Optimal convergence of local discontinuous Galerkin methods for convection-diffusion equations

This paper bridges the gap between theoretical estimates and numerical evidence for the $hp$ local discontinuous Galerkin method applied to convection-diffusion equations by establishing new approximation results for Gauss-Radau projections, thereby proving optimal convergence rates even for solutions with limited spatial regularity.

The Gist

Imagine you are trying to paint a perfect portrait of a mountain range using a set of digital brushes. The mountain has smooth slopes, but it also has a jagged, sharp peak.

In the world of computer simulations, this "painting" is solving complex math problems (specifically, equations that describe how things move and spread, like heat or pollution). The "brushes" are mathematical tools called Local Discontinuous Galerkin (LDG) methods.

Here is the story of this paper, broken down into simple concepts:

1. The Problem: The "Blurry" Peak

For a long time, scientists knew that if the mountain (the solution to the equation) was perfectly smooth, their digital brushes worked beautifully. They could make the picture incredibly sharp just by using more complex, higher-quality brushes (increasing the "polynomial degree," or $p$).

However, when the mountain had a sharp, jagged peak (a "singularity" or a point where the math gets messy), something strange happened.

• The Theory: The old math books said, "If you use a super-complex brush on a jagged peak, you will only get a slightly better picture. You lose one step of quality."
• The Reality: When scientists actually ran the computer simulations, they saw something different. The picture got much sharper than the old math predicted. It was as if the theory was underestimating the power of the brushes.

There was a gap between what the math said should happen and what the computers actually did.

2. The Solution: A New Way to Look at the Jagged Peak

The authors of this paper decided to fix this gap. They realized the old way of measuring the "jaggedness" of the peak was too blunt. It was like trying to measure the sharpness of a knife with a ruler meant for measuring a log.

They introduced a new, more sensitive tool: Fractional Calculus.

• The Analogy: Imagine you have a smooth curve that suddenly breaks. The old math looked at the break and said, "It's broken, so we can't do much."
• The New Insight: The authors looked closer and realized that even though the curve is broken, the way it breaks follows a very specific, predictable pattern (like a fractal). They used a special kind of math (involving "fractional derivatives") to describe exactly how the curve behaves right at that sharp point.

3. The Magic Trick: The "Gauss-Radau Projection"

To fix the math, they had to improve a specific step in the painting process called the Gauss-Radau projection.

• The Metaphor: Imagine you are trying to fit a smooth, curved piece of clay onto a jagged rock. The "projection" is the technique you use to mold the clay so it fits the rock as perfectly as possible.
• The Old Way: The old technique assumed the rock was rough in a generic way, so it left a gap.
• The New Way: The authors developed a new technique that understands the exact shape of the jagged rock. By using their new "fractional" understanding of the sharp peak, they could mold the clay (the math approximation) to fit the rock perfectly.

4. The Results: Two Types of Jagged Peaks

The paper also discovered that where the jagged peak is located matters:

• Case A: The Peak is on a Grid Line (Fitted). Imagine the jagged peak sits exactly on a line where your grid of pixels meets. The new method showed that if the peak is here, the computer can get an optimal result. The picture becomes incredibly sharp, matching the best possible performance.
• Case B: The Peak is Inside a Pixel (Unfitted). Imagine the jagged peak is floating right in the middle of a single pixel, not on a line. Here, the result is slightly less perfect (you lose a tiny bit of sharpness), but it is still much better than the old theory predicted.

5. Why This Matters

Before this paper, if a scientist saw a computer simulation working better than the math predicted, they might have thought, "Oh, the math is just a lower bound; it's safe, but maybe not the whole story."

This paper says: "No, the math was actually wrong about the limit."

By creating this new framework, the authors proved that the computer simulations were actually doing the best possible job all along. They closed the gap between the "theory" (what we think should happen) and "evidence" (what actually happens).

Summary

Think of it like this:

• The Old View: "Trying to draw a sharp corner with a smooth brush is impossible; you'll always lose some detail."
• The New View: "Actually, if you understand the geometry of the corner perfectly, you can draw it with amazing precision. We just needed a new set of glasses (fractional calculus) to see how the brush interacts with the corner."

This work ensures that engineers and scientists can trust their computer models more, knowing that the math finally catches up to the reality of the simulations.

Technical Summary

Here is a detailed technical summary of the paper "Optimal Convergence of Local Discontinuous Galerkin Methods for Convection-Diffusion Equations" by Liu, Xie, Wang, and Zhang.

1. Problem Statement

The paper addresses a long-standing discrepancy between theoretical error estimates and numerical observations regarding the Local Discontinuous Galerkin (LDG) method for convection-diffusion equations, specifically in the context of $p$-version convergence (increasing polynomial degree $p$ with fixed mesh size $h$).

• The Gap: Previous theoretical analysis (notably Castillo et al., 2002) predicted suboptimal convergence in $p$ for solutions with limited spatial regularity (singularities). Specifically, the theory suggested a loss of one order of convergence compared to numerical experiments.
• The Specific Issue: For a solution with a singularity of order $\alpha$ (e.g., $u \sim x^\alpha$), the existing theory predicted a convergence rate of $O(p^{-2\alpha})$ (hyperbolic case) or $O(p^{-2\alpha-2})$ (convection-diffusion case), whereas numerical experiments consistently showed optimal rates of $O(p^{-2\alpha-1})$ and $O(p^{-2\alpha+1})$ respectively.
• Goal: The authors aim to close this theoretical gap by establishing a new approximation framework that accurately characterizes the regularity of singular solutions, thereby proving optimal $p$-convergence rates that match numerical evidence.

2. Methodology

The core of the methodology involves a refined analysis of Gauss-Radau projections applied to functions with fractional regularity.

A. New Function Spaces for Singular Solutions

The authors introduce a specialized function space, $U^{\alpha, m}_{\hat{a}+}(\Lambda)$, to characterize solutions with algebraic singularities.

• Definition: A function $u$ belongs to this space if $u \in W^{k,1}$ and its Caputo fractional derivative of order $\alpha$ (denoted $CD^\alpha_{\hat{a}+}u$) belongs to $W^{m,1}$.
• Rationale: Standard Sobolev spaces fail to capture the precise regularity of singularities like $x^\alpha$. By utilizing the Caputo derivative, the authors can define a semi-norm that naturally arises from the exact formula of Legendre expansion coefficients derived via fractional integration by parts.

B. Analysis of Gauss-Radau Projections

The LDG error analysis relies heavily on the properties of Gauss-Radau projections ($\Pi^\pm_p$). The authors derive new, sharp error bounds for these projections acting on functions in the newly defined fractional spaces.

• Key Insight: The error bounds depend on the interplay between the singularity exponent $\alpha$, the additional regularity index $m$, and the polynomial degree $p$.
• Technique: The proof utilizes the exact formula for Legendre expansion coefficients involving generalized Gegenbauer functions of fractional degree (GGF-F). This allows for precise estimation of the projection error in $L^2$ and at the boundaries.

C. Case Distinctions

The analysis is extended to cover three distinct scenarios regarding the location of the singularity:

1. Left-Endpoint Singularity: The singularity is at the boundary of the domain (e.g., $x=a$).
2. Fitted Interior Singularity: The singularity lies exactly on a mesh node ($x = x_j$).
3. Unfitted Interior Singularity: The singularity lies strictly inside an element ($x \in (x_{j-1}, x_j)$).

3. Key Contributions

1. Resolution of the Suboptimality Gap: The paper provides the first theoretical proof that the LDG method achieves optimal $p$-convergence for singular solutions, correcting the previous "loss of one order" prediction.
2. Novel Approximation Framework: The introduction of the space $U^{\alpha, m}$ based on Caputo fractional derivatives provides a natural and rigorous way to characterize singular solutions for DG methods.
3. Sharp Error Estimates: The authors derive explicit convergence rates for both the hyperbolic case ($d=0$) and the convection-diffusion case ($d \neq 0$), distinguishing between fitted and unfitted grids.
4. Unified Treatment: The framework handles various types of singularities (endpoint and interior) and different regularity levels within a single theoretical structure.

4. Main Results

The paper establishes the following optimal convergence rates for the error $\|u - u_h^p\|$ in the energy norm, where $\alpha$ is the singularity exponent and $m$ is the regularity of the fractional derivative:

A. Endpoint Singularities (Theorem 3.1)

For a singularity at the left endpoint:

• Purely Convective ($d=0$): Convergence order is $O(p^{-\min\{2\alpha+1, \alpha+m-1/2\}})$.
  • Note: If $m$ is large, this recovers the optimal $O(p^{-(2\alpha+1)})$.
• Convection-Diffusion ($d \neq 0$): Convergence order is $O(p^{-\min\{2\alpha-3/2, \alpha+m-3/2\}})$.
  • Correction: This corrects the previous theoretical prediction of $O(p^{-(2\alpha-2)})$ to the observed $O(p^{-(2\alpha-3/2)})$.

B. Fitted Interior Singularities (Theorem 4.1)

When the singularity aligns with a mesh node:

• Purely Convective ($d=0$): Order is $O(p^{-\min\{2\alpha+1/2, \alpha+m-1/2\}})$.
• Convection-Diffusion ($d \neq 0$): Order matches the endpoint case ($O(p^{-\min\{2\alpha-3/2, \alpha+m-3/2\}})$).

C. Unfitted Interior Singularities (Theorem 4.2)

When the singularity is inside an element (not on a node):

• Purely Convective ($d=0$): Order degrades to $O(p^{-(\alpha+1/2)})$.
• Convection-Diffusion ($d \neq 0$): Order degrades to $O(p^{-(\alpha-1/2)})$.
• Significance: This explains why fitting the mesh to the singularity is crucial for maintaining high-order accuracy in the $p$-version.

5. Significance and Impact

• Theoretical Validation: The work validates the empirical success of the LDG method for singular problems, confirming that the method is indeed capable of optimal convergence if the solution's regularity is properly characterized.
• Guidance for Mesh Design: The results highlight the critical importance of mesh fitting for interior singularities. If a singularity is not aligned with a grid node, the convergence rate drops significantly (by roughly half an order or more), providing a clear directive for adaptive mesh refinement strategies.
• Broader Applicability: The authors suggest that the proposed framework (using fractional Sobolev spaces and refined Gauss-Radau analysis) can be applied to resolve similar suboptimality issues in other Discontinuous Galerkin (DG) schemes for hyperbolic and elliptic problems.
• Numerical Consistency: The paper includes extensive numerical experiments that perfectly align with the new theoretical bounds, demonstrating the tightness of the derived estimates.

In summary, this paper resolves a decade-old theoretical inconsistency in the analysis of LDG methods by introducing a sophisticated fractional-regularity framework, proving that the method is optimal for singular solutions provided the mesh is appropriately aligned with the singularities.