Abstract
This paper investigates the application of Balancing Domain Decomposition by Constraints (BDDC) methods to the saddle point problems arising from Hybridizable Discontinuous Galerkin (HDG) discretizations of the incompressible Stokes equations. The authors provide a rigorous analysis of the condition number for the preconditioned system, accounting for various local stabilization parameters common in HDG methods. Theoretical estimates and numerical experiments demonstrate that the algorithm is highly scalable, with a condition number that grows only polylogarithmically with the local problem size.
Introduction
Simulating fluid flow via the Stokes equations is computationally demanding because it results in a "saddle point" problem—a mathematical system that is indefinite and difficult to solve efficiently. To enable large-scale simulations, researchers use Domain Decomposition, which breaks the global problem into smaller subdomains that can be processed in parallel.
A particularly effective approach is the BDDC algorithm. This paper focuses on pairing BDDC with HDG discretization, a method favored for its stability and reduced global coupling. A key challenge addressed here is the impact of the "stabilization parameter" ($\tau_\kappa$) used in HDG, which can significantly alter the numerical properties of the resulting operators.
Methodology
The authors' approach involves several critical numerical strategies:
- System Reduction: The global saddle point problem is reduced to a smaller interface problem involving only the numerical traces of velocity and mean pressure.
- The Benign Subspace: Because the Stokes operator is indefinite, standard Conjugate Gradient (CG) methods are not directly applicable. The authors use specific constraints (no-net-flux) to force the CG iterates into a "benign subspace" where the operator behaves as if it were positive definite, ensuring convergence.
- Coarse Problem Constraints: The BDDC preconditioner utilizes a coarse problem built from edge and face average constraints across subdomain interfaces. These constraints facilitate the rapid global transfer of information across the domain.
Key Insights and Theoretical Bounds
A major contribution of this work is the observation of stabilization invariance. The authors found that while different stabilization parameters change the resulting matrices, they do not change the underlying divergence-free space of the Stokes equation. This allowed them to relate the complex Stokes problem to simpler second-order elliptic problems, leveraging existing mathematical tools to derive condition number estimates.
The authors proved (and numerically verified) that the condition number ($\kappa$) follows the bound: $$\kappa(M^{-1}\hat{S}) \leq C \left(1 + \log\left(\frac{H}{h}\right)\right)^2$$ where $H$ is the subdomain size and $h$ is the mesh size. This "log-squared" growth is the hallmark of an optimal, scalable domain decomposition algorithm.
Numerical Results
The theory was tested using a standard benchmark on a unit square domain with known exact solutions. The results confirmed several key properties:
- Scalability: When the number of subdomains was increased (from $4 \times 4$ up to $32 \times 32$) while keeping the local problem size constant, the number of iterations required for convergence remained nearly constant.
- Parameter Robustness: The algorithm performed consistently well across three popular choices for the stabilization parameter ($\tau_\kappa = h_\kappa, 1, \text{ and } 1/h_\kappa$).
- Polynomial Orders: Tests across different finite element approximation orders ($k=0, 1, 2$) supported the theoretical findings, though the authors noted that the dependence on $k$ warrants further formal analysis.
Limitations and Future Work
While the BDDC algorithm is highly effective, the authors identify two primary areas for future research:
- Inexact Solvers: For extreme-scale parallel computing, the global coarse problem can become a bottleneck. Implementing "inexact solvers" for this step is a necessary progression.
- Complex Geometries: The current analysis assumes quasi-uniform meshes and polygonal domains. Extending these proofs to the highly stretched meshes and curved geometries found in industrial aerospace or automotive design remains an open challenge.
Summary updated