A convergence analysis of GMRES and FOM methods for Sylvester equations (Q1604691)

The authors discuss the convergence properties of the block generalized minimal residual (GMRES) method and the block full orthogonal method (FOM) for solving Sylvester equations \(S(X) = C\), where \(S(X) = AX - XB\), \(A \in R^{N \times N}\), \(B \in R^{p \times p}\), \(C \in R^{N \times p}\), \(N \gg p\). At first, the relationship between the Krylov subspaces constructed with \(S\) and \(A\) as well as the application of the block Arnoldi method to the Sylvester operator \(S\) and to the matrix \(A\) are discussed. Then, it is described how GMRES and FOM methods reduce the computation of an approximate solution of the Sylvester equation \(S(X) = C\) to solving a small least squares Sylvester problem and a small Sylvester equation, respectively. It is shown that the GMRES algorithm used for solving the Sylvester equation \(S(X) = C\) converges, when the fields of values \({\mathcal F}(A)\) and \({\mathcal F}(B)\) are disjoint. Furthermore, the stagnation in GMRES and its consequence on FOM are discussed. The breakdown in the block Arnoldi method and the consequence on FOM are analysed. Finally, the authors present some numerical experiments illustrating the theoretical results.