Search

We obtain upper bounds for the estimation error of Kernel Ridge Regression (KRR) for all non-negative regularization parameters, offering a geometric perspective on various phenomena in KRR. As applications: 1. We address the multiple descent problem, unifying the proofs of arxiv:1908.10292 and arxiv:1904.12191 for polynomial kernels and we establish multiple descent for the upper bound of estimation error of KRR under sub-Gaussian design and non-asymptotic regimes. 2. For a sub-Gaussian design vector and for non-asymptotic scenario, we prove a one-sided isomorphic version of the Gaussian Equivalent Conjecture. 3. We offer a novel perspective on the linearization of kernel matrices of non-linear kernel, extending it to the power regime for polynomial kernels. 4. Our theory is applicable to data-dependent kernels, providing a convenient and accurate tool for the feature learning regime in deep learning theory. 5. Our theory extends the results in arxiv:2009.14286 under weak moment assumption. Our proof is based on three mathematical tools developed in this paper that can be of independent interest: 1. Dvoretzky-Milman theorem for ellipsoids under (very) weak moment assumptions. 2. Restricted Isomorphic Property in Reproducing Kernel Hilbert Spaces with embedding index conditions. 3. A concentration inequality for finite-degree polynomial kernel functions.

Motivated by several examples, we consider a general framework of learning with linear loss functions. In this context, we provide excess risk and estimation bounds that hold with large probability for four estimators: ERM, minmax MOM and their regularized versions. These general bounds are applied for the problem of robustness in sparse PCA. In particular, we improve the state of the art result for this this problems, obtain results under weak moment assumptions as well as for adversarial contaminated data.

In the linear regression model, the minimum l2-norm interpolant estimator has received much attention since it was proved to be consistent even though it fits noisy data perfectly under some condition on the covariance matrix $\Sigma$ of the input vector, known as benign overfitting. Motivated by this phenomenon, we study the generalization property of this estimator from a geometrical viewpoint. Our main results extend and improve the convergence rates as well as the deviation probability from [Tsigler and Bartlett]. Our proof differs from the classical bias/variance analysis and is based on the self-induced regularization property introduced in [Bartlett, Montanari and Rakhlin]: the minimum l2-norm interpolant estimator can be written as a sum of a ridge estimator and an overfitting component. The two geometrical properties of random Gaussian matrices at the heart of our analysis are the Dvoretsky-Milman theorem and isomorphic and restricted isomorphic properties. In particular, the Dvoretsky dimension appearing naturally in our geometrical viewpoint, coincides with the effective rank and is the key tool for handling the behavior of the design matrix restricted to the sub-space where overfitting happens. We extend these results to heavy-tailed scenarii proving the universality of this phenomenon beyond exponential moment assumptions. This phenomenon is unknown before and is widely believed to be a significant challenge. This follows from an anistropic version of the probabilistic Dvoretsky-Milman theorem that holds for heavy-tailed vectors which is of independent interest.

We consider the problem of robust mean and location estimation w.r.t. any pseudo-norm of the form $x\in\mathbb{R}^d\to ||x||_S = \sup_{v\in S}$ where $S$ is any symmetric subset of $\mathbb{R}^d$. We show that the deviation-optimal minimax subgaussian rate for confidence $1-\delta$ is $$ \max\left(\frac{l^*(\Sigma^{1/2}S)}{\sqrt{N}}, \sup_{v\in S}||\Sigma^{1/2}v||_2\sqrt{\frac{\log(1/\delta)}{N}}\right) $$where $l^*(\Sigma^{1/2}S)$ is the Gaussian mean width of $\Sigma^{1/2}S$ and $\Sigma$ the covariance of the data (in the benchmark i.i.d. Gaussian case). This improves the entropic minimax lower bound from [Lugosi and Mendelson, 2019] and closes the gap characterized by Sudakov's inequality between the entropy and the Gaussian mean width for this problem. This shows that the right statistical complexity measure for the mean estimation problem is the Gaussian mean width. We also show that this rate can be achieved by a solution to a convex optimization problem in the adversarial and $L_2$ heavy-tailed setup by considering minimum of some Fenchel-Legendre transforms constructed using the Median-of-means principle. We finally show that this rate may also be achieved in situations where there is not even a first moment but a location parameter exists.

We consider median of means (MOM) versions of the Stahel-Donoho outlyingness (SDO) [stahel 1981, donoho 1982] and of Median Absolute Deviation (MAD) functions to construct subgaussian estimators of a mean vector under adversarial contamination and heavy-tailed data. We develop a single analysis of the MOM version of the SDO which covers all cases ranging from the Gaussian case to the L2 case. It is based on isomorphic and almost isometric properties of the MOM versions of SDO and MAD. This analysis also covers cases where the mean does not even exist but a location parameter does; in those cases we still recover the same subgaussian rates and the same price for adversarial contamination even though there is not even a first moment. These properties are achieved by the classical SDO median and are therefore the first non-asymptotic statistical bounds on the Stahel-Donoho median complementing the $\sqrt{n}$-consistency [maronna 1995] and asymptotic normality [Zuo, Cui, He, 2004] of the Stahel-Donoho estimators. We also show that the MOM version of MAD can be used to construct an estimator of the covariance matrix under only a L2-moment assumption or of a scale parameter if a second moment does not exist.

Many statistical learning problems have recently been shown to be amenable to Semi-Definite Programming (SDP), with community detection and clustering in Gaussian mixture models as the most striking instances [javanmard et al., 2016]. Given the growing range of applications of SDP-based techniques to machine learning problems, and the rapid progress in the design of efficient algorithms for solving SDPs, an intriguing question is to understand how the recent advances from empirical process theory can be put to work in order to provide a precise statistical analysis of SDP estimators. In the present paper, we borrow cutting edge techniques and concepts from the learning theory literature, such as fixed point equations and excess risk curvature arguments, which yield general estimation and prediction results for a wide class of SDP estimators. From this perspective, we revisit some classical results in community detection from [gu\'edon et al.,2016] and [chen et al., 2016], and we obtain statistical guarantees for SDP estimators used in signed clustering, group synchronization and MAXCUT.

We construct an algorithm, running in time $\tilde{\mathcal O}(N d + uK d)$, which is robust to outliers and heavy-tailed data and which achieves the subgaussian rate from [Lugosi, Mendelson] \begin{equation}\label{eq:intro_subgaus_rate} \sqrt{\frac{{\rm Tr}(\Sigma)}{N}}+\sqrt{\frac{||\Sigma||_{op}K}{N}} \end{equation}with probability at least $1-\exp(-c_0K)-\exp(-c_1 u)$ where $\Sigma$ is the covariance matrix of the informative data, $K\in\{1, \ldots, K\}$ is some parameter (number of block means) and $u>0$ is another parameter of the algorithm. This rate is achieved when $K\geq c_1 |\mathcal O|$ where $|\mathcal O|$ is the number of outliers in the database and under the only assumption that the informative data have a second moment. The algorithm is fully data-dependent and does not use in its construction the proportion of outliers nor the rate above. Its construction combines recently developed tools for Median-of-Means estimators and covering-Semi-definite Programming [Chen, Diakonikolas, Ge] and [Peng, Tangwongsan, Zhang].

We establish risk bounds for Regularized Empirical Risk Minimizers (RERM) when the loss is Lipschitz and convex and the regularization function is a norm. In a first part, we obtain these results in the i.i.d. setup under subgaussian assumptions on the design. In a second part, a more general framework where the design might have heavier tails and data may be corrupted by outliers both in the design and the response variables is considered. In this situation, RERM performs poorly in general. We analyse an alternative procedure based on median-of-means principles and called minmax MOM. We show optimal subgaussian deviation rates for these estimators in the relaxed setting. The main results are meta-theorems allowing a wide-range of applications to various problems in learning theory. To show a non-exhaustive sample of these potential applications, it is applied to classification problems with logistic loss functions regularized by LASSO and SLOPE, to regression problems with Huber loss regularized by Group LASSO and Total Variation. Another advantage of the minmax MOM formulation is that it suggests a systematic way to slightly modify descent based algorithms used in high-dimensional statistics to make them robust to outliers. We illustrate this principle in a Simulations section where a minmax MOM version of classical proximal descent algorithms are turned into robust to outliers algorithms.

Hyperparameters tuning and model selection are important steps in machine learning. Unfortunately, classical hyperparameter calibration and model selection procedures are sensitive to outliers and heavy-tailed data. In this work, we construct a selection procedure which can be seen as a robust alternative to cross-validation and is based on a median-of-means principle. Using this procedure, we also build an ensemble method which, trained with algorithms and corrupted heavy-tailed data, selects an algorithm, trains it with a large uncorrupted subsample and automatically tune its hyperparameters. The construction relies on a divide-and-conquer methodology, making this method easily scalable for autoML given a corrupted database. This method is tested with the LASSO which is known to be highly sensitive to outliers., Comment: 17 pages, 3 figures

We present an extension of Vapnik's classical empirical risk minimizer (ERM) where the empirical risk is replaced by a median-of-means (MOM) estimator, the new estimators are called MOM minimizers. While ERM is sensitive to corruption of the dataset for many classical loss functions used in classification, we show that MOM minimizers behave well in theory, in the sense that it achieves Vapnik's (slow) rates of convergence under weak assumptions: data are only required to have a finite second moment and some outliers may also have corrupted the dataset. We propose an algorithm inspired by MOM minimizers. These algorithms can be analyzed using arguments quite similar to those used for Stochastic Block Gradient descent. As a proof of concept, we show how to modify a proof of consistency for a descent algorithm to prove consistency of its MOM version. As MOM algorithms perform a smart subsampling, our procedure can also help to reduce substantially time computations and memory ressources when applied to non linear algorithms. These empirical performances are illustrated on both simulated and real datasets.

Classical approach to regularization is to design norms enhancing smoothness or sparsity and then to use this norm or some power of this norm as a regularization function. The choice of the regularization function (for instance a power function) in terms of the norm is mostly dictated by computational purpose rather than theoretical considerations. In this work, we design regularization functions that are motivated by theoretical arguments. To that end we introduce a concept of optimal regularization called "minimax regularization" and, as a proof of concept, we show how to construct such a regularization function for the $\ell_1^d$ norm for the random design setup. We develop a similar construction for the deterministic design setup. It appears that the resulting regularized procedures are different from the one used in the LASSO in both setups.

Mean embeddings provide an extremely flexible and powerful tool in machine learning and statistics to represent probability distributions and define a semi-metric (MMD, maximum mean discrepancy; also called N-distance or energy distance), with numerous successful applications. The representation is constructed as the expectation of the feature map defined by a kernel. As a mean, its classical empirical estimator, however, can be arbitrary severely affected even by a single outlier in case of unbounded features. To the best of our knowledge, unfortunately even the consistency of the existing few techniques trying to alleviate this serious sensitivity bottleneck is unknown. In this paper, we show how the recently emerged principle of median-of-means can be used to design estimators for kernel mean embedding and MMD with excessive resistance properties to outliers, and optimal sub-Gaussian deviation bounds under mild assumptions., Comment: ICML-2019: camera-ready paper. Code: https://bitbucket.org/TimotheeMathieu/monk-mmd

We introduce new estimators for robust machine learning based on median-of-means (MOM) estimators of the mean of real valued random variables. These estimators achieve optimal rates of convergence under minimal assumptions on the dataset. The dataset may also have been corrupted by outliers on which no assumption is granted. We also analyze these new estimators with standard tools from robust statistics. In particular, we revisit the concept of breakdown point. We modify the original definition by studying the number of outliers that a dataset can contain without deteriorating the estimation properties of a given estimator. This new notion of breakdown number, that takes into account the statistical performances of the estimators, is non-asymptotic in nature and adapted for machine learning purposes. We proved that the breakdown number of our estimator is of the order of (number of observations)*(rate of convergence). For instance, the breakdown number of our estimators for the problem of estimation of a d-dimensional vector with a noise variance sigma^2 is sigma^2d and it becomes sigma^2 s log(d/s) when this vector has only s non-zero component. Beyond this breakdown point, we proved that the rate of convergence achieved by our estimator is (number of outliers) divided by (number of observation). Besides these theoretical guarantees, the major improvement brought by these new estimators is that they are easily computable in practice. In fact, basically any algorithm used to approximate the standard Empirical Risk Minimizer (or its regularized versions) has a robust version approximating our estimators. As a proof of concept, we study many algorithms for the classical LASSO estimator. A byproduct of the MOM algorithms is a measure of depth of data that can be used to detect outliers., Comment: 48 pages, 6 figures

We obtain estimation error rates and sharp oracle inequalities for regularization procedures of the form \begin{equation*} \hat f \in argmin_{f\in F}\left(\frac{1}{N}\sum_{i=1}^N\ell(f(X_i), Y_i)+\lambda \|f\|\right) \end{equation*} when $\|\cdot\|$ is any norm, $F$ is a convex class of functions and $\ell$ is a Lipschitz loss function satisfying a Bernstein condition over $F$. We explore both the bounded and subgaussian stochastic frameworks for the distribution of the $f(X_i)$'s, with no assumption on the distribution of the $Y_i$'s. The general results rely on two main objects: a complexity function, and a sparsity equation, that depend on the specific setting in hand (loss $\ell$ and norm $\|\cdot\|$). As a proof of concept, we obtain minimax rates of convergence in the following problems: 1) matrix completion with any Lipschitz loss function, including the hinge and logistic loss for the so-called 1-bit matrix completion instance of the problem, and quantile losses for the general case, which enables to estimate any quantile on the entries of the matrix; 2) logistic LASSO and variants such as the logistic SLOPE; 3) kernel methods, where the loss is the hinge loss, and the regularization function is the RKHS norm.

Penalized least squares estimation is a popular technique in high-dimensional statistics. It includes such methods as the LASSO, the group LASSO, and the nuclear norm penalized least squares. The existing theory of these methods is not fully satisfying since it allows one to prove oracle inequalities with fixed high probability only for the estimators depending on this probability. Furthermore, the control of compatibility factors appearing in the oracle bounds is often not explicit. Some very recent developments suggest that the theory of oracle inequalities can be revised in an improved way. In this paper, we provide an overview of ideas and tools leading to such an improved theory. We show that, along with overcoming the disadvantages mentioned above, the methodology extends to the hilbertian framework and it applies to a large class of convex penalties. This paper is partly expository. In particular, we provide adapted proofs of some results from other recent work.

For a convex class of functions $F$, a regularization functions $\Psi(\cdot)$ and given the random data $(X_i, Y_i)_{i=1}^N$, we study estimation properties of regularization procedures of the form \begin{equation*} \hat f \in {\rm argmin}_{f\in F}\Big(\frac{1}{N}\sum_{i=1}^N\big(Y_i-f(X_i)\big)^2+\lambda \Psi(f)\Big) \end{equation*} for some well chosen regularization parameter $\lambda$. We obtain bounds on the $L_2$ estimation error rate that depend on the complexity of the "true model" $F^*:=\{f\in F: \Psi(f)\leq\Psi(f^*)\}$, where $f^*\in {\rm argmin}_{f\in F}\mathbb{E}(Y-f(X))^2$ and the $(X_i,Y_i)$'s are independent and distributed as $(X,Y)$. Our estimate holds under weak stochastic assumptions -- one of which being a small-ball condition satisfied by $F$ -- and for rather flexible choices of regularization functions $\Psi(\cdot)$. Moreover, the result holds in the learning theory framework: we do not assume any a-priori connection between the output $Y$ and the input $X$. As a proof of concept, we apply our general estimation bound to various choices of $\Psi$, for example, the $\ell_p$ and $S_p$-norms (for $p\geq1$), weak-$\ell_p$, atomic norms, max-norm and SLOPE. In many cases, the estimation rate almost coincides with the minimax rate in the class $F^*$.

We show that two polynomial time methods, a Lasso estimator with adaptively chosen tuning parameter and a Slope estimator, adaptively achieve the exact minimax prediction and $\ell_2$ estimation rate $(s/n)\log (p/s)$ in high-dimensional linear regression on the class of $s$-sparse target vectors in $\mathbb R^p$. This is done under the Restricted Eigenvalue (RE) condition for the Lasso and under a slightly more constraining assumption on the design for the Slope. The main results have the form of sharp oracle inequalities accounting for the model misspecification error. The minimax optimal bounds are also obtained for the $\ell_q$ estimation errors with $1\le q\le 2$ when the model is well-specified. The results are non-asymptotic, and hold both in probability and in expectation. The assumptions that we impose on the design are satisfied with high probability for a large class of random matrices with independent and possibly anisotropically distributed rows. We give a comparative analysis of conditions, under which oracle bounds for the Lasso and Slope estimators can be obtained. In particular, we show that several known conditions, such as the RE condition and the sparse eigenvalue condition are equivalent if the $\ell_2$-norms of regressors are uniformly bounded.

We obtain bounds on estimation error rates for regularization procedures of the form \begin{equation*} \hat f \in {\rm argmin}_{f\in F}\left(\frac{1}{N}\sum_{i=1}^N\left(Y_i-f(X_i)\right)^2+\lambda \Psi(f)\right) \end{equation*} when $\Psi$ is a norm and $F$ is convex. Our approach gives a common framework that may be used in the analysis of learning problems and regularization problems alike. In particular, it sheds some light on the role various notions of sparsity have in regularization and on their connection with the size of subdifferentials of $\Psi$ in a neighbourhood of the true minimizer. As `proof of concept' we extend the known estimates for the LASSO, SLOPE and trace norm regularization.

We consider the problem of recovering sparse vectors from underdetermined linear measurements via $\ell_p$-constrained basis pursuit. Previous analyses of this problem based on generalized restricted isometry properties have suggested that two phenomena occur if $p\neq 2$. First, one may need substantially more than $s \log(en/s)$ measurements (optimal for $p=2$) for uniform recovery of all $s$-sparse vectors. Second, the matrix that achieves recovery with the optimal number of measurements may not be Gaussian (as for $p=2$). We present a new, direct analysis which shows that in fact neither of these phenomena occur. Via a suitable version of the null space property we show that a standard Gaussian matrix provides $\ell_q/\ell_1$-recovery guarantees for $\ell_p$-constrained basis pursuit in the optimal measurement regime. Our result extends to several heavier-tailed measurement matrices. As an application, we show that one can obtain a consistent reconstruction from uniform scalar quantized measurements in the optimal measurement regime.

Let $X=(x_1,...,x_n)$ be a random vector that satisfies a weak small ball property and whose coordinates $x_i$ satisfy that $\|x_i\|_{L_p} \lesssim \sqrt{p} \|x_i\|_{L_2}$ for $p \sim \log n$. In \cite{LM_compressed}, it was shown that $N$ independent copies of $X$ can be used as measurement vectors in Compressed Sensing (using the basis pursuit algorithm) to reconstruct any $d$-sparse vector with the optimal number of measurements $N\gtrsim d \log\big(e n/d\big)$. In this note we show that the result is almost optimal. We construct a random vector $X$ with iid, mean-zero, variance one coordinates that satisfies the same weak small ball property and whose coordinates satisfy that $\|x_i\|_{L_p} \lesssim \sqrt{p} \|x_i\|_{L_2}$ for $p \sim (\log n)/(\log N)$, but the basis pursuit algorithm fails to recover even $1$-sparse vectors. The construction shows that `spiky' measurement vectors may lead to a poor performance by the basis pursuit algorithm, but on the other hand may still perform in an optimal way if one chooses a different reconstruction algorithm (like $\ell_0$-minimization). This exhibits the fact that the convex relaxation of $\ell_0$-minimization comes at a significant cost when using `spiky' measurement vectors.

We study conditions under which, given a dictionary $F=\{f_1,\ldots ,f_M\}$ and an i.i.d. sample $(X_i,Y_i)_{i=1}^N$, the empirical minimizer in $\operatorname {span}(F)$ relative to the squared loss, satisfies that with high probability \[R\bigl(\tilde{f}^{\mathrm{ERM}}\bigr)\leq\inf_{f\in\operatorname {span}(F)}R(f)+r_N(M),\] where $R(\cdot)$ is the squared risk and $r_N(M)$ is of the order of $M/N$. Among other results, we prove that a uniform small-ball estimate for functions in $\operatorname {span}(F)$ is enough to achieve that goal when the noise is independent of the design., Comment: Published at http://dx.doi.org/10.3150/15-BEJ701 in the Bernoulli (http://isi.cbs.nl/bernoulli/) by the International Statistical Institute/Bernoulli Society (http://isi.cbs.nl/BS/bshome.htm)

We prove that iid random vectors that satisfy a rather weak moment assumption can be used as measurement vectors in Compressed Sensing, and the number of measurements required for exact reconstruction is the same as the best possible estimate -- exhibited by a random gaussian matrix. We also prove that this moment condition is necessary, up to a $\log \log $ factor. Applications to the Compatibility Condition and the Restricted Eigenvalue Condition in the noisy setup and to properties of neighbourly random polytopes are also discussed.

Let $F$ be a finite model of cardinality $M$ and denote by $\operatorname {conv}(F)$ its convex hull. The problem of convex aggregation is to construct a procedure having a risk as close as possible to the minimal risk over $\operatorname {conv}(F)$. Consider the bounded regression model with respect to the squared risk denoted by $R(\cdot)$. If ${\widehat{f}}_n^{\mathit{ERM-C}}$ denotes the empirical risk minimization procedure over $\operatorname {conv}(F)$, then we prove that for any $x>0$, with probability greater than $1-4\exp(-x)$, \[R({\widehat{f}}_n^{\mathit{ERM-C}})\leq\min_{f\in \operatorname {conv}(F)}R(f)+c_0\max \biggl(\psi_n^{(C)}(M),\frac{x}{n}\biggr),\] where $c_0>0$ is an absolute constant and $\psi_n^{(C)}(M)$ is the optimal rate of convex aggregation defined in (In Computational Learning Theory and Kernel Machines (COLT-2003) (2003) 303-313 Springer) by $\psi_n^{(C)}(M)=M/n$ when $M\leq \sqrt{n}$ and $\psi_n^{(C)}(M)=\sqrt{\log (\mathrm{e}M/\sqrt{n})/n}$ when $M>\sqrt{n}$., Comment: Published in at http://dx.doi.org/10.3150/12-BEJ447 the Bernoulli (http://isi.cbs.nl/bernoulli/) by the International Statistical Institute/Bernoulli Society (http://isi.cbs.nl/BS/bshome.htm)

We study the performance of Empirical Risk Minimization in noisy phase retrieval problems, indexed by subsets of $\R^n$ and relative to subgaussian sampling; that is, when the given data is $y_i=\inr{a_i,x_0}^2+w_i$ for a subgaussian random vector $a$, independent noise $w$ and a fixed but unknown $x_0$ that belongs to a given subset of $\R^n$. We show that ERM produces $\hat{x}$ whose Euclidean distance to either $x_0$ or $-x_0$ depends on the gaussian mean-width of the indexing set and on the signal-to-noise ratio of the problem. The bound coincides with the one for linear regression when $\|x_0\|_2$ is of the order of a constant. In addition, we obtain a minimax lower bound for the problem and identify sets for which ERM is a minimax procedure. As examples, we study the class of $d$-sparse vectors in $\R^n$ and the unit ball in $\ell_1^n$.

We obtain sharp oracle inequalities for the empirical risk minimization procedure in the regression model under the assumption that the target Y and the model F are subgaussian. The bound we obtain is sharp in the minimax sense if F is convex. Moreover, under mild assumptions on F, the error rate of ERM remains optimal even if the procedure is allowed to perform with constant probability. A part of our analysis is a new proof of minimax results for the gaussian regression model., Comment: learning theory, empirical process, minimax rates, Topics in Learning Theory - Societe Mathematique de France, (S. Boucheron and N. Vayatis Eds.). 2016

Given a finite class of functions F, the problem of aggregation is to construct a procedure with a risk as close as possible to the risk of the best element in the class. A classical procedure (PAC-Bayesian statistical learning theory (2004) Paris 6, Statistical Learning Theory and Stochastic Optimization (2001) Springer, Ann. Statist. 28 (2000) 75-87) is the aggregate with exponential weights (AEW), defined by \[\tilde{f}^{\mathrm{AEW}}=\sum_{f\in F}\hat{\theta}(f)f,\qquad where \hat{\theta}(f)=\frac{\exp(-({n}/{T})R_n(f))}{\sum_{g\in F}\exp(-({n}/{T})R_n(g))},\] where $T>0$ is called the temperature parameter and $R_n(\cdot)$ is an empirical risk. In this article, we study the optimality of the AEW in the regression model with random design and in the low-temperature regime. We prove three properties of AEW. First, we show that AEW is a suboptimal aggregation procedure in expectation with respect to the quadratic risk when $T\leq c_1$, where $c_1$ is an absolute positive constant (the low-temperature regime), and that it is suboptimal in probability even for high temperatures. Second, we show that as the cardinality of the dictionary grows, the behavior of AEW might deteriorate, namely, that in the low-temperature regime it might concentrate with high probability around elements in the dictionary with risk greater than the risk of the best function in the dictionary by at least an order of $1/\sqrt{n}$. Third, we prove that if a geometric condition on the dictionary (the so-called "Bernstein condition) is assumed, then AEW is indeed optimal both in high probability and in expectation in the low-temperature regime. Moreover, under that assumption, the complexity term is essentially the logarithm of the cardinality of the set of "almost minimizers" rather than the logarithm of the cardinality of the entire dictionary. This result holds for small values of the temperature parameter, thus complementing an analogous result for high temperatures., Comment: Published in at http://dx.doi.org/10.3150/11-BEJ408 the Bernoulli (http://isi.cbs.nl/bernoulli/) by the International Statistical Institute/Bernoulli Society (http://isi.cbs.nl/BS/bshome.htm)

We consider a general supervised learning problem with strongly convex and Lipschitz loss and study the problem of model selection aggregation. In particular, given a finite dictionary functions (learners) together with the prior, we generalize the results obtained by Dai, Rigollet and Zhang [Ann. Statist. 40 (2012) 1878-1905] for Gaussian regression with squared loss and fixed design to this learning setup. Specifically, we prove that the $Q$-aggregation procedure outputs an estimator that satisfies optimal oracle inequalities both in expectation and with high probability. Our proof techniques somewhat depart from traditional proofs by making most of the standard arguments on the Laplace transform of the empirical process to be controlled., Comment: Published in at http://dx.doi.org/10.1214/13-AOS1190 the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)

We show that empirical risk minimization procedures and regularized empirical risk minimization procedures satisfy nonexact oracle inequalities in an unbounded framework, under the assumption that the class has a subexponential envelope function. The main novelty, in addition to the boundedness assumption free setup, is that those inequalities can yield fast rates even in situations in which exact oracle inequalities only hold with slower rates. We apply these results to show that procedures based on $\ell_1$ and nuclear norms regularization functions satisfy oracle inequalities with a residual term that decreases like $1/n$ for every $L_q$-loss functions ($q\geq2$), while only assuming that the tail behavior of the input and output variables are well behaved. In particular, no RIP type of assumption or "incoherence condition" are needed to obtain fast residual terms in those setups. We also apply these results to the problems of convex aggregation and model selection., Comment: Published in at http://dx.doi.org/10.1214/11-AOS965 the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)

This paper is about iteratively reweighted basis-pursuit algorithms for compressed sensing and matrix completion problems. In a first part, we give a theoretical explanation of the fact that reweighted basis pursuit can improve a lot upon basis pursuit for exact recovery in compressed sensing. We exhibit a condition that links the accuracy of the weights to the RIP and incoherency constants, which ensures exact recovery. In a second part, we introduce a new algorithm for matrix completion, based on the idea of iterative reweighting. Since a weighted nuclear "norm" is typically non-convex, it cannot be used easily as an objective function. So, we define a new estimator based on a fixed-point equation. We give empirical evidences of the fact that this new algorithm leads to strong improvements over nuclear norm minimization on simulated and real matrix completion problems.

We present an argument based on the multidimensional and the uniform central limit theorems, proving that, under some geometrical assumptions between the target function $T$ and the learning class $F$, the excess risk of the empirical risk minimization algorithm is lower bounded by \[\frac{\mathbb{E}\sup_{q\in Q}G_q}{\sqrt{n}}\delta,\] where $(G_q)_{q\in Q}$ is a canonical Gaussian process associated with $Q$ (a well chosen subset of $F$) and $\delta$ is a parameter governing the oscillations of the empirical excess risk function over a small ball in $F$., Comment: Published in at http://dx.doi.org/10.3150/09-BEJ225 the Bernoulli (http://isi.cbs.nl/bernoulli/) by the International Statistical Institute/Bernoulli Society (http://isi.cbs.nl/BS/bshome.htm)

We give estimates on the logarithmic Sobolev constant of some finite lamplighter graphs in terms of the spectral gap of the underlying base. Also, we give examples of application.

We observe $(X_i,Y_i)_{i=1}^n$ where the $Y_i$'s are real valued outputs and the $X_i$'s are $m\times T$ matrices. We observe a new entry $X$ and we want to predict the output $Y$ associated with it. We focus on the high-dimensional setting, where $m T \gg n$. This includes the matrix completion problem with noise, as well as other problems. We consider linear prediction procedures based on different penalizations, involving a mixture of several norms: the nuclear norm, the Frobenius norm and the $\ell_1$-norm. For these procedures, we prove sharp oracle inequalities, using a statistical learning theory point of view. A surprising fact in our results is that the rates of convergence do not depend on $m$ and $T$ directly. The analysis is conducted without the usually considered incoherency condition on the unknown matrix or restricted isometry condition on the sampling operator. Moreover, our results are the first to give for this problem an analysis of penalization (such nuclear norm penalization) as a regularization algorithm: our oracle inequalities prove that these procedures have a prediction accuracy close to the deterministic oracle one, given that the reguralization parameters are well-chosen.

In this paper, we consider the problem of "hyper-sparse aggregation". Namely, given a dictionary $F = \{f_1, ..., f_M \}$ of functions, we look for an optimal aggregation algorithm that writes $\tilde f = \sum_{j=1}^M \theta_j f_j$ with as many zero coefficients $\theta_j$ as possible. This problem is of particular interest when $F$ contains many irrelevant functions that should not appear in $\tilde{f}$. We provide an exact oracle inequality for $\tilde f$, where only two coefficients are non-zero, that entails $\tilde f$ to be an optimal aggregation algorithm. Since selectors are suboptimal aggregation procedures, this proves that 2 is the minimal number of elements of $F$ required for the construction of an optimal aggregation procedures in every situations. A simulated example of this algorithm is proposed on a dictionary obtained using LARS, for the problem of selection of the regularization parameter of the LASSO. We also give an example of use of aggregation to achieve minimax adaptation over anisotropic Besov spaces, which was not previously known in minimax theory (in regression on a random design)., Comment: 33 pages

We consider a $l_1$-penalization procedure in the non-parametric Gaussian regression model. In many concrete examples, the dimension $d$ of the input variable $X$ is very large (sometimes depending on the number of observations). Estimation of a $\beta$-regular regression function $f$ cannot be faster than the slow rate $n^{-2\beta/(2\beta+d)}$. Hopefully, in some situations, $f$ depends only on a few numbers of the coordinates of $X$. In this paper, we construct two procedures. The first one selects, with high probability, these coordinates. Then, using this subset selection method, we run a local polynomial estimator (on the set of interesting coordinates) to estimate the regression function at the rate $n^{-2\beta/(2\beta+d^*)}$, where $d^*$, the "real" dimension of the problem (exact number of variables whom $f$ depends on), has replaced the dimension $d$ of the design. To achieve this result, we used a $l_1$ penalization method in this non-parametric setup., Comment: Published in at http://dx.doi.org/10.1214/08-EJS327 the Electronic Journal of Statistics (http://www.i-journals.org/ejs/) by the Institute of Mathematical Statistics (http://www.imstat.org)

Let $\cF$ be a set of $M$ classification procedures with values in $[-1,1]$. Given a loss function, we want to construct a procedure which mimics at the best possible rate the best procedure in $\cF$. This fastest rate is called optimal rate of aggregation. Considering a continuous scale of loss functions with various types of convexity, we prove that optimal rates of aggregation can be either $((\log M)/n)^{1/2}$ or $(\log M)/n$. We prove that, if all the $M$ classifiers are binary, the (penalized) Empirical Risk Minimization procedures are suboptimal (even under the margin/low noise condition) when the loss function is somewhat more than convex, whereas, in that case, aggregation procedures with exponential weights achieve the optimal rate of aggregation., Comment: 15 pages

We want to recover the regression function in the single-index model. Using an aggregation algorithm with local polynomial estimators, we answer in particular to the second part of Question~2 from Stone (1982) on the optimal convergence rate. The procedure constructed here has strong adaptation properties: it adapts both to the smoothness of the link function and to the unknown index. Moreover, the procedure locally adapts to the distribution of the design. We propose new upper bounds for the local polynomial estimator (which are results of independent interest) that allows a fairly general design. The behavior of this algorithm is studied through numerical simulations. In particular, we show empirically that it improves strongly over empirical risk minimization., Comment: 36 pages

We study the performances of an adaptive procedure based on a convex combination, with data-driven weights, of term-by-term thresholded wavelet estimators. For the bounded regression model, with random uniform design, and the nonparametric density model, we show that the resulting estimator is optimal in the minimax sense over all Besov balls under the $L^2$ risk, without any logarithm factor.

We construct a classifier which attains the rate of convergence $\log n/n$ under sparsity and margin assumptions. An approach close to the one met in approximation theory for the estimation of function is used to obtain this result. The idea is to develop the Bayes rule in a fundamental system of $L^2([0,1]^d)$ made of indicator of dyadic sets and to assume that coefficients, equal to $-1,0 {or} 1$, belong to a kind of $L^1-$ball. This assumption can be seen as a sparsity assumption, in the sense that the proportion of coefficients non equal to zero decreases as "frequency" grows. Finally, rates of convergence are obtained by using an usual trade-off between a bias term and a variance term.

We consider the problem of optimality, in a minimax sense, and adaptivity to the margin and to regularity in binary classification. We prove an oracle inequality, under the margin assumption (low noise condition), satisfied by an aggregation procedure which uses exponential weights. This oracle inequality has an optimal residual: $(\log M/n)^{\kappa/(2\kappa-1)}$ where $\kappa$ is the margin parameter, $M$ the number of classifiers to aggregate and $n$ the number of observations. We use this inequality first to construct minimax classifiers under margin and regularity assumptions and second to aggregate them to obtain a classifier which is adaptive both to the margin and regularity. Moreover, by aggregating plug-in classifiers (only $\log n$), we provide an easily implementable classifier adaptive both to the margin and to regularity., Comment: accepted to COLT 2006

In the same spirit as Tsybakov (2003), we define the optimality of an aggregation procedure in the problem of classification. Using an aggregate with exponential weights, we obtain an optimal rate of convex aggregation for the hinge risk under the margin assumption. Moreover we obtain an optimal rate of model selection aggregation under the margin assumption for the excess Bayes risk.