FR3.R5.2

Asymptotic Estimates for Spectral Estimators of Rotationally Invariant Matrices

Zhuohang He, University of Electronic Science and Technology of China, China; Junjie Ma, Chinese Academy of Sciences, China; Xiaojun Yuan, University of Electronic Science and Technology of China, China

Session:
Estimation 1

Track:
11: Information Theory and Statistics

Location:
Omikron I

Presentation Time:
Fri, 12 Jul, 14:55 - 15:15

Session Chair:
Shao-Lun Huang,
Abstract
In this paper, we consider the recovery of low-rank matrices from noisy observations using spectral denoisers, where the singular values are denoised through an identical scalar smoothing function. We explore the asymptotic mean squared error (AMSE) of these denoisers within a framework where the rank of the matrix to be recovered grows linearly with the matrix size. We demonstrate that, under arbitrary i.i.d. noise and some mild regularity assumptions, the AMSE converges in probability to a deterministic function of the noise power. Our results are applicable to commonly used denoisers, including the best-rank-r denoiser, the singular-value soft-threshold denoiser, and the singular-value hard-threshold denoiser.
Resources