Existence of \(\mathcal H\)-matrix approximants to the inverse FE-matrix of elliptic operators with \(L^\infty\)-coefficients (Q1402177)

The authors state, ``This article deals with the existence of blockwise low-rank approximants -- so-called \(\mathcal H\)-matrices -- to inverses of finite element (FE) matrices in the case of uniformly elliptic operators with \(L^{\infty}\)-coefficients.'' Suppose \(L\) denotes a uniformly elliptic partial differential operator defined on a domain \(\Omega\), and \(\phi_i\) is a sequence of finite element shape functions. The Green's function \(G(x,y)\) for the Dirichlet problem for \(L\) on \(\Omega\) could be used to construct a matrix \(B\) as the finite element approximation to \(L^{-1}\) via the expression \((B)_{ij}=\int_\Omega \phi_i G(x,y)\phi_j\). Because the matrix \(B\) is full, however, this representation is not practical for solving the finite element equations. Instead, the finite element mass (\(M\)) and stiffness (\(K\)) matrices are constructed and inverted numerically. Suppose \(I\) and \(J\) are index sets chosen so that \(\bigcup_{i\in I} \text{ supp}(\phi_i)\) and \(\bigcup_{j\in J} \text{ supp}(\phi_j)\) are separated by a finite minimum difference. If the coefficients of \(L\) are smooth, a sub-block of \(B\) can be approximated by a sum of low-rank \(\mathcal H\)-matrices by approximating \(G(x,y)\) by a polynomial, whose existence is guaranteed because of smoothness. Diagnonal entries of \(B\) can be generated using \(G(x,y)\) itself. The result is an approximation of \(B\) as a sum of blockwise low-rank matrices. This paper focusses on showing that \(B\) can be approximated as a sum of low-rank \(\mathcal H\)-matrices in the case that the coefficients of \(L\) are not smooth, but merely in \(L^\infty\). This is proved by first proving that for \(x\) and \(y\) chosen from separated subdomains of \(\Omega\), for each \(\epsilon\in(0,1)\), \(G(x,y)\) can be approximated by the separable sum \(G_k(x,y)=\sum_1^k u_i(x)v_i(y)\) with \(k\) growing as a power of the logarithm of \(\epsilon\). In addition, the authors show that the inverse of the finite element mass matrix \(M^{-1}\) can also be approximated as an \(\mathcal H\)-matrix. Further, the inverse of the finite element stiffness matrix \(K^{-1}\) can be approximated by \(M^{-1}BM^{-1}\), with \(B\) as above, so that \(K^{-1}\) is also an \(\mathcal H\)-matrix. Finally, the authors present two numerical examples to show that their results can be practically implemented. One example has large coefficient jumps near boundaries of the unit square and the other has coefficients with large numbers of jumps in the interior of the square.