Bounds on condition number of a singular matrix and its applications (Q1763253): Difference between revisions

The condition number of a nonsingular matrix is defined as \(P(A)=\| A \| \| A^{-1}\| \). Similarly, when \(A\) is singular, then the Drazin inverse \(A^D\) is used and we can define for a matrix \(A \in \mathbb C^{n \times n}\) the condition number as \(P_D(A)=\| A\| \| A^D\| \). Let \(U\) be the set of all norms defined on \(\mathbb C^{n \times n}\). It is shown that for a singular matrix \(A \in \mathbb C^{n \times n}\), \(A \neq \alpha I\) and with at least one nonzero eigenvalue there is no finite upper bound of \(P_D(A)\) while the norm \(\| .\| \) varies on \(U\) and that \(\inf_{\| .\| \in U}\| A\| \| A^D\| = \rho(A) \rho(A^D)\), where, e. g., \(\rho(A)\) denotes the spectral radius of \(A\). This represents an extension of the result of \textit{H. C. Huang} [J. Comput. Math. 2, 356--360 (1984; Zbl 0569.65033)] for a nonsingular matrix \(A \in \mathbb C^{n \times n}\). A similar result, i. e. \(\| A\| \| A^D\| = \rho(A) \rho(A^D)\), is proved for a diagonal matrix with the Jordan canonical form \(A=P\text{Diag}(\lambda_1, \dots, \lambda_n)P^{-1}\) with eigenvalues \(\lambda_i\). Finally, the minimum property saying that \(\min\| AX- I\| _P\) can be attained when \(X=A^g\), where \(A^g\) is called the group inverse of \(A\) when the index \(k=1\) and the \(P\)-norm is \(\| A\| ^P=\| P^{-1}AP\| _2\).