Multistep matrix splitting iteration preconditioning for singular linear systems (Q2360679): Difference between revisions

To solve a singular system \(Ax=b\), the matrix \(A\) is split as \(A=M-N\) with \(M\) nonsingular. When iterated, this gives an iteration matrix \(H=M^{-1}N\) and after \(\ell\) steps, a preconditioner \(C^{(\ell)}=\sum_{i=0}^{\ell-1} H^iM^{-1}\) is obtained. If \(H\) is semiconvergent (i.e. \(\lim_{i\to\infty}H^i\) exists), then it is proved that the generalized minimal residual (GMRES) method with this preconditioner converges to a solution without breakdown. The spectral properties of the preconditioned matrix, a bound for the residual norm, and an analysis of the complexity is given. In flexible GMRES, the \(\ell\) may differ for every iteration. Also, here conditions for convergence without breakdown are given. Numerical experiments compare the methods with generalized shift-splitting and (skew-)Hermitian matrix splitting.