A note on preconditioning for \(M\)-matrix (Q812679)

The authors consider a linear system \(Tx = b\) with the following symmetric positive definite \(M\)-matrix: \(T = [t_{ij}]\); \(t_{ij} = t_{ji}\); \(t_{ij}< 0\), for \(i \neq j\), and \(\sum_{k=1}^n t_{ik} > 0\), for \(i = 1,\dots,n\). Introducing a good approximating inverse of \(T\) \textit{G. Simons} and \textit{Y. Yao} [Linear Algebra Appl. 281, No. 1--3, 97--103 (1998; Zbl 0936.15005)] and applying the Sherman-Morrison-Woodbury formula \textit{G. Golub} and \textit{C. F. van Loan} [Matrix computations. 3rd ed., Johns Hopkins Univ. Press, Baltimore (1996; Zbl 0865.65009)], a matrix \(P\) is defined which is used as preconditioner in the conjugate gradient method to solve the system \(Tx = b\). A theorem proved shows that a fast convergence rate can be obtained. Special examples for matrices with different, but small sizes are given. Compared with a preconditioner [\textit{T. Chan}, SIAM J. Sci. Stat. Comput. 9, No.~4, 766--771 (1988; Zbl 0646.65042)], which is known to be a good preconditioner for solving a large class of structured systems, the new preconditioner need about half the number of iterations.