Restricted rank modification of the symmetric eigenvalue problem: Theoretical considerations (Q1106912): Difference between revisions

Let A be a symmetric matrix with known spectral decomposition. Two modified eigenvalue problems are treated: a) Perturbation by a matrix of low rank, i.e \(\tilde A=A+VV^ T\) where V has rank r. It is shown by establishing relations between the inertias of \(\lambda-A\), \(\lambda-\tilde A\) and \(I_ r-V^ T(\lambda-A)^{-1}V\), that the eigenvalues of \(\tilde A\) can be located to any desired accuracy by means of the last one. b) Restricted eigenvalue problem \(PAPx=\lambda x\), \(Px=x\), where P is an orthogonal projection on a \(n-r\)-dimensional subspace. Very similar inclusions are possible. The proof apply ideas from the Weinstein-Aronszajn methods for selfadjoit operators to the finite-dimensional case. The numerical implementation of a bisection method based on the results given here will be discussed later.