Variants of the accelerated parameterized inexact Uzawa method for saddle-point problems (Q291897)

The authors consider the usual augmented system (with zero block entry in position 2,2) and start by splitting the system matrix into block-diagonal, block-lower triangular and block-upper triangular matrices containing three iteration parameters. The splitting gives rise to an iteration in which the matrix to be solved in every step is block-lower triangular. For the latter matrix, they consider two variants: a successive overrelaxation (SOR) method and a symmetric SOR (SSOR). Next, they prove convergence (resp. semi-convergence) of both variants for the case that the system matrix is non-singular (resp. singular) -- in case the iteration parameters satisfy certain inequalities. These inequalities, for the SSOR variant, suppose information about certain minimal and maximal eigenvalues. The authors provide detailed results for test examples (2D Stokes problems) of their methods including now a preconditioned generalized minimal residual (GMRES) method. Compared with other methods, those proposed here are shown to be viable competitors. Only the method proposed by \textit{Z.-Z. Bai} et al. [Numer. Math. 102, No. 1, 1--38 (2005; Zbl 1083.65034)], and generalized to the case of singular system matrices by \textit{B. Zheng} et al. [Linear Algebra Appl. 431, No. 5--7, 808--817 (2009; Zbl 1173.65026)], turns out to be sometimes better and has the advantage of known optimal iteration parameters. In turn, the parameters in the present paper are tuned numerically but the new methods are anticipated to be better suited for large sparse 1,1 blocks in the system matrix.