The optimal perturbation bounds of the Moore-Penrose inverse under the Frobenius norm (Q846324): Difference between revisions

An interesting problem in the literature is the estimation of the perturbation bounds of the Moore-Penrose inverse. Among many authors, \textit{P.-A. Wedin} [BIT, Nord. Tidskr. Inf.-behandl. 13, 217--232 (1973; Zbl 0263.65047)] answered this problem by giving perturbation bounds of the Moore-Penrose inverse under general unitarily invariant norm, the spectral norm and the Frobenius norm, respectively. More specifically, it is proved by Wedin [loc. cit.] that in case we have two matrices \(A\in\mathbb{C}^{m\times n}\) and \(B=A+E\), then \(\left \| B^{\dagger }-A^{\dagger }\right \| \leq \mu \max \left \{ \left \| A^{\dagger }\right \| _{2}^{2},\left \| B^{\dagger }\right \| _{2}^{2}\right \} \left \| E\right \| \) where \(\mu \) is equal to \thinspace \(3,\frac{1+\sqrt{5}}{2}, \sqrt{2}\) for the unitarily invariant norm, the spectral norm and the Frobenius norm, respectively. The main purpose of this paper is to develop optimal constants \(\mu \) for the Frobenius norm, by using singular value decomposition. Two perturbation models are studied: the additive perturbation model where \(A\in\mathbb{C}^{m\times n}\) and \(B=A+E\), and the multiplicative perturbation model where \( A\in\mathbb{C}^{m\times n}\) and \(B=D_{1}^{\ast }AD_{2}\) with \(D_{1}\in\mathbb{C}^{m\times m},D_{2}\in\mathbb{C} ^{m\times m}\) nonsingular matrices (* denotes the conjugate transpose). The multiplicative perturbation model is a special case of the additive model since \(B=D_{1}^{\ast }AD_{2}\) can be rewritten as \(B=A+E\) with \( E=-\left( I_{m}-D_{1}^{\ast }\right) A-D_{1}^{\ast }A\left( I_{n}-D_{2}\right) \) or \(E=-A\left( I_{n}-D_{2}\right) -\left( I_{m}-D_{1}^{\ast }\right) AD_{2}\).