The elimination of Lanczos ghosting effects by MINRES filter diagonalization (Q1287186)

The paper is dealt with the elimination of Lanczos ghosting effects which appear during evaluation of the eigenvalues of large sparse real-symmetric or complex-Hermitian matrices. These effects duplicates of true eigenvalues which are stable with respect to the size of Lanczos subspace, and spurious eigenvalues which are unstable with respect to this size. Moreover, the subspace loses strict orthogonality. One of the ways to eliminate the ghosting effects is to use a filter diagonalization (FD) technique. In the paper the authors investigate their FD approach, based on the Lanczos algorithm connected with the minimum residual (MINRES) algorithm of \textit{C. C. Paige} and \textit{M. A. Saunders} [SIAM J. Numer. Anal. 12, 617-629 (1975; Zbl 0319.65025)] (presented in previous work of the authors), with reference to one- and two-dimensional applications. Their MINRES filter diagonalization (MFD) method, which has the properties that at high energies, where the Lanczos eigenvalues are not convergent, it gives improved estimates of the true eigenvalues and eigenvectors, and that at lower energies, where the Lanczos algorithm rapidly converges, the eigenstates and duplicates eigenvalues, it produces only true eigenvalues and eigenvectors, is applied to a one-dimensional Morse oscillator problem and to the Wyatt test matrix. Appropriate computer calculations are presented in both cases, and the authors show that their MFD algorithm applied to these problems eliminates the spurious and duplicate eigenvalues which complicate the interpretation of the regular Lanczos spectrum. Hence the authors conclude their MFD technique can be successfully used for solving similar problems in the future.