Abstract
This paper introduces a discretization-accurate stopping criterion of symmetric iterative methods for solving systems of algebraic equations resulting from the finite element approximation. The stopping criterion consists of the evaluations of the discretization and the algebraic error estimators, that are based on the respective duality error estimator and the difference of two consecutive iterates. Iterations are terminated when the algebraic estimator is of the same magnitude as the discretization estimator. Numerical results for multigrid $V(1,1)$-cycle and symmetric Gauss-Seidel iterative methods are presented for the linear finite element approximation to the Poisson equations. A large reduction in computational cost is observed compared to the standard residual-based stopping criterion.
Introduction
In Finite Element Analysis (FEA), solving the resulting linear system $Au=f$ is often the most computationally expensive step. Iterative solvers like Multigrid or Gauss-Seidel are standard tools for navigating these massive systems. However, a common pitfall in engineering practice is "over-solving"—continuing iterations until the algebraic residual reaches an arbitrarily small tolerance (e.g., $10^{-8}$) even when the underlying mesh discretization error is orders of magnitude larger (e.g., $10^{-2}$).
This paper addresses the inefficiency of traditional stopping criteria by proposing a method to stop iterations precisely when the algebraic error no longer significantly impacts the total error. By accurately estimating both discretization and algebraic errors on the fly, the authors demonstrate that computational costs can be slashed—sometimes by over 90%—without sacrificing the quality of the final physical solution.
Methodology
The core of the proposed method lies in decomposing the total error into two distinct components: Discretization Error (the gap between the exact physical solution and the exact solution of the linear system) and Algebraic Error (the gap between the current iteration and the exact solution of the linear system).
1. Discretization Error Estimator ($\eta_d$)
The paper utilizes an "equilibrated flux" approach. This involves locally post-processing the flux on vertex patches (small groups of elements). Unlike standard a posteriori estimators that assume the linear system is solved perfectly, the authors modify these local problems to account for the remaining algebraic residual. This ensures the estimator remains reliable even during the early stages of the iterative process.
2. Algebraic Error Estimator ($\eta_a$)
The algebraic estimator is derived from the difference between two consecutive iterates ($u^{(k+1)} - u^{(k)}$). To ensure this serves as a true upper bound, it is scaled by a factor involving the spectral radius (the rate of convergence) of the solver.
3. Adaptive Stopping Criterion
A key innovation is the on-the-fly approximation of the spectral radius using the ratio of consecutive residuals. The authors also introduce a conservative correction factor ($e^{1/k}$) to maintain the reliability of the upper bound before the solver reaches its steady-state convergence rate. The solver is terminated when: $$\eta_a \leq \gamma \eta_d$$ where $\gamma$ is a user-defined safety parameter, typically set near 1.
Experimental Evaluation
The method was evaluated using two primary 2D benchmark problems:
- The Poisson Equation: A standard test case on a uniform triangular mesh with a smooth solution.
- The Kellogg Intersecting Interface Problem: A challenging benchmark involving a "checkerboard" of varying diffusion coefficients. This tests the method's robustness against jumping coefficients and non-uniform, graded meshes.
Key Results
The results demonstrate massive efficiency gains compared to standard residual-based stopping:
- Multigrid Solver (Poisson): The new criterion stopped the solver after just 2 iterations (Total Error: 0.0821). In contrast, the standard method required 15 iterations to reach a Total Error of 0.0741. The extra 13 iterations provided less than a 10% improvement in the final result, representing significant wasted effort.
- Symmetric Gauss-Seidel (Poisson): The gains were even more pronounced for this slower solver. The proposed method stopped at 31 iterations, while the conventional method continued for 289 iterations—nearly 10 times the work for no meaningful gain in physical accuracy.
- Robustness: In the Kellogg problem, the estimators remained accurate despite complex interfaces, proving the method's applicability to non-trivial engineering problems.
Discussion and Limitations
The primary insight of this work is that over-solving is a silent killer of performance in high-performance computing. By using "constant-free" bounds that do not depend on mesh-specific hidden constants, the authors provide a robust tool for practical use.
However, some limitations remain:
- Symmetry: The current theoretical framework is developed for symmetric iterative methods (e.g., Multigrid V-cycles, Conjugate Gradient). Non-symmetric methods like GMRES would require further extension.
- Data Complexity: The analysis currently assumes piecewise constant source terms and diffusion coefficients to avoid data oscillation terms.
- Linearity: The focus is on linear diffusion. Extending this logic to non-linear solvers (e.g., Newton's method) could potentially offer even greater computational savings in fields like fluid dynamics.
Conclusion
This discretization-accurate stopping criterion offers a mathematically rigorous way to optimize iterative solvers. By aligning the solver's precision with the mesh's inherent accuracy, researchers can achieve significant speedups without compromising the integrity of their numerical simulations.
Summary updated