The main aim of a posteriori error estimation is to provide computable bounds for the error commited solving PDEs using finite elements methods. A crucial property for an a posteriori estimator is to be local, in other words to be able to bound the error on each cell of the mesh. Using this property we are able to determine on which part of the domain the error is mainly distributed. This information allows us to refine the mesh in this special zone and try to reduce the global error while maintaining a reasonable computational cost. Among all a posteriori error estimators we can cite the residual estimator described in [1] or the ZZ estimator introduced by O.C. Zienkiewicz and J.Z. Zhu in [2]. Here we will focus on an estimator introduced by R.E. Bank and A. Weiser in [3] and applied on elliptic PDEs.

Let us call true error the difference between the solution of our PDE and its Galerkin approximation. We build the Bank-Weiser estimator (BW estimator) from a specific discontinuous polynomial function called residual solution. This function is defined locally on each cell of the mesh as the elliptic projection (in the sense of Céa’s lemma, see [4]) of the (local) true error on a specific discontinuous finite elements (FE) space.

 

This FE space is chosen in order to ensure that this residual solution is well-defined. Indeed, the computation of this residual solution implies to solve local Neumann problems on each cell and specific assumptions must be made to ensure their well-posedness.

 

Since this solution is defined as a projection, it is the best approximation of the true error we can find on this subspace. Finally, we define the BW estimator as the sum over the cells of the norm of the energy of the residual solution.

 

The first advantage of the BW estimator is its accuracy for problems admitting a rough solution (H¹) as well as when the solution is more regular (H²). Indeed, numerical results has been done in the case H¹ in [3] when the solution admits a singularity. Similarily, the estimator has been generalized to the Stokes problem in [5] and the numerical results suggest a good accuracy in this case as well.

 

Another advantage of the BW estimator is the (relative) rapidity of computation, due to the choice of the FE space containing the residual solution.

However, the main disadvantage of the BW estimator lays in the lack of theoretical guaranties. Indeed, the proof of the equivalence between this estimator and the true error has been proved (in [3]) but only under a constraining

assumption called saturation assumption. This assumption can be related to the regularity of the data (and then the regularity of the solution). It is possible to find very simple cases where the saturation assumption fails, see [6]. Also, the numerical examples shown in [3, 5] do not satisfy this assumption but the BW estimator remains accurate.

For the time being the theory can not fully explain this accuracy except in the case of linear FE with a proof from R. H. Nochetto in [7]. Even if the  numerical results seem to suggest it, the extension of this proof to higher order FE is still an open question.

 

 

Our aim is to implement an efficient algorithm computing the BW estimator using the FEniCS software. By definition the computation of the BW estimator is fully parallelisable which is very interesting in order to save computational time. Then we will be able to test it on mechanics problems and compare it to other estimators like the residual estimator or ZZ estimator and corroborate the results of the papers cited above.