🔢 mathematics

A $C^0$ weak Galerkin method with preconditioning for constrained optimal control problems with general tracking

This paper proposes a $C^0$ weak Galerkin method combined with an additive Schwarz preconditioner to efficiently and robustly solve constrained optimal control problems governed by partial differential equations with general tracking cost functionals and pointwise state constraints.

Original authors: SeongHee Jeong, Seulip Lee, Kening Wang

Published 2026-04-02

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: SeongHee Jeong, Seulip Lee, Kening Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the director of a massive, complex city. Your goal is to control the flow of traffic (the control) to ensure that the city's temperature or pollution levels (the state) match a perfect, ideal blueprint you have in mind. However, you have two major headaches:

The "General Tracking" Problem: You don't just want the whole city to be perfect on average. You have specific, critical spots—like a hospital, a school, or a park—where the conditions must be exact. If the temperature is off by a degree at the hospital, it's a disaster.
The "State Constraints" Problem: You can't just let the city do whatever it wants. There are hard limits. For example, the water level in the reservoir cannot go below a certain line (drought) or above a certain line (flood).

Mathematically, solving this is like trying to balance a wobbly, four-legged table on a trampoline while someone is pushing it from different angles. The equations are incredibly stiff, the numbers get huge and messy (ill-conditioned), and the solution is often "rough" or jagged because of those hard limits.

This paper introduces a new, smarter way to solve these problems using two main tools: a new way to build the math model and a new way to speed up the computer.

1. The New Building Blocks: The "C0 Weak Galerkin" Method

Traditionally, mathematicians build models for these problems using a method called C0 Interior Penalty (C0-IP).

The Old Way (C0-IP): Imagine you are building a wall out of bricks. To make sure the wall is strong and smooth, you have to use a special, heavy-duty mortar (a "penalty parameter") to glue the bricks together. You have to guess exactly how much mortar to use. Too little, and the wall falls; too much, and the wall cracks. Also, you have to check every single edge where two bricks touch to make sure they align perfectly. This is slow and requires a lot of fiddling.
The New Way (C0-WG): The authors propose a new method called C0 Weak Galerkin.
- The Analogy: Instead of gluing bricks with heavy mortar, imagine you are building a structure using magnetic blocks. The blocks naturally snap together in the right way without needing a specific amount of glue.
- How it works: This method uses standard, simple blocks (quadratic Lagrange elements) but calculates the "smoothness" of the wall in a clever, indirect way (using "weak" derivatives).
- The Benefit: It is parameter-free. You don't have to guess how much "mortar" to use. It automatically figures out the right balance. It's also faster to assemble because you don't have to check every single edge connection manually; the math handles it locally, block by block.

2. The Speed Booster: The "Additive Schwarz Preconditioner"

Even with the new building blocks, solving the equations is like trying to push a giant, rusty boulder. The computer gets stuck because the numbers are so huge and tangled that it takes forever to find the answer. This is called an ill-conditioned system.

The Analogy: Imagine you are trying to move a giant boulder. If you push it alone, it won't budge.
The Solution: The authors introduce an Additive Schwarz Preconditioner.
- How it works: Instead of one person pushing the whole boulder, you cut the boulder into smaller, overlapping pieces. You assign a team to push each piece simultaneously. Then, you combine their efforts to move the whole thing.
- The Result: This breaks the giant, messy problem into smaller, manageable chunks that computers can solve very quickly. It turns a task that might take hours into one that takes minutes.

Why Does This Matter?

In the real world, this means we can simulate complex scenarios much faster and more accurately.

For Engineers: You can design better heat shields for rockets or optimize wind farms without waiting days for a computer to finish the calculation.
For City Planners: You can model traffic or pollution control with specific, strict rules (like "no pollution at the school") and get a reliable answer quickly.

Summary in a Nutshell

The paper says: "We found a way to build our math models using magnetic blocks instead of heavy glue (making them easier and faster to build), and we found a way to split the heavy lifting among a team (making the computer solve the problem much faster). This allows us to solve difficult control problems with strict rules and specific targets that were previously too hard or too slow to handle."

It's a win-win: Simpler construction + Faster solving = Better solutions for complex real-world problems.

1. Problem Statement

The paper addresses Optimal Control Problems (OCPs) governed by elliptic partial differential equations (PDEs). Specifically, it focuses on problems characterized by:

General Tracking Cost Functionals: The objective is to minimize the difference between the state $y$ and a desired state $y_d$ , where $y_d$ can be defined over points, curves, or subregions (not just the whole domain).
Pointwise State Constraints: The state variable $y(x)$ must satisfy lower and upper bounds ( $\psi_-(x) \leq y(x) \leq \psi_+(x)$ ) almost everywhere in the domain.
Mathematical Formulation: By eliminating the control variable via the PDE constraint, the problem is reduced to a fourth-order variational inequality involving the state variable.
Challenges:
- Reduced Regularity: The presence of general tracking terms (measures) and pointwise constraints degrades the regularity of the optimal solution and Lagrange multipliers, often resulting in measure-valued multipliers.
- Numerical Complexity: Solving fourth-order problems typically requires $C^1$ -continuous elements, which are complex to implement. Existing $C^0$ -Interior Penalty (C0-IP) methods require careful tuning of penalty parameters and involve trace terms that complicate element-wise assembly.
- Ill-Conditioning: The resulting linear systems are severely ill-conditioned, especially when general tracking terms are involved, hindering the performance of iterative solvers.

2. Methodology

A. The C0 Weak Galerkin (C0-WG) Method

The authors propose a C0 Weak Galerkin method using globally continuous quadratic Lagrange elements ( $P_2$ ).

Weak Laplacian: Instead of using strong derivatives, the method employs a discrete weak Laplacian ( $\Delta_w$ $Δ_{w}$ ). For a function $v_h$ $v_{h}$ , the weak Laplacian on an element $T$ $T$ is computed as a piecewise constant ( $P_0$ $P_{0}$ ) function.
- It is derived explicitly using function values at the barycenters of shared edges/faces and basic geometric data (edge lengths, element areas).
- Formula: $\Delta_w v_h|_T = \frac{1}{|T|} \sum_{e \subset \partial T} \frac{|e|}{2} (\nabla v_h^+ + \nabla v_h^-) \cdot n_T$ .
Bilinear Form: The discrete bilinear form $a_w(\cdot, \cdot)$ $a_{w} (\cdot, \cdot)$ consists of:
1. The $L^2$ inner product of the weak Laplacians: $(\Delta_w u, \Delta_w v)_{\mathcal{T}_h}$ .
2. A parameter-free stabilization term: $\frac{1}{4} \langle h_e^{-1} [\nabla u] \cdot n_e, [\nabla v] \cdot n_e \rangle_{\mathcal{E}_h^o}$ .
Advantages over C0-IP:
- No Penalty Parameter: Unlike the C0-IP method, which requires a large, tuned penalty parameter $\rho$ to ensure stability, the C0-WG method is parameter-free.
- Simplified Assembly: The formulation eliminates the complex trace terms (involving averages of gradients) found in C0-IP, allowing for efficient element-wise assembly of the stiffness matrix.

B. Additive Schwarz Preconditioner (ASP)

To address the ill-conditioning of the linear systems arising from the fourth-order formulation:

Approach: A one-level additive Schwarz preconditioner is developed.
Mechanism: The global domain is partitioned into overlapping subdomains. The global problem is decomposed into local subdomain problems solved independently, and their solutions are combined additively.
Theoretical Bound: The condition number $\kappa(B_h A_h)$ of the preconditioned system is proven to be bounded by $O(\delta^{-4})$ , where $\delta$ is the overlap size. With shape-regular subdomains, this can be improved to $O(H^{-1}\delta^{-3})$ .

3. Key Contributions

Novel Discretization (C0-WG):
- Development of a C0-WG method specifically for constrained OCPs with general tracking.
- Utilization of globally continuous quadratic elements to simplify the weak Laplacian computation.
- Elimination of penalty parameters and trace terms, leading to a parameter-free, robust, and computationally efficient implementation.
- Rigorous error analysis establishing well-posedness and convergence rates.
Specialized Preconditioning:
- Design of an Additive Schwarz preconditioner tailored to the C0-WG formulation.
- Demonstration that the preconditioner significantly reduces the condition number, enabling scalable iterative solvers for the ill-conditioned systems typical of fourth-order OCPs.
Rigorous Convergence Analysis:
- Proof of optimal convergence orders in mesh-dependent norms.
- Establishment of norm equivalence between the C0-WG norm and the standard broken Sobolev norms used in C0-IP analysis.

4. Numerical Results

The authors conducted numerical experiments in 2D settings for both biharmonic problems and constrained OCPs.

Biharmonic Problem:
- The C0-WG method achieved first-order convergence (and observed superconvergence) in the mesh-dependent norm, consistent with theoretical predictions and comparable to C0-IP.
Constrained OCP with General Tracking:
- Accuracy: The C0-WG method consistently produced more accurate approximations than the C0-IP method across $L^\infty$ , $L^2$ , and $H^1$ norms.
- Convergence Rates: C0-WG exhibited higher convergence orders in $L^\infty$ , $L^2$ , and $H^1$ norms compared to C0-IP.
Preconditioning Performance:
- Condition Numbers: Without preconditioning, the condition number grew as $O(h^{-4})$ .
- With ASP: The condition number growth was significantly mitigated (reduced by more than two orders of magnitude), scaling roughly as $O(h^{-3})$ or better depending on the overlap $\delta$ .
- Overlap Effect: Increasing the overlap size ( $\delta = 2h$ vs. $\delta = h$ ) further reduced the condition number, confirming the effectiveness of the domain decomposition strategy.

5. Significance

This paper presents a significant advancement in the numerical solution of high-order constrained optimal control problems:

Efficiency: By removing the need for penalty parameter tuning and simplifying matrix assembly, the C0-WG method reduces implementation complexity and computational overhead.
Robustness: The method handles the reduced regularity of solutions (caused by pointwise constraints and general tracking) effectively without sacrificing stability.
Scalability: The integration of the Additive Schwarz preconditioner solves the critical bottleneck of ill-conditioning, making the method suitable for large-scale simulations and parallel computing environments.
Framework: The work establishes a robust computational framework that can be extended to more complex scenarios, including time-dependent and nonlinear OCPs.