Accelerated simultaneous iterations for large finite element eigenproblems (Q1115098)

The problem of computing some few of the smallest eigenvalues and corresponding eigenvectors of the generalized eigenvalue problem \(Ax=\lambda B_ x\) for symmetric, positive definite and sparse matrices A and B of high order is considered. An attractive idea and an algorithm are described that iterates simultaneously a set of vectors in order to find the smallest stationary values of the Rayleigh quotient by means of the method of conjugate gradients. This basic algorithm with known bad convergence properties is combined with a preconditioning of the eigenvalue problem by using an incomplete factorization of A to improve substantially the convergence. Numerical experiments indicate the superiority of the preconditioned simultaneous Rayleigh quotient minimization algorithm over existing methods that exploit the sparsity to a full extent.