Geometric Sparsification of Closeness Relations: Eigenvalue Clustering for Computing Matrix Functions

Abstract: We show how to efficiently solve a clustering problem that arises in a method to evaluate functions of matrices. The problem requires finding the connected components of a graph whose vertices are eigenvalues of a real or complex matrix and whose edges are pairs of eigenvalues that are at most \delta away from each other. Davies and Higham proposed solving this problem by enumerating the edges of the graph, which requires at least $\Omega(n^{2})$ work. We show that the problem can be solved by computing the Delaunay triangulation of the eigenvalues, removing from it long edges, and computing the connected components of the remaining edges in the triangulation. This leads to an $O(n\log n)$ algorithm. We have implemented both algorithms using CGAL, a mature and sophisticated computational-geometry software library, and we demonstrate that the new algorithm is much faster in practice than the naive algorithm. We also present a tight analysis of the naive algorithm, showing that it performs $\Theta(n^{2})$ work, and correct a misrepresentation in the original statement of the problem. To the best of our knowledge, this is the first application of computational geometry to solve a real-world problem in numerical linear algebra.