Matrix perturbation bounds via contour bootstrapping

Phuc Tran,Van Vu

2024-10-21

Abstract:Matrix perturbation bounds play an essential role in the design and analysis of spectral algorithms. In this paper, we use a "contour bootstrapping" argument to derive several new perturbation bounds. As applications, we discuss new bounds on the error occurring when one uses matrix sparsification to speed up the computation of spectral parameters. Another potential application is the estimation of the trade-off in computing with privacy.

Numerical Analysis,Statistics Theory

What problem does this paper attempt to address?

This paper attempts to solve the key problems in matrix perturbation theory, that is, when a matrix \(A\) is perturbed by another matrix \(E\), how to quantify the influence of this perturbation on the eigenvalues and eigenvectors of the matrix. Specifically, the paper uses a method called "contour bootstrapping" to derive new matrix perturbation bounds.

### Main problem description of the paper

1. **Eigensubspace perturbation**: Consider a matrix \(A\) and its perturbed matrix \(\tilde{A}=A + E\), where \(E\) is a noise matrix. The author studies the influence of the perturbation on the subspace projection operator \(\Pi_p\) spanned by the first \(p\) eigenvectors and gives new perturbation bounds.

2. **Low - rank approximation perturbation**: For the best rank - \(p\) approximation \(A_p\) of matrix \(A\), the author studies the influence of the perturbed matrix \(\tilde{A}_p\) on \(A_p\) and derives new perturbation bounds.

3. **Matrix perturbation of general function \(f\)**: The author also studies the change of the more general matrix function \(f(A)\) under perturbation and provides the corresponding perturbation bounds.

### Core method of the paper

The core method of the paper is "contour bootstrapping", which is a technique based on Cauchy's integral theorem in complex analysis. By choosing an appropriate contour \(\Gamma\), the matrix function \(f(A)\) can be expressed in the form of an integral along \(\Gamma\), thus simplifying the derivation of perturbation bounds.

### Application scenarios

1. **Random sparsification to accelerate calculation**: In practical applications, in order to accelerate matrix calculations, a common method is to sparsify the input matrix (i.e., set some elements to zero). The paper uses the newly - derived perturbation bounds to estimate the error brought by this sparsification operation.

2. **Privacy - protecting calculation**: In applications involving privacy protection, matrix perturbation theory can help evaluate the influence of data perturbation on the results, so as to ensure privacy while maintaining the accuracy of calculation.

### Formula summary

- **Eigensubspace perturbation bound**:
\[
\|\tilde{\Pi}_p-\Pi_p\|\leq24\left(\frac{\|E\|}{|\lambda_p|}\log\left(\frac{6\sigma_1}{\delta_p}\right)+\frac{r^2x}{\delta_p}\right)
\]
where \(\delta_p = \lambda_p-\lambda_{p + 1}\), \(r\) is the smallest integer satisfying \(\frac{|\lambda_p|}{2}\leq|\lambda_p-\lambda_{r+1}|\), and \(x=\max_{i,j\leq r}|u_i^TEu_j|\)

- **Low - rank approximation perturbation bound**:
\[
\|\tilde{A}_p - A_p\|\leq30\left(\frac{\|E\|+r^2\bar{x}}{\log\left(\frac{6\sigma_1}{\delta_k}\right)+\log\left(\frac{6\sigma_1}{\delta_{n-(p - k)}}\right)}+\frac{r^2\bar{x}}{\frac{\lambda_k}{\delta_k}+\frac{|\lambda_{n-(p - k)+1}|}{\delta_{n-(p - k)}}}\right)
\]

These formulas show the key perturbation bounds derived in the paper and help to understand the influence of perturbation on the eigenvalues and eigenvectors of the matrix.