A rational Arnoldi approach for ill-conditioned linear systems (Q2428121)

A rational Arnoldi method for the solution of ill-posed full rank linear systems \(Ax=b\) is derived by reformulating the problem as \(x = f_\lambda(Z_\lambda)b\), \(f_\lambda(z) = (1/z - \lambda)^ {-1}\), \(Z_\lambda = (A+\lambda I)^{-1}\) and applying the involved matrix-function \(f_\lambda\) in a suitable way to the Hessenberg-matrices obtained in the Arnoldi-process. A-priori and a-posteriori error representations are proven. Increasing the parameter \(\lambda\) improves the accuracy of the application of \(Z_\lambda\) (which means solution of a linear system with matrix \(A+\lambda I\)) in each Arnoldi-step while at the same time the speed of convergence of the Arnoldi-process decreases. By balancing the condition numbers of \(f_\lambda\) and \(Z_\lambda\), a strategy for the choice of \(\lambda\) is obtained. The performance of the method for unperturbed systems is demonstrated by several numerical examples. Finally, the method is extended to the case of noisy right hand sides \(\tilde{b}\) by adapting it to the Tikhonov system \((A^TA + \lambda H^T H)x_\lambda = A^T \tilde{b}\).