Enhanced RFB method (Q2574947)

The authors explain that, ``The underlying principle behind the use of bubble methods is that of enriching the Galerkin finite element space with functions (bubbles) having compact support on every element of the given partition. The bubbles are successively eliminated through static condensation\dots''. The residual-free bubble (RFB) method is constructed by choosing bubbles that satisfy the underlying differential equation on each element and with boundary values orthogonal to traces of the finite element shape functions. In the current work, the authors focus on two space dimensions and enhance the RFB method by adding bubble functions with support on pairs of elements sharing an edge. They term the resulting method the ``RFBe'' method. These additional bubble functions cannot be eliminated through static condensation, but if the number of additional bubble functions is kept small then computational effort should not be adversely affected. The presence of the additional functions extends the benefit of the RFBe method from improved accuracy in the presence of subgrid-scale features to improved accuracy in the presence of features such as boundary layers that are smaller than grid size in one direction but extend over many elements in the transverse direction. For the convection-diffusion equation on an axi-parallel mesh, additional bubble functions can be added for those elements adjacent to the part of boundary on which boundary layers arise. In the case that the mesh size is larger than the width of the boundary layer, the authors show a priori error estimates that indicate the RFBe error is much smaller than the RFB error. The bubble functions depend on the differential equation, so they must be computed numerically. The authors note that the error estimates do not consider the numerical errors introduced during this step. Three numerical examples are presented, all based on the convection-diffusion equation. The first two, with Shishkin-type meshes, show that the convergence of RFBe initially appears to be of higher order and then, as mesh size becomes small enough to resolve the boundary layer, convergence becomes similar to RFB. The third example includes an internal layer and employs a uniform mesh, and illustrates that the RFBe can be worthwhile in this case, unsupported by theoretical results.