Nonstationary multisplittings with general weighting matrices (Q2706308)

During an iterative solution of a system of linear algebraic equations with a symmetric positive definite matrix \(A\), one may split the matrix \(A = M - N\) repeatedly, i.e., using some \(p\) different splittings with non-singular \(M\)'s and with related scalar weights \(E\). A parallelized version of such an algorithm requires a matrix form of \(E\). The paper offers the corresponding generalization of the convergence theorems. Thus, one can make several iterations (with a ``non-stationary'' block-dependence of their number) in each processor. The authors also extend their analysis to two-stage parallelized iterative treatment of a semi-definite matrix \(A\).