Séminaire de mathématiques appliquées (archives)

We consider the problem of learning a graph modeling the statistical relations of the d variables of a dataset with n samples. Standard approaches amount to searching for a precision matrix representative of a Gaussian graphical model that adequately explains the data. However, most maximum likelihood-based estimators usually require storing the d^2 values of the empirical covariance matrix, which can become prohibitive in a high-dimensional setting. In this talk, we adopt a “compressive” viewpoint and aim to estimate a sparse precision matrix from a sketch of the data, i.e., a low-dimensional vector of size m≪d^2 carefully designed from the data using nonlinear random features (e.g., rank-one projections). Under certain spectral assumptions, we show that it is possible to recover the precision matrix from a sketch of size m=Ω((d+2k)log(d)), where k is the maximal number of edges of the underlying graph. These information-theoretic guarantees are inspired by the compressed sensing theory. We investigate the possibility of achieving practical recovery with an iterative algorithm based on the graphical lasso, viewed as a specific denoiser. We compare our approach and the graphical lasso on synthetic datasets, demonstrating its favorable performance even when the dataset is compressed. Joint work with : Titouan Vayer, Rémi Gribonval and Paulo Gonçalves.

Gaussian fields with logarithmically decaying correlations, such as branching Brownian motion and the two-dimensional Gaussian free field, are conjectured to form universality class of extreme value statistics (notably in the work of Carpentier & Le Doussal and Fyodorov & Bouchaud). This class is the borderline case between the class of IID random variables, and models where correlations start to affect the statistics. In this talk, I will describe a general approach based on rigorous works in spin glass theory to describe features of the Gibbs measure of these Gaussian fields. I will focus on the two-dimensional discrete Gaussian free field. At low temperature, we show that the normalized covariance of two points sampled from the Gibbs measure is either 0 or 1. This is used to prove that the joint distribution of the Gibbs weights converges in a suitable sense to that of a Poisson-Dirichlet variable. (with L.-P. Arguin, 2015).

In a second work (with M. Pain, 2021), we prove absence of temperature chaos for the two-dimensional discrete Gaussian free field using the convergence of the full extremal process, which has been obtained by Biskup and Louidor. This means that the overlap of two points chosen under Gibbs measures at different temperatures has a nontrivial distribution. Whereas this distribution is the same as for the random energy model when the two points are sampled at the same temperature, we point out here that they are different when temperatures are distinct: more precisely, we prove that the mean overlap of two points chosen under Gibbs measures at different temperatures for the DGFF is strictly smaller than the REM's one. Therefore, although neither of these models exhibits temperature chaos, one could say that the DGFF is more chaotic in temperature than the REM.

Finally, I will discuss in detail (depending on the time left) recent works with B. Bonnefont (ex-PhD student, now Post-Doc at University of Geneva) and M. Pain (CR @Toulouse), on questions suggested by B. Derrida.

This talk is about learning the parameter-to-solution map for systems of partial differential equations (PDEs) that depend on a potentially large number of parameters covering all PDE types for which a stable variational formulation (SVF) can be found. A central constituent is the notion of variationally correct residual loss function, meaning that its value is always uniformly proportional to the squared solution error in the norm determined by the SVF, hence facilitating rigorous a posteriori accuracy control. It is based on a single variational problem, associated with the family of parameter dependent fiber problems, employing the notion of direct integrals of Hilbert spaces. Since in its original form, the loss function is given as a dual test norm of the residual, a central objective is to develop equivalent computable expressions. A first critical role is played by hybrid hypothesis classes, whose elements are piecewise polynomial in (low-dimensional) spatio-temporal variables with parameter-dependent coefficients that can be represented, for example, by neural networks. Second, working with first order SVFs, we distinguish two scenarios: (i) the test space can be chosen as an L2-space (such as for elliptic or parabolic problems) so that residuals can be evaluated directly as elements of L2; (ii) when trial and test spaces for the fiber problems depend on the parameters (as for transport equations), we use ultraweak formulations. In combination with discontinuous Petrov-Galerkin concepts, the hybrid format is then instrumental to arrive at variationally correct computable residual loss functions. Our findings are illustrated by numerical experiments representing (i) and (ii), namely elliptic boundary value problems with piecewise constant diffusion coefficients and pure transport equations with parameter-dependent convection fields.

This presentation focuses on the Vlasov-Poisson system with and without collisions. This kinetic model encodes the multiple scales that arise in a plasma, ranging from fluid-like behavior when collisions dominate, to wave interactions in weakly collisional regimes. We present a numerical method for the Vlasov-Poisson system which preserves its structure in both collision and collisionless regimes. We explain the key ideas in order to preserve the energy structure of the system and its large time behavior in collisional settings. We also show that the method adapts to higher dimensional frameworks.

Dans cet exposé, j'étudierai un modèle discret d'endommagement brutal dans différents régimes où la zone endommagée se concentre sur des ensembles infiniment petits. Nous identifierons la nature des modèles limites obtenus au moyen d'une analyse asymptotique basée sur la Gamma-convergence des énergies totales. Je commencerai par rappeler un modèle mécanique d'endommagement brutal introduit par Francfort et Marigo (1993), spécifié dans le cadre discret (bidimensionnel) où les énergies totales sont restreintes à des déplacements continus et affines par morceaux. Plus précisément, dans le contexte de l'endommagement brutal, nous considérons un matériau linéairement élastique composé de deux phases pures : une phase endommagée dont les propriétés élastiques sont affaiblies, et une phase saine. En introduisant des petits paramètres dans l'énergie totale de Francfort et Marigo, nous forçons les propriétés élastiques de l'état endommagé à dégénérer vers 0 tandis que les zones endommagées se concentrent sur des ensembles Lebesgue-négligeables. Selon les différents régimes asymptotiques de ces petits paramètres, nous obtenons cinq modèles mécaniques que je vous présenterai. J'essaierai de vous montrer les différences et idées clés de leurs démonstrations.

On s'intéresse aux marches aléatoires en milieu aléatoire. La difficulté principale dans l'étude de ces systèmes est l'interaction très forte entre la marche et son environnement. Dans cet exposé, on propose un critère qui, lorsqu'il est satisfait, permet de décomposer la trajectoire de la marche en incréments iid, et à terme de démontrer des théorèmes limites. Le critère porte sur l'environnement et implique la construction d'un champ aléatoire qui satisfait une certaine propriété de Markov ainsi que des estimées de décorrélation. On applique ce critère à des environnements corrélés en temps comme la percolation booléenne sur Z^d x N et des chaînes de renouvellement iid en espace. Basé sur https://arxiv.org/abs/2409.12515 (travail réalisé avec J. Allasia, R. Baldasso et A. Teixeira).

Le problème de concentration spectrale est un problème posé dans les années 1960 par Slepian, Landau et Pollak. Ils se sont intéressés aux fonctions qui ont une norme L^2([-1, 1]) maximale et pour lesquelles la transformée de Fourier est supportée dans [-c, c], c>0. Leurs résultats ont été ensuite utilisés dans de nombreux domaines d'applications. Dans cet exposé nous parlerons d'une généralisation du problème de concentration spectrale qui utilises des masques en espace et en Fourier, et nous en donnerons quelques propriétés basiques. En s'attardant sur le cas des masques binaires, nous étendrons le résultat de commutation au coeur des travaux de Slepian, Landau et Pollak. Le cas des masques gaussiens sera aussi abordé, et nous verrons alors comment décrire exactement les solutions du problème. Enfin, nous passerons à la résolution numérique dans le cas des masques binaires, puisque la présence de "quasi-clusters" de valeurs propres rend difficile la recherche de vecteurs propres. Nous donnerons une procédure numérique alternative, et comparerons ses résultats sur plusieurs exemples numériques.

We present an extension of the results obtained by Colombo and Perrollaz regarding the set of inverse designs for a class of scalar conservation laws with compact space dependency. The key ingredients are the notion of generalized characteristics of Dafermos and the correspondence with the associated Hamilton-Jacobi equation. Numerical simulations are presented to highlight the differences with the homogeneous case.

Dans cet exposé, nous nous intéresserons aux méthodes de décomposition de domaine appliquées à un problème de contrôle optimal linéaire quadratique. Après un rapide retour sur les méthodes de résolution numérique pour les problèmes de contrôle optimal, je présenterai deux approches pour faire de la décomposition de domaine pour ces problèmes en faisant un parallèle avec les méthodes numériques classiques.

We consider the simple random walk on the Euclidean lattice in transient dimensions. It is known that the number of distinct visited sites is asymptotically linear in time. The probability of visiting a smaller number of sites, with a difference of the order of the mean, was evaluated asymptotically by Phetpradap in 2010, taking up the seminal work of van den Berg, Bolthausen and den Hollander in 2001 concerning the volume of a Wiener sausage. We consider the random walk conditioned to such a rare event and prove that the occupation measure of a certain random walk skeleton converges to a unique optimal profile modulo space shift, provided the deviation from the mean is large enough if dimension is four or higher. Our proof of this so-called tube property relies on the recent compactification of the space of measures introduced by Mukherjee and Varadhan, and it is a first step in the rigourous proof of the Swiss cheese picture proposed by van den Berg, Bolthausen and den Hollander. This is joint work with Dirk Erhard (Salvador de Bahia, Brazil).

References : M. van den Berg, E. Bolthausen and F. den Hollander, Moderate deviations for the volume of the Wiener sausage, Ann. of Math. (2) 153 (2001), no.~2, 355--406; MR1829754 C. Mukherjee and S. R. S. Varadhan, Brownian occupation measures, compactness and large deviations, Ann. Probab. 44 (2016), no.~6, 3934--3964; MR3572328