On the validity of "A proof that the discrete singular convolution (DSC)/Lagrange-distributed approximation function (LDAF) method is inferior to high order finite differences"

Guo-Wei Wei, Guo-Wei Wei 0001, Shan ZhaoJournal of Computational Physics 2007 · DOI

TL;DR

This paper rigorously challenges a previous claim that Discrete Singular Convolution (DSC) methods are inferior to high-order finite differences for approximating derivatives. Through extensive numerical experiments across various problems, the authors demonstrate that DSC methods, particularly the regularized Shannon kernel (DSC-RSK) variant, often significantly outperform traditional finite difference schemes, especially for problems involving high wavenumbers.

What Problem Does This Paper Address?

The paper directly addresses a specific claim made in a 2006 paper by J.P. Boyd, titled "A proof that the discrete singular convolution (DSC)/Lagrange-distributed approximation function (LDAF) method is inferior to high order finite differences." Boyd's paper asserted that DSC/LDAF methods are generally less accurate than high-order finite differences for approximating derivatives. This paper aims to analyze the validity of Boyd's claims by providing a wide variety of counterexamples and detailed numerical comparisons. The authors highlight that Boyd's original analysis relied heavily on Fourier analysis of the function $e^{ikx}$ and informal arguments, which they argue are insufficient for drawing general conclusions about numerical method performance.

Methods

The authors employ a comparative numerical analysis approach, evaluating six different numerical methods for derivative approximation and solving differential equations. These methods include:

  • Discrete Singular Convolution (DSC) with Regularized Shannon Kernel (DSC-RSK): The primary method of interest, which the authors demonstrate to be highly accurate.
  • Standard Finite Difference (FD): A baseline method, often implemented using Fornberg's code.
  • Boyd's Finite Difference (Boyd's FD): A sum-accelerated finite difference method proposed by Boyd.
  • Boyd's Euler-accelerated Sinc Algorithm (Euler): Another sum-accelerated pseudospectral method.
  • Boyd's Modified Euler-accelerated Sinc Algorithm (MEuler): A variation of the Euler method.
  • Sech Method (Spectrally-weighted): A spectrally-weighted difference method proposed by Boyd.

The core of their analysis involves applying these methods to various problems, including differentiating $e^{ikx}$, solving boundary value problems, and tackling the Helmholtz and unsteady hyperbolic equations. They meticulously compare the $L_\infty$ errors across different stencil widths (M) and report execution times for weight generation.

Datasets

Not discussed in the paper. The paper focuses on synthetic problems designed to test specific aspects of numerical methods, rather than using pre-existing datasets.

Key Insights and Ideas

  • Fourier analysis of $e^{ikx}$ is insufficient: The paper strongly argues that analyzing the performance of numerical methods solely on the function $e^{ikx}$ is not a reliable way to assess their general accuracy, especially for solving differential equations.
  • DSC-RSK excels in high-wavenumber problems: Contrary to Boyd's claims, the DSC-RSK method consistently outperforms other methods, including finite differences, for problems involving high wavenumbers.
  • Context matters for method selection: The paper emphasizes that the suitability of a numerical method is highly problem-dependent. A method that performs well for one type of function or problem may not be optimal for another.
  • Boyd's claims are overly general: The authors provide numerous counterexamples that contradict Boyd's broad assertions about the inferiority of DSC/LDAF methods.
  • The "spectrally-weighted" advantage is not universal: While spectrally-weighted methods like Sech are claimed to be superior for high wavenumbers, the DSC-RSK method often surpasses them in these scenarios.

Key Results

  • Differentiating $e^{ikx}$: In many cases, the DSC-RSK method outperforms the FD method, sometimes by orders of magnitude, even for small wavenumbers. For higher wavenumbers, DSC-RSK is significantly more accurate than FD and Sech methods.
  • Boundary Value Problems: For Boyd's boundary value problem, DSC-RSK significantly outperforms all other methods for large stencil widths. For a boundary value problem with a confined distribution of wavenumbers, DSC-RSK is the most accurate method across all stencil widths.
  • Helmholtz Equation: For high-frequency problems, DSC-RSK is the only method capable of delivering extremely accurate solutions at high grid densities, outperforming others by orders of magnitude.
  • Unsteady Hyperbolic Equations: DSC-RSK is shown to be the best method for highly oscillatory solutions, being over $10^6$ times more accurate than other approaches. For a wave comb solution, DSC-RSK maintains its accuracy advantage.
  • Navier-Stokes Equation: For a high-frequency problem, DSC-RSK is significantly more accurate than Sech and FD methods.
  • Quantum Eigenvalue Problem: While Sech shows surprising accuracy in this specific problem, DSC-RSK generally performs better than FD, Euler, MEuler, and Sech methods.

Limitations and Future Work

  • LDAF Confusion: The authors note that Boyd's paper uses "Lagrange-distributed approximating function" (LDAF) and "linear distributed approximating functional" interchangeably, and that the latter term appears to not exist in the literature, making the analysis of LDAF claims unclear.
  • Parameter Optimization: The DSC-RSK method has an adjustable parameter 'r' that needs to be optimized for specific applications. While the authors use near-optimal values, they acknowledge that this requires some experience.
  • Scope of "Proof": The authors focus on refuting specific claims from Boyd's paper and do not aim to claim the absolute superiority of DSC-RSK across all possible problems. They acknowledge that other methods might outperform DSC-RSK in specific, perhaps low-order, scenarios.
  • Future Work: The paper implicitly suggests that further detailed analysis and comparison of numerical methods for various problem classes are always warranted, as method suitability is problem-dependent. The authors also mention that the DSC algorithm has been applied to many practical problems, implying its utility in real-world applications.

Summary updated

BibTeX
@article{wei2007validity,
  author = {Guo-Wei Wei and Guo-Wei Wei and Shan Zhao},
  title = {{On the validity of "A proof that the discrete singular convolution (DSC)/Lagrange-distributed approximation function (LDAF) method is inferior to high order finite differences"}},
  journal = {Journal of Computational Physics},
  year = {2007},
  doi = {10.1016/j.jcp.2007.05.036},
}

Loading related content...

Figure