The trace minimization method for the symmetric generalized eigenvalue problem (Q1591180): Difference between revisions

With solutions of a generalized eigenvalue problem (for the lowest part of the spectrum) kept in mind, the authors start from a Rayleigh-Ritz process on a subspace and from its (favorable) comparison with the method of minimization of trace. Still they notice, step-by-step, several merits of the latter approach (first of all, its capability of avoidance of stagnation in larger dimensions) and remove some of its shortcomings (in particular, using an elaborate shifting strategy). In order not to weaken the robustness of such an approach (endangered by the unstable convergence), they finally propose the use of the expanding subspaces in the Lanczos and Davidson spirit. This gives their final proposal which can be understood as an improved Jacobi-Davidson scheme and compares very well with it standard block version, especially when the initial subspace is not optimal.