No label defined (Q4394865)

A method for assessing the quality of an error estimator and improving it is presented. The originality of the analysis is to use a randomized error function in order to a priori evaluate the average behavior of the estimator. The systematic bias introduced in the estimate can, hence, be corrected and an unbiased estimate is obtained. This analysis requires to introduce mathematical tools which are not usual in the domain of the finite element error estimation. Thus, we develop a set of hyperspherical coordinates allowing to integrate in the hyperspheres. We also introduce the generalized concept of angle between subspaces and we use it to synthetically describe the relative position of two subspaces. The average efficiency of the error estimator is proved to depend on this synthetic information. Moreover, we give a simple expression that approximately describes this average behavior. The application to a simple example of error estimation in a bidimensional Poisson equation with analytical solution shows that this theory can be used in both the analysis and the practical use of the error estimator. For the entire collection see [Zbl 0884.00048].