Abstract:Aasen's algorithm factorizes a symmetric indefinite matrix $A$ as $A = P^TLTL^TP$, where $P$ is a permutation matrix, $L$ is unit lower triangular with its first column being the first column of the identity matrix, and $T$ is tridiagonal. In this note, we provide a growth factor upper bound for Aasen's algorithm which is much smaller than that given by Higham. We also show that the upper bound we have given is not tight when the matrix dimension is greater than or equal to $6$.
What problem does this paper attempt to address?
This paper attempts to solve the problem of the upper bound of the growth factor in the Aasen algorithm when decomposing symmetric indefinite matrices. Specifically, the Aasen algorithm decomposes a symmetric indefinite matrix \(A\) as follows:
\[ PAP^T = LTL^T \]
where:
- \(P\) is a permutation matrix,
- \(L\) is a unit lower triangular matrix, and its first column is the first column of the identity matrix,
- \(T\) is a symmetric tridiagonal matrix.
### Main Problems
1. **Estimation of the Upper Bound of the Growth Factor**:
- Higham previously gave the upper bound of the growth factor of the Aasen algorithm as \(4n - 2\), but whether this upper bound can be reached for \(n\geq4\) is an open problem.
- Cheng showed by constructing examples that when \(n = 3\), the growth factor can reach 4, and obtained the growth factors of 7.99 and 14.61 when \(n = 4\) and \(n = 5\) respectively by the direct search method.
2. **Improved Upper Bound of the Growth Factor**:
- Through detailed mathematical derivations, the author of this paper provides a smaller upper bound of the growth factor, \(2n - 1\), and proves that this upper bound is not tight (that is, the actual growth factor will be less than this upper bound) when the matrix dimension \(n\geq6\).
### Key Conclusions
- **New Upper Bound of the Growth Factor**: The main contribution of the paper is to give the upper bound of the growth factor of the Aasen algorithm as \(2n - 1\), which is smaller than \(4n - 2\) given by Higham.
- **Non - tightness of the Upper Bound**: It is further proved that when the matrix dimension \(n\geq6\), this upper bound \(2n - 1\) is not tight, that is, the actual growth factor will be less than \(2n - 1\).
- **Verification Examples**: Some specific matrix examples (such as matrices with \(n = 4, 5, 6\)) are also constructed in the paper to verify the correctness of the theoretical analysis.
### Formula Summary
- **Aasen Algorithm Decomposition Formula**:
\[ PAP^T = LTL^T \]
- **Upper Bound Formula of the Growth Factor**:
\[ \text{Upper bound of growth factor}=2n - 1 \]
These results are helpful for better understanding the numerical stability of the Aasen algorithm when processing symmetric indefinite matrices and provide more accurate theoretical support for practical applications.