracket (k_2'-1)2^{K_0-3},\\ k_2'2^{K_0-3} \\right\\rrbracket \\right). \\end{align*}\n\nThe latter result together with the definition of $K_0$ yield the lower bound on the cardinality of $\\mathcal{K}_0(k_1,K_1,K_2,k_2)$:\n\n\\begin{align*} \\left| \\mathcal{K}_0(K_1,k_1,K_2,k_2) \\right| &\\geqslant 2^{K_0}-2^{K_0-3}-2^{K_0-3}=\\frac{3}{4}2^{K_0}\\geqslant \\frac{3}{4}2^{\\log_2(V/3)+2}=V. \\end{align*}\n\n## Proof of Lemma 4.4\n\nLet $(m,m')\\in \\mathcal{M}^2$ and $v\\in [V]$. By construction of $T_v^{(m,m')}$, there exists $k\\in \\left[2^{K_0} \\right]$ such that $T_v^{(m,m')}=B_k^{(K_0)}$. Then, it follows from (6) that $2^{K_0}\\leqslant 8V/3$, so that we can write:\n\n\\begin{equation*} \\left| T_v^{(m,m')} \\right| \\geqslant \\left\\lfloor \\frac{N}{2^{K_0}} \\right\\rfloor\\geqslant \\left\\lfloor \\frac{3N}{8V} \\right\\rfloor= \\left\\lfloor \\frac{N}{4V}+\\frac{N}{8V} \\right\\rfloor \\geqslant \\left\\lfloor \\frac{N}{4V}+1 \\right\\rfloor \\geqslant \\frac{N}{4V}, \\end{equation*}\n\nwhere we used the fact that $V\\leqslant N/8$ by assumption.", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "the process above could be formulated as:\n\n$$\n\\begin{aligned}\n\t\t& \\bm{\\mu}_v = {\\rm{MLM}\\text{-}\\rm{head}}\\mathopen{}\\mathclose\\bgroup\\originalleft({\\rm{ReLU}}(\\bm{v} \\bm{W}_{\\mu} + \\bm{b}_{\\mu}) \\aftergroup\\egroup\\originalright) \\\\\n\t\t& \\bm{\\sigma}^2_v = {\\rm{exp}} \\mathopen{}\\mathclose\\bgroup\\originalleft( {\\rm{MLM}\\text{-}\\rm{head}} \\mathopen{}\\mathclose\\bgroup\\originalleft({\\rm{ReLU}}(\\bm{v} \\bm{W}_{\\sigma} + \\bm{b}_{\\sigma}) \\aftergroup\\egroup\\originalright) \\aftergroup\\egroup\\originalright),\n\t\\end{aligned}\n$$\n\nwhere $\\bm{W}_{\\mu}$, $\\bm{W}_{\\sigma}$, $\\bm{b}_{\\mu}$, $\\bm{b}_{\\sigma}$ are trainable parameters. $\\mathcal{N} (\\bm{\\mu}_v, \\bm{\\sigma}_v^2)$ represents the label token $v(y)$ with a soft region in the $\\mathcal{V}$ ’s semantic space.\n\n## 4.2.2. Uncertainty-aware Semantics Aggregation\n\nDue to the semantic limitation of one single token to express task labels [25], we employ multiple tokens to construct the verbalizer for the complicated semantic relation between candidate events and the event chain. We therefore design this module to aggregate their semantics.\n\nSupposing that a combination of $l$ tokens collaboratively serve as the label tokens, we first follow Eq. 4 to acquire their Gaussian embeddings as $\\mathcal{N}_1 (\\bm{\\mu}_{v,1}, \\bm{\\sigma}_{v,1}^2) , ... , \\mathcal{N}_l (\\bm{\\mu}_{v,l}, \\bm{\\sigma}_{v,l}^2)$. As prior research [5, 20] suggested, a token with larger uncertainty may carry some unnecessary noises, we thus employ a smooth exponential function to balance the contribution of these label tokens on weight $\\kappa_i$ as follows:\n\n$$\n\\kappa_i = {\\rm{exp}} (- \\lambda \\bm{\\sigma}_i),\n$$\n\nwhere $\\kappa_i$ is the weight for the $i_{\\rm{th}}$ label token and $\\lambda$ is a hyper-parameter. Consequently, we aggregate Gaussian embeddings of these tokens as follows:\n\n$$\n% \n\t\\hat{\\bm{\\mu}}_v =\\sum_{i=1}^{l} (\\kappa_i \\odot \\bm{\\mu}_i) \\quad \\hat{\\bm{\\sigma}}_v^2 =\\sum_{i=1}^{l} (\\kappa_i \\odot \\kappa_i \\odot \\bm{\\sigma}_i^2) ,\n$$\n\nwhere $\\odot$ denotes the element-wise product. Finally, $\\mathcal{N}(\\hat{\\bm{\\mu}_v}, \\hat{\\bm{\\sigma}}_v^2)$ is Gaussian embedding of the final label token $v(y)$.\n\n## 4.3. Prediction\n\nTo predict the subsequent event, we compute each candidate event’s score $\\verb|score|(e_{c_j})$ to justify whether it is the correct subsequent event. According to Eq. 1, $\\verb|score|(e_{c_j})$ equals to the probability of how $v(y)$ could fill the masked token $\\mathtt{[MASK]}$ in the pattern $f(e_{c_j})$ as follows:\n\n$$\n{\\rm{score}}(e_{c_j}) = P\\mathopen{}\\mathclose\\bgroup\\originalleft(\\mathtt{[MASK]}=v(y)| f(e_{c_j})\\aftergroup\\egroup\\originalright).\n$$\n\nSince all prompt tokens are represented as Gaussian distributional embeddings, we obtain their explicit represen- tations using Monte Carlo Sampling [22] technique. To make the sampling process differentiable, we further adopt the re-parameterization trick [27]. We illustrate this process by taking the first prompt token $\\mathtt{[t]}_{j,1}$, whose Gaussian embedding is $\\mathcal{N}(\\bm{\\mu}_{j, 1}, \\bm{\\sigma}^2_{j, 1})$ for $j_{th}$ candidate event, as an example. Specifically, its representation could be obtained as follows:\n\n$$\n\\bm{z}_{j, 1} = \\bm{\\mu}_{j, 1} + \\varepsilon \\bm{\\sigma}_{j, 1}, \\text{where} ~ \\varepsilon \\sim \\mathcal{N} (0, 1),\n$$\n\nwhere $(\\bm{\\mu}_{j, 1}, \\bm{\\sigma}^2_{j, 1})$ are acquired from Eq. 3 and $\\varepsilon$ is a random noise from the standard Gaussian distribution. The representations of other prompt tokens are derived in the same process above. After obtaining the representations of prompt tokens, we feed $f(e_{c_j})$ into $\\mathcal{M}$, where the $\\rm{MLM}\\text{-}\\rm{head}$ of $\\mathcal{M}$ could produce the logits, $\\mathcal{M}(f(e_{c_j}))$, towards the vocabulary. Next, a softmax layer over $\\mathcal{M}(f(e_{c_j}))$ yields the probability distribution $\\bm{p}_j \\in \\mathbb{R}^{1 \\times |\\mathcal{V}|}$ over the vocabulary. The process could be formulated as follows:\n\n$$\n\\bm{p}_j = {\\rm{softmax}} \\mathopen{}\\mathclose\\bgroup\\originalleft(\\mathcal{M}\\mathopen{}\\mathclose\\bgroup\\originalleft(f(e_{c_j}) \\aftergroup\\egroup\\originalright) \\aftergroup\\egroup\\originalright),\n$$\n\nwhere $\\bm{p}_j$ indicates how each token in $\\mathcal{V}$ is likely to fill $\\mathtt{[MASK]}$.", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "\\begin{tabular}{|c|c|c|}\n\\hline\nBending model & \\multicolumn{2}{c|}{Parameter set} \\hline\n& & [-3.5mm]\n3D linear elasticity & $E\\, = \\,480\\,E_0$ & $\\nu\\,=\\,0.38$ [0.5mm] \\hline\n& & [-3.5mm]\nKoiter & \\begin{tabular}{c}$\\mu\\,=\\,65.217\\,E_0L_0$$\\Lambda\\,=\\,79.944\\,E_0\\,L_0$ \\end{tabular} & $\\nu\\,=\\,0.38$ [0.5mm] \\hline\n& & [-3.5mm]\nHelfrich & \\begin{tabular}{c}$k\\,=\\,4.9308\\,E_0L^3$ $\\bar{k}\\,=\\,-1.5285\\,E_0L^3$$H_0\\,=\\,0$\\end{tabular} & $\\nu\\,=\\,0.38$ [0.5mm] \\hline\n& & [-3.5mm]\napH & $\\mu\\, =\\,60\\,E_0L_0$ & $\\nu\\,=\\,0.5$ [0.5mm] \\hline\n& & [-3.5mm]\nnew & \\begin{tabular}{c} $c_{1}\\,=\\,c_{2}\\,=\\,2.4654\\,E_0\\,L_0^3$$c_{12}\\,=\\,0.9368\\,E_0\\,L_0^3$$c_{3}\\,=\\,0.7643\\,E_0\\,L_0^3$\\end{tabular} & $\\nu\\,=\\,0.38$ [0.5mm] \\hline\n\\end{tabular}\n\nTable8: Simply supported plate: Parameter set used for different bending models according to Eqs. (53), (62) and (76).\n\nSimply supported plate: Problem setup with (a) regular mesh and (b) skew mesh. (c) Deformed configuration (scaled by $400$) for the skew mesh colored by the vertical displacement. (d) Difference of the center displacement between analytical and numerical solution for the new bending model. The solid lines show the results of the regular mesh, while the dashed lines show the results of the skew mesh. The displacements obtained for the other bending models are similar to the new bending model and hence omitted here.\n\n## Pinched cylindrical shell – linear case\n\nThis example analyzes a cylinder with rigid diaphragms at its ends and subjected to point forces as shown in Fig. 5. The cylinder is of dimension $R\\,=\\,300L_0$, $L\\,=\\,600L_0$ and has thickness $T\\,=\\,3L_0$. The rigid diaphragm is realized by fixing the $x$ - and $z$ - degrees-of-freedom of the nodes lying at the cylinder ends. The material parameters are Young’s modulus $E \\,=\\,3\\, E_0$ and Poisson’s ratio $\\nu\\,=\\,0.3$. The corresponding material parameters for the different bending models are shown in Tab. 9. The magnitude of the two opposing pinching forces is $F\\,=\\,10^{-6}\\,E_0L^2_0$.", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "# I NTERACTIVE M ODEL WITH S TRUCTURAL L OSS FOR L ANGUAGE - BASED A BDUCTIVE R EASONING\n\nLinhao Li 1, Ming Xu 1, Yongfeng Dong 1, Xin Li 2, Ao Wang 1 1 The School of Artificial Intellgence and the Hebei Provincial Key Laboratory of Big Data Computing, Hebei University of Technology 2 The Lane Department of Computer Science and Electrical Engineering, West Virginia University {lilinhao,dongyf} @hebut.edu.cn, [email protected], [email protected],[email protected]\n\nThe abductive natural language inference task ($\\alpha$ NLI) is proposed to infer the most plausible ex- planation between the cause and the event. In the $\\alpha$ NLI task, two observations are given, and the most plausible hypothesis is asked to pick out from the candidates. Existing methods model the relation between each candidate hypothesis separately and penalize the inference network uniformly. In this paper, we argue that it is unnecessary to distinguish the reasoning abilities among correct hypotheses; and similarly, all wrong hypotheses contribute the same when explaining the reasons of the observations. Therefore, we propose to group instead of ranking the hypotheses and design a structural loss called “ joint softmax focal loss ” in this paper. Based on the observation that the hypotheses are generally semantically related, we have designed a novel interactive language model aiming at exploiting the rich interaction among competing hypotheses. We name this new model for $\\alpha$ NLI: Interactive Model with Structural Loss (IMSL). The experimental results show that our IMSL has achieved the highest performance on the RoBERTa-large pretrained model, with ACC and AUC results increased by about 1% and 5% respectively.\n\n## Introduction\n\nAbductive natural language inference ($\\alpha$ NLI) (Bhagavatulaetal.,2020) is a newly established branch of natural language inference (NLI) and is an interesting task in the area of natural language processing (NLP) based commonsense reasoning. Originating from NLI which targets at the semantic relationship between the two sentences, $\\alpha$ NLI further estimates the abductive reasoning of each sentence by explicitly deducing its cause. In the past years, $\\alpha$ NLI has attracted increasing attentions as it makes NLP tools more explainable and comprehensible. As of today, typical applications of $\\alpha$ NLI include knowledge graph Completion (Yuetal.,2020)(Bauer&Bansal,2021), question answering (Maetal.,2021), sentence in-filling (Huangetal.,2020), knowledge integration (Zhouetal.,2021) and so on.\n\nTo better motivate this work, we have shown a comparison between NLI and $\\alpha$ NLI in Table 1. For NLI, the task is to judge the relationship between the premise statement $\\rm{P}$ and the hypothetical sentence $\\rm{H}$ based on the given information in $\\rm{P}$. Options of the answer can be implication, neutrality, or contradiction. For $\\alpha$ NLI, a pair of observations ($\\rm O_1$ and $\\rm O_2$) and some hypotheses (e.g., two competing hypotheses $\\rm H^1$ and $\\rm H^2$ in the example) are given. The task of $\\alpha$ NLI is to deduce the more plausible reason between $\\rm H^1$ and $\\rm H^2$ that can explain the situational change from $\\rm O_1$ to $\\rm O_2$. In", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Figure 14: Increasing SAV CO 2 sensitivity to sea-level $\\ensuremath{\\gamma_{\\textrm{SAV}}}$ with other CO 2 effects turned off: $\\ensuremath{\\gamma_{\\textrm{T}}} , \\ensuremath{\\gamma_{\\textrm{MOR}}}\n =0$Mt CO 2 /yr per cm/yr. Panel (a) is ice volume in metres sea level equivalent. Panel (b) is CO 2 concentration in the atmosphere. The values in the legend are roughly equivalent to the peak increase in SAV CO 2 emissions flux caused by a deglaciation where$\\dot{V}=1$ cm/yr.\n\nFigure 15: Increasing MOR CO 2 sensitivity to sea-level $\\ensuremath{\\gamma_{\\textrm{MOR}}}$ with other CO 2 effects turned off: $\\ensuremath{\\gamma_{\\textrm{T}}} , \\ensuremath{\\gamma_{\\textrm{SAV}}} =0$. Panel (a) is ice volume in metres sea level equivalent. Panel (b) is CO 2 concentration in the atmosphere. The values in the legend are roughly equivalent to the peak decrease in MOR CO 2 emissions flux caused by a deglaciation where $\\dot{V}=1$ cm/yr.\n\nincreases the amplitude of $C$ cycles. However, the response in ice volume is more complex; as $\\ensuremath{\\gamma_{\\textrm{MOR}}}$ is increased, different maxima and minima in $V$ become more extreme or suppressed. This occurs because, as in section 4.1, the MOR-based variability in $C$ opposes some insolation- driven changes in $V$ (and reinforces others). As $\\ensuremath{\\gamma_{\\textrm{MOR}}}$ increases further, the model moves towards glacial cycles at a multiple of the obliquity cycle: for $\\ensuremath{\\gamma_{\\textrm{MOR}}}=80$ Mt CO 2 /yr per cm/yr a $\\sim$ 100 kyr oscillation dominates in the final 300 kyrs of $C$ (fig 15 b).\n\nThese results for varying $C$ terms in eqn (8) are consistent with our suggestion that feedbacks need at least a $30$ kyr lag time to disrupt 40 kyr glacial cycles. This reinforces the conclusions from exploring varying MOR lag times in section 4.1.\n\nCollectively, the results presented so far imply that variable MOR CO 2 emissions can change both the amplitude and periodicity of glacial cycles in $C$ and $V$, and the short-timescale carbon feedbacks from subaerial volcanism and the surface system only change the amplitude of glacial cycles. However, the increased amplitude of sea level change causes increased amplitude of cumulative MOR CO $_2$ emissions, therefore when all CO 2 feedbacks are active in the model the short-timescale feedbacks will affect when the model generates $\\sim$ 100 kyr cycles.", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "- Non-trivial to sample data. Given a physical process, it is natural to consider constructing a synthetic paired training dataset by sampling inputs first and then generating outputs. However, an explicit prior of input parameters is normally difficult to obtain, while random sampling can be computationally infeasible and may end up with meaningless samples.\n\n- Generalization issues. Even when priors or samples of input parameters are available, a trained model may suffer the generalization problem when applied to unseen observations.\n\nTo tackle these challenges, we propose a novel analysis-by-synthesis procedure by iteratively sampling and training. At the sampling step, given an observed data, a neural network is used to approximate the intractable posterior of input parameters, then parameters are sampled from the posterior to generate outputs in the observational space through the physical process; At the training step, the same network is trained on the sampled paired data to predict input parameters from observations. These two steps operate iteratively and boost each other, similar to the guess-try-feedback process in human learning. The entire learning procedure integrates the physical generation process to obtain meaningful parameter estimates in a self-supervised manner. The proposed method can address the above-mentioned three challenges as follows:\n\n- The physical process is only used for generating outputs, without passing gradients through.\n\n- No need of priors on input parameters, the model learns from unlabelled data to approximate the posterior of input parameters for efficiently sampling.\n\n- The self-supervised nature of the training procedure means a trained model can adapt itself to unseen observations.\n\nWe verify the proposed method by tackling the acoustic-to-articulatory inversion problem to extract articulatory kinematics information from speech. We adopt the Tube Resonance Model (TRM) [14] to simulate the human vocalization mechanism. Given unlabelled reference utterances and starting completely from scratch with random initialization, a network can learn to infer how to control the TRM model to synthesize sound similar to the reference utterances by alternately sampling data and training itself. Experiments show that our proposed algorithm can converge steadily. The synthesized sounds achieve a signal to noise ratio of around 16dB on both single-speaker and multi-speaker datasets. Further experiments show that trained models can generalize well to unseen speakers or even a new language, and performance can be further promoted through self-adaptation.\n\n## Related Work\n\nInverseProblem Inverse problems aim to infer causes from observations. Traditionally, analytical methods have been used to solve inverse problems such as image super-resolution (SR) [25], computed tomography (CT) [23]. In recent years, data-driven methods especially the deep learning approach has achieved state-of-the-art results [4,21]. For those problems with a differentiable or invertible forward process, methods have been proposed to solve problems like SR [28]. Otherwise, the key to solve an inverse problem in a data-driven way is to obtain supervision data. When a physical forward process is available, paired training data can be synthetically generated by applying the forward process to samples of input parameters (e.g., adding noise to clean images for denoising training). However, appropriate priors or samples of the input parameters are not always available, while the parameter space can be high dimensional, making input parameters sampling non-trivial.\n\nTo solve an inverse problem for a given observed data, sampling from the posterior distribution can be efficient. Markov chain Monte Carlo (MCMC) is a widely used sampling method, but it requires an explicit prior and can be computationally infeasible in a high-dimensional space. As for deep sampling methods, [1] proposed a method to directly sample from the posterior with a conditional generative adversarial network (CGAN), but this method requires supervised training. Variational auto-encoders (VAEs) [9] directly approximate the posterior of latent variables with a neural network. Similar to that, we use a neural network to approximate the posterior of parameters and sample from it to generate paired training data. However, since the forward process (corresponding to the decoder in VAE) can be non-differentiable, we turn to minimize the reconstruction error of latent input parameters rather than observed data. Acoustic-to-ArticulatoryInversion(AAI) The motor theory [20] indicates that when perceiving speech, except for acoustic features, human also perceive articulatory information. There has been", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "working environment already prior to the COVID-19 pandemic 2, the discus- sion has now gained further traction.\n\nMore specifically, the discussion is subject on the newly introduced Di- versity & Inclusion (D & I) score by Refinitiv (formerly ASSET4) (Refinitiv, 2022a). The score is built using 24 metrics defined on four main pillars linked to workplace well-being (Refinitiv, 2022a). These are: diversity (e.g., gender diversity and diversity process), inclusion (e.g., flexible working hours and daycare services), people development (e.g., internal promotion, employee satisfaction, and career development process), and news and controversies (e.g., diversity and opportunity and wages and working conditions) (Refini- tiv, 2022a). The input data is gathered from publicly available sources, and an overall D & I score is computed. The D & I score ranges from 0 to 100, 0 indicating poor D & I performance. According to Refinitiv (2022a), this score shall provide insights into two main issues employers face: cultural diversity and workplace transition.\n\nDifferent scholars have already shown that diversity can positively affect economic performance and firm value. 3 Others find mixed effects using a sample of Bangladesh companies (Dutta and Bose, 2006) and Italian compa- nies (Gordini and Rancati, 2017). Additionally, diversity and inclusion have been linked to workplace happiness, productivity, and better performance in challenging situations (Vasa and Thatta, 2018; Arenas, 2017; Frey and Stutzer, 2010; Warr, 2011).\n\nWhile this topic has been explored from a company management and hu- man resource perspective, happiness and satisfaction can technically trans- late into higher financial performance measures through different channels (Kessleretal.,2020;Refinitiv,2022b), including increased productivity (Utami etal.,2018) or employee well-being (Krekeletal.,2019). However, measuring and predicting workplace happiness can be complex and different measures\n\n2 These beneficial effects can take place either directly (Krapivin, 2018; Refinitiv, 2022b), indirectly through customer satisfaction (Chi and Gursoy, 2009), or in an in- vestment setting (Mishra, 2018), and also help the organization to grow (Spreitzer and Porath, 2012).\n\n3 See for diversity in general in the US (Erhardt et al., 2003), diversity on boards in Spain (CampbellandM´ınguez-Vera,2008;Reguera-Alvaradoetal.,2017) and in Malaysia (Lee-Kuen et al., 2017)), diversity on top management teams in Denmark (Opstrup and Villadsen, 2015)).", "role": "assistant" } ]

[ { "content": "<image>Extract the exact text from this image of the correct order, reading left to right, top to bottom. Include all figure names, math equations, newline characters (\\n), and any other essential symbols. Don't output 'subsection{}' or 'section{}', instead, output plain text. Include all headers and section headers, all footers, page numbers, figure captions and table captions. Pay special attention to punctuations and make sure they are correct. Make sure you output all the texts in the image. When there are two columns in the image, make sure to output texts in each column. Omit any descriptions, additional punctuations, or symbols that are not part of the visible text. Don't output latex format unless it's math equations.", "role": "user" }, { "content": "Xmakefnmark makefnmark Xthefnmark thefnmark makefnmark thefnmark footnotetext AMS 2010 sub- ject classification: 60H15; 65C30 footnotetext Keywords and phrases: stochastic partial differential equations, exponential moments, stochastic Burgers equations, tamed euler approximations, strong convergence rates makefnmark Xmakefnmark thefnmark Xthefnmark", "role": "assistant" } ]