MiS Preprint Repository

The discretization of integral operators corresponding to non-local kernel functions typically gives rise to densely populated matrices. In order to be able to treat these matrices in an efficient manner, they have to be compressed, e.g., by panel clustering algorithms, multipole expansions or wavelet techniques.

By choosing the correct panel clustering approach, the resulting approximation of the matrix can be written in the form of a so-called ${\mathcal{H}}^2$-matrix. The ${\mathcal{H}}^2$-matrix representation can be computed for fairly general kernel functions by a black box algorithm that requires only pointwise evaluations of the kernel function.

Although this technique leads to good results, the expansion system tends to contain a certain level of redundancy that leads to an unnecessarily high complexity for the memory requirements and the matrix-vector multiplication. We present two variants of the original method that can compress the matrix even further. Both methods work on the fly, i.e., it is not necessary to keep the original ${\mathcal{H}}^2$-matrix in memory, and both methods perform an algebraic compression, so that the black box character of the algorithm is preserved.