A reciprocal preconditioner for structured matrices arising from elliptic problems with jumping coefficients (Q417417): Difference between revisions

The authors present a preconditioner for a class of structured and sparse matrices \(\Gamma(a)\)'s, which are functions of \(a\). The matrices in the form of \(\Gamma(a)\) may arise after discretizing elliptic problems with a discontinuous diffusion coefficient \(a\). The authors show that the preconditioner \(P\) is of the form \(P = \Gamma^{-1}(1) \Gamma(1/a) \Gamma^{-1}(1)\) and can be considered as an approximation to \(\Gamma^{-1}(a)\) since \(P\) and \(\Gamma^{-1}(a)\) are spectrally equivalent. The authors also report their observation that \(\Gamma(a)P\) has a cluster at unity. In the one-level (1D) case, the authors show that the preconditioned matrix is exactly the identity plus a matrix of rank one. For the multi-level case (specifically, the 2D stratified case), a rigorous proof is provided and some hypotheses and conjectures are proposed. Numerical tests on 2D and 3D problems for the proposed preconditioner are performed with iterative methods such as the preconditioned conjugate gradient method and the generalized minimal residual method, and the authors show that the proposed precoditioner provides fast convergence.