Factor-Based Large Covariance Matrix Estimation: Nonconvex Optimization and Statistical Guarantees
Lists
Only you see this
Abstract
In this paper, we study the problem of large covariance matrix estimation based on factor analysis models. In this setting, the covariance matrix is decomposed into a low-rank component and a sparse component. We formulate the estimation problem as a nonconvex optimization problem and propose an iterative algorithm to solve it. Our algorithm starts with the <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</u>rincipal <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">o</u>rthogonal compl<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e</u>ment <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">t</u>hresholding (POET) estimator, a widely used optimization-free method for factor-based covariance estimation. We then refine this initial estimator through alternating optimization of the low-rank and sparse components. The resulting nonconvex estimator is referred to as POET with <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</u>efining iterativel<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">y</u> (POETRY). We theoretically prove that the algorithm converges linearly to the ground truth with an improved statistical rate compared to POET, achieving the minimax rate of convergence for factor-based covariance matrix estimation. This matches the statistical rate achieved by existing convex relaxation methods but with significantly reduced per-iteration computational complexity. We further establish the statistical rate of the inverse covariance matrix estimator. Additionally, by introducing the notion of effective rank, we are able to further reduce the sample size requirement and enhance the statistical rate. Numerical experiments validate our theoretical findings and demonstrate the superiority of the proposed algorithm compared to state-of-the-art methods.
BibTeX
@article{zou2025factor,
author = {Shanshan Zou and Ziping Zhao},
title = {{Factor-Based Large Covariance Matrix Estimation: Nonconvex Optimization and Statistical Guarantees}},
journal = {IEEE Transactions on Signal Processing},
year = {2025},
doi = {10.1109/tsp.2025.3626934},
}Loading related content...