Séminaire Probabilités et Statistique

Point-cloud datasets are ubiquitous in many science and non-science fields. These data are usually coming along with unique patterns that some algorithms are meant to extract and that are linked with the underlying phenomenon that generated the data.
In this presentation, motivated by cosmological problematics, we will focus on two kinds of spatially structured datasets. First, clustered-type patterns in which the datapoints are separated in the input space into multiple groups. We will show that the unsupervised clustering procedure performed with a Gaussian Mixture Model can be formulated in terms of a statistical physics optimisation problem. This formulation enables the unsupervised extraction of many key information about the dataset itself, like the number of clusters, their size and how they are embedded in space, particularly interesting for high-dimensional input spaces where visualisation is not possible.
On the other hand, we will study spatially continuous datasets assuming as standing on an underlying 1D structure that we aim to learn. To this end, we resort to a regularisation of the Gaussian Mixture Model in which a spatial graph is used as a prior to approximate the underlying 1D structure. The overall graph is efficiently learnt by means of the Expectation-Maximisation algorithm with guaranteed convergence and comes together with the learning of the local width of the structure. We then illustrate applications of the algorithm to model and identify the filamentary pattern drawn by the galaxy distribution of the Universe in cosmological datasets.

In this talk, we will discuss some results on the estimation of the invariant density associated to a multivariate diffusion X = (X_t)_t≥0, assuming that a continuous record of observations (X_t)_0≤t≤T is available. We will see that, when X = (X_t)_t≥0 is the solution of a stochastic differential equation with Levy-type jumps, it is possible to find the parametric convergence rate 1/T in the monodimensional case and log(T)/T when the dimension d is equal to 2. For d ≥ 3 we find the convergence rate (log(T)/T)^γ, where γ is an explicit exponent depending on the dimension d and on β₃, the harmonic mean of the smoothness of the invariant density over the d directions after having removed β₁ and β₂, which are the smallest. Moreover, we obtain a lower bound on the L²-risk for pointwise estimation, with the rate (1/T)^γ. In order to fill the logarithmic gap we consider then X = (X_t)_t≥0 as a solution to a continuous stochastic differential equation. One (surprising) finding is that the convergence rate depends on the fact that β₂ < β₃ or β₂ = β₃. In particular, we show that kernel density estimators achieve the rate (log(T)/T)^γ in the first case and (1/T)^γ in the second. Finally, we prove a minimax lower bound on the L²-risk for the pointwise estimation with the same rates (log(T)/T)^γ or (1/T)^γ, depending on the value of β₂ and β₃.

One of the seemingly innocent reformulations of the terrifying Riemann Hypothesis (RH) is the Nyman-Beurling criterion: The indicator function of (0,1) can be linearly approximated in a L^2 space by dilations of the fractional part function. Randomizing these dilations generates new structures and criteria for RH, regularizing very intricate ones. One other possible nice feature is to consider polynomials instead of Dirichlet polynomials for the approximations. How then are the huge difficulties reallocated? The answers are quite surprising!

The talk will be very accessible, especially for graduate students.
Joint work with F. Alouges and E. Hillion.

Dans cet exposé, on construit, pour tout automate cellulaire probabiliste uni-dimensionnel exponentiellement ergodique et possédant une propriété de taux positifs, un flot CFTP (« coupling from the past ») localement défini. Plusieurs conséquences de cette construction sont discutées. (Travail exposé dans l’article arXiv:2106.07219).

Je montre comment à partir d’une étude approfondie statistique et probabiliste d’une base de données de températures moyennes mondiales, j’ai découvert des comportements climatologiques sans doute très difficilement accessibles par les techniques usuelles utilisées en climatologie.

In a recent paper, Leonid Petrov showed that the up-down chains associated to the Chinese Restaurant Process (CRP) have a scaling limit – namely, a two-parameter family of diffusions that extend the one-parameter infinitely-many-neutral-alleles diffusions of Ethier and Kurtz. There has since been considerable interest in constructing ordered analogues of Petrov’s diffusions, and it is conjectured that an ordered analogue of the up-down chains will give rise to such an object. In this talk, I’ll discuss my resolution of this conjecture (joint with Douglas Rizzolo). Our approach is mainly inspired by Petrov’s work, and involves using quasisymmetric functions to describe the transition operators.

This presentation is a summary of my PhD work. I focus on the topic of hypothesis testing, extensively studied in statistics and theoretical computer science.

I start with presenting the classical identity testing problem, in which an independent sample set X ~ q is given and one would like to determine whether q=p for some fixed known p. This problem is very related to that of estimating a distribution from a given sample set. The study of testing is relevant, because for the same fixed sample size, it is possible to test against a distribution up to a smaller separation distance than what is possible in estimation. This will give me the opportunity to describe the minimax framework which proves the theoretical optimality of statistical methods in the worst case.

I will refine the study of minimax identity testing by adding a local differential privacy condition and the interest will be in the quantitative effect of ensuring privacy. The presentation will largely be on the topic of privacy, because it bears similarities with ensuring fairness conditions.

We will also shortly consider the neighboring problem of closeness testing, where the goal remains to determine whether p=q, but only an independent sample set Y ~ p is given instead of p directly. In this context, we will go beyond a simple worst-case analysis and develop instance optimal results instead. This will highlight the interplay between one-sample testing and two-sample testing, the latter being a harder problem.

Télécharger le résumé (pdf)

Differentiation is the process whereby a cell acquires a specific phenotype, by differential gene expression as a function of time. This is thought to result from the dynamical functioning of an underlying Gene Regulatory Network (GRN). The precise path from the stochastic GRN behavior to the resulting cell state is still an open question. In this presentation, we detail a methodology to reduce a mechanistic model characterizing the evolution of a cell by a system of piecewise deterministic Markov processes (PDMP), to a discrete coarse-grained model on a limited number of cell types, defined as the basins of attraction of the deterministic limit. The transitions between the basins in the weak noise limit can be determined by the unique solution of an Hamilton-Jacobi equation under a particular constraint, which corresponds to the rate function associated to a Large Deviations Principle for the PDMP. We develop a numerical method for approximating the coarse-grained model parameters, and show its accuracy for a toggle-switch network. We deduce from the reduced model an analytical approximation of the stationary distribution of the PDMP system, which appears as a Beta mixture.

Les modèles d’urnes sont utilisés dans de nombreuses applications et sont un exemple fondamental de processus stochastiques de renforcement. En partant de ces modèles, nous nous intéresserons à plusieurs familles de systèmes (finis) de processus de renforcement. Différents résultats sur le comportement collectif en temps long seront présentés. La présence/absence de synchronisation sera discutée, ainsi que les vitesses de convergence en fonction de différents régimes de paramètres. Cet exposé se fonde sur des travaux en collaboration avec I. Crimaldi, P. Dai Pra, I. Minelli et M. Mirebrahimi.