On the convergence of generalized parallel multisplitting iterative methods for semidefinite linear systems (Q1940970)

This paper studies generalized multisplitting methods for large sparse linear systems \(Ax = b\) denoted by \(x^{(k+1)} = (I-RA)^{(k)} + Rb\). Here \(R\) may differ from the classical uses of \(R = M^{-1}\), \(R = M^\dagger\) or \(R\) as the Drazin inverse of \(M\) for an additive splitting \(A = M - N\). The main results are: If \(A\) is positive semidefinite, if all \(R_j^T+R_j-R_j^TAR_j\) are positive semidefinite on the column space of \(A\) in a generalized stationary parallel multisplitting method then the quotient convergence and convergence are equivalent. This is generalized to generalized non-stationary parallel multisplitting methods and under relevant conditions, convergence or quotient convergence is proved. A 2 by 2 example shows the value of this method for a system with singular semidefinite \(A\).