Preconditioned minimal residual methods for Chebyshev spectral calculations (Q1820572)

Preconditioned iterative methods are considered for the linear systems of algebraic equations which arise when elliptic problems of second order are discretized by spectral methods. While these methods often are very accurate, they lead to dense, rather illconditioned coefficient matrices. The techniques considered here are based on preconditioning the operator with low accuracy finite difference and finite element approximations or by incomplete Cholesky factorizations of the corresponding sparse matrices. It should be noted the fast Fourier transforms can be used to find the matrix vector products required to compute the residuals related to the original models. In this quite carefully prepared paper, a number of iterative methods are studied. Of these is a normal equation version of the conjugate gradient method. The best results are obtained by version of a stationary second- degree method called the DuFort-Frankel method by the authors. A minimal residual strategy is used in which the two parameters are determined dynamically. A number of numerical results are given.