Perturbation of the SVD in the presence of small singular values (Q854854)

The author presents singular value decomposition perturbation bounds and expansions that are of use when an \(m \times n, m > n\), matrix \(A\) has small singular values. The main goal of the paper is to describe the difference between the sets of perturbed and unperturbed singular values, the difference between the left singular subspaces and the difference between the right singular subspaces. The motivation for presenting a new analysis is to derive theorems that are more accurate when the blocks of the given complex matrix (that are diagonal with positive diagonal elements), have small singular values. The first part of the paper gives subspace bounds that are closely related to those of \textit{P.-B. Wedin} [BIT 12, 99--111 (1972; Zbl 0239.15015)] but are stated so as to isolate the effect of any small singular values to the left singular subspace. In the second part first and second order approximations are given for perturbed singular values. The subspace bounds are used to show that all approximations retain accuracy when applied to small singular values. The paper concludes by deriving a subspace bound for multiplicative perturbations and using that bound to give a simple approximation to a singular value perturbed by a multiplicative perturbation.