Block-adaptive cross approximation of discrete integral operators (Q2246155)

The authors discuss and analyse the hierarchical matrix approximation in a block-adaptive way. As model problem the Laplace equation in a domain \(\Omega \in \mathrm{R}^d\) or in \(\Omega^c = \mathbb{R}^d \setminus \overline{\Omega}\) with Dirichlet boundary conditions \(u = g\) on \(\partial \Omega\) is considered. The solution of this problem is given by the representation formula \(u(x) = \mathcal{V}\psi(x) - \mathcal{K}g(x)\) with the single layer potential operator \(\mathcal{V}\) and the double layer potential operator \(\mathcal{K}\). The Neumann data \(\psi\) is the solution of the boundary integral equation \(\mathcal{V}\psi = (\frac{1}{2} \mathcal{I} + \mathcal{K}) g\) on \(\partial \Omega\). For finding this solution numerically, a Galerkin discretisation with piecewise constant functions is used which leads to a system of linear equations \(Ax = b\). A short overview on the structure of hierarchical matrices and on the adaptive cross approximation (ACA) method are given. Whereas in the ACA method all blocks of the matrix partition are approximated with the same prescribed accuracy, the authors propose to adapt the accuracy of each block to the error of the solution. This adaptive strategy is based on ideas of residual error estimators. In this way one obtains a sequence \(\{ A_k\}\) of approximations of the matrix \(A\) and a sequence of solutions \(\{x_k\}\) of the systems \(A_kx_k = b\). It is proved that the residuals \(\| b-Ax_k\|_2\) of the constructed sequence of solutions converge to zero. In the numerical examples situations are considered where the solution contains structural differences or parts of the mesh are over-refined. The examples show that in these situations the proposed method is more efficient with respect to the computational time and the required storage than the ACA method.