On the preconditioning of matrices with skew-symmetric splittings (Q5934363)

This paper studies the Hermitian splitting \(A=H+S\) of a system matrix \(A\) in order to solve a linear system \(Ax=b\) iteratively. If \(A\) and \(H\) are nonsingular, then \(A=H+S=H(I+H^{-1}S)\) and \(A^{-1} = (I+H^{-1}S)^{-1}H^{-1} \approx M^{-1}:= (I-H^{-1}S)H^{-1}\) by truncated Taylor expansion. Thus \(M^{-1}\) could be used as a preconditioner. If the skew part \(S\) is large, however, compared to the Hermitian part \(H\), this approximation and preconditioner is not very good and the paper suggests to use the shifted matrices \(S(\alpha) = S + \alpha I,\;H(\alpha) = H-\alpha I\) with \(A= H(\alpha) + S(\alpha)\) instead and to approximate \(A^{-1}\) by \(M^{-1}(\alpha)=(I-S(\alpha)^{-1}H(\alpha))S(\alpha)^{-1}\). An analysis of optimal choices for \(\alpha \in \) {\textbf{R}}\(^+\) is given, as well as numerical experiments and comparisons with other incomplete factorizations.