Can a finite element method perform arbitrarily badly?

From MaRDI portal





Suppose \(\{\chi_k\}_{k=0}^{\infty}\) is a sequence of positive real numbers converging monotonically to zero, no matter how slowly. A boundary value problem can be constructed whose finite element approximate solutions converge to the exact solution precisely as fast as \(\chi_k\searrow 0\). Based on the sequence \(\{\chi_k\}_{k=0}^\infty\), the authors construct a measurable coefficient function \(a(x)\) satisfying \(0.9<a(x)<1.1\). Denote by \(u\) the solution to the boundary value problem \[ -\tfrac{d}{dx}\bigl(a(x)\tfrac{du}{dx}\bigr) =0, \qquad 0<x<1, \] \[ u(0)=0, \qquad au'(1)=1, \] and by \(\{u_k\}_{k=0}^{\infty}\) its sequence of approximations established using standard piecewise linear finite elements on a uniform mesh. The error estimate \[ (0.03)\chi_k\leq \|u-u_k \|_E \leq (1.9)\chi_k \] holds for all \(k\), in the energy norm. The authors also treat pointwise and \(L_2\) norms. Furthermore, the error is distributed equally over the mesh elements so that adaptive mesh strategies based on error equidistribution will not improve the convergence rate. The function \(a(x)\) is constructed as a gap series in functions \(g_n(x)\) that alternate between the values \(\pm 1\) on intervals of length \(1/2^n\). The gap sizes and series coefficients depend on \(\chi_k\). Both the solution \(u(x)\) and its approximates \(u_k(x)\) can be constructed explicitly from \(a(x)\), and the error estimates follow. The authors include remarks pointing out the relationships between their results and more standard ones involving a smooth coefficient function \(a(x)\).




Cited in
(77)








This page was built for publication: Can a finite element method perform arbitrarily badly?

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4942773)