In this paper we propose a new method to stabilize nonsymmetric indefinite problems. The idea is to solve a forward and an adjoint problem simultaneously using a suitable stabilized finite element method. Both stabilization of the element residual and of the jumps of certain derivatives of the discrete solution over element faces may be used. Under the assumption of well-posedness of the partial differential equation and its associated adjoint problem we prove optimal error estimates in H 1 and L 2 norms in an abstract framework. Some examples of problems that are neither symmetric nor coercive but that enter the abstract framework are given. First we treat indefinite convectiondiffusion equations with nonsolenoidal transport velocity and either pure Dirichlet conditions or pure Neumann conditions and then a Cauchy problem for the Helmholtz operator. Some numerical illustrations are given.
Introduction.The computation of indefinite elliptic problems often involves certain conditions on the mesh size h for the system to be well-posed and for the derivation of error estimates. The first results on this problem are due to Schatz [19]. The conditions on the mesh parameter can be avoided if a stabilized finite element method is used. Such methods have been proposed by Bramble, Lazarov, and Pasciak [4] and Ku [16] or more recently the continuous interior penalty (CIP) method for the Helmholtz equation suggested by Wu [21], and Zhu, Burman, and Wu [20]. The method proposed herein has some common features with both these methods but appears to have a wider field of applicability. We may treat not only symmetric indefinite problems such as the (real valued) Helmholtz equation but also nonsymmetric indefinite problems such as convection-diffusion problems with nonsolenoidal convection velocity or the Cauchy problem. The latter problem is known to be illposed in general [1] and will mainly be explored numerically herein. For all these cases we show that if the primal and adjoint problems admit a unique solution with sufficient smoothness the proposed algorithm converges with optimal order. The case of hyperbolic problems is treated in the companion paper [5].The idea of this work is to assume ill-posedness of the discrete form of the PDE and regularize it in the form of an optimization problem under constraints. Indeed we seek to minimize the size of the stabilization operator under the constraint of the discrete variational form. The regularization terms are then chosen from well-known stabilized methods respecting certain design criteria given in an abstract analysis. This leads to an extended method where simultaneously both a primal and a dual