New perturbation analysis for generalized saddle point systems (Q835745)

The saddle point linear system \[ \left(\begin{matrix} A&B^T\cr B&C\end{matrix}\right) \left(\begin{matrix} x\cr y\end{matrix}\right) = \left(\begin{matrix} f\cr g\end{matrix}\right) \] with \(A,B,C\) matrices is perturbed with \(\|\Delta A\|_F\leq \epsilon d_1\), \(\|\Delta B\|_F\leq \epsilon d_2\), \(\|\Delta C\|_F\leq \epsilon d_3\), \(\|\Delta f\|_2\leq \epsilon d_4\), \(\|\Delta g\|_2\leq \epsilon d_5\) and \(\epsilon\) small. Using tensor product techniques, upper bounds for \(\|\Delta x\|_2\) and \(\|\Delta y\|_2\) are given with first and second order terms in \(\epsilon\). For \(\|(\Delta x,\Delta y)\|_2\) only a first order upper bound is obtained. A condition number gives the maximal relative perturbation of the solution in an \(\epsilon\) neighborhood of the original data when \(\epsilon\) tends to zero. Depending on whether the neighborhood is defined in terms of a global joint norm or in terms of the individual norms of the different perturbations, different condition numbers can be defined. An exact expression is derived for the condition number in the first case and an upper bound in the second case. The numerical results are comparable with what was obtained by \textit{H.~Xiang, Y.~Wei} and \textit{H.~Diao} [Linear Algebra Appl. 419, No.~1, 8--23 (2006; Zbl 1109.65043)].