Structured Sketching for Linear Systems

Johannes J Brust,Michael A Saunders
2024-07-01
Abstract:For linear systems $Ax=b$ we develop iterative algorithms based on a sketch-and-project approach. By using judicious choices for the sketch, such as the history of residuals, we develop weighting strategies that enable short recursive formulas. The proposed algorithms have a low memory footprint and iteration complexity compared to regular sketch-and-project methods. In a set of numerical experiments the new methods compare well to GMRES, SYMMLQ and state-of-the-art randomized solvers.
Numerical Analysis,Mathematical Software,Computation
What problem does this paper attempt to address?
This paper attempts to solve several key problems in linear system solving, especially for large - scale sparse matrices. Specifically, the paper focuses on how to improve the iterative solving efficiency of linear systems through an effective "Sketch - and - Project" method. The following are the main problems that the paper attempts to solve: 1. **Improve solving efficiency**: - The paper proposes an iterative algorithm based on "Sketch - and - Project". By selecting an appropriate sketch matrix (such as the history of residuals), a weighting strategy is developed, making the algorithm have low memory usage and iterative complexity. - Compared with the traditional "Sketch - and - Project" method, the new method has significant improvements in both memory usage and the number of iterations. 2. **Handle different types of problems**: - The paper not only considers the case of symmetric positive definite matrices (Symmetric Positive Definite, SPD), but also discusses the cases of symmetric indefinite matrices and general square matrices. - For symmetric indefinite matrices, a new method is proposed, whose performance is better than the existing random or deterministic methods. - For general square matrices, a nested algorithm is developed, which can show good performance in numerical experiments. 3. **Optimize convergence speed**: - The paper analyzes the influence of different weight matrices \( B \) on the convergence speed of the algorithm, and proposes several effective weight selection strategies. - For example, when \( A \) is a symmetric matrix, \( B = A \) or \( B = A^{-1} \) can be selected, and these selections can significantly improve the convergence speed of the algorithm. 4. **Finite termination property**: - The paper proves that under specific conditions, the proposed algorithm can converge to the solution within a finite number of steps, even when using a random sketch. - This property is very important for efficient solving in practical applications because it ensures that the algorithm can find the solution within a finite time. 5. **Numerical experiment verification**: - The paper verifies the effectiveness of the new method through a series of numerical experiments, especially its performance in handling large - scale sparse matrices. - The experimental results show that the new method is superior to existing methods such as GMRES and SYMMLQ in terms of solving time. In conclusion, by introducing new "Sketch - and - Project" methods and weighting strategies, this paper aims to improve the efficiency and robustness of linear system solving, especially when dealing with large - scale sparse matrices.