The streamline-diffusion method for conforming and nonconforming finite elements of lowest order applied to convection-diffusion problems (Q5945714)

The authors investigate stability and convergence properties of the streamline diffusion finite element method for both conforming and nonconforming low order elements. In particular, they are interested in: (1) conforming linear triangular elements, (2) conforming bilinear quadrilateral elements, (3) nonconforming linear elements of Crouzeix and Raviart, and (4) nonconforming rotated bilinear quadrilateral elements. At first, the authors summarize results in the so-called streamline diffusion norm for the streamline diffusion method applied to the Dirichlet problem for the equation \[ -\varepsilon\Delta u+b\cdot\nabla u +cu=f. \] This discussion is followed by a superconvergence result for nonconforming elements. \[ \|u-u_h\|_{H_0} \leq Ch^2\|u\|_{H_3} \] where \(u_h\) is the approximate solution on a mesh of size \(h\). This superconvergence result requires that both \(\varepsilon\) and the streamline upwind parameters are \(O(h^2)\) and the mesh is based on square elements. The authors also demonstrate that one fundamental lemma that is used to prove superconvergence is sharp and cannot be used to prove superconvergence when \(\varepsilon\) or the streamline upwind parameter is \(O(h)\). The final section contains two numerical examples, both using meshes based on square elements. The first example, with a smooth solution, exhibits \(O(h^{3/2})\) convergence in the streamline diffusion norm except in the conforming quadrilateral case, where it is \(O(h^2)\). Both of these rates agree with theory. Convergence in the \(H_0\) norm is also \(O(h^2)\), illustrating the superconvergence result proved earlier. The second example has a solution with boundary layers. Convergence in the \(H_0\) norm is only \(O(h^{1/2})\) because of these layers. Interestingly, on a subdomain that excludes the boundary layers, convergence exhibits the same trends as the smooth example exhibits, including superconvergence.