A study of the method of approximate factorization for solving grid elliptic boundary-value problems (Q1337876)

The paper concerns the preconditioning of an SPD matrix \(A\) by an approximate factorization of the form \(B = (G^{-1} + U)^ T G(G^{-1} + U)\) where \(U\) is the strictly upper triangular part of \(A\) and \(G\) is a diagonal matrix determined by the condition of equal rowsums of \(A\) and \(B\). The investigation of the preconditioning is performed on model problems with \(A\) arising from regular five-point discretization of the Poisson or Helmholtz equation on a rectangle. Especially, three examples with Dirichlet and Neumann boundary conditions and with discontinuous coefficients are studied in detail. The main result of the paper is a new estimate of the condition number \(\text{cond} (B^{-1}A)\). This estimate can be used for computing a realistic (quantitative) upper bound to \(\text{cond} (B^{-1} A)\) as well as for showing that \(\text{cond} (B^{-1} A) = O(h^{-1})\) without necessity to introduce some perturbed rowsum condition as e.g. rowsums of \(B\) are equal to rowsums of \(A + h^ 2 \xi \text{diag} (A)\) where \(h\) is the meshsize and \(\xi\) is a parameter which is \(\xi = O(h^{-1})\) at grid nodes corresponding to the Neumann boundary conditions and \(\xi = O(1)\) otherwise.