A refinement method for maximal deflating bases of regular pencils (Q1330474)

Generalized eigenvalues \(\lambda\) of a regular matrix pencil \((A,B)\) and the associated maximal deflating subspace \(M\) are computed numerically. A nonlinear operator \(F(X) = AX - BX (W \Phi_ 0)^*AX\) \((A, B, W \in \mathbb{C}^{n \times n}\); \(\Phi_ 0, X\in \mathbb{C}^{n \times m})\) is introduced and the solution \(\Phi\) of the equation \(F(X) = 0\) is a basis of \(M\) normalized by suitable linear norms. Local convergence methods such as Newton or inexact Newton methods are used to solve the equation \(F(X) = 0\). For a Hermitian positive definite matrix pencil a new algorithm, called Right-Lanczos method, is developed to compute the starting basis \(\Phi_ 0\). This algorithm contains a projective step where the contributions of all undesired directions are progressively cut off. Numerical refinement of an approximate eigenbasis by iterations, the effect of defective eigenvalues on the computed maximal deflating basis, and the computation of some natural frequencies of a girder embedded at one of its ends are discussed.