On level-2 condition number for the weighted Moore-Penrose inverse (Q2426899)

This paper deals with condition numbers for the weighted Moore-Penrose inverse and bounds of level-2 condition numbers. The condition number measures the sensitivity of the output of a problem with respect to small perturbations of the input data. A problem \(\Pi\) induces a new problem, namely the computation of its condition number \(\text{cond}_{\Pi}(d)\) for a given input \(d\). In general, condition numbers can not be computed exactly, and hence it is of interest to know the sensitivity of the problem to compute the condition number, that is, the condition number of the condition number. This level-2 condition number was defined by \textit{J. W. Demmel} [Numer. Math. 51, 251--289 (1987; Zbl 0597.65036)] and denoted by \(\text{cond}_{\Pi}^{[2]}(d)\). For an arbitrary matrix \(A \in \mathbb C^{m \times n}\) and Hermitian positive definite matrices \(M\) and \(N\) of order \(m\) and \(n\), respectively, the weighted Moore-Penrose inverse of \(A\) is denoted by \(A_{MN}^{\dag}\). The authors use in this work the weighted matrix norm \[ \| A\| _{MN}=\max_{\| y\| _N=1} \| Ay \| _M=\max_{\| y\| _N=1} \| M^{1/2}Ay\| _2. \] They present characterizations of condition numbers for the weighted Moore-Penrose inverse under the weaker condition, namely \(\kappa_{MN}(A)\), defined as \[ \kappa_{MN}(A)=\lim_{\varepsilon \rightarrow 0} \sup_{\|\Delta A\|_{MN}\leq \varepsilon, A+\Delta A \in \Sigma(r)} \frac{\|(A+\Delta A)_{MN}^{\dag}-A_{MN}^{\dag}\|_{NM}\|A\|_{MN}} {\|A_{MN}^{\dag}\|_{NM}\|\Delta A\|_{MN}}, \] where \(\Delta A\) is a perturbation of \(A\) and \(\Sigma(r)\) denotes the set of \(m \times n\) matrices with rank \(r=\text{rank}(A)\). Specifically, the authors prove that \[ \text{cond}_{MN}(A) \leq \kappa_{MN}(A) \leq \lambda \text{cond}_{MN}(A) \] where the constant value \(\lambda\) depend on \(\text{rank}(A)\) and \(\text{cond}_{MN}(A)=\| A\| _{MN}\| A_{MN}^{\dag}\| _{NM}\). Finally, they also prove that for all \(A \in\mathbb C^{m \times n}\), if \(\text{rank}(A+\Delta A)=\text{rank}(A)\), then \[ \text{cond}_{MN}(A) \leq \text{cond}_{MN}^{[2]}(A) \leq \text{cond}_{MN}(A)+1. \]