Abstract:In this work, we focus on CholeskyQR-type algorithms for sparse matrices. We introduce a new model for sparse matrices and categorize them into two types: $T_{1}$ matrices and $T_{2}$ matrices, based on the presence of dense columns. We provide alternative choices for the shifted parameter $s$ in Shifted CholeskyQR3 \cite{Shifted}, specifically tailored for sparse matrices, and conduct a rounding error analysis for the algorithm using this $s$. The alternative $s$ can enhance the applicability of Shifted CholeskyQR3 for $T_{1}$ matrices under appropriate element-norm conditions (ENCs). Our analytical approach utilizes the properties of the $g$-norm of the matrix, as outlined in \cite{New}. Numerical experiments demonstrate that our alternative $s$ significantly improves the applicability of Shifted CholeskyQR3 for $T_{1}$ matrices compared to the $s$ from \cite{New} under suitable ENCs. Shifted CholeskyQR3 with our alternative $s$ is also applicable to $T_{2}$ matrices, which are more ill-conditioned than the dense cases. Furthermore, Shifted CholeskyQR3 with our alternative $s$ shows good efficiency in these sparse cases.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is how to improve the CholeskyQR - type algorithm for sparse matrices, especially the Shifted CholeskyQR3 (SCholeskyQR3) algorithm. Specifically, the author focuses on the following points:
1. **Classifying sparse matrices**: Sparse matrices are divided into two types (T1 matrices and T2 matrices) according to whether there are dense columns.
2. **Selecting shift parameters**: Provide alternative shift parameters \( s \) applicable to sparse matrices for the SCholeskyQR3 algorithm, and ensure that these parameters can enhance the applicability of the algorithm under appropriate element - norm conditions (ENCs).
3. **Numerical stability analysis**: Conduct a rounding - error analysis of the algorithm by introducing a new g - norm to ensure numerical stability, especially when dealing with ill - conditioned matrices.
### Main contributions
1. **New model and classification**:
- Propose a new model of sparse matrices based on column sparsity, and define T1 matrices and T2 matrices.
- For T1 matrices, propose a new shift parameter \( s \), whose value depends on the element with the largest absolute value in the matrix and its structure.
2. **Theoretical results**:
- Prove that when the sparse matrix is a T1 matrix, the new shift parameter \( s \) can prevent numerical collapse and ensure the numerical stability of the SCholeskyQR3 algorithm.
- Provide a detailed theoretical analysis, including the relationship between κ2(X) and κ2(Q), rounding - error analysis, and the optimal selection of the shift parameter \( s \) under appropriate conditions.
3. **Numerical experiments**:
- Verify the effectiveness and superiority of the improved shift parameter \( s \) in dealing with sparse matrices through numerical experiments, especially its performance in dealing with ill - conditioned matrices is better than existing methods.
### Formula summary
- **Shift parameter \( s \)**:
\[
s = 11 \cdot (mu+(n + 1)u) \cdot (vt_1+nt_2)c^2
\]
where \( c=\max |x_{ij}| \), \( t_1 \) and \( t_2 \) are the maximum numbers of non - zero elements in dense columns and sparse columns respectively, and \( v \) is the number of dense columns.
- **Relationship between κ2(X) and κ2(Q)**:
\[
\kappa_2(Q)\leq 2h\cdot\sqrt{1+\alpha_0 (\kappa_2(X))^2}
\]
where \( \alpha_0=\frac{s}{\|X\|^2_2} \), \( h = \sqrt{2.3+0.37r + 0.015r^2} \), \( r=\frac{n\sqrt{n}}{m\sqrt{v}} \).
- **Rounding - error analysis**:
\[
\|Q^\top_2 Q_2 - I\|_F\leq 6(mnu + n(n + 1)u)
\]
\[
\|Q_2 R_4 - X\|_F\leq (2.79+3.97l)hn^2 u \|X\|_2
\]
Through these improvements, the paper not only improves the stability and efficiency of the SCholeskyQR3 algorithm in dealing with sparse matrices, but also provides a new theoretical basis and method for follow - up research.