Numerical methods for eigenvalues of singular polynomial eigenvalue problems (Q6985251)

This paper is about the solution of the singular PEP (polynomial eigenvalue problem) \(P(\lambda):=\sum_{k=0}^d A_n\lambda^k=0\) where \(d\ge2\), \(A_k\in\mathbb{C}^{m\times n}\), \(A_d\ne0\). Singular means here that \(m\ne n\) or \(m=n\) but \(\det(P(z))\equiv0\).\N\NIn their previous papers [\textit{M.E. Hochstenbach} et al., SIAM J. Matrix Anal. Appl. 40, No.~3, 1022--1046 (2019, Zbl 1435.65056); SIAM J. Matrix Anal. Appl. 44, No.~4, 1589--1618 (2023, Zbl 1526.65011)], these authors considered the solution of a singular matrix pencil (i.e., when \(d=1\)) by transforming it into a regular (full rank) problem by applying one of three perturbation techniques: rank completion, rank-projection, or augmentation. By using the (left and right) eigenvectors it is possible to identify the true eigenvalues of the original problem among the eigenvalues of the perturbed system.\N\NIn this paper the authors generalize these three perturbation methods to solve the singular polynomial eigenvalue problem. Especially identifying the true eigenvalues requires a modified analysis for \(d\ge2\).\N\NThe three methods are formulated as formal algorithms, they are compared to a classical linearization technique (i.e.\ first perturb and then linearize is proposed here, unlike the classical approach to first linearize and then perturb) and they are tested numerically by applying them to (1) the solution of bivariate polynomial systems, (2) the compution of zero group velicity (ZGV) points in wave analysis and (3) to solve several quadratic eigenvalue problems (QEP) from [\textit{D. Kressner} et al. BIT 63, No. 1, Paper No. 18, 25 p. (2023; Zbl 1508.65033)].