Perturbation of linear forms of singular vectors under Gaussian noise

TL;DR

This paper derives sharp probabilistic bounds on the deviations of singular vectors of a matrix from its noisy observation, with implications for understanding the stability of singular vectors under Gaussian noise.

Contribution

It provides new concentration bounds for linear and bilinear forms of perturbed singular vectors in the presence of Gaussian noise, extending previous results to more general settings.

Findings

Bounds of order O(√(log(m+n)/(m∨n))) for deviations of singular vectors

High-probability bounds on maximum coordinate deviations of singular vectors

Characterization of bias in empirical singular vectors

Abstract

Let $A\in\mathbb{R}^{m\times n}$ be a matrix of rank $r$ with singular value decomposition (SVD) $A=\sum_{k=1}^r\sigma_k (u_k\otimes v_k),$ where $\{\sigma_k, k=1,\ldots,r\}$ are singular values of $A$ (arranged in a non-increasing order) and $u_k\in {\mathbb R}^m, v_k\in {\mathbb R}^n, k=1,\ldots, r$ are the corresponding left and right orthonormal singular vectors. Let $\tilde{A}=A+X$ be a noisy observation of $A,$ where $X\in\mathbb{R}^{m\times n}$ is a random matrix with i.i.d. Gaussian entries, $X_{ij}\sim\mathcal{N}(0,\tau^2),$ and consider its SVD $\tilde{A}=\sum_{k=1}^{m\wedge n}\tilde{\sigma}_k(\tilde{u}_k\otimes\tilde{v}_k)$ with singular values $\tilde{\sigma}_1\geq\ldots\geq\tilde{\sigma}_{m\wedge n}$ and singular vectors $\tilde{u}_k,\tilde{v}_k,k=1,\ldots, m\wedge n.$ The goal of this paper is to develop sharp concentration bounds for linear forms $\langle \tilde u_k,x\rangle, x\in {\mathbb R}^m$ and $\langle \tilde v_k,y\rangle, y\in {\mathbb R}^n$ of the perturbed (empirical) singular vectors in the case when the singular values of $A$ are distinct and, more generally, concentration bounds for bilinear forms of projection operators associated with SVD. In particular, the results imply upper bounds of the order $O\biggl(\sqrt{\frac{\log(m+n)}{m\vee n}}\biggr)$ (holding with a high probability) on $$\max_{1\leq i\leq m}\big|\big<\tilde{u}_k-\sqrt{1+b_k}u_k,e_i^m\big>\big|\ \ {\rm and} \ \ \max_{1\leq j\leq n}\big|\big<\tilde{v}_k-\sqrt{1+b_k}v_k,e_j^n\big>\big|,$$ where $b_k$ are properly chosen constants characterizing the bias of empirical singular vectors $\tilde u_k, \tilde v_k$ and $\{e_i^m,i=1,\ldots,m\}, \{e_j^n,j=1,\ldots,n\}$ are the canonical bases of $\mathbb{R}^m, {\mathbb R}^n,$ respectively.