Year: 2026
Type: preprint
Abstract: In this work, we develop a quantitative resolvent calculus for sample covariance matrices tailored to applications in high-dimensional statistics. We treat single resolvents, their derivatives, and products of two resolvents with different spectral parameters in the proportional-dimensional regime. Our bounds hold uniformly over compact subsets of $\mathbb C\setminus\mathbb R_{\geq0}$ with approximation errors of order $O_{\mathbb P}(\|Θ_n\|_F/\sqrt n)$ for deterministic test matrices $Θ_n$. These bounds require only uniformly bounded fourth moments and a uniformly bounded population covariance operator norm. Neither regular spectral edges nor convergence of the population spectral distribution is required. As a key technical ingredient, we develop an averaging argument for centered quadratic forms involving one or two leave-one-out resolvents. Applications of our results include precision-matrix shrinkage, prediction risk in ridge regression, and generalized cross-validation. For these problems, we obtain quantitative guarantees for estimation and data-driven parameter selection.
Source: arXiv (Cornell University)
Authors: Stark, Thomas, Dörnemann, Nina
Institution: Aarhus University
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
Awards: G6513393930, G8991901542
OpenAlex ID: W7224017493