Deflation for the Symmetric Arrowhead and Diagonal-Plus-Rank-One Eigenvalue Problems

Abstract

We discuss the eigenproblem for the symmetric arrowhead matrix $C = (\begin{smallmatrix} D \& {z} {z}^T & \alpha \end{smallmatrix})$, where $D \in \mathbb{R}^{n \times n}$ is diagonal, ${z} \in \mathbb{R}^n$, and $\alpha \in \mathbb{R}$, in order to examine criteria for when components of ${z}$ may be set to zero. We show that whenever two eigenvalues of $C$ are sufficiently close, some component of ${z}$ may be deflated to zero, without significantly perturbing the eigenvalues of $C$, by either substituting zero for that component or performing a Givens rotation on each side of $C$. The strategy for this deflation requires ${\mathcal{O}(n^2)}$ comparisons. Although it is too costly for many applications, when we use it as a benchmark, we can analyze the effectiveness of ${{O}(n)}$ heuristics that are more practical approaches to deflation. We show that one such ${\mathcal{O}(n)}$ heuristic finds all sets of three or more nearby eigenvalues, misses sets of two or more nearby eigenvalues under limited circumstances, and produces a reduced matrix whose eigenvalues are distinct in double the working precision. Using the ${\mathcal{O}(n)}$ heuristic, we develop a more aggressive method for finding converged eigenvalues in the symmetric Lanczos algorithm. It is shown that except for pathological exceptions, the ${\mathcal{O}(n)}$ heuristic finds nearly as much deflation as the ${\mathcal{O}(n^2)}$ algorithm that reduces an arrowhead matrix to one that cannot be deflated further. The deflation algorithms and their analysis are extended to the symmetric diagonal-plus-rank-one eigenvalue problem and lead to a better deflation strategy for the LAPACK routine dstedc.f.