Applying GMRES to the Helmholtz equation with shifted Laplacian preconditioning: What is the largest shift for which wavenumber-independent convergence is guaranteed? (Q500366): Difference between revisions

To solve the problem \((\Delta +k^2)u=-f\) in \(\Omega\), \(\partial_n u-iku=g\), on its boundary, using Galerkin, leads to a system \(Au=f\). A preconditioner for \(A\) is obtained by considering the shifted operator \(\Delta+k^2+i\epsilon\) leading to a matrix \(A_\epsilon\). In practice one uses an approximation \(B_\epsilon\) that is easier to compute. So one should try to consolidate two conflicting conditions: (1) \(A_\epsilon\) should be a good preconditioner for \(A\) and (2) \(B_\epsilon\) should be a good preconditioner for \(A_\epsilon\). (1) requires \(\epsilon\) to be small and (2) needs it to be large. The paper gives conditions on how small \(\epsilon\) should be to satisfy (1) and get \(k\)-independent convergence for the generalized minimal residual (GMRES) method. This is investigated for the interior impedance problem given above and also for the more complex truncated sound-soft scattering problem (which has a different domain inside the domain where \(u\) has to be solved). The results hold for left and right preconditioners. It boils essentially down to the fact that \(\epsilon/k\) should be small enough besides conditions on the domain. A follow-up paper has been prepared and is available on arXiv: [\textit{I. G. Graham, E. A. Spence} and \textit{E. Vainikko}, ``Domain decomposition preconditioning for high-frequency Helmholtz problems using absorption'', \url{arXiv:1507.02097}].