Abstract:The generalized Golub-Kahan bidiagonalization has been used to solve saddle-point systems where the leading block is symmetric and positive definite. We extend this iterative method for the case where the symmetry condition no longer holds. We do so by relying on the known connection the algorithm has with the Conjugate Gradient method and following the line of reasoning that adapts the latter into the Full Orthogonalization Method. We propose appropriate stopping criteria based on the residual and an estimate of the energy norm for the error associated with the primal variable. Numerical comparison with GMRES highlights the advantages of our proposed strategy regarding its low memory requirements and the associated implications.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the solution of nonsymmetric saddle point systems. Specifically, the traditional generalized Golub - Kahan bidiagonalization method has been used to solve saddle point systems where the main block is symmetric and positive definite. However, when the symmetry condition no longer holds, that is, when the main block is nonsymmetric, how to effectively solve such systems is an unsolved problem.
To solve this problem, the authors propose a new iterative method, which extends the generalized Golub - Kahan bidiagonalization algorithm to adapt to nonsymmetric saddle point systems. They draw on the connection between the Conjugate Gradient method and the Full Orthogonalization Method (FOM), propose appropriate stopping criteria, and evaluate convergence based on residual and energy - norm error estimates.
### Main contributions
1. **Extended algorithm**: Extend the generalized Golub - Kahan bidiagonalization algorithm from symmetric saddle point systems to nonsymmetric saddle point systems.
2. **Theoretical basis**: Establish the equivalence with the application of the Arnoldi process to the Schur complement, thereby deriving an algorithm applicable to nonsymmetric systems.
3. **Stopping criteria**: Propose stopping criteria based on residual and energy - norm error estimates to ensure the termination of iteration with moderate precision.
4. **Numerical comparison**: Through numerical comparison with existing methods such as GMRES, show the advantages of the new method in terms of memory requirements and computational efficiency.
### Summary of mathematical formulas
- Matrix form of the saddle point system:
\[
\begin{bmatrix}
M & A \\
A^T & 0
\end{bmatrix}
\begin{bmatrix}
u \\
p
\end{bmatrix}
=
\begin{bmatrix}
b_1 \\
b_2
\end{bmatrix}
\]
where \(M\in\mathbb{R}^{m\times m}\) is nonsymmetric and positive definite, \(A\in\mathbb{R}^{m\times n}\), \(b_1\in\mathbb{R}^m\), \(b_2\in\mathbb{R}^n\).
- Generalized Golub - Kahan bidiagonalization:
\[
\begin{cases}
AQ = MV B \\
V^T M V = I \\
A^T V = Q B^T \\
Q^T Q = I
\end{cases}
\]
- Residual norm:
\[
\|\ b - A^T u_k\ \|_2=\beta_{k + 1}\chi_k
\]
where \(\chi_1=\frac{\beta_1}{\alpha_1}, \chi_k = -\frac{\beta_k}{\alpha_k}\chi_{k - 1}\).
- Energy - norm error estimate:
\[
\|\ e^{(k)}\ \|_M=\sqrt{x_{n - k}^T L^{-T}_{n - k}x_{n - k}}
\]
Through these improvements, the paper provides a new method for effectively solving nonsymmetric saddle point systems and verifies its superiority through numerical experiments.