English

A deterministic Kaczmarz algorithm for solving linear systems

Numerical Analysis 2023-05-17 v5 Numerical Analysis

Abstract

We propose a new deterministic Kaczmarz algorithm for solving consistent linear systems Ax=bA\mathbf{x}=\mathbf{b}. Basically, the algorithm replaces orthogonal projections with reflections in the original scheme of Stefan Kaczmarz. Building on this, we give a geometric description of solutions of linear systems. Suppose AA is m×nm\times n, we show that the algorithm generates a series of points distributed with patterns on an (n1)(n-1)-sphere centered on a solution. These points lie evenly on 2m2m lower-dimensional spheres {§k0,§k1}k=1m\{\S_{k0},\S_{k1}\}_{k=1}^m, with the property that for any kk, the midpoint of the centers of §k0,§k1\S_{k0},\S_{k1} is exactly a solution of Ax=bA\mathbf{x}=\mathbf{b}. With this discovery, we prove that taking the average of O(η(A)log(1/ε))O(\eta(A)\log(1/\varepsilon)) points on any §k0§k1\S_{k0}\cup\S_{k1} effectively approximates a solution up to relative error ε\varepsilon, where η(A)\eta(A) characterizes the eigengap of the orthogonal matrix produced by the product of mm reflections generated by the rows of AA. We also analyze the connection between η(A)\eta(A) and κ(A)\kappa(A), the condition number of AA. In the worst case η(A)=O(κ2(A)logm)\eta(A)=O(\kappa^2(A)\log m), while for random matrices η(A)=O(κ(A))\eta(A)=O(\kappa(A)) on average. Finally, we prove that the algorithm indeed solves the linear system ATW1Ax=ATW1bA^T W^{-1}A \mathbf{x} = A^T W^{-1} \mathbf{b}, where WW is the lower-triangular matrix such that W+WT=2AATW+W^T = 2AA^T. The connection between this linear system and the original one is studied. The numerical tests indicate that this new Kaczmarz algorithm has comparable performance to randomized (block) Kaczmarz algorithms.

Keywords

Cite

@article{arxiv.2105.07736,
  title  = {A deterministic Kaczmarz algorithm for solving linear systems},
  author = {Changpeng Shao},
  journal= {arXiv preprint arXiv:2105.07736},
  year   = {2023}
}

Comments

30 pages, more numerical experiments are added, some typos are fixed