Regularized Estimation of High-dimensional Covariance Matrices.

Abstract

Many signal processing methods are fundamentally related to the estimation of covariance matrices. In cases where there are a large number of covariates the dimension of covariance matrices is much larger than the number of available data samples. This is especially true in applications where data acquisition is constrained by limited resources such as time, energy, storage and bandwidth. This dissertation attempts to develop necessary components for covariance estimation in the high-dimensional setting. The dissertation makes contributions in two main areas of covariance estimation: (1) high dimensional shrinkage regularized covariance estimation and (2) recursive online complexity regularized estimation with applications of anomaly detection, graph tracking, and compressive sensing.

New shrinkage covariance estimation methods are proposed that significantly outperform previous approaches in terms of mean squared error. Two multivariate data scenarios are considered: (1) independently Gaussian distributed data; and (2) heavy tailed elliptically contoured data. For the former scenario we improve on the Ledoit-Wolf (LW) shrinkage estimator using the principle of Rao-Blackwell conditioning and iterative approximation of the clairvoyant estimator. In the latter scenario, we apply a variance normalizing transformation and propose an iterative robust LW shrinkage estimator that is distribution-free within the elliptical family. The proposed robustified estimator is implemented via fixed point iterations with provable convergence and unique limit.

A recursive online covariance estimator is proposed for tracking changes in an underlying time-varying graphical model. Covariance estimation is decomposed into multiple decoupled adaptive regression problems. A recursive recursive group lasso is derived using a homotopy approach that generalizes online lasso methods to group sparse system identification. By reducing the memory of the objective function this leads to a group lasso regularized LMS that provably dominates standard LMS. Finally, we introduce a state-of-the-art sampling system, the Modulated Wideband Converter (MWC) which is based on recently developed analog compressive sensing theory. By inferring the block-sparse structures of the high-dimensional covariance matrix from a set of random projections, the MWC is capable of achieving sub-Nyquist sampling for multiband signals with arbitrary carrier frequency over a wide bandwidth.