Sine transform based preconditioners for elliptic problems (Q4383435)

The paper deals with a preconditioning method for the system \(Ax=b\) where \(A\) is a symmetric and positive definite matrix of tridiagonal block form with diagonal blocks \(D_i\) and off-diagonal blocks \(A_i\). For a general square matrix \(Q\), the optimal sine transform approximation \(s(Q)\) to \(Q\) is defined. Then the preconditioner \(\widehat s(A)\) for the preconditioned conjugate gradient (PCG) method is constructed by a block-by-block way. It is demonstrated that \(\widehat s(A)\) coincides with a matrix \(G\) which has the same block form as \(A\). The respective blocks of \(G\), however, are \(s(D_i)\) and \(s(A_i)\). For two-dimensional domains, in this article the authors show that the cost for each PCG iteration is \(\mathcal O(n^2\log n)\) including the construction of \(\widehat s(A)\), and, for rectangular regions, the condition number for the preconditioned system \((\widehat s(A))^{-1}A\) is of order \(\mathcal O(1)\). Some extensions to irregular or three-dimensional domains are also given. The method is numerically tested and compared with the modified incomplete LU and the modified incomplete block Cholesky factorization. Though the results seem to be favourable for the proposed method, the reader might ask what they would be if irregular meshes, more general domains, and another operator (different from a perturbed Laplacian one) were considered.