On the subspace projected approximate matrix method. (Q499032)

The paper discusses a modification of the iterative method of Davidson for computing a few eigenvalues of a Hermitian matrix that is called SPAM (the subspace projected approximate matrix method). The method strives to reduce the costs to solve the eigenvalue problem using an inner loop over a sequence of matrices that increasingly better approximate the given matrix \(A\). A crucial step of the SPAM is the choice of the initial matrix \(A_0\) of the sequence. The authors show that special choices of \(A_0\) lead to well-known eigenvalue methods. A more sophisticated choice by approximations from below then leads to a variant of the Lanczos method that could be considered as a source of good starting approximations to the Jacobi-Davidson method. Numerical experiments provided by the authors then demonstrate various aspects of the SPAM method and provide a real comparative study of its various aspects.