🔢 mathematics

Shape optimisation for adaptive $r$ -refinement: the one-dimensional case with residual based error estimators

This paper proposes and validates a shape optimisation algorithm for adaptive $r$ -refinement in one-dimensional Poisson problems, demonstrating that minimising a differentiable residual-based error estimator effectively repositions mesh nodes to improve finite element approximations.

Original authors: Philip J. Herbert

Published 2026-07-27

📖 3 min read🧠 Deep dive

CC BY 4.0

Original authors: Philip J. Herbert

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 trying to draw a perfect map of a mountainous landscape, but you only have a limited number of pegs and rubber bands to create a grid over the terrain. In the world of computer science and engineering, this is exactly what happens when scientists try to solve complex equations that describe how heat flows, how fluids move, or how structures bend. These equations, known as Partial Differential Equations (PDEs), are too hard to solve with a pencil and paper, so computers break the world down into tiny shapes—like a net made of triangles or squares—to approximate the answer. This process is called "discretization."

The problem is that a uniform grid, where every square is the same size, is often a waste of time. If you are mapping a flat valley, you don't need a million tiny pegs; if you are mapping a jagged cliff, a few big pegs won't cut it. Usually, computers fix this by adding more pegs where the action is (making the net denser), a method called "h-refinement." But there's another way: instead of adding more pegs, you could just move the ones you already have to the spots where they are needed most. This is called "r-refinement." It's like rearranging the furniture in a room to make the space work better without buying a single new chair. The question scientists have been asking is: how do we know exactly where to move those pegs to get the most accurate map possible?

This paper by Philip J. Herbert tackles that exact question, but with a clever twist. Instead of just guessing where to move the mesh nodes (the pegs), the author treats the movement of the mesh as a "shape optimization" problem. Think of it like a sculptor who doesn't just chip away at stone, but actively reshapes the entire block to fit a specific form perfectly. The paper proposes an algorithm that mathematically calculates the best way to slide the nodes of a computer mesh to minimize the error in the solution.

The author proves that this method works. In a simplified one-dimensional world (imagine a straight line instead of a 2D map), the paper demonstrates that a specific algorithm, which uses a "residual error estimator" (a clever way of guessing how wrong the current map is without knowing the true answer), can successfully rearrange the mesh. The computer runs through a series of steps, moving the nodes slightly, checking if the error went down, and repeating until it can't get any better. The paper shows that this process converges, meaning it reliably finds a better arrangement.

However, the author is careful not to overpromise. The proof of convergence is solid for the one-dimensional case and for an idealized version where the computer knows the exact answer. For the more practical, real-world scenario where the computer only has the error estimator, the paper provides strong numerical evidence (simulations) that the method works well. When tested against standard methods that simply add more grid points, this "moving mesh" approach performed just as well, and in some cases, better, using the same number of computational resources. The paper doesn't claim to have solved every problem in the universe; it specifically focuses on a type of equation called the Poisson problem and shows that moving the mesh nodes is a viable, efficient strategy that deserves a spot in the toolbox of modern scientific computing.

Technical Summary: Shape Optimisation for Adaptive r-Refinement

Problem Statement
The paper addresses the challenge of accurately and efficiently discretizing partial differential equations (PDEs), specifically focusing on the Poisson problem with Dirichlet boundary conditions. While standard adaptive strategies often rely on $h$ -refinement (locally refining the mesh) or $p$ -refinement (increasing polynomial order), this work investigates adaptive $r$ -refinement. The goal of $r$ -refinement is to improve the quality of the finite element discretisation by repositioning the vertices of a computational mesh without altering the number of degrees of freedom or the mesh topology. The central difficulty lies in determining how to move these nodes to minimize the discretisation error, particularly when the exact solution is unknown and a posteriori error estimators must be used.

Methodology
The authors frame the mesh movement problem within the context of computational shape optimisation.

Mathematical Framework: The domain $\Omega$ is treated as a reference shape with a reference triangulation $\hat{T}$ . Mesh movement is modeled via a parametrisation map $\Phi$ , which is a bi-Lipschitz map from the reference domain to the physical domain. The space of admissible maps $U$ ensures that the image of the mesh remains a valid, non-degenerate triangulation.
Optimisation Formulation: The problem is cast as minimizing a cost functional $J(\Phi)$ $J (Φ)$ .
- Theoretical Case: Initially, the functional is defined as the true finite element error (e.g., $\|y - y_\Phi\|_{H^1_0}^2$ ), where $y$ is the exact solution and $y_\Phi$ is the discrete solution on the mesh defined by $\Phi$ .
- Practical Case: Since the exact solution is rarely available, the authors propose using a residual-based a posteriori error estimator as the cost functional. In one dimension, this estimator is simplified to depend on the local mesh size $h_T$ and the residual of the PDE.
Algorithm: The authors employ a standard Armijo-Goldstein descent method (Algorithm 1). This involves:
1. Calculating the shape derivative (or gradient) of the cost functional with respect to the mesh parametrisation.
2. Finding a descent direction $U$ that minimizes the directional derivative subject to constraints (e.g., $|DU| \leq 1$ ).
3. Performing a line search to determine an admissible step size $\alpha_n$ that ensures sufficient decrease in the cost functional.
Differentiability: A significant technical hurdle addressed is the differentiability of the residual estimator, which involves mesh sizes ( $h_T$ ) that are typically only Lipschitz continuous. The authors restrict the analysis to the one-dimensional case to demonstrate that the estimator is differentiable and that the Armijo condition can be satisfied.

Key Contributions

Formulation of r-Refinement as Shape Optimisation: The paper explicitly links adaptive $r$ -refinement to shape optimisation theory, utilizing bi-Lipschitz maps to ensure mesh validity during movement.
Convergence of Idealised Method (Theorem 1): The authors prove that for an objective function based on the true error (in arbitrary dimensions), the Armijo-type descent method generates a sequence of meshes that converges to a stationary point.
Convergence of Practical Method in 1D (Theorem 2): The most novel contribution is the establishment of convergence for a fully computable, residual-based adaptive $r$ -refinement strategy in one dimension. This involves proving the differentiability of the residual estimator and demonstrating that the Armijo step size exists, ensuring the algorithm terminates with a sequence of meshes where the gradient of the estimator vanishes.
Numerical Verification: The paper presents numerical experiments using linear finite elements on the Poisson problem. These experiments verify that the proposed method improves the finite element approximation compared to uniform meshes and performs comparably to greedy $h$ -refinement strategies in the tested 1D scenarios.

Results

Convergence: Theoretical results confirm that the sequence of meshes generated by the algorithm converges, with the gradient of the objective functional approaching zero.
Error Reduction: Numerical experiments in one dimension show that meshes optimised using the residual estimator yield lower errors than uniform meshes.
Comparison with h-Refinement: In the 1D comparison, the $r$ -refinement method (using both the true error and the residual estimator) performs comparably to a greedy $h$ -refinement strategy, with all methods outperforming the initial uniform mesh.
2D Extension: While the theoretical convergence proof for the residual estimator is limited to 1D, the authors present 2D numerical experiments using the true error functional, demonstrating that the optimisation process successfully reduces the error and produces non-uniform meshes that better capture the solution features.

Significance and Claims
The paper claims to introduce a new strategy for adaptive $r$ -refinement by placing mesh movement within a computational shape optimisation framework. The primary significance lies in the rigorous proof of convergence for a practical, residual-based algorithm in one dimension. The authors note that while $h$ - and $p$ -refinement are well-understood and widely implemented, $r$ -refinement offers the advantage of a fixed mesh topology, which can be more efficient for highly parallel machines. The work demonstrates that standard shape optimisation tools can be successfully adapted to move mesh nodes to minimize discretisation error, verified by numerical experiments that show the method's efficacy in improving finite element approximations. The authors remain modest, noting that the theoretical convergence of the residual-based estimator is currently established only for the one-dimensional case, with future work planned for higher dimensions and more general equations.

More like this