Searcharxiv⌕ Search

arXiv subjects

Sara van de Geer

Publications and source records attributed to Sara van de Geer.

At least 55 records · Page 3Linked to original sources
arXiv2014

Statistical Theory for High-Dimensional Models

These lecture notes consist of three chapters. In the first chapter we present oracle inequalities for the prediction error of the Lasso and square-root Lasso and briefly describe the scaled Lasso. In the second chapter we establish asymptotic linearity of a de-sparsified Lasso. This implies asymptotic normality under certain conditions and therefore can be used to construct confidence intervals for parameters of interest. A similar line of reasoning can be invoked to derive bounds in sup-norm for the Lasso and asymptotic linearity of de-sparsified estimators of a precision matrix. In the third chapter we consider chaining and the more general generic chaining method developed by Talagrand. This allows one to bound suprema of random processes. Concentration inequalities are refined probability inequalities, mostly again for suprema of random processes. We combine the two. We prove a deviation inequality directly using (generic) chaining.

math.ST↗
arXiv2014

On asymptotically optimal confidence regions and tests for high-dimensional models

We propose a general method for constructing confidence intervals and statistical tests for single or low-dimensional components of a large parameter vector in a high-dimensional model. It can be easily adjusted for multiplicity taking dependence among tests into account. For linear models, our method is essentially the same as in Zhang and Zhang [J. R. Stat. Soc. Ser. B Stat. Methodol. 76 (2014) 217-242]: we analyze its asymptotic properties and establish its asymptotic optimality in terms of semiparametric efficiency. Our method naturally extends to generalized linear models with convex loss functions. We develop the corresponding theory which includes a careful analysis for Gaussian, sub-Gaussian and bounded correlated designs.

math.ST↗
arXiv2014

The additive model with different smoothness for the components

We consider an additive regression model consisting of two components $f^0$ and $g^0$, where the first component $f^0$ is in some sense "smoother" than the second $g^0$. Smoothness is here described in terms of a semi-norm on the class of regression functions. We use a penalized least squares estimator $(\hat f, \hat g)$ of $(f^0, g^0)$ and show that the rate of convergence for $\hat f $ is faster than the rate of convergence for $\hat g$. In fact, both rates are generally as fast as in the case where one of the two components is known. The theory is illustrated by a simulation study. Our proofs rely on recent results from empirical process theory.

math.ST↗
arXiv2014

Censored linear model in high dimensions

Censored data are quite common in statistics and have been studied in depth in the last years. In this paper we consider censored high-dimensional data. High-dimensional models are in some way more complex than their low-dimensional versions, therefore some different techniques are required. For the linear case appropriate estimators based on penalized regression, have been developed in the last years. In particular in sparse contexts the $l_1$-penalised regression (also known as LASSO) performs very well. Only few theoretical work was done in order to analyse censored linear models in a high-dimensional context. We therefore consider a high-dimensional censored linear model, where the response variable is left-censored. We propose a new estimator, which aims to work with high-dimensional linear censored data. Theoretical non-asymptotic oracle inequalities are derived.

math.ST↗
arXiv2014

Worst possible sub-directions in high-dimensional models

We examine the rate of convergence of the Lasso estimator of lower dimensional components of the high-dimensional parameter. Under bounds on the $\ell_1$-norm on the worst possible sub-direction these rates are of order $\sqrt {|J| \log p / n }$ where $p$ is the total number of parameters, $J \subset \{ 1, \ldots, p \}$ represents a subset of the parameters and $n$ is the number of observations. We also derive rates in sup-norm in terms of the rate of convergence in $\ell_1$-norm. The irrepresentable condition on a set $J$ requires that the $\ell_1$-norm of the worst possible sub-direction is sufficiently smaller than one. In that case sharp oracle results can be obtained. Moreover, if the coefficients in $J$ are small enough the Lasso will put these coefficients to zero. This extends known results which say that the irrepresentable condition on the inactive set (the set where coefficients are exactly zero) implies no false positives. We further show that by de-sparsifying one obtains fast rates in supremum norm without conditions on the worst possible sub-direction. The main assumption here is that approximate sparsity is of order $o (\sqrt n / \log p )$. The results are extended to M-estimation with $\ell_1$-penalty for generalized linear models and exponential families for example. For the graphical Lasso this leads to an extension of known results to the case where the precision matrix is only approximately sparse. The bounds we provide are non-asymptotic but we also present asymptotic formulations for ease of interpretation.

math.ST↗
arXiv2013

Confidence sets in sparse regression

The problem of constructing confidence sets in the high-dimensional linear model with $n$ response variables and $p$ parameters, possibly $p\ge n$, is considered. Full honest adaptive inference is possible if the rate of sparse estimation does not exceed $n^{-1/4}$, otherwise sparse adaptive confidence sets exist only over strict subsets of the parameter spaces for which sparse estimators exist. Necessary and sufficient conditions for the existence of confidence sets that adapt to a fixed sparsity level of the parameter vector are given in terms of minimal $\ell^2$-separation conditions on the parameter space. The design conditions cover common coherence assumptions used in models for sparsity, including (possibly correlated) sub-Gaussian designs.

math.ST↗
arXiv2013

On the uniform convergence of empirical norms and inner products, with application to causal inference

Uniform convergence of empirical norms - empirical measures of squared functions - is a topic which has received considerable attention in the literature on empirical processes. The results are relevant as empirical norms occur due to symmetrization. They also play a prominent role in statistical applications. The contraction inequality has been a main tool but recently other approaches have shown to lead to better results in important cases. We present an overview including the linear (anisotropic) case, and give new results for inner products of functions. Our main application will be the estimation of the parental structure in a directed acyclic graph. As intermediate result we establish convergence of the least squares estimator when the model is wrong.

math.ST↗
arXiv2013

The partial linear model in high dimensions

Partial linear models have been widely used as flexible method for modelling linear components in conjunction with non-parametric ones. Despite the presence of the non-parametric part, the linear, parametric part can under certain conditions be estimated with parametric rate. In this paper, we consider a high-dimensional linear part. We show that it can be estimated with oracle rates, using the LASSO penalty for the linear part and a smoothness penalty for the nonparametric part.

math.ST↗
arXiv2013

$\ell_0$-penalized maximum likelihood for sparse directed acyclic graphs

We consider the problem of regularized maximum likelihood estimation for the structure and parameters of a high-dimensional, sparse directed acyclic graphical (DAG) model with Gaussian distribution, or equivalently, of a Gaussian structural equation model. We show that the $\ell_0$-penalized maximum likelihood estimator of a DAG has about the same number of edges as the minimal-edge I-MAP (a DAG with minimal number of edges representing the distribution), and that it converges in Frobenius norm. We allow the number of nodes p to be much larger than sample size n but assume a sparsity condition and that any representation of the true DAG has at least a fixed proportion of its nonzero edge weights above the noise level. Our results do not rely on the faithfulness assumption nor on the restrictive strong faithfulness condition which are required for methods based on conditional independence testing such as the PC-algorithm.

math.ST↗
arXiv2013

Quasi-Likelihood and/or Robust Estimation in High Dimensions

We consider the theory for the high-dimensional generalized linear model with the Lasso. After a short review on theoretical results in literature, we present an extension of the oracle results to the case of quasi-likelihood loss. We prove bounds for the prediction error and $\ell_1$-error. The results are derived under fourth moment conditions on the error distribution. The case of robust loss is also given. We moreover show that under an irrepresentable condition, the $\ell_1$-penalized quasi-likelihood estimator has no false positives.

math.ST↗
arXiv2012

Weakly decomposable regularization penalties and structured sparsity

It has been shown in literature that the Lasso estimator, or l1-penalized least squares estimator, enjoys good oracle properties. This paper examines which special properties of the l1-penalty allow for sharp oracle results, and then extends the situation to general norm-based penalties that satisfy a weak decomposability condition.

math.ST↗
arXiv2012

Correlated variables in regression: clustering and sparse estimation

We consider estimation in a high-dimensional linear model with strongly correlated variables. We propose to cluster the variables first and do subsequent sparse estimation such as the Lasso for cluster-representatives or the group Lasso based on the structure from the clusters. Regarding the first step, we present a novel and bottom-up agglomerative clustering algorithm based on canonical correlations, and we show that it finds an optimal solution and is statistically consistent. We also present some theoretical arguments that canonical correlation based clustering leads to a better-posed compatibility constant for the design matrix which ensures identifiability and an oracle inequality for the group Lasso. Furthermore, we discuss circumstances where cluster-representatives and using the Lasso as subsequent estimator leads to improved results for prediction and detection of variables. We complement the theoretical analysis with various empirical results.

stat.ME↗
arXiv2012

Generic chaining and the l1-penalty

We address the choice of the tuning parameter $λ$ in $\ell_1$-penalized M-estimation. Our main concern is models which are highly nonlinear, such as the Gaussian mixture model. The number of parameters $p$ is moreover large, possibly larger than the number of observations $n$. The generic chaining technique of Talagrand[2005] is tailored for this problem. It leads to the choice $λ\asymp \sqrt {\log p / n}$, as in the standard Lasso procedure (which concerns the linear model and least squares loss).

math.ST↗
arXiv2012

L1-Penalization for Mixture Regression Models

We consider a finite mixture of regressions (FMR) model for high-dimensional inhomogeneous data where the number of covariates may be much larger than sample size. We propose an l1-penalized maximum likelihood estimator in an appropriate parameterization. This kind of estimation belongs to a class of problems where optimization and theory for non-convex functions is needed. This distinguishes itself very clearly from high-dimensional estimation with convex loss- or objective functions, as for example with the Lasso in linear or generalized linear models. Mixture models represent a prime and important example where non-convexity arises. For FMR models, we develop an efficient EM algorithm for numerical optimization with provable convergence properties. Our penalized estimator is numerically better posed (e.g., boundedness of the criterion function) than unpenalized maximum likelihood estimation, and it allows for effective statistical regularization including variable selection. We also present some asymptotic theory and oracle inequalities: due to non-convexity of the negative log-likelihood function, different mathematical arguments are needed than for problems with convex losses. Finally, we apply the new method to both simulated and real data.

stat.ME↗
arXiv2011

The Bernstein-Orlicz norm and deviation inequalities

We introduce two new concepts designed for the study of empirical processes. First, we introduce a new Orlicz norm which we call the Bernstein-Orlicz norm. This new norm interpolates sub-Gaussian and sub-exponential tail behavior. In particular, we show how this norm can be used to simplify the derivation of deviation inequalities for suprema of collections of random variables. Secondly, we introduce chaining and generic chaining along a tree. These simplify the well-known concepts of chaining and generic chaining. The supremum of the empirical process is then studied as a special case. We show that chaining along a tree can be done using entropy with bracketing. Finally, we establish a deviation inequality for the empirical process for the unbounded case.

math.PR↗
arXiv2011

The Lasso, correlated design, and improved oracle inequalities

We study high-dimensional linear models and the $\ell_1$-penalized least squares estimator, also known as the Lasso estimator. In literature, oracle inequalities have been derived under restricted eigenvalue or compatibility conditions. In this paper, we complement this with entropy conditions which allow one to improve the dual norm bound, and demonstrate how this leads to new oracle inequalities. The new oracle inequalities show that a smaller choice for the tuning parameter and a trade-off between $\ell_1$-norms and small compatibility constants are possible. This implies, in particular for correlated design, improved bounds for the prediction error of the Lasso estimator as compared to the methods based on restricted eigenvalue or compatibility conditions only.

stat.ME↗
arXiv2010

Estimation for High-Dimensional Linear Mixed-Effects Models Using $\ell_1$-Penalization

We propose an $\ell_1$-penalized estimation procedure for high-dimensional linear mixed-effects models. The models are useful whenever there is a grouping structure among high-dimensional observations, i.e. for clustered data. We prove a consistency and an oracle optimality result and we develop an algorithm with provable numerical convergence. Furthermore, we demonstrate the performance of the method on simulated and a real high-dimensional data set.

stat.ME↗