1: \begin{abstract}
2: Rational filter functions can be used to improve convergence of
3: contour-based eigensolvers, a popular family of algorithms for the
4: solution of the interior eigenvalue problem. We present a framework
5: for the optimization of rational filters based on a non-convex
6: weighted Least-Squares scheme. When used in combination with the
7: FEAST library, our filters out-perform existing ones on a large
8: and representative set of benchmark problems. % We provide a framework
9: % for (non-convex) weighted Least-Squares optimization of rational
10: % filter functions
11: This work provides a detailed description of % our non-convex
12: % optimization. In particular we discuss
13: : (1) a set up of the
14: optimization process that exploits symmetries of the filter function
15: for Hermitian eigenproblems, (2) a formulation of the gradient
16: descent and Levenberg-Marquardt algorithms that exploits the
17: symmetries, (3) a method to select the starting position for the
18: optimization algorithms that reliably produces effective filters,
19: (4) a constrained optimization scheme that produces filter functions
20: with specific properties that may be beneficial to the performance
21: of the eigensolver that employs them.
22: % makes them particularly apt to be used
23: % in conjunction with Krylov-based linear system solves.% , poles of
24:
25: % the rational filter function with large imaginary parts which
26: % benefit convergence of Krylov based solvers for the arising linear
27: % system solves.
28: \end{abstract}
29: