Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy

TL;DR

This paper develops new spectral-norm perturbation bounds for low-rank matrix approximations, improving understanding of noise effects in privacy-preserving PCA and providing sharper utility guarantees.

Contribution

It introduces refined spectral perturbation bounds that explicitly account for matrix-perturbation interactions, improving prior results and resolving open problems in differentially private PCA.

Findings

Bounds improve accuracy by up to a factor of √n

New contour bootstrapping method extends spectral analysis tools

Empirical results confirm bounds closely match actual spectral errors

Abstract

A central challenge in machine learning is to understand how noise or measurement errors affect low-rank approximations, particularly in the spectral norm. This question is especially important in differentially private low-rank approximation, where one aims to preserve the top-$p$ structure of a data-derived matrix while ensuring privacy. Prior work often analyzes Frobenius norm error or changes in reconstruction quality, but these metrics can over- or under-estimate true subspace distortion. The spectral norm, by contrast, captures worst-case directional error and provides the strongest utility guarantees. We establish new high-probability spectral-norm perturbation bounds for symmetric matrices that refine the classical Eckart--Young--Mirsky theorem and explicitly capture interactions between a matrix $A \in \mathbb{R}^{n \times n}$ and an arbitrary symmetric perturbation $E$. Under mild eigengap and norm conditions, our bounds yield sharp estimates for $\|(A + E)_p - A_p\|$, where $A_p$ is the best rank-$p$ approximation of $A$, with improvements of up to a factor of $\sqrt{n}$. As an application, we derive improved utility guarantees for differentially private PCA, resolving an open problem in the literature. Our analysis relies on a novel contour bootstrapping method from complex analysis and extends it to a broad class of spectral functionals, including polynomials and matrix exponentials. Empirical results on real-world datasets confirm that our bounds closely track the actual spectral error under diverse perturbation regimes.