## Abstract

Convergence rates for finite element discretizations of elliptic eigenvalue problems in the literature usually are of the following form: If the mesh width h is fine enough, then the eigenvalues, resp., eigenfunctions, converge at some well-defined rate. In this paper, we will determine the maximal mesh width h(0)-more precisely the minimal dimension of a finite element space-so that the asymptotic convergence estimates hold for h <= h(0). This mesh width will depend on the size and spacing of the exact eigenvalues, the spatial dimension, and the local polynomial degree of the finite element space. For example, in the one-dimensional case, the condition lambda(3/4)h(0) less than or similar to 1 is sufficient for piecewise linear finite elements to compute an eigenvalue lambda with optimal convergence rates as h(0) >= h -> 0. It will turn out that the condition for eigenfunctions is slightly more restrictive. Furthermore, we will analyze the dependence of the ratio of the errors of the Galerkin approximation and of the best approximation of an eigenfunction on lambda and h. In this paper, the error estimates for the eigenvalue/-function are limited to the selfadjoint case. However, the regularity theory and approximation property cover also the nonselfadjoint case and, hence, pave the way towards the error analysis of nonselfadjoint eigenvalue/-function problems.