Responsable : Paul Thévenin
Séminaires à venir
Je présenterai de nouvelles formules explicites pour toutes les densités de volume des faces et tous les volumes des faces typiques de l'IPVT de l'espace hyperbolique réel de dimension $d$, pour $d \geq 2$. Celles-ci découlent d'une nouvelle formule de type Blaschke–Petkantschin pour les mosaïques de Poisson–Voronoi d'intensité positive, exprimant les densités de volume des faces via les moments du facteur de croissance de surface hyperbolique évalué au rayon du plus proche voisin d'un point typique. Ces dernières formules généralisent largement des résultats antérieurs d'Isokawa, qui se limitaient aux dimensions deux et trois. Exposé basé sur un travail commun avec Christoph Thäle (2606.26049, soumis).
TBD
TBD
TBD
TBA
TBD
TBA
Séminaires passés
Les surfaces browniennes sont des surfaces aléatoires continues qui apparaissent comme limites de grandes surfaces discrètes choisies au hasard. Elles constituent aujourd’hui des modèles fondamentaux de géométrie aléatoire. Cet exposé retracera certaines propriétés des surfaces browniennes. On verra notamment qu'elles satisfont une propriété de Markov spatiale. Aucun prérequis particulier n’est nécessaire : l’objectif sera surtout de donner une intuition des objets et des méthodes, plutôt que d’entrer dans les détails techniques.
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<
Consensus-Based Optimization (CBO) is a derivative-free method for finding the global minimum of a possibly non-convex function. It uses a population of particles that explore the search space together. At each step, the particles compute a consensus point, a smooth approximation of the best particle at the previous iteration. Each particle is then pulled towards this consensus point and perturbed by random noise. The pull concentrates the population around promising regions, while the noise keeps it exploring and helps it escape poor local minima. Most existing results establish the convergence of CBO only over a finite time horizon. In this talk, we study its convergence in the long-time regime and establish theoretical guarantees for the finite-particle algorithm. In particular, we quantify the different sources of error and study the interplay between the number of particles, the number of iterations, the discretization step and the smoothing parameter. We then introduce a new variant, closely linked to CBO, that adds some noise to the CBO dynamics. Instead of collapsing onto a minimizer, the particle system then explores the target distribution and becomes a gradient-free sampler.
Un cladogramme est un arbre binaire (non planaire et non enraciné) dont les feuilles sont étiquetées. Couper une feuille uniformément au hasard et la réinsérer au milieu d’une arête uniforme définit une dynamique, dont nous avons montré que le temps de mélange est O(n²), conformément à une conjecture d’Aldous. L’exposé présentera un couplage multicolore et d’autres arguments, assez peu techniques, qui ont mené à ce résultat. Travail réalisé en collaboration avec Valentin Féray.
We introduce a novel approach, called E-local stationarity, that bridges extreme value theory and inference for heavy-tailed, multivariate, and nonstationary processes. For heavy-tailed nonstationary processes, conventional notions of local stationarity may be inadequate. First, the process may lack finite moments, preventing certain distance measures from being well-defined. Second, standard measures of proximity may fail to capture the behavior of extremes, which are crucial in the heavy-tailed setting. To overcome these shortcomings, our definition of E-local stationarity focuses exclusively on the limiting behavior of tail probabilities of nonstationary multivariate time series. We show that several widely used processes satisfy this definition, including heavy-tailed linear processes, M4 processes, and stochastic volatility models. We build upon this framework to develop nonparametric kernel regression estimators of the (local) tail dependence measure for a nonstationary multivariate time series, and for its extreme quantiles when it is univariate. The asymptotic normality of the estimators is established under a list of reasonable conditions. Finally, we demonstrate the practical relevance of the proposed theory through simulations and an application to data on European stock market indices.
We introduce and study V-monotone independence, which can be considered as a combination of two twin models of independence, monotone independence and antimonotone independence, into one model. We investigate the combinatorics of mixed moments of V-monotone independent random variables and prove the central and Poisson limit theorem. We obtain a combinatorial formula for the limit moments and we find the limit measures. [1] Adrian Dacko. V-monotone independence. Colloq. Math., 162(1):77–107, 2020. [2] Adrian Dacko. Central limit theorem for V-monotone independence. Complex Anal. Oper. Th., 19(128), 2025. [3] Adrian Dacko. Poisson limit theorems for V-monotone independence. in preparation, 2026. [4] Adrian Dacko and Lahcen Oussi. Distribution for nonsymmetric V-monotone position operators. in preparation, 2026.
Federated learning on Riemannian manifolds enables collaborative training without centralized data pooling when model parameters are intrinsically constrained. Existing methods either rely on geometric operations lacking closed-form expressions on key manifolds (such as the Stiefel manifold), employ optimizer-specific gradient streams that incur information loss through successive transports, or -- even when computationally cheap -- require drift correction terms and full client participation. We propose two aggregation strategies, RFedProj and RFedRL, that are optimizer-agnostic, lightweight (requiring only standard projections and retractions), and support partial participation without auxiliary correction. Both achieve identical convergence rates under these relaxed conditions and data heterogeneity, with bounds explicitly characterizing how participation ratio and heterogeneity interact -- mirroring classical Euclidean federated guarantees. Experiments on EEG motor imagery classification with SPDNet on compact Stiefel manifolds validate competitive performance against centralized and Euclidean baselines.
We consider the framework of least-squares penalized estimation where the penalty term is given by a polyhedral norm, or more generally, a polyhedral gauge, which encompasses methods such as LASSO and. generalized LASSO, SLOPE, OSCAR, PACS and others. Each of these estimators can uncover a different structure or “pattern”of the unknown parameter vector. We define a novel and general notion of patterns based on subdifferentials and formalize an approach to measure pattern complexity. For pattern recovery, we provide a minimal condition for a particular pattern to be detected with positive probability, the so-called accessibility condition. We make the connection to estimation uniqueness by showing that uniqueness holds if and only if no pattern with complexity exceeding the rank of the X-matrix is accessible. Subsequently, we introduce the noiseless recovery condition which is a stronger requirement than accessibility and which can be shown to play exactly the same role as the well-known irrepresentability condition for the LASSO – in that the probability of pattern recovery is bounded by 1/2 if the condition is not satisfied. Through this, we unify and extend the irrepresentability condition to a broad class of penalized estimators using an interpretable criterion. We also look at the “gap” between accessibility and the noiseless recovery condition and discuss that our criteria show that it is more pronounced for simple patterns. Finally, we prove that the noiseless recovery condition can indeed be relaxed when turning to so-called thresholded penalized estimation: in this setting, the accessibility condition is already sucient (and necessary) for sure pattern recovery provided that the signal of the pattern is large enough. We demonstrate how our findings can be interpreted through a geometrical lens throughout the talk and illustrate our results for the Lasso as well as other estimation procedures. Related references: P. Graczyk, U. Schneider, T. Skalski and P. Tardivel, A Unified Framework for Pattern Recovery in Penalized and Thresholded Estimation and its Geometry; https://doi.org/10.1007/s10957-025-02863-6 U. Schneider and P. Tardivel, The Geometry of Uniqueness, Sparsity and Clustering in Penalized Estimation; https://jmlr.org/papers/v23/21-0420.html K. Ewald and U. Schneider, On the Distribution, Model Selection Properties and Uniqueness of the Lasso Estimator in Low and High Dimensions; https://doi.org/10.1214/20-EJS1687
Je présenterai une toute nouvelle méthode pour montrer, à partir de 0, le comportement champ moyen de la percolation critique en dimension d>6. L’idée principale est pour p
In this talk, we address the stability problem of the famous Brascamp-Lieb inequality for strictly log-concave probability measures on the Euclidean space. More precisely, if a given function almost satisfies the equality in the BL inequality, is it true that it is close in some sense to the underlying extremal functions? Using a spectral interpretation of the BL inequality, we prove that the distance to the extremal functions in quadratic norm is of order square root of the deficit parameter, and involves the second positive eigenvalue of a convenient diffusion operator we wish to estimate. Our results are illustrated by some examples for which the usual uniform convexity assumption on the potential is relaxed. This is a joint work with M. Bonnefont (Institut de Mathématiques de Bordeaux) and J. Serres (Sorbonne Université).
The current surge in data has led to a significant increase in the size of networks used to model relationships between different objects represented as nodes. Therefore, summarizing network information is a crucial task that can be conducted using node clustering methods. Additionally, to ensure interpretable results, it is essential to employ relevant visualisation techniques to depict the network. To tackle both issues, we propose a new methodology called the deep latent position block model (Deep LPBM). This simultaneously provides a network visualisation coherent with block modelling, allowing a clustering more general than community detection methods, as well as a continuous representation of nodes in a latent space given by partial membership vectors. Deep LPBM is based on a variational autoencoder strategy, relying on a graph convolutional network, with a specifically designed decoder. The inference involves the construction of an approximation of the marginal likelihood of Deep LPBM through the expected lower bound (ELBO). A gradient-descent algorithm based on Monte-Carlo approximations is used to optimize the ELBO with respect to its parameters. To select the number of clusters, we compare three model selection criteria. A node clustering benchmark comprising positional community detection as well as model methods is conducted. We also compare the quality of Deep LPBM node partial membership estimation with other methodologies. We conclude with an analysis of the French political blogosphere network and a comparison with another methodology to illustrate the novelty provided by Deep LPBM results.
Conformal Prediction is a principled framework for quantifying uncertainty in blackbox learning models, by constructing prediction sets with finite-sample coverage guarantees. We introduce a novel procedure for multi-output settings, through the lens of optimal transport. Specifically, we replace univariate quantiles by Monge-Kantorovich vector quantiles to construct prediction regions with flexible shapes better suited to multivariate learning tasks. We prove that our approach ensures finite-sample, distribution-free coverage properties, similar to typical methods. We then apply our method for multi-output regression and multiclass classification. In the regression case, we use multivariate quantile regression to generate prediction regions adaptive to the inputs with asymptotic conditional coverage guarantees. Finally, we evaluate our method on practical regression and classification problems, illustrating its advantages in terms of (conditional) coverage or efficiency.
Ce séminaire explore le comportement universel des systèmes de particules en interaction unidimensionnels avec plusieurs quantités conservées, à travers la théorie de l'hydrodynamique fluctuante non linéaire (NLFHT). On commence par le processus d'exclusion simple (SEP) et sa transition d'un comportement diffusif à un comportement non-linéaire KPZ. On développera ensuite la NLFH comme un cadre systématique pour l'étude des systèmes avec nombreuses quantités conservées, en dérivant les équations de couplage de modes et en expliquant leur signification physique. Une classification complète des classes d'universalité pour deux quantités conservées est présentée, suivie d'une application détaillée au modèle ABC.
Les jeux de sortie sont des jeux dynamiques en temps discret, qui se déroulent de la façon suivante : à chaque étape du jeu, chaque joueur n'a que deux actions possibles : rester ou sortir. Tant que personne ne sort, rien ne se passe. Puis, lorsqu'au moins un des joueurs choisit de s'en aller, le jeu s'arrête, et tout le monde obtient un paiement qui ne dépend que de qui est sorti. Une question centrale en théorie des jeux est l'existence et caractérisation d'équilibres de Nash. Or, pour les jeux de sortie, malgré leur simplicité apparente, la question de l'existence est toujours ouverte, dès qu'il y a plus de trois joueurs. Dans une série de travaux en collaboration avec Galit Ashkenazi-Golan, Ilia Krasikov et Eilon Solan, nous nous focalisons sur une famille d'équilibres qu'il est possible d'abord de traduire dans un cadre continu puis de trouver explicitement grâce à un algorithme sur des graphes.
Le Théorème de Kaluza donne une condition suffisante en terme de log-convexité pour que l'inverse d'une série entière à coefficients positifs ait elle aussi des coefficients positifs. Une inteprétation probabiliste fait apparaitre des marches aléatoires en dualité, l'équivalent en temps continu faisant intervenir la notion de subordinateur spécial. Nous discuterons de ces questions en lien avec la classification de ces subordinateurs.
Dans les années 70, Tutte a développé une méthode permettant de résoudre certaines équations fonctionnelles. Après avoir évoqué le contexte combinatoire dans lequel il l’a d’abord appliquer, je montrerai comment l’adapter à l’étude de processus stochastiques continus, en déterminant la nature différentielle et algébrique de la transformée de Laplace de la mesure invariante d’un mouvement brownien dégénéré, réfléchi dans un quart de plan. Les résultats évoqués sont issus d’une récente collaboration avec Thomas Dreyfus et Sandro Franceschi.
Dans ce travail, nous étudions un modèle de régression visant à décrire le comportement des valeurs extrêmes d’une variable Y à partir de covariables X. Nous proposons une méthode de réduction de dimension spécialement conçue pour les queues de distribution, permettant de surmonter le fléau de la grande dimension et d’améliorer l’estimation de l’indice des valeurs extrêmes conditionnel.
Random fields are popular models in statistics and machine learning for spatially dependent data on Euclidian domains. However, in many applications, data is observed on non-Euclidian domains such as street networks, or river networks. In this case, it is much more difficult to construct valid random field models. In this talk, we discuss some recent approaches to modeling data in this setting, and in particular define a new class of Gaussian processes on compact metric graphs. The proposed models, the Whittle-Matérn fields, are defined via a stochastic partial differential equation on the compact metric graph and are a natural extension of Gaussian fields with Matérn covariance functions on Euclidean domains to the non-Euclidean metric graph setting.
We will start by a short presentation of Gaussian graphical models in statistics, without and with colorings. In a common research with P. Graczyk, H. Ishi and B. Kolodziejek, we study Bayesian model selection in colored Gaussian graphical models (CGGMs), which combine sparsity of conditional independencies with symmetry constraints, encoded by vertex- and edge-colored graphs. A key computational bottleneck in Bayesian inference for CGGMs is the evaluation of the Diaconis–Ylvisaker normalizing constants, given by Gamma-type Laplace integrals over cones of precision matrices with prescribed zeros and equality constraints. We introduce a new class of models for which these normalizing constants admit closed-form expressions. On the algebraic side, we identify conditions on the space of precision matrices that guarantee tractability of the associated integrals, leading to the notions of Block-Cholesky (BC) spaces. On the combinatorial side, we characterize the colored graphs inducing such spaces via a color perfect elimination ordering and a 2-path regularity condition. This class strictly extends decomposable graphs in the uncolored setting and contains all permutation invariant models associated with decomposable graphs. In the one-color case, our framework reveals a close connection with Bose–Mesner algebras. For models defined on BC spaces, we derive explicit closed-form formulas for the normalizing constants in terms of a finite collection of structure constants and propose an efficient method for computing these quantities in the commutative case. Our results substantially broaden the range of CGGMs amenable to Bayesian structure learning in Big Data applications.
Mixed-phenotype acute leukemia (MPAL) is a rare disease with poor prognosis. So far, no standard approach has been established as the “know-how” of MPAL is based only on retrospective analyses performed on small groups of patients. In this talk we describe the Bayesian technique combined with profile likelihood: a statistical technique whose main advantage is the ability to precisely examine the uncertainty of one selected parameter in complex models. We will show the results obtained with fruitful collaboration with the Polish Adult Leukemia Group.
Imaginons qu'on joue au démineur sur une grand grille, et qu’on place des mines au hasard avec une densité prescrite. On observe alors une transition de phase (grossière), c’est à dire que si notre densité est en dessous d’un ordre de grandeur critique, alors on peut toujours gagner (et avec un algoithme de complexité linéaire), alors qu’au dessus de cet ordre de grandeur critique, on ne peut jamais gagner, ceci étant dû à l’apparition de motifs ambigus.
The Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is a stochastic derivative-free optimization algorithm designed to solve difficult (e.g., ill-conditioned, multi-modal) optimization problems. In this presentation, I will talk about the recent convergence results that are obtained by studying the stability of a Markov chain defined by the normalization of the state variables of CMA-ES. Even so CMA-ES has been the state-of-the-art algorithm for black-box optimization in moderate dimensions for more than 20 years, this is its first convergence proof. Joint work with Anne Auger and Nikolaus Hansen.
In this talk, we address the stability problem of the famous Brascamp-Lieb inequality for strictly log-concave probability measures on the Euclidean space. More precisely, if a given function almost satisfies the equality in the BL inequality, is it true that it is close in some sense to the underlying extremal functions? Using a spectral interpretation of the BL inequality, we prove that the distance to the extremal functions in quadratic norm is of order square root of the deficit parameter, and involves the second positive eigenvalue of a convenient diffusion operator we wish to estimate. Our results are illustrated by some examples for which the usual uniform convexity assumption on the potential is relaxed. This is a joint work with M. Bonnefont (Institut de Mathématiques de Bordeaux) and J. Serres (Sorbonne Université).



