A simple method for error bounds of eigenvalues of symmetric matrices (Q5940698): Difference between revisions

If \(\rho_1\leq\rho_2\leq\cdots\leq\rho_n\) are the eigenvalues of a symmetric matrix \(A\) and \(\widetilde{\rho}\) is an approximate eigenvalue, then, given \(\delta_1,\delta_2\geq 0\), it is shown how to compute an interval \([\widetilde{\rho}-\varepsilon(\delta_1),\widetilde{\rho}+\varepsilon(\delta_2))\) that contains the eigenvalues \(\rho_k,\ldots,\rho_{k+r}\), where \(k-1\) and \(k+r\) are the number of negative eigenvalues of \(Y_1=A-(\widetilde{\rho}-\delta_1)I\) and \(Y_2=A-(\widetilde{\rho}+\delta_2)I\), respectively. The \(\varepsilon(\delta_i)\) are computed from the numerical error in the Cholesky decompositions of \(Y_i\). This technique can also be applied to the generalized eigenvalue problem \(Ax=\rho Bx\) if \(A\) and \(B\) are positive definite.