A hierarchical random compression method for kernel matrices
This work addresses the computational bottleneck of kernel matrix operations for large-scale problems in scientific computing and machine learning, offering a faster alternative to existing methods.
The paper proposes a hierarchical random compression method (HRCM) for kernel matrices that achieves O(N log N) complexity for fast kernel summations, with numerical results showing a cross-over matrix size in the order of thousands for 3-4 digits relative accuracy compared to direct O(N^2) summations.
In this paper, we propose a hierarchical random compression method (HRCM) for kernel matrices in fast kernel summations. The HRCM combines the hierarchical framework of the H-matrix and a randomized sampling technique of the column and row spaces for far-field interaction kernel matrices. We show that a uniform column/row sampling (with a given sample size) of a far-field kernel matrix, with- out the need and associated cost to pre-compute a costly sampling distribution, will give a low-rank compression of such low-rank matrices, independent of the matrix sizes and only dependent on the separation of the source and target locations. This far-field random compression technique is then implemented at each level of the hierarchical decomposition for general kernel matrices, resulting in an O(N logN) random compression method. Error and complexity analysis for the HRCM are included. Numerical results for electrostatic and Helmholtz wave kernels have vali- dated the efficiency and accuracy of the proposed method with a cross-over matrix size, in comparison of direct O(N^2) summations, in the order of thousands for a 3-4 digits relative accuracy.