\(\mathcal H\)-matrices for convection-diffusion problems with constant convection (Q1404548)

This paper presents an \(\mathcal H\)-matrix formulation for the inverse of the stiffness matrix arising from convection-diffusion equations with specified convection direction. This inverse matrix is full, in general. The author shows that when the mesh is aligned with the convection direction, an optimal \(\mathcal H\)-matrix approximation can be constructed for the inverse of the stiffness matrix. This approximation requires \(O(n\log n)\) storage, for \(n\) degrees of freedom, with similar operation count for matrix-vector products. Constructing an \(\mathcal H\)-matrix requires both a hierarchical decomposition of the mesh and an admissibility condition to identify an appropriate block subdivision of the matrix. A counterexample shows the standard approach to decomposition and admissiblity fails in the case of pure convection. Explicit computation of the singular values of the resulting blocks shows the singular values do not decrease exponentially. An alternative decomposition is based on preferential decomposition in the convection direction in terms of a weighted norm based on the size of diffusion coefficient. This approach is shown to yield a strategy for optimal \(\mathcal H\)-matrix decomposition. Numerical examples are provided that illustrate: (1) the failure of the standard approach in the convection-dominated case; (2) the success of the proposed approach when the mesh is aligned with the convection direction; and, unfortunately, (3) the failure of the proposed approach when the mesh and convection direction are not aligned.