Minimization principles and computation for the generalized linear response eigenvalue problem (Q2450894)

Minimization principles and Cauchy-like interlacing inequalities for the generalized linear response (LR) eigenvalue problem are analyzed. The generalized LR eigenvalue problem can be recast in the form \[ H z = \left(\begin{matrix} 0 & K \\ M & 0 \end{matrix} \right) \left( \begin{matrix} y \\ x \end{matrix} \right) = \lambda \left( \begin{matrix} E_+ & 0 \\ 0 & E_- \end{matrix} \right) \left( \begin{matrix} y \\ x \end{matrix} \right) = \lambda E z, \tag{1} \] where \(K\) and \(M\) are \(n \times n\) real symmetric positive semi-definite matrices, one of them being definite, and \(E_{\pm}\) are nonsingular and \(E_+^T = E_-\). All the eigenvalues of (1) are real and such that \(-\lambda_n \leq \cdots \leq -\lambda_1 \leq +\lambda_1 \leq \cdots \leq +\lambda_n\), \(\lambda_1\) verifies an extension of Thouless' minimization principle, which can be further generalized (by means of a subspace version) to the sum of the first \(k\) eigenvalues. With the help of Cauchy-like interlacing inequalities involving the eigenvalues \(\lambda_i\) and the eigenvalues of another related matrix it is possible to use a variation of the classical conjugate gradient method to design an efficient algorithm which allows the authors to simultaneously compute several of the smallest eigenvalues with the positive sign and the corresponding eigenvectors.