Multisplitting methods: Optimal schemes for the unknowns in a given overlap (Q2706274)

The author presents an optimal alternative of the weighted sums which are formed in the traditional multisplitting (MS) methods when a linear algebraic problem (\(Au=b\)), where the set of unknowns is an union of subsets, has to be solved. The convergence of this new form of the MS method is studied for both symmetric positive definite (SPD) matrices and M-matrices. NEWLINENEWLINENEWLINEOptimal is interpreted in three ways. First, the spectral radius of the iteration matrix \(H\) has to be minimal. Second, for the chosen iteration matrix the residual \(R(u)=r^T(u)r(u)\) has to be minimal for \(u=u^+\), where \(u^+\) is the the next value of the solution \(u\). Third, only when the coefficient matrix is SPD, \(J(u^+)\) has to be a minimum, where \(J(u) = \frac{1}{2}u^TAu - u^Tb\). NEWLINENEWLINENEWLINETwo variants of the weighted MS algorithm are considered: MinR MS and MinJ MS. Numerical results obtained for two test problems: (i) a steady state boundary value problem in \(\mathbb R^2\); (ii) fluid flow in driven cavity problem. In the driven cavity problem the new MS algorithm (MinR MS) is used as a preconditioner in the restarted GMRES algorithm.