A truncated-CG style method for symmetric generalized eigenvalue problems (Q818183): Difference between revisions

The authors are concerned with the computation of the largest (rightmost) and smallest (leftmost) generalized eigenvalues of \(Ax=\lambda Bx\) and their corresponding eigenvectors. Here \(A,B\) are real symmetric matrices and \(B\) is positive definite. Since \(\frac{v_{1}^{t}Av_{1}}{v_{1}^{t}Bv_{1}}< \frac{y^{t}Ay}{y^{t}By}<\frac{v_{n}^{t}Av_{n}}{v_{n}^{t}Bv_{n}}\) the authors apply the general Riemannian trust-region algorithm to minimize the Rayleigh quotient on the compact manifold \(y^{t}By=1.\) Finding \(v_{1}\) yields the leftmost eigenpair. By looking at \(-Ax=\lambda Bx\) the rightmost eigenpair becomes the leftmost eigenpair. It is also proven that convergence is superlinear.