Restarted GMRES preconditioned by deflation (Q1919484)

The authors present a prospective preconditioning technique for the restarted GMRES algorithm. For that purpose a new restarted GMRES scheme for solving the linear system \(AM^{-1}_i \widehat{x} = b\), \(\widehat{x} = M_i x_i\), is given, where \(M_i\) is the preconditioner in the \(i\)th GMRES cycle, \(M_1 = I\). Based on an invariant subspace approach, after each restart the smallest eigenvalues of the current matrix \(AM^{-1}_i\) are estimated and the new preconditioner \(M_{i+1}\) is updated so that these eigenvalues are removed and replaced by eigenvalues equal to the largest modulus of eigenvalues of \(A\). The eigenvalues of \(AM^{-1}_{i+1}\) will be a multiple eigenvalue equal to this largest modulus and the eigenvalues of \(AM^{-1}_i\) which are not removed. Implementation issues and several numerical results are given.