Hidden Accuracy and Superconvergence Analysis of Central Discontinuous Galerkin Methods on Overlapping Meshes
Source: arXiv:2606.31931 · Published 2026-06-30 · By Manting Peng, Kailiang Wu
TL;DR
This paper addresses the long-standing open problem of rigorously establishing superconvergence for central discontinuous Galerkin (CDG) methods applied to linear hyperbolic equations on overlapping meshes. Prior to this work, optimal L2 convergence of Q^k CDG schemes on Cartesian meshes was known, but the numerically observed higher-order pointwise superconvergence of order O(h^{k+2}) lacked a rigorous theoretical explanation. The difficulty arises from the loss of standard Galerkin orthogonality due to the overlapping primal and dual mesh structure in CDG, which couples primal-dual residuals and invalidates classical single-mesh DG analysis techniques. To overcome this fundamental barrier, the authors develop a projection-correction framework which identifies a hidden superconvergent residual cancellation mechanism in 1D and a novel high-order cancellation-by-aggregation (HOCA) principle in multiple dimensions. This framework enables a first rigorous proof of the O(h^{k+2}) pointwise superconvergence in the discrete ℓ^∞ norm across all superconvergent points. Furthermore, under a corrected initialization scheme, they prove a stronger cell-average superconvergence estimate of O(h^{min{2k+1,k+3}}). The analysis also extends to fully discrete explicit Runge–Kutta CDG schemes, showing superconvergence up to temporal truncation errors, and leads to a stable reconstruction-based postprocessing method. Extensive numerical experiments in one and two spatial dimensions confirm the sharpness of the theoretical rates and demonstrate the practical relevance of the discovered hidden structures and correction functions.
Key findings
- The semidiscrete Q^k CDG method achieves O(h^{min{2k+1,k+3}}) cell-average superconvergence under a systematically corrected initialization (Theorem 3.1 and Corollary 3.2).
- Pointwise superconvergence of order O(h^{k+2}) in the discrete ℓ^∞ norm across all superconvergent points is rigorously proven under the Liu–Shu–Zhang projection initialization (Theorem 4.17).
- The high-order cancellation-by-aggregation (HOCA) mechanism enables recovering the sharp residual cancellation order in multiple dimensions by aggregating directional projection and correction residuals (Proposition 4.10).
- The loss of single-mesh Galerkin orthogonality caused by the overlapping mesh structure is circumvented through the projection-correction framework that reveals asymptotic weak residual cancellation in 1D (Lemma 3.12) and HOCA in higher dimensions.
- Only two correction levels are needed in the CDG framework for optimal superconvergence, since higher corrections do not improve the rate limited by residual cancellation (Remark 3.15).
- Fully discrete explicit Runge–Kutta CDG methods preserve the spatial superconvergence rates up to temporal truncation errors, supported by stagewise corrected error constructions (Section 5).
- Numerical experiments confirm the theoretical superconvergence orders in 1D and 2D, validating the projection-correction and HOCA framework (Section 6).
- The cell-average superconvergent accuracy can be converted into higher-order polynomial postprocessed approximations via a stable reconstruction procedure (Section 5).
Methodology — deep read
The authors tackle the superconvergence analysis for semidiscrete and fully discrete CDG schemes applied to the linear advection equation on periodic domains with uniform Cartesian overlapping meshes. The core methodology consists of the following steps:
Threat Model & Assumptions: The setting involves linear constant-coefficient hyperbolic equations discretized by CDG on two overlapping meshes (primal and dual). The key difficulty arises from the coupled primal-dual residuals due to the overlapping structure, which causes the loss of classical Galerkin orthogonality. The analysis assumes sufficiently smooth exact solutions in H^{k+2} or higher Sobolev spaces and periodic boundary conditions. Two initializations are considered: the standard Liu–Shu–Zhang projection and a systematically corrected initialization that facilitates stronger cell-average estimates.
Data & Setup: The meshes are uniform Cartesian grids with mesh size h, with polynomial spaces Q^k on primal and dual cells forming the approximation spaces V_h and W_h. The CDG scheme is formulated as a coupled system on these spaces, with carefully defined spatial operators (ˆB for primal, ˜B for dual) involving residual coupling terms scaled by the maximum allowed timestep τ_max.
Projection and Correction: The authors introduce specialized projection operators P* mapping continuous functions into V_h×W_h, preserving cell averages and satisfying variational properties adapted to the overlapping mesh CDG operator. Correction functions R_l u for l=1,2 are defined recursively to cancel the leading primal-dual projection residuals. These corrections belong to zero-cell-average subspaces and satisfy shifted finite difference structures. Their norm bounds are established inductively.
Error Decomposition and Weak Cancellation: The error is decomposed into the difference between numerical solution and projection (ζ), further corrected by R_l u to obtain the corrected error ˆζ. The analysis exploits a key asymptotic weak cancellation in 1D for piecewise constant test functions (Lemma 3.12 and Theorem 3.19) whereby residuals vanish or reduce sharply after correction, despite the lack of full Galerkin orthogonality. In multiple dimensions, directional corrections and the high-order cancellation-by-aggregation (HOCA) principle aggregate several directional residuals to recover the higher-order cancellation.
Stability and Grönwall Argument: Using known L2 stability of CDG (Proposition 2.1), the corrected error equation is analyzed to bound the residual terms C1_i, C2_i, C3_i arising from time derivatives and correction residuals. The residual norms are bounded by O(h^{σ_k}) with σ_k = min{2k+1, k+3}. Grönwall's inequality then yields the superconvergence bounds for ˆζ in L2 norm.
Pointwise and Cell-Average Estimates: The 1D result is extended to pointwise estimates in ℓ^∞ via discrete Sobolev inequalities applied to the superconvergent points. Cell-average superconvergence follows from properties of the projections and zero-average corrections. Multidimensional generalizations rely on tensor-product spaces and directional corrections.
Fully Discrete Analysis: The theory extends to explicit Runge-Kutta CDG schemes by constructing stagewise corrected errors preserving spatial superconvergence apart from temporal truncation errors constrained by CFL stability conditions.
Numerical Experiments: The authors perform extensive testing in 1D and 2D, confirming O(h^{k+2}) pointwise superconvergence and the stronger O(h^{min{2k+1,k+3}}) cell-average convergence. They also verify the effectiveness of the reconstruction-based postprocessing that leverages cell-average accuracy for improved polynomial approximations.
Reproducibility: While the paper does not explicitly state a public code release, all operators, projections, corrections, and test functions are explicitly defined mathematically. The uniform Cartesian mesh setting and polynomial spaces Q^k enable straightforward numerical replication. The detailed constructions and lemma proofs provide a complete roadmap for independent implementation.
Technical innovations
- Identification and rigorous proof of asymptotic weak residual cancellation in 1D CDG despite the loss of classical Galerkin orthogonality, revealing a hidden superconvergent mechanism.
- Formulation of the high-order cancellation-by-aggregation (HOCA) principle to recover superconvergence in multiple dimensions by aggregating directional projection and correction residuals, overcoming primal-dual coupling.
- Introduction of a projection-correction framework with two correction levels that specifically targets coupled primal-dual residuals and enables rigorous superconvergence proofs for CDG on overlapping meshes.
- Extension of the projection-correction superconvergence analysis to fully discrete explicit Runge–Kutta CDG schemes via stagewise corrected errors preserving spatial superconvergence up to temporal truncation error.
- Discovery of a systematically corrected initialization that yields a stronger cell-average superconvergence estimate of order O(h^{min{2k+1, k+3}}), a previously unknown result for CDG.
Baselines vs proposed
- Liu–Shu–Zhang projection initialization: pointwise ℓ∞ superconvergence rate = O(h^{k+2}) (proved in this paper) vs previous: observed but not proven
- Corrected initialization: cell-average superconvergence rate = O(h^{min{2k+1, k+3}}) (newly established) vs standard initialization: no such stronger estimate
Limitations
- The theory and proofs are restricted to uniform Cartesian overlapping meshes; extensions to non-Cartesian or triangular meshes remain open and challenging.
- The superconvergence analysis assumes smooth exact solutions with high Sobolev regularity (H^{k+2} or higher), limiting applicability to rough or discontinuous solutions.
- The correction hierarchy is truncated at two levels due to the limiting order of asymptotic weak cancellation; improvements beyond O(h^{k+3}) for low-order elements are not possible in this framework.
- The fully discrete analysis addresses only explicit Runge–Kutta schemes under CFL stability conditions; implicit or other time integrators are not studied here.
- No adversarial or perturbed data evaluations are conducted, so robustness of superconvergence under mesh or equation perturbations is unknown.
Open questions / follow-ons
- Can the projection-correction and HOCA superconvergence framework be extended to non-Cartesian or general unstructured overlapping meshes for CDG methods?
- Is it possible to develop superconvergence theory for CDG methods applied to nonlinear hyperbolic equations or systems with discontinuous solutions?
- How robust are the discovered residual cancellations and superconvergence properties under mesh perturbations, variable coefficients, or less smooth data?
- Can the reconstruction-based postprocessing leveraging cell-average superconvergence be optimized or generalized for practical adaptive CDG schemes?
Why it matters for bot defense
Though this paper does not directly address bot-defense or CAPTCHA systems, the advanced numerical analysis techniques it develops for uncovering hidden accuracy and subtle cancellation mechanisms could inspire approaches in CAPTCHA adversarial robustness and bot detection modeled by hyperbolic PDEs or overlapping-data structures. The methods to rigorously characterize and exploit superconvergence points and averaged quantities might analogously inform feature extraction or anomaly detection in bot-defense ML models that rely on high-accuracy sampling of complex data. Furthermore, the projection-correction strategies for error mitigation could inform the design of layered checks or correction functions to improve robustness of CAPTCHA challenge evaluations under adversarial inputs. However, the technical content remains specialized to numerical PDE discretizations rather than direct security application, so application would require significant adaptation and cross-domain translation.
Cite
@article{arxiv2606_31931,
title={ Hidden Accuracy and Superconvergence Analysis of Central Discontinuous Galerkin Methods on Overlapping Meshes },
author={ Manting Peng and Kailiang Wu },
journal={arXiv preprint arXiv:2606.31931},
year={ 2026 },
url={https://arxiv.org/abs/2606.31931}
}