Dirichlet control problems with energy regularization governed by non-coercive elliptic equations

This paper investigates linear-quadratic Dirichlet control problems governed by non-coercive elliptic equations on non-convex polygonal domains using energy regularization, establishing solution regularity in weighted Sobolev spaces and deriving optimal error estimates for finite element discretizations that employ graded meshes and a specialized discrete projection.

The Gist

Here is an explanation of the paper "Dirichlet control problems with energy regularization governed by non-coercive elliptic equations," translated into simple, everyday language with creative analogies.

The Big Picture: Steering a Wobbly Boat in a Rocky Harbor

Imagine you are the captain of a boat (the system) trying to reach a specific destination (the target). However, your boat is in a very tricky harbor with jagged, non-convex rocks (a non-convex polygonal domain). The water currents are unpredictable and sometimes push the boat in ways that make it unstable (a non-coercive equation).

Your goal is to adjust the rudder (the control) at the edge of the harbor to steer the boat as close as possible to your target spot, but you want to do it without jerking the rudder too violently (which costs energy).

This paper is about finding the perfect way to move that rudder, even when the harbor is weirdly shaped and the physics of the water are tricky. The authors prove that they can calculate this perfect steering method on a computer with high precision.

---

The Main Characters and Concepts

1. The "Wobbly" Physics (Non-Coercive Equations)

Usually, in math problems like this, the laws of physics are "nice" and predictable. If you push a little, the boat moves a little. This is called being coercive.

• The Problem: In this paper, the physics are "wobbly." The boat might not move exactly as expected, or the math might suggest there are infinite ways to steer it. It's like trying to balance a pencil on its tip; it's unstable.
• The Solution: The authors use a special mathematical "safety net" called Energy Regularization. Think of this as adding a heavy anchor to the rudder. It doesn't stop the boat from moving, but it prevents the rudder from shaking wildly. It forces the solution to be smooth and stable.

2. The Jagged Harbor (Non-Convex Domains)

Imagine a harbor shaped like an "L" or a star. These shapes have sharp corners.

• The Problem: In sharp corners, the water flow (or the solution to the equation) gets chaotic and "spiky." If you try to map this harbor with a standard grid (like graph paper), you miss the details near the corners, and your steering instructions become inaccurate.
• The Solution: The authors use Graded Meshes. Imagine taking your graph paper and stretching the squares so they become tiny, microscopic squares near the sharp corners and larger squares in the open water. This allows the computer to see the chaos in the corners clearly without needing a million tiny squares everywhere else.

3. The "Ghost" Rudder (The Control)

In this problem, you don't control the boat inside the water; you only control the water at the edge (the boundary).

• The Challenge: You can't just pick a point on the edge and say "move here." You have to define a smooth curve along the entire edge.
• The Innovation: The authors introduce a special Projection. Imagine you have a rough sketch of a curve (your computer's guess). This projection is like a "smoothing filter" that forces your sketch to fit perfectly onto the allowed shapes, but it does it in a way that respects the energy of the curve, not just its shape. This is a new trick they developed because older methods were too "blurry" for these tricky harbors.

---

How They Solved It (The Journey)

Step 1: Proving the Boat Can Be Steered
First, they had to prove that even with the wobbly physics, there is actually one best way to steer the boat. Usually, math proofs rely on the physics being "nice" (coercive). Since this wasn't the case, they had to invent a new proof showing that the "energy cost" of steering is always positive, guaranteeing a unique best solution.

Step 2: The Digital Map (Finite Element Discretization)
They broke the harbor down into small triangles (a mesh) to solve the equations on a computer.

• The Trick: They used the Graded Meshes mentioned earlier. They proved that if you make the triangles small enough near the corners (using a specific grading rule), the computer's answer gets closer to the true answer at the fastest possible speed.

Step 3: The "Double-Check" (Optimality)
They showed that the computer's version of the problem is also "convex" (stable). This means that if the computer finds a solution, it's definitely the best one, not just a local "good enough" one. It's like ensuring that when you find the bottom of a valley, it's the bottom of the entire mountain range, not just a small dip.

Step 4: The Final Result (Error Estimates)
They calculated exactly how much the computer's answer differs from the real answer.

• The Result: They proved that with their special "microscopic squares" near the corners, the error shrinks perfectly as the grid gets finer. It's the "Goldilocks" result: not too slow, not too fast, but exactly the optimal speed.

---

Why This Matters (The "So What?")

Imagine you are designing a bridge, a heat shield for a spaceship, or a medical device. These shapes often have sharp corners, and the materials inside might behave in complex, unstable ways.

• Before this paper: Engineers might have used standard computer grids, leading to inaccurate predictions near the corners, or they might have avoided these complex shapes entirely.
• After this paper: We have a rigorous mathematical guarantee that we can simulate these complex, unstable systems with high precision. We know exactly how to build the computer grid to get the best possible answer without wasting computing power.

Summary Analogy

Think of this paper as a new GPS navigation system for driving through a city with terrible, jagged streets and slippery roads.

1. Old GPS: Used a standard grid map. It got lost in the sharp turns and gave you bad directions.
2. This Paper: Invented a map that zooms in automatically on the sharp turns (Graded Meshes) and added a "stabilizer" to the car so it doesn't skid (Energy Regularization).
3. The Proof: They mathematically proved that this new GPS will always get you to your destination with the shortest possible error, no matter how weird the city streets are.

The authors didn't just say "it works"; they built the engine, proved the math, and showed the speedometer (numerical examples) to confirm it runs perfectly.

Technical Summary

Here is a detailed technical summary of the paper "Dirichlet control problems with energy regularization governed by non-coercive elliptic equations" by Apel, Mateos, and Rösch.

1. Problem Statement

The paper investigates a linear-quadratic Dirichlet boundary optimal control problem governed by a general second-order elliptic partial differential equation (PDE) on a possibly non-convex polygonal domain $\Omega \subset \mathbb{R}^2$.

• Objective Functional: The goal is to minimize the cost function:
  $$J(u) = \frac{1}{2}\|y_u - y_d\|_{L^2(\Omega)}^2 + \frac{\kappa}{2}|u - u_d|_{H^{1/2}(\Gamma)}^2$$
  where $y_u$ is the state, $u$ is the control (Dirichlet boundary condition), $y_d$ is a desired state, $u_d$ is a desired control, and $\kappa > 0$ is a regularization parameter.
• State Equation: The state $y_u$ satisfies:
  $$-\nabla \cdot (A(x)\nabla y) + b(x) \cdot \nabla y + a_0(x)y = 0 \quad \text{in } \Omega, \quad y = u \quad \text{on } \Gamma.$$
• Key Challenge: Unlike previous studies (e.g., [21], [22]), the associated bilinear form $a(\cdot, \cdot)$ is not assumed to be coercive in $H^1_0(\Omega) \times H^1_0(\Omega)$. This allows for a broader class of operators (e.g., those with indefinite zero-order terms or specific drift terms) but complicates the existence and uniqueness proofs.
• Domain Geometry: The domain may be non-convex, leading to singularities in the solution at re-entrant corners. This necessitates the use of graded meshes to recover optimal convergence rates.

2. Methodology

A. Continuous Analysis

1. Existence and Uniqueness:
   • The authors establish the existence and uniqueness of the state solution $y_u$ using results from previous works [2, 7] regarding non-coercive operators.
   • Crucially, they prove that the objective functional $J(u)$ is strictly convex and coercive on $H^{1/2}(\Gamma)$, even without the coercivity of the underlying PDE operator. This is achieved by analyzing the second derivative of $J$ (Lemma 3.3), showing it is bounded below by a positive constant times the $H^{1/2}(\Gamma)$ norm.
2. Regularity Theory:
   • The optimal control $\bar{u}$ and adjoint state $\bar{\phi}$ are analyzed in weighted Sobolev spaces $W^{k,2}_{\vec{\beta}}(\Omega)$ and $V^{k,2}_{\vec{\beta}}(\Omega)$.
   • The weights $\vec{\beta}$ depend on the singular exponents $\lambda_j$ of the operator at the corners of the polygon.
   • The authors derive that $\bar{u} \in W^{3/2,2}_{\vec{\beta}}(\Gamma)$, which is essential for establishing error estimates on graded meshes.
3. Optimality System:
   • The first-order optimality conditions involve the state equation, the adjoint equation, and a gradient equation linking the control to the normal derivative of the adjoint state and the target control.

B. Numerical Discretization

1. Mesh Strategy: The authors employ graded meshes (refining towards corners) with grading parameters $\mu_j$. This is required to handle the singularities in non-convex domains and achieve optimal convergence orders.
2. Discrete Control Space: The control space $U_h$ consists of continuous piecewise linear functions on the boundary.
3. Key Innovation: $H^{1/2}(\Gamma)$ Projection:
   • Standard approaches often use $L^2(\Gamma)$ projections for boundary controls. The authors argue this is insufficient for obtaining optimal error estimates in weighted spaces with graded meshes.
   • They introduce a discrete projection $Q_h$ onto $U_h$ in the sense of the $H^{1/2}(\Gamma)$ norm. This projection is defined via a bilinear form $p(u,v) = (u,v)_\Gamma + (\nabla Hu, \nabla Hv)_\Omega$.
   • This choice ensures the necessary orthogonality properties in the correct energy space, allowing for sharp error bounds.
4. Discrete Harmonic Extension:
   • A discrete harmonic extension operator $H_h$ is defined using the $Q_h$ projection. The authors prove that this operator is stable uniformly with respect to the mesh size $h$, even on graded meshes (Lemma 5.7).
5. Discrete Optimality System:
   • The discrete problem $(P_h)$ is formulated using a discrete Dirichlet-to-Neumann operator $D_h$.
   • The authors prove that the discrete second derivative of the functional is uniformly coercive with respect to $h$ (Theorem 6.3). This is a non-trivial result requiring convergence analysis of the finite element approximations.

3. Key Contributions

1. Relaxation of Coercivity Assumptions: The paper extends Dirichlet control theory to non-coercive elliptic equations. This is significant because many physical problems (e.g., convection-diffusion with large drift or indefinite reaction terms) do not satisfy standard coercivity conditions.
2. Optimal Convergence on Non-Convex Domains: The authors provide a rigorous proof of optimal order of convergence ($O(h^s)$) for problems on non-convex polygonal domains using graded meshes. Previous works often assumed convexity or quasi-uniform meshes, which yield suboptimal rates in the presence of singularities.
3. Novel Projection Technique: The introduction of the $H^{1/2}(\Gamma)$-projection ($Q_h$) is a major theoretical and practical advancement. It overcomes the limitations of $L^2$-projections in the context of energy regularization and graded meshes, enabling the derivation of optimal error estimates.
4. Uniform Coercivity of the Discrete Functional: The proof that the discrete Hessian is uniformly coercive (Theorem 6.3) ensures the stability of the discrete optimization problem and the uniqueness of the discrete solution, independent of the mesh size.

4. Main Results

• Error Estimates: Under appropriate regularity assumptions on the data ($y_d \in L^2(\Omega)$, $u_d \in W^{3/2,2}_{\vec{\beta}}(\Gamma)$), the authors prove the following error estimate for the optimal control $\bar{u}$ and state $\bar{y}$:
  $$\|\bar{u} - \bar{u}_h\|_{H^{1/2}(\Gamma)} + \|\bar{y} - \bar{y}_h\|_{H^1(\Omega)} + \|\bar{\phi} - \bar{\phi}_h\|_{H^1_0(\Omega)} \leq C h^s$$
  where $s$ depends on the grading parameters $\mu_j$ and the singular exponents $\lambda_j$. With optimal grading, $s$ can approach 1.
• Stability: The discrete solution operator is shown to be bounded uniformly in $h$.
• Numerical Validation: Two numerical examples are presented:
  1. A coercive case (Poisson-like) on an L-shaped domain to verify the theory against known results.
  2. A non-coercive case with specific coefficients ($b(x)$ and $a_0(x)$) designed to violate coercivity.
  • Both examples demonstrate that the experimental order of convergence matches the theoretical predictions ($s \approx 1$) when using graded meshes with appropriate parameters ($\mu < 2/3$).

5. Significance

This paper represents a significant step forward in the numerical analysis of boundary control problems. By removing the coercivity requirement and rigorously handling non-convex geometries with energy regularization, the authors broaden the applicability of finite element methods to a wider class of control problems found in engineering and physics.

The specific contribution of the $H^{1/2}$-projection resolves a long-standing technical hurdle in discretizing Dirichlet controls with energy regularization, ensuring that the discrete problem retains the stability and convergence properties of the continuous one. The results provide a solid theoretical foundation for implementing efficient solvers for complex, non-coercive control problems on realistic, non-convex domains.