On the minimum convergence factor of a class of GSOR-like methods for augmented systems (Q745227)

The authors consider the usual augmented \(2\times 2\) block system of linear equations characterized by matrices \(A\) and \(B\) and propose a general iterative method for the numerical solution of the system. This method contains a preconditioning matrix \(Q\) and two functions of parameters and specializes to 5 previously investigated generalized successive overrelaxation (GSOR)-like methods. For some of the latter, the optimal iteration parameters were not obtained earlier. Under a number of conditions (like the reality of all eigenvalues of \(Q^{-1}B^TA^{-1}B\)), they establish the convergence of their general method and find also the optimal parameters of all the mentioned 5 methods which methods turn out to be equivalent. Analyzing the behaviour of the convergence factor in a vicinity of the optimal parameters for two often considered cases of augmented systems (one of them corresponding to a discretized Stokes system), they distinguish one of the 5 methods to be less sensitive to a change of the iteration parameters near the optimal ones.