Multilevel Evaluation of Multidimensional Integral Transforms with Asymptotically Smooth Kernels

TL;DR

This paper extends a multilevel, adaptive grid algorithm for fast evaluation of multidimensional integral transforms with asymptotically smooth kernels, demonstrating improved efficiency over previous methods.

Contribution

It develops and analyzes a multidimensional extension of a multilevel integral transform algorithm based on asymptotic smoothness, including implementation details and efficiency analysis.

Findings

Multidimensional extension is feasible despite nonlocal corrections.

The algorithm achieves optimal work estimates for multidimensional transforms.

Numerical results show improved efficiency over previous algorithms.

Abstract

In many practical applications of numerical methods a substantial increase in efficiency can be obtained by using local grid refinement, since the solution is generally smooth in large parts of the domain and large gradients occur only locally. Fast evaluation of integral transforms on such an adaptive grid requires an algorithm that relies on the smoothness of the continuum kernel only, independent of its discrete form. A multilevel algorithm with this property was presented in [A. Brandt and C.H. Venner, SIAM J. Sci. Stat. Comput. 19 (1998) pp.468-492] [Bra1998]. Ref. [Bra1998] shows that already on a uniform grid the new algorithm is more efficient than earlier fast evaluation algorithms, and elaborates the application to one-dimensional transforms. The present work analyses the extension and implementation of the algorithm for multidimensional transforms. The analysis conveys that the multidimensional extension is nontrivial, on account of the occurence of nonlocal corrections. However, by virtue of the asymptotic smoothness properties of the continuum kernel, these corrections can again be evaluated fast. By recursion, it is then possible to obtain the optimal work estimates indicated in [Bra1998]. Currently, only uniform grids are considered. Detailed numerical results will be presented for a two dimensional model problem. The results demonstrate that with the new algorithm the evaluation of multidimensional transforms is also more efficient than with previous algorithms.