arXiv subjects

Véronique Gayrard

Publications and source records attributed to Véronique Gayrard.

14 recordsLinked to original sources

arXiv2026

Mixed memories in Hopfield networks

We consider the class of Hopfield models of associative memory with activation function $F$ and state space $\{-1,1\}^N$, where each vertex of the cube describes a configuration of $N$ binary neurons. $M$ randomly chosen configurations, called patterns, are stored using an energy function designed to make them local minima. If they are, which is known to depend on how $M$ scales with $N$, then they can be retrieved using a dynamics that decreases the energy. However, storing the patterns in the energy function also creates unintended local minima, and thus false memories. Although this has been known since the earliest work on the subject, it has only been supported by numerical simulations and non-rigorous calculations, except in elementary cases. Our results are twofold. For a generic function $F$, we explicitly construct a set of configurations, called mixed memories, whose properties are intended to characterise the local minima of the energy function. For three prominent models, namely the classical, the dense and the modern Hopfield models, obtained for quadratic, polynomial and exponential functions $F$ respectively, we give conditions on the growth rate of $M$ which guarantee that, as $N$ diverges, mixed memories are fixed points of the retrieval dynamics and thus exact minima of the energy. We conjecture that in this regime, all local minima are mixed memories.

math.PR↗

arXiv2026

Integral Means Spectrum for the Random Riemann Zeta Function

We study the integral means spectrum associated with the analytic function whose derivative is the so-called randomized Riemann zeta-function, introduced some time ago by Bagchi. The randomized $ζ$-function, $ζ_{\mathrm{rand}}(σ+ih)$, is known to represent the asymptotic statistical behaviour of the random vertical shifts of the actual $ζ$-function in the critical strip, $1/2 <σ\leq 1, h\in \mathbb R$, and appears in a number of recent works on the asymptotic behavior of the moments and maxima of the $ζ$-function on short intervals along the critical axis $σ=1/2$. Using probability and basic analytic number theory, we show that the complex integral means spectrum of the primitive of $ζ_{\mathrm{rand}}$ is almost surely of the form conjectured 30 years ago by Kraetzer, for the so-called universal integral means spectrum of univalent functions in the disc. The Riemann $ζ$-function and its random version have recently been rigorously related to the so-called Gaussian multiplicative chaos (GMC), initiated by Kahane 40 years ago. In the case of the holomorphic multiplicative chaos on the unit disc -- an important stochastic object closely related to Liouville quantum gravity on the unit circle -- we prove that the integral means spectrum of the primitive is almost surely also of the same Kraetzer form. However, we establish that neither the primitive of the random function $ζ_{\mathrm{rand}}$, nor that of the holomorphic GMC are injective. Building on earlier work by one of the authors and Webb on the convergence of Riemann $ζ$-function on the critical line to a holomorphic GMC distribution, we finally provide an alternative derivation of the integral means spectrum for the random Riemann $ζ$-function.

math.CV↗

arXiv2023

Emergence of near-TAP free energy functional in the SK model at high temperature

We study the SK model at inverse temperature $β>0$ and strictly positive field $h>0$ in the region of $(β,h)$ where the replica-symmetric formula is valid. An integral representation of the partition function derived from the Hubbard-Stratonovitch transformation combined with a duality formula is used to prove that the infinite volume free energy of the SK model can be expressed as a variational formula on the space of magnetisations, $m$. The resulting free energy functional differs from that of Thouless, Anderson and Palmer (TAP) by the term $-\frac{β^2}{4}\left(q-q_{\text{EA}}(m)\right)^2$ where $q_{\text{EA}}(m)$ is the Edwards-Anderson parameter and $q$ is the minimiser of the replica-symmetric formula. Thus, both functionals have the same critical points and take the same value on the subspace of magnetisations satisfying $q_{\text{EA}}(m)=q$. This result is based on an in-depth study of the global maximum of this near-TAP free energy functional using Bolthausen's solutions of the TAP equations, Bandeira & van Handel's bounds on the spectral norm of non-homogeneous Wigner-type random matrices, and Gaussian comparison techniques. It holds for $(β,h)$ in a large subregion of the de Almeida and Thouless high-temperature stability region.

math.PR↗

arXiv2019

Dynamic phase diagram of the REM

By studying the two-time overlap correlation function, we give a comprehensive analysis of the phase diagram of the Random Hopping Dynamics of the Random Energy Model (REM) on time-scales that are exponential in the volume. These results are derived from the convergence properties of the clock process associated to the dynamics and fine properties of the simple random walk in the $n$-dimensional discrete cube.

math.PR↗

arXiv2018

Asymptotic behavior and aging of a low temperature cascading 2-GREM dynamics at extreme time scales

We derive scaling limit results for the Random Hopping Dynamics for the cascading two-level GREM at low temperature at extreme time scales. It is known that in the cascading regime there are two static critical temperatures. We show that there exists a (narrow) set of fine tuning temperatures; when they lie below the static lowest critical temperature, three distinct dynamical phases emerge below the lowest critical temperature, with three different types of limiting dynamics depending on whether the temperature is (well) above or below, or at a fine tuning temperature, all of which are given in terms of K processes. We also derive scaling limit results for temperatures between the lowest and he highest critical ones, as well as aging results for all the limiting processes mentioned above, by taking a second small time limit.

math.PR↗

arXiv2016

Aging in Metropolis dynamics of the REM: a proof

We study the aging behavior of the Random Energy Model (REM) evolving under Metropolis dynamics. We prove that a classical two-time correlation function converges almost surely to the arcsine law distribution function that characterizes activated aging, as predicted in the physics literature, in the optimal domain of the time-scale and temperature parameters where this result can be expected to hold. In the course of the proof we establish that a certain continuous time clock process, after proper rescaling, converges almost surely to a stable subordinator, improving upon the result of Cerny and Wassmer (2015) where a closely related clock is shown to converge in probability only, and in a restricted region of the time-scale and temperature parameters. The random rescaling involved in this convergence is controlled at the fine level of fluctuations. As a byproduct, we refine and prove a conjecture made in Cerny and Wassmer (2015).

math.PR↗

arXiv2015

Convergence of clock processes on infinite graphs and aging in Bouchaud's asymmetric trap model on ${\Bbb Z}^d$

Using a method developed by Durrett and Resnick [22] we establish general criteria for the convergence of properly rescaled clock processes of random dynamics in random environments on infinite graphs. This complements the results of [26], [19], and [20]: put together these results provide a unified framework for proving convergence of clock processes. As a first application we prove that Bouchaud's asymmetric trap model on ${\Bbb Z}^d$ exhibits a normal aging behavior for all $d\geq 2$. Namely, we show that certain two-time correlation functions, among which the classical probability to find the process at the same site at two time points, converge, as the age of the process diverges, to the distribution function of the arcsine law. As a byproduct we prove that the fractional kinetics process ages.

math.PR↗

arXiv2015

Aging in the GREM-like trap model

The GREM-like trap model is a continuous time Markov jump process on the leaves of a finite volume $L$-level tree whose transition rates depend on a trapping landscape built on the vertices of the whole tree. We prove that the natural two-time correlation function of the dynamics ages in the infinite volume limit and identify the limiting function. Moreover, we take the limit $L\to\infty$ of the two-time correlation function of the infinite volume $L$-level tree. The aging behavior of the dynamics is characterized by a collection of clock processes, one for each level of the tree. We show that for any $L$, the joint law of the clock processes converges. Furthermore, any such limit can be expressed through Neveu's continuous state branching process. Hence, the latter contains all the information needed to describe aging in the GREM-like trap model both for finite and infinite levels.

math.PR↗

arXiv2014

Convergence of clock processes and aging in Metropolis dynamics of a truncated REM

We study the aging behavior of a truncated version of the Random Energy Model evolving under Metropolis dynamics. We prove that the natural time-time correlation function defined through the overlap function converges to an arcsine law distribution function, almost surely in the random environment and in the full range of time scales and temperatures for which such a result can be expected to hold. This establishes that the dynamics ages in the same way as Bouchaud's REM-like trap model, thus extending the universality class of the latter model. The proof relies on a clock process convergence result of a new type where the number of summands is itself a clock process. This reflects the fact that the exploration process of Metropolis dynamics is itself an aging process, governed by its own clock. Both clock processes are shown to converge to stable subordinators below certain critical lines in their time-scale and temperature domains, almost surely in the random environment.

math.PR↗

arXiv2013

Convergence of clock processes in random environments and ageing in the p-spin SK model

We derive a general criterion for the convergence of clock processes in random dynamics in random environments that is applicable in cases when correlations are not negligible, extending recent results by Gayrard [(2010), (2011), forthcoming], based on general criterion for convergence of sums of dependent random variables due to Durrett and Resnick [Ann. Probab. 6 (1978) 829-846]. We demonstrate the power of this criterion by applying it to the case of random hopping time dynamics of the p-spin SK model. We prove that on a wide range of time scales, the clock process converges to a stable subordinator almost surely with respect to the environment. We also show that a time-time correlation function converges to the arcsine law for this subordinator, almost surely. This improves recent results of Ben Arous, Bovier and Cerny [Comm. Math. Phys. 282 (2008) 663-695] that obtained similar convergence results in law, with respect to the random environment.

math.PR↗

arXiv2010

Aging in reversible dynamics of disordered systems. II. Emergence of the arcsine law in the random hopping time dynamics of the REM

Applying the new tools developed in [G1], we investigate the arcsine aging regime of the random hopping time dynamics of the REM. Our results are optimal in several ways. They cover the full time-scale and temperature domain where this phenomenon occurs. On this domain the limiting clock process and associated time correlation function are explicitly constructed. Finally, all convergence statements w.r.t. the law of the random environment are obtained in the strongest sense possible, except perhaps on the very last scales before equilibrium.

math.PR↗

arXiv2010

Aging in reversible dynamics of disordered systems. I. Emergence of the arcsine law in Bouchaud's asymmetric trap model on the complete graph

In this paper the celebrated arcsine aging scheme of G. Ben Arous and J. Cerný is taken up. Using a brand new approach based on point processes and weak convergence techniques, this scheme is implemented in a wide class of Markov processes that can best be described as Glauber dynamics of discrete disordered systems. More specifically, conditions are given for the underlying clock process (a partial sum process that measures the total time elapsed along paths of a given length) to converge to a subordinator, and this subordinator is constructed explicitly. This approach is illustrated on Bouchaud's asymmetric trap model on the complete graph for which aging is for the first time proved, and the full, optimal picture, obtained.

math.PR↗

arXiv1995

The Retrieval Phase of the Hopfield Model: A Rigorous Analysis of the Overlap Distribution

Standard large deviation estimates or the use of the Hubbard-Stratonovich transformation reduce the analysis of the distribution of the overlap parameters essentially to that of an explicitly known random function $Φ_{N,\b}$ on $\R^M$. In this article we present a rather careful study of the structure of the minima of this random function related to the retrieval of the stored patterns. We denote by $m^*(\b)$ the modulus of the spontaneous magnetization in the Curie-Weiss model and by $\a$ the ratio between the number of the stored patterns and the system size. We show that there exist strictly positive numbers $0<\g_a<\g_c$ such that 1) If $\sqrt\a\leq \g_a (m^*(\b))^2$, then the absolute minima of $Φ$ are located within small balls around the points $\pm m^*e^μ$, where $e^μ$ denotes the $μ$-th unit vector while 2) if $\sqrt\a\leq \g_c (m^*(\b))^2$ at least a local minimum surrounded by extensive energy barriers exists near these points. The random location of these minima is given within precise bounds. These are used to prove sharp estimates on the support of the Gibbs measures. KEYWORDS: Hopfield model, neural networks, storage capacity, Gibbs measures, self-averaging, random matrices

cond-mat↗

arXiv1995

An almost sure large deviation principle for the Hopfield model

We prove a large deviation principle for the finite dimensional marginals of the Gibbs distribution of the macroscopic `overlap'-parameters in the Hopfield model in the case where the number of random patterns, $M$, as a function of the system size $N$ satisfies $\limsup M(N)/N=0$. In this case the rate function (or free energy as a function of the overlap parameters) is independent of the disorder for almost all realization of the patterns and given by an explicit variational formula.

cond-mat↗