Computation of eigenvalues by numerical upscaling (Q888523)

The paper presents a two-level technique for the computation of lowermost eigenvalues of self-adjoint linear elliptic second-order partial differential equations. In order to avoid the solution of the eigenvalue problem for the fine scale discretization, the authors construct a generalized finite element space on a coarser discretization scale \(H\) with dimension \(N_H\). The two-scale decomposition uses weighted Clement-type quasi-interpolation operators and involves the approximate solution of \(N_H\) independent linear equations on the fine scale. Then the solutions of the corresponding eigenvalue problem on this low-dimensional space approximate the first \(N_H\) eigenpairs of the fine scale problem in a superconvergent way. It is proved that without any assumptions on the smoothness of the eigenfunctions the error between the eigenvalues of the low-dimensional problem and the first \(N_H\) eigenvalues of the fine scale problem is at least of order \(H^4\). Numerical experiments for \(L\)-shaped domains, problems with rough coefficients, and unstructured meshes illustrate the theoretical results.