A new iterative method for solving a class of complex symmetric system of linear equations (Q501962)

To avoid complex arithmetic, the authors consider an equivalent \(2\times 2\) real block system (with diagonal blocks \(W\) and \(T\)) to a given complex linear system and propose a one-parameter iterative method called scaled-splitted iteration (SCSP). The eigenvalues of the iteration matrix of this method are investigated for the case of \(W,T\) being symmetric positive semi-definite, its spectral radius is shown to be less than \(1\) when the iteration parameter obeys certain inequalities depending on the extreme generalized eigenvalues of the pair \(W,T\). Similarly, the optimal iteration parameter is obtained in dependence of those extreme eigenvalues. The splitting also gives rise to a preconditioner, and the preconditioned matrix is shown to have eigenvalues on a straight line orthogonal to the real axis and with center \((1,0)\). Similar results are obtained for an iteration (called PGSOR) proposed by the authors in [Numer. Linear Algebra Appl. 22, No. 4, 761--776 (2015; Zbl 1363.65049)]. Here, the eigenvalues of the preconditioned matrix are contained in a circle around the same centre \((1,0)\). They show also results of extensive numerical experiments comparing their methods to GMRES and its restarted or ILU-preconditioned variants on 4 example systems (including a discrete 3D Helmholtz equation). Whereas for small systems the Matlab backslash operator is faster, for large systems the preconditioned forms of GMRES and also a further method proposed by \textit{Z.-Z. Bai} [Sci. China, Math. 56, No. 12, 2523--2538 (2013; Zbl 1305.65114)], perform well and produce computing times not too far from each other.