Backward errors for multiple eigenpairs in structured and unstructured nonlinear eigenvalue problems

Abstract: Given a nonlinear matrix-valued function $F(\lambda)$ and approximate eigenpairs $(\lambda_i, v_i)$, we discuss how to determine the smallest perturbation $\delta F$ such that $F + \delta F v_i = 0$; we call the distance between the $F$ and $F + \delta F$ the backward error for this set of approximate eigenpairs. We focus on the case where $F(\lambda)$ is given as a linear combination of scalar functions multiplying matrix coefficients $F_i$, and the perturbation is done on the matrix coefficients. We provide inexpensive upper bounds, and a way to accurately compute the backward error by means of direct computations or through Riemannian optimization. We also discuss how the backward error can be determined when the $F_i$ have particular structures (such as symmetry, sparsity, or low-rank), and the perturbations are required to preserve them. For special cases (such as for symmetric coefficients), explicit and inexpensive formulas to compute the $\delta F_i$ are also given.