A New Matrix Truncation Method for Improving Approximate Factorisation Preconditioners

Andreas A. Bock,Martin S. Andersen

2024-10-09

Abstract:In this experimental work, we present a general framework based on the Bregman log determinant divergence for preconditioning Hermitian positive definite linear systems. We explore this divergence as a measure of discrepancy between a preconditioner and a matrix. Given an approximate factorisation of a given matrix, the proposed framework informs the construction of a low-rank approximation of the typically indefinite factorisation error. The resulting preconditioner is therefore a sum of a Hermitian positive definite matrix given by an approximate factorisation plus a low-rank matrix. Notably, the low-rank term is not generally obtained as a truncated singular value decomposition (TSVD). This framework leads to a new truncation where principal directions are not based on the magnitude of the singular values, and we prove that such truncations are minimisers of the aforementioned divergence. We present several numerical examples showing that the proposed preconditioner can reduce the number of PCG iterations compared to a preconditioner constructed using a TSVD for the same rank. We also propose a heuristic to approximate the proposed preconditioner in the case where exact truncations cannot be computed explicitly (e.g. in a large-scale setting) and demonstrate its effectiveness over TSVD-based approaches.

Numerical Analysis

What problem does this paper attempt to address?

### Problems the paper attempts to solve

The paper aims to propose a new matrix truncation method to improve the approximate factorization preconditioning technique, so as to solve Hermitian positive - definite linear systems more effectively. Specifically, the paper constructs a general framework based on the Bregman log - determinant divergence to measure the difference between the preconditioning matrix and the target matrix. Given an approximate factorization, this framework can construct a low - rank approximation to approximate the factorization error which is usually indefinite. The final preconditioning matrix consists of an approximately factorized Hermitian positive - definite matrix and a low - rank matrix.

### Main contributions

1. **New truncation method**: The paper proposes a new truncation method, called Bregman log - determinant (BLD) truncation. Different from the traditional truncated singular value decomposition (TSVD), BLD truncation does not depend on the magnitude of singular values, but minimizes the Bregman log - determinant divergence.
2. **Theoretical analysis**: The paper proves that BLD truncation is the solution that minimizes the Bregman log - determinant divergence and provides detailed mathematical derivations.
3. **Numerical experiments**: The paper shows the effectiveness of the proposed preconditioning technique in reducing the number of PCG iterations through multiple numerical experiments, especially in large - scale problems.
4. **Approximation method**: For cases where the exact truncation cannot be calculated explicitly, the paper proposes a heuristic method to approximate BLD truncation and verifies its effectiveness in practical applications.

### Key concepts

- **Bregman log - determinant divergence**: Defined as \( D_{\text{LD}}(X, Y)=\text{trace}(XY^{- 1})-\log\det(XY^{-1})-n \), which is used to measure the difference between two matrices.
- **Low - rank approximation**: Construct a low - rank matrix by selecting specific eigenvalues and eigenvectors to approximate the original matrix.
- **Preconditioning technique**: By constructing a preconditioning matrix \( P \), the original linear system \( Sx = b \) is transformed into \( Q^{-1}SQ^{-*}z = Q^{-1}b \) to accelerate the iterative solution process.

### Application background

This research has broad application prospects in scientific computing and engineering fields, especially in scenarios where efficient solution of large - scale linear systems is required. For example, in finite element analysis, image processing, data assimilation and other fields, preconditioning techniques can significantly improve computational efficiency and accuracy.

### Conclusion

The paper proposes a new truncation method based on the Bregman log - determinant divergence to improve the approximate factorization preconditioning technique. Through theoretical analysis and numerical experiments, the effectiveness and superiority of this method in reducing the number of PCG iterations are proved. This provides a new tool and idea for solving large - scale linear systems.