Divergence of FEM: Babuška-Aziz triangulations revisited. (Q499034)

The author re-examines the arguments and counterexamples presented by \textit{I. Babuška} and \textit{A. K. Aziz} [SIAM J. Numer. Anal. 13, 214--226 (1976; Zbl 0324.65046)] and concerning the well-known maximum angle condition. He shows that the convergence of a finite element method (FEM) solution can be arbitrarily slow or even lost on a sequence of obtuse triangulations. A Poisson problem with a polynomial solution \(u\) is considered on the unit square \(\Omega\). Triangulations \(\mathcal T_{n,m}\), known from the Babuška-Aziz paper, comprise right angle triangles and triangles with an obtuse angle that converges to \(\pi\) if \(m\) goes to infinity faster than \(n\). Continuous piecewise linear functions form \(V_{n,m}\), the finite element space over \(\mathcal T_{n,m}\). It is proved that the best \(H^1(\Omega)\)-approximation error of \(u\) w.r.t. \(V_{n,m}\) is greater than a positive constant if \(m/n^5\geq c>0\) and \(n\to\infty\). The main result states that the FE Galerkin solutions determined by \(\mathcal T_{n,m}\) and \(V_{n,m}\) converge to \(u\) in the energy norm if and only if \(m/n^2=o(1)\). This paper seems to provide the first formal proof of the divergence of the FEM for a standard elliptic problem with a smooth solution. Six remarks also offer valuable contributions to the subject.