An exact solver for simple ${\mathcal H}$-matrix systems

Abstract

Hierarchical matrices (usually abbreviated ${\mathcal H}$-matrices) are frequently used to construct preconditioners for systems of linear equations. Since it is possible to compute approximate inverses or $LU$ factorizations in ${\mathcal H}$-matrix representation using only ${\mathcal O}(n \log^2 n)$ operations, these preconditioners can be very efficient. Here we consider an algorithm that allows us to solve a linear system of equations given in a simple ${\mathcal H}$-matrix format \emph{exactly} using ${\mathcal O}(n \log^2 n)$ operations. The central idea of our approach is to avoid computing the inverse and instead use an efficient representation of the $LU$ factorization based on low-rank updates performed with the well-known Sherman-Morrison-Woodbury equation.