Minimax Estimation of Linear Functions of Eigenvectors in the Face of Small Eigen-Gaps

  • 2021-04-07 17:55:10
  • Gen Li, Changxiao Cai, Yuantao Gu, H. Vincent Poor, Yuxin Chen
  • 0

Abstract

Eigenvector perturbation analysis plays a vital role in various statisticaldata science applications. A large body of prior works, however, focused onestablishing $\ell_{2}$ eigenvector perturbation bounds, which are often highlyinadequate in addressing tasks that rely on fine-grained behavior of aneigenvector. This paper makes progress on this by studying the perturbation oflinear functions of an unknown eigenvector. Focusing on two fundamentalproblems -- matrix denoising and principal component analysis -- in thepresence of Gaussian noise, we develop a suite of statistical theory thatcharacterizes the perturbation of arbitrary linear functions of an unknowneigenvector. In order to mitigate a non-negligible bias issue inherent to thenatural "plug-in" estimator, we develop de-biased estimators that (1) achieveminimax lower bounds for a family of scenarios (modulo some logarithmicfactor), and (2) can be computed in a data-driven manner without samplesplitting. Noteworthily, the proposed estimators are nearly minimax optimaleven when the associated eigen-gap is substantially smaller than what isrequired in prior theory.

 

Quick Read (beta)

loading the full paper ...