Searcharxiv⌕ Search

arXiv subjects

Olivier Catoni

Publications and source records attributed to Olivier Catoni.

15 recordsLinked to original sources
arXiv2022

Constant payoff in zero-sum stochastic games

In a zero-sum stochastic game, at each stage, two adversary players take decisions and receive a stage payoff determined by them and by a controlled random variable representing the state of nature. The total payoff is the normalized discounted sum of the stage payoffs. In this paper we solve the "constant payoff" conjecture formulated by Sorin, Vigeral and Venel (2010): if both players use optimal strategies, then for any alpha>0, the expected discounted payoff between stage 1 and stage alpha/lambda tends to the limit discounted value of the game, as the discount rate lambda goes to 0.

math.OC↗
arXiv2021

New bounds for $k$-means and information $k$-means

In this paper, we derive a new dimension-free non-asymptotic upper bound for the quadratic $k$-means excess risk related to the quantization of an i.i.d sample in a separable Hilbert space. We improve the bound of order $\mathcal{O} \bigl( k / \sqrt{n} \bigr)$ of Biau, Devroye and Lugosi, recovering the rate $\sqrt{k/n}$ that has already been proved by Fefferman, Mitter, and Narayanan and by Klochkov, Kroshnin and Zhivotovskiy but with worse log factors and constants. More precisely, we bound the mean excess risk of an empirical minimizer by the explicit upper bound $16 B^2 \log(n/k) \sqrt{k \log(k) / n}$, in the bounded case when $\mathbb{P}( \lVert X \rVert \leq B) = 1$. This is essentially optimal up to logarithmic factors since a lower bound of order $\mathcal{O} \bigl( \sqrt{k^{1 - 4/d}/n} \bigr)$ is known in dimension $d$. Our technique of proof is based on the linearization of the $k$-means criterion through a kernel trick and on PAC-Bayesian inequalities. To get a $1 / \sqrt{n}$ speed, we introduce a new PAC-Bayesian chaining method replacing the concept of $δ$-net with the perturbation of the parameter by an infinite dimensional Gaussian process. In the meantime, we embed the usual $k$-means criterion into a broader family built upon the Kullback divergence and its underlying properties. This results in a new algorithm that we named information $k$-means, well suited to the clustering of bags of words. Based on considerations from information theory, we also introduce a new bounded $k$-means criterion that uses a scale parameter but satisfies a generalization bound that does not require any boundedness or even integrability conditions on the sample. We describe the counterpart of Lloyd's algorithm and prove generalization bounds for these new $k$-means criteria.

math.ST↗
arXiv2017

Dimension-free PAC-Bayesian bounds for matrices, vectors, and linear least squares regression

This paper is focused on dimension-free PAC-Bayesian bounds, under weak polynomial moment assumptions, allowing for heavy tailed sample distributions. It covers the estimation of the mean of a vector or a matrix, with applications to least squares linear regression. Special efforts are devoted to the estimation of Gram matrices, due to their prominent role in high-dimension data analysis.

math.ST↗
arXiv2016

PAC-Bayesian bounds for the Gram matrix and least squares regression with a random design

The topics dicussed in this paper take their origin inthe estimation of the Gram matrix of a random vector from a sample made of n independent copies. They comprise the estimation of the covariance matrix and the study of least squares regression with a random design. We propose four types of results, based on non-asymptotic PAC-Bayesian generalization bounds: a new robust estimator of the Gram matrix and of the covariance matrix, new results on the empirical Gram matrix, new robust least squares estimators and new results on the ordinary least squares estimator, including its exact rate of convergence under polynomial moment assumptions.

math.ST↗
arXiv2013

Toric grammars: a new statistical approach to natural language modeling

We propose a new statistical model for computational linguistics. Rather than trying to estimate directly the probability distribution of a random sentence of the language, we define a Markov chain on finite sets of sentences with many finite recurrent communicating classes and define our language model as the invariant probability measures of the chain on each recurrent communicating class. This Markov chain, that we call a communication model, recombines at each step randomly the set of sentences forming its current state, using some grammar rules. When the grammar rules are fixed and known in advance instead of being estimated on the fly, we can prove supplementary mathematical properties. In particular, we can prove in this case that all states are recurrent states, so that the chain defines a partition of its state space into finite recurrent communicating classes. We show that our approach is a decisive departure from Markov models at the sentence level and discuss its relationships with Context Free Grammars. Although the toric grammars we use are closely related to Context Free Grammars, the way we generate the language from the grammar is qualitatively different. Our communication model has two purposes. On the one hand, it is used to define indirectly the probability distribution of a random sentence of the language. On the other hand it can serve as a (crude) model of language transmission from one speaker to another speaker through the communication of a (large) set of sentences.

stat.ML↗
arXiv2012

Robust linear least squares regression

We consider the problem of robustly predicting as well as the best linear combination of $d$ given functions in least squares regression, and variants of this problem including constraints on the parameters of the linear combination. For the ridge estimator and the ordinary least squares estimator, and their variants, we provide new risk bounds of order $d/n$ without logarithmic factor unlike some standard results, where $n$ is the size of the training data. We also provide a new estimator with better deviations in the presence of heavy-tailed noise. It is based on truncating differences of losses in a min--max framework and satisfies a $d/n$ risk bound both in expectation and in deviations. The key common surprising factor of these results is the absence of exponential moment condition on the output distribution while achieving exponential deviations. All risk bounds are obtained through a PAC-Bayesian analysis on truncated differences of losses. Experimental results strongly back up our truncated min--max estimator.

math.ST↗
arXiv2011

Linear regression through PAC-Bayesian truncation

We consider the problem of predicting as well as the best linear combination of d given functions in least squares regression under L^\infty constraints on the linear combination. When the input distribution is known, there already exists an algorithm having an expected excess risk of order d/n, where n is the size of the training data. Without this strong assumption, standard results often contain a multiplicative log(n) factor, complex constants involving the conditioning of the Gram matrix of the covariates, kurtosis coefficients or some geometric quantity characterizing the relation between L^2 and L^\infty-balls and require some additional assumptions like exponential moments of the output. This work provides a PAC-Bayesian shrinkage procedure with a simple excess risk bound of order d/n holding in expectation and in deviations, under various assumptions. The common surprising factor of these results is their simplicity and the absence of exponential moment condition on the output distribution while achieving exponential deviations. The risk bounds are obtained through a PAC-Bayesian analysis on truncated differences of losses. We also show that these results can be generalized to other strongly convex loss functions.

math.ST↗
arXiv2011

Challenging the empirical mean and empirical variance: a deviation study

We present new M-estimators of the mean and variance of real valued random variables, based on PAC-Bayes bounds. We analyze the non-asymptotic minimax properties of the deviations of those estimators for sample distributions having either a bounded variance or a bounded variance and a bounded kurtosis. Under those weak hypotheses, allowing for heavy-tailed distributions, we show that the worst case deviations of the empirical mean are suboptimal. We prove indeed that for any confidence level, there is some M-estimator whose deviations are of the same order as the deviations of the empirical mean of a Gaussian statistical sample, even when the statistical sample is instead heavy-tailed. Experiments reveal that these new estimators perform even better than predicted by our bounds, showing deviation quantile functions uniformly lower at all probability levels than the empirical mean for non Gaussian sample distributions as simple as the mixture of two Gaussian measures.

math.ST↗
arXiv2010

Risk bounds in linear regression through PAC-Bayesian truncation

We consider the problem of predicting as well as the best linear combination of d given functions in least squares regression, and variants of this problem including constraints on the parameters of the linear combination. When the input distribution is known, there already exists an algorithm having an expected excess risk of order d/n, where n is the size of the training data. Without this strong assumption, standard results often contain a multiplicative log n factor, and require some additional assumptions like uniform boundedness of the d-dimensional input representation and exponential moments of the output. This work provides new risk bounds for the ridge estimator and the ordinary least squares estimator, and their variants. It also provides shrinkage procedures with convergence rate d/n (i.e., without the logarithmic factor) in expectation and in deviations, under various assumptions. The key common surprising factor of these results is the absence of exponential moment condition on the output distribution while achieving exponential deviations. All risk bounds are obtained through a PAC-Bayesian analysis on truncated differences of losses. Finally, we show that some of these results are not particular to the least squares loss, but can be generalized to similar strongly convex loss functions.

stat.ML↗
arXiv2009

High confidence estimates of the mean of heavy-tailed real random variables

We present new estimators of the mean of a real valued random variable, based on PAC-Bayesian iterative truncation. We analyze the non-asymptotic minimax properties of the deviations of estimators for distributions having either a bounded variance or a bounded kurtosis. It turns out that these minimax deviations are of the same order as the deviations of the empirical mean estimator of a Gaussian distribution. Nevertheless, the empirical mean itself performs poorly at high confidence levels for the worst distribution with a given variance or kurtosis (which turns out to be heavy tailed). To obtain (nearly) minimax deviations in these broad class of distributions, it is necessary to use some more robust estimator, and we describe an iterated truncation scheme whose deviations are close to minimax. In order to calibrate the truncation and obtain explicit confidence intervals, it is necessary to dispose of a prior bound either on the variance or the kurtosis. When a prior bound on the kurtosis is available, we obtain as a by-product a new variance estimator with good large deviation properties. When no prior bound is available, it is still possible to use Lepski's approach to adapt to the unknown variance, although it is no more possible to obtain observable confidence intervals.

math.ST↗
arXiv2007

Pac-Bayesian Supervised Classification: The Thermodynamics of Statistical Learning

This monograph deals with adaptive supervised classification, using tools borrowed from statistical mechanics and information theory, stemming from the PACBayesian approach pioneered by David McAllester and applied to a conception of statistical learning theory forged by Vladimir Vapnik. Using convex analysis on the set of posterior probability measures, we show how to get local measures of the complexity of the classification model involving the relative entropy of posterior distributions with respect to Gibbs posterior measures. We then discuss relative bounds, comparing the generalization error of two classification rules, showing how the margin assumption of Mammen and Tsybakov can be replaced with some empirical measure of the covariance structure of the classification model.We show how to associate to any posterior distribution an effective temperature relating it to the Gibbs prior distribution with the same level of expected error rate, and how to estimate this effective temperature from data, resulting in an estimator whose expected error rate converges according to the best possible power of the sample size adaptively under any margin and parametric complexity assumptions. We describe and study an alternative selection scheme based on relative bounds between estimators, and present a two step localization technique which can handle the selection of a parametric model from a family of those. We show how to extend systematically all the results obtained in the inductive setting to transductive learning, and use this to improve Vapnik's generalization bounds, extending them to the case when the sample is made of independent non-identically distributed pairs of patterns and labels. Finally we review briefly the construction of Support Vector Machines and show how to derive generalization bounds for them, measuring the complexity either through the number of support vectors or through the value of the transductive or inductive margin.

stat.ML↗
arXiv2006

PAC-Bayesian inductive and transductive learning

We present here a PAC-Bayesian point of view on adaptive supervised classification. Using convex analysis, we show how to get local measures of the complexity of the classification model involving the relative entropy of posterior distributions with respect to Gibbs posterior measures. We discuss relative bounds, comparing two classification rules, to show how the margin assumption of Mammen and Tsybakov can be replaced with some empirical measure of the covariance structure of the classification model. We also show how to associate to any posterior distribution an {\em effective temperature} relating it to the Gibbs prior distribution with the same level of expected error rate, and how to estimate this effective temperature from data, resulting in an estimator whose expected error rate adaptively converges according to the best possible power of the sample size. Then we introduce a PAC-Bayesian point of view on transductive learning and use it to improve on known Vapnik's generalization bounds, extending them to the case when the sample is independent but not identically distributed. Eventually we review briefly the construction of Support Vector Machines and show how to derive generalization bounds for them, measuring the complexity either through the number of support vectors or through transductive or inductive margin estimates.

math.ST↗
arXiv2004

Improved Vapnik Cervonenkis bounds

We give a new proof of VC bounds where we avoid the use of symmetrization and use a shadow sample of arbitrary size. We also improve on the variance term. This results in better constants, as shown on numerical examples. Moreover our bounds still hold for non identically distributed independent random variables. Keywords: Statistical learning theory, PAC-Bayesian theorems, VC dimension.

math.ST↗