SoftFEM: revisiting the spectral finite element approximation of second-order elliptic operators

TL;DR

This paper introduces softFEM, a novel finite element method that reduces stiffness by penalizing gradient jumps, leading to improved eigenvalue approximation and lower condition numbers, with proven optimal convergence and practical benefits demonstrated numerically.

Contribution

The paper presents a new softFEM approach that stabilizes eigenvalue approximation by damping high frequencies, maintaining optimal convergence rates, and providing a practical parameter choice for 1D problems.

Findings

SoftFEM improves eigenvalue approximation in the upper spectrum.

SoftFEM reduces the condition number of the stiffness matrix.

Numerical experiments confirm the advantages of softFEM over standard FEM.

Abstract

We propose, analyze mathematically, and study numerically a novel approach for the finite element approximation of the spectrum of second-order elliptic operators. The main idea is to reduce the stiffness of the problem by subtracting a least-squares penalty on the gradient jumps across the mesh interfaces from the standard stiffness bilinear form. This penalty bilinear form is similar to the known technique used to stabilize finite element approximations in various contexts. The penalty term is designed to dampen the high frequencies in the spectrum and so it is weighted here by a negative coefficient. The resulting approximation technique is called softFEM since it reduces the stiffness of the problem. The two key advantages of softFEM over the standard Galerkin FEM are to improve the approximation of the eigenvalues in the upper part of the discrete spectrum and to reduce the condition number of the stiffness matrix. We derive a sharp upper bound on the softness parameter weighting the stabilization bilinear form so as to maintain coercivity for the softFEM bilinear form. Then we prove that softFEM delivers the same optimal convergence rates as the standard Galerkin FEM approximation for the eigenvalues and the eigenvectors. We next compare the discrete eigenvalues obtained when using Galerkin FEM and softFEM. Finally, a detailed analysis of linear softFEM for the 1D Laplace eigenvalue problem delivers a sensible choice for the softness parameter. With this choice, the stiffness reduction ratio scales linearly with the polynomial degree. Various numerical experiments illustrate the benefits of using softFEM over Galerkin FEM.