On Stewart's Perturbation Theorem for SVD

Ren-Cang Li,Ninoslav Truhar,Lei-Hong Zhang

2024-06-11

Abstract:This paper establishes a variant of Stewart's theorem (Theorem~6.4 of Stewart, {\em SIAM Rev.}, 15:727--764, 1973) for the singular subspaces associated with the SVD of a matrix subject to perturbations. Stewart's original version uses both the Frobenius and spectral norms, whereas the new variant uses the spectral norm and any unitarily invariant norm that offer choices per convenience of particular applications and lead to sharper bounds than that straightforwardly derived from Stewart's original theorem with the help of the well-known equivalence inequalities between matrix norms. Of interest in their own right, bounds on the solution to two couple Sylvester equations are established for a few different circumstances.

Numerical Analysis

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve Stewart's singular value decomposition (SVD) perturbation theorem. Specifically, the original Stewart theorem uses the Frobenius norm and the spectral norm to measure the influence of matrix perturbations on singular subspaces. The goal of this paper is to establish a more general and more accurate version, which provides more flexible applications under the spectral norm and any unitarily invariant norm, and can derive more stringent bounds than those directly derived from Stewart's original theorem.

### Main problem description

1. **Limitations of existing theories**:
- Stewart's original theorem (Theorem 6.4) uses both the Frobenius norm and the spectral norm.
- This method has overly strict conditions in some cases and the results are not accurate enough.

2. **Directions for improvement**:
- Use the spectral norm and any unitarily invariant norm, rather than being limited to the Frobenius norm.
- Provide more relaxed conditions and more stringent bounds to improve the applicability and accuracy of the theory.

3. **Specific goals**:
- Establish a new variant theorem so that it has better flexibility and applicability under the spectral norm and any unitarily invariant norm.
- In specific cases (for example, when the unitarily invariant norm is also set to the spectral norm), this variant is more convenient for analyzing the quality of reduced - order models.

### Mathematical formula representation

- Let \( G\in\mathbb{C}^{m\times n} \) be a complex matrix, and its singular value decomposition is:
\[
U^{H}GV=\begin{bmatrix}
\Sigma_{1}&0\\
0&\Sigma_{2}
\end{bmatrix}
\]
where \( U \) and \( V \) are unitary matrices, and \( \Sigma_{1} \) and \( \Sigma_{2} \) are diagonal matrices composed of the first \( r \) and the last \( \min(m, n)-r \) singular values of \( G \), respectively.

- When \( G \) is perturbed to become \( \tilde{G}=G + E \), we hope to find new unitary matrices \( \tilde{U} \) and \( \tilde{V} \) such that:
\[
\tilde{U}^{H}\tilde{G}\tilde{V}=\begin{bmatrix}
\tilde{\Sigma}_{1}&0\\
0&\tilde{\Sigma}_{2}
\end{bmatrix}
\]
and this decomposition is close enough to the original decomposition.

- Define some key parameters:
\[
\hat{\varepsilon}=\sqrt{\|E_{12}\|_{F}^{2}+\|E_{21}\|_{F}^{2}}
\]
\[
\delta=\min_{\mu\in\text{sv}(G_{1}),\nu\in\text{svext}(G_{2})}|\mu - \nu|
\]
\[
\bar{\delta}=\delta-\|E_{11}\|_{2}-\|E_{22}\|_{2}
\]

- If the following conditions are met:
\[
\bar{\delta}>0\quad\text{and}\quad\frac{\hat{\varepsilon}}{\bar{\delta}}<\frac{1}{2}
\]
then there exist \( \Omega \) and \( \Gamma \) such that:
\[
\sqrt{\|\Gamma\|_{F}^{2}+\|\Omega\|_{F}^{2}}\leq\frac{1 + \sqrt{1 - 4(\hat{\varepsilon}/\bar{\delta})^{2}}}{1 - }