Search

We study estimators with generalized lasso penalties within the computational sufficiency framework introduced by Vu (2018, arXiv:1807.05985). By representing these penalties as support functions of zonotopes and more generally Minkowski sums of line segments and rays, we show that there is a natural reflection group associated with the underlying optimization problem. A consequence of this point of view is that for large classes of estimators sharing the same penalty, the penalized least squares estimator is computationally minimal sufficient. This means that all such estimators can be computed by refining the output of any algorithm for the least squares case. An interesting technical component is our analysis of coordinate descent on the dual problem. A key insight is that the iterates are obtained by reflecting and averaging, so they converge to an element of the dual feasible set that is minimal with respect to a ordering induced by the group associated with the penalty. Our main application is fused lasso/total variation denoising and isotonic regression on arbitrary graphs. In those cases the associated group is a permutation group., Comment: 28 page, 3 figures

Statistical sufficiency formalizes the notion of data reduction. In the decision theoretic interpretation, once a model is chosen all inferences should be based on a sufficient statistic. However, suppose we start with a set of procedures rather than a specific model. Is it possible to reduce the data and yet still be able to compute all of the procedures? In other words, what functions of the data contain all of the information sufficient for computing these procedures? This article presents some progress towards a theory of "computational sufficiency" and shows that strong reductions can be made for large classes of penalized $M$-estimators by exploiting hidden symmetries in the underlying optimization problems. These reductions can (1) reveal hidden connections between seemingly disparate methods, (2) enable efficient computation, (3) give a different perspective on understanding procedures in a model-free setting. As a main example, the theory provides a surprising answer to the following question: "What do the Graphical Lasso, sparse PCA, single-linkage clustering, and L1 penalized Ising model selection all have in common?", Comment: Fixed mistakes in Theorem 2 and proof of Corollary 2; added Proposition 1; corrected typos (particularly in definition of computational sufficiency); added references; revised for readability and clarity

The presence of a sparse "truth" has been a constant assumption in the theoretical analysis of sparse PCA and is often implicit in its methodological development. This naturally raises questions about the properties of sparse PCA methods and how they depend on the assumption of sparsity. Under what conditions can the relevant variables be selected consistently if the truth is assumed to be sparse? What can be said about the results of sparse PCA without assuming a sparse and unique truth? We answer these questions by investigating the properties of the recently proposed Fantope projection and selection (FPS) method in the high-dimensional setting. Our results provide general sufficient conditions for sparsistency of the FPS estimator. These conditions are weak and can hold in situations where other estimators are known to fail. On the other hand, without assuming sparsity or identifiability, we show that FPS provides a sparse, linear dimension-reducing transformation that is close to the best possible in terms of maximizing the predictive covariance., Comment: Published in at http://dx.doi.org/10.1214/14-AOS1273 the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)

This note extends a recent result of Mendelson on the supremum of a quadratic process to squared norms of functions taking values in a Banach space. Our method of proof is a reduction by a symmetrization argument and observation about the subadditivity of the generic chaining functional. We provide an application to the supremum of a linear process in the sample covariance matrix indexed by finite rank, positive definite matrices., Comment: This note was first written in 2012 and a version has been available on the first author's website. It is now posted on arXiv to provide a universal identifier for citations elsewhere. It is not intended for publication in its current form, but a revision incorporating later work will be posted separately

We study sparse principal components analysis in high dimensions, where $p$ (the number of variables) can be much larger than $n$ (the number of observations), and analyze the problem of estimating the subspace spanned by the principal eigenvectors of the population covariance matrix. We introduce two complementary notions of $\ell_q$ subspace sparsity: row sparsity and column sparsity. We prove nonasymptotic lower and upper bounds on the minimax subspace estimation error for $0\leq q\leq1$. The bounds are optimal for row sparse subspaces and nearly optimal for column sparse subspaces, they apply to general classes of covariance matrices, and they show that $\ell_q$ constrained estimates can achieve optimal minimax rates without restrictive spiked covariance conditions. Interestingly, the form of the rates matches known results for sparse regression when the effective noise variance is defined appropriately. Our proof employs a novel variational $\sin\Theta$ theorem that may be useful in other regularized spectral estimation problems., Comment: Published in at http://dx.doi.org/10.1214/13-AOS1151 the Annals of Statistics (http://www.imstat.org/aos/) by the Institute of Mathematical Statistics (http://www.imstat.org)

We study sparse principal components analysis in the high-dimensional setting, where $p$ (the number of variables) can be much larger than $n$ (the number of observations). We prove optimal, non-asymptotic lower and upper bounds on the minimax estimation error for the leading eigenvector when it belongs to an $\ell_q$ ball for $q \in [0,1]$. Our bounds are sharp in $p$ and $n$ for all $q \in [0, 1]$ over a wide class of distributions. The upper bound is obtained by analyzing the performance of $\ell_q$-constrained PCA. In particular, our results provide convergence rates for $\ell_1$-constrained PCA., Comment: To appear in Proceedings of the 15th International Conference on Artificial Intelligence and Statistics (AISTATS) 2012, La Palma, Canary Islands. Volume 22 of JMLR: W&CP 22

Functional MRI (fMRI) has become the most common method for investigating the human brain. However, fMRI data present some complications for statistical analysis and modeling. One recently developed approach to these data focuses on estimation of computational encoding models that describe how stimuli are transformed into brain activity measured in individual voxels. Here we aim at building encoding models for fMRI signals recorded in the primary visual cortex of the human brain. We use residual analyses to reveal systematic nonlinearity across voxels not taken into account by previous models. We then show how a sparse nonparametric method [J. Roy. Statist. Soc. Ser. B 71 (2009b) 1009-1030] can be used together with correlation screening to estimate nonlinear encoding models effectively. Our approach produces encoding models that predict about 25% more accurately than models estimated using other methods [Nature 452 (2008a) 352--355]. The estimated nonlinearity impacts the inferred properties of individual voxels, and it has a plausible biological interpretation. One benefit of quantitative encoding models is that estimated models can be used to decode brain activity, in order to identify which specific image was seen by an observer. Encoding models estimated by our approach also improve such image identification by about 12% when the correct image is one of 11,500 possible images., Comment: Published in at http://dx.doi.org/10.1214/11-AOAS476 the Annals of Applied Statistics (http://www.imstat.org/aoas/) by the Institute of Mathematical Statistics (http://www.imstat.org)

Information estimates such as the ``direct method'' of Strong et al. (1998) sidestep the difficult problem of estimating the joint distribution of response and stimulus by instead estimating the difference between the marginal and conditional entropies of the response. While this is an effective estimation strategy, it tempts the practitioner to ignore the role of the stimulus and the meaning of mutual information. We show here that, as the number of trials increases indefinitely, the direct (or ``plug-in'') estimate of marginal entropy converges (with probability 1) to the entropy of the time-averaged conditional distribution of the response, and the direct estimate of the conditional entropy converges to the time-averaged entropy of the conditional distribution of the response. Under joint stationarity and ergodicity of the response and stimulus, the difference of these quantities converges to the mutual information. When the stimulus is deterministic or non-stationary the direct estimate of information no longer estimates mutual information, which is no longer meaningful, but it remains a measure of variability of the response distribution across time.