On the infinite-dimensional QR algorithm (Q2315193)

The determination of spectra of operators over infinite-dimensional spaces is well known to be a very hard computational problem. In spite of several investigations over the past many decades, one cannot determine whether a given class of operators is amenable towards computation of its spectra and eigenvectors, supplemented by results on convergence rates and error estimates. One may observe that such problems arise frequently in one application problem or the other warranting a theoretical study as well as computational procedures. The authors of this very interesting work address the question as to which classes of operators allow for computations of spectra along with a study on convergence rates and error control. As an answer, they prove a generalization of the QR algorithm for infinite matrices satisfying the condition that each column has finitely many nonzero entries. The proposed algorithm is executable on a finite machine and it is shown that the one may recover the extremal parts of the spectrum and the corresponding eigenvectors, together with results on convergence rates and error control. A variety of numerical illustrations are presented and a comparative study is made. Examples include Toeplitz/Laurent operators and their perturbations, PT-symmetry in quantum mechanics, Hopping sign model in sparse neural networks and NSA Anderson model in superconductors. The proposed infinite dimensional QR algorithm is shown to perform better than what is predicted theoretically, in some cases.