Source-linked AI summary
Quadrature by Expansion: A New Method for the Evaluation of Layer Potentials
Andreas Klöckner, Alexander Barnett, Leslie Greengard, Michael O'Neil
TL;DR
Singular and nearly singular boundary integrals hinder the practical use of integral equation methods. The paper introduces QBX, which exploits local interior or exterior smoothness to evaluate layer potentials accurately, with convergence theory and compatibility with fast hierarchical algorithms.
Problem
Integral equation methods require accurate evaluation of boundary integrals with singular or nearly singular kernels, where standard high-order quadrature is difficult.
Method
QBX evaluates layer potentials using expansions centered near the boundary, relying on local smoothness in the interior or exterior and a smooth underlying quadrature rule.
Results
QBX permits rapid, high-order accurate evaluation of layer potentials and is supported by a complete convergence theory.
Takeaways & Limitations
The scheme is easy to implement and compatible with fast hierarchical algorithms, including applications requiring off-surface evaluation arbitrarily close to the boundary.
Takeaways & Limitations
For operators beyond the weakly singular single layer, QBX computes a one-sided limit determined by the expansion-center location, so principal values may require two-sided treatment.
Abstract
from arXiv · showhide
Integral equation methods for the solution of partial differential equations, when coupled with suitable fast algorithms, yield geometrically flexible, asymptotically optimal and well-conditioned schemes in either interior or exterior domains. The practical application of these methods, however, requires the accurate evaluation of boundary integrals with singular, weakly singular or nearly singular kernels. Historically, these issues have been handled either by low-order product integration rules (computed semi-analytically), by singularity subtraction/cancellation, by kernel regularization and asymptotic analysis, or by the construction of special purpose "generalized Gaussian quadrature" rules. In this paper, we present a systematic, high-order approach that works for any singularity (including hypersingular kernels), based only on the assumption that the field induced by the integral operator is locally smooth when restricted to either the interior or the exterior. Discontinuities in the field across the boundary are permitted. The scheme, denoted QBX (quadrature by expansion), is easy to implement and compatible with fast hierarchical algorithms such as the fast multipole method. We include accuracy tests for a variety of integral operators in two dimensions on smooth and corner domains.
1. Introduction
The paper addresses the difficulty of accurately evaluating singular, weakly singular, nearly singular, and hypersingular layer-potential integrals. It introduces QBX, which exploits local smoothness of induced fields and uses near-boundary expansions with smooth high-order quadrature.
- Motivation: Singular and nearly singular boundary integrals are a central practical difficulty in applying integral equation methods to complicated domains.The challenge includes layer potentials with singular or weakly singular kernels and principal-value or finite-part integrals.
- Prior approaches: Existing approaches include product integration, singularity subtraction, changes of variables, kernel regularization, and special-purpose high-order quadratures.Classical quadrature schemes are especially difficult for hypersingular integrals, although integration by parts can help.
- QBX approach: QBX evaluates layer potentials by expanding fields that are locally smooth on either the interior or exterior side of the boundary, despite possible jumps across it.The method uses a smooth underlying quadrature rule, which may be global, panel-based, or adaptive.
- Scope and implementation: The approach is designed to be easy to implement, high-order accurate, compatible with fast hierarchical algorithms, and essentially dimension-independent.The numerical experiments in this paper are restricted to two-dimensional problems.
- QBX approach: Unlike global Taylor-expansion methods for analytic Cauchy integrals, QBX uses near-boundary expansion centers and error estimates based on local smoothness and finite boundary differentiability.It does not require assumptions about the location of the nearest singularity.
2. Smooth, high-order quadrature
The section shows that direct smooth quadrature loses accuracy near the boundary, while QBX uses off-surface expansions to recover high-order on-surface evaluation. Accuracy depends on expansion order and sufficient quadrature resolution.
- Direct quadrature near the boundary: Direct trapezoidal quadrature is accurate away from Γ but loses accuracy as targets approach the boundary.The high-accuracy region excludes a neighborhood of Γ whose extent depends on grid spacing h and requested precision ϵ.
- Direct quadrature near the boundary: Choosing an off-surface center c within O(h) of a boundary target gives only first-order on-surface accuracy when evaluating Sσ(c) directly.Under trapezoidal-rule error ϵ, the resulting error is O(ϵ + h).
- Local expansions: QBX expands Sσ about an off-surface center c and evaluates the local Bessel expansion at boundary targets.The expansion coefficients are computed from smooth integrands using the underlying trapezoidal rule and Graf’s addition theorem.
- Local expansions: An expansion of order p = 3 provides about four digits uniformly, while p = 6 increases accuracy near and on Γ from four to six digits.These results rely on one-sided smoothness of the induced field in the interior or exterior.
- Resolution requirements: Increasing p without refining the quadrature can significantly reduce accuracy because higher-order coefficient integrands become more oscillatory and sharply peaked.Increasing the trapezoidal-rule nodes restored more than ten digits of accuracy for p = 12.
- Resolution requirements: The method extends beyond the trapezoidal rule: composite Gauss-Legendre quadrature produces analogous behavior, and the inaccurate region shrinks roughly with h.The underlying smooth rule may therefore be global or panel-based and adaptive.
QBX as a regularization scheme
The section interprets QBX as replacing the original Green’s function near the boundary with a truncated, regularizing kernel. This filtered kernel enables high-order accuracy without additional correction, although increasing expansion order sharpens the integrand.
- Kernel interpretation: QBX can be viewed as substituting the original Green’s function G with a truncated kernel Gp for targets on or near the boundary.This interpretation follows by interchanging summation and integration in the local expansion.
- Kernel interpretation: The substituted kernel Gp acts as a filter that regularizes the kernel and achieves high-order accuracy without additional correction.The regularization is the kernel-level interpretation of evaluating the local expansion instead of the original layer-potential integral.
- Resolution trade-off: As p increases, Gp becomes more sharply peaked, explaining why the underlying quadrature must use smaller h to resolve higher-order expansions.The same resolution trade-off appears when computing higher-order expansion coefficients.
3. Mathematical Foundations of QBX
QBX approximates layer potentials by truncated local expansions whose coefficients are computed with smooth quadrature, yielding separate truncation and coefficient-approximation errors. The method handles one-sided, principal-value, derivative, and hypersingular evaluations, with accuracy and spectral behavior depending on center placement and limit processing.
- Error analysis: Theorem 1 establishes QBX error bounds for smooth curves, panelized quadrature, expansion order p, and q-node Gaussian rules.The assumptions include a center whose radius-r ball meets the curve at the target point and a density in specified Hölder spaces.
- Error analysis: QBX error separates into truncation from the p-term Bessel expansion and numerical error from approximating its coefficients.The truncation term is O(r^(p+1)) under the stated density regularity, while the second term arises from quadrature of the coefficients.
- Error analysis: With expansion-center distance r = O(h), truncation error is O(h^(p+1)); coefficient accuracy additionally requires h/(4r) < 1, or r > h/4.Centers placed too close to the boundary relative to the discretization lose accuracy; controlled precision is obtained when the error components are kept separate.
- Error analysis: Classical convergence can be recovered by refining r more slowly than h, while adaptive discretizations require a related local error analysis.The adaptive version is omitted because its difficulty is local and its estimates are similar to the equal-panel case.
- Derivatives, jumps, and principal value integrals: QBX directly computes one-sided limits; principal-value and finite-part quantities require jump-condition subtraction or appropriately averaged two-sided limits.Derivatives of layer potentials can be obtained by analytically differentiating the local expansion, with higher derivatives incurring additional powers of r^-1 in the error estimate.
- Spectral structure of operators approximated by QBX: Two-sided averaging roughly doubles cost and is governed by the less accurate side, but improves the spectral behavior of the Nyström approximation for integral-equation solves.For compact cases, the averaged operator has spectrum accumulating at zero and supports iterative convergence near machine precision; this advantage does not generally apply to hypersingular or Hilbert-Riesz operators.
- Adaptive discretization: Adaptive source grids with abrupt mesh changes require control over expansion-center placement to avoid local-expansion errors.The implementation description assumes a smooth curve subdivided into panels with potentially varying arc lengths.
- Spectral structure of operators approximated by QBX: Truncated expansions attenuate or alias high-frequency density components, with attenuation empirically dominant and potentially erroneous high-frequency data treated benignly.For the single-layer potential, the QBX one-sided limit faithfully reproduces the continuous operator’s spectrum accumulating at zero.
4. Numerical experiments
The experiments evaluate QBX across diverse layer-potential operators, smooth and cornered geometries, and integral-equation boundary-value problems. Results show high-order accuracy, with the expected one-order loss for hypersingular Neumann operators and rapid GMRES convergence.
- Test geometries: The test suite uses ellipses, a starfish, and a teardrop with a corner, with panelized curve representations.Smooth curves are adaptively subdivided using Legendre expansions; panels are also refined when geometric resolution criteria fail.
- Layer Potential Evaluation: QBX tests cover target derivatives of single- and double-layer potentials, tangential dipoles, and Helmholtz parameter k = 0.5.The operators are applied to σ(t) = sin(10πt), using local expansion order p = 16 and L2 and L∞ error comparisons.
- Integral equation solvers: Integral-equation tests construct exact solutions from exterior or interior monopole point charges, then solve Dirichlet and Neumann problems with QBX.Computed fields at observation points are compared with the exact solution to obtain relative errors.
- Integral equation solvers: Neumann operators show a one-order accuracy loss because QBX is applied to a hypersingular kernel.The results are described as somewhat more erratic, but the loss is reported as expected.
- Integral equation solvers: GMRES reaches a residual of 10^-14 with a modest number of iterations in nearly all cases, even for low-order discretizations.This behavior is identified as making QBX-based solvers particularly robust.
- Non-smooth geometries: Corner-focused refinement produces accurate teardrop results, with the apparent p = 5 convergence drop attributed to stopping refinement at 10^-8.The experiments report no significant obstacle to applying QBX in this setting.
5. Generalizations and implementation issues
The paper argues that QBX extends beyond its two-dimensional Helmholtz setting to broader kernels and dimensions, while noting that computational cost was not addressed in detail.
- Generalizations: QBX is presented as independent of dimension and applicable to kernels not directly connected to partial differential equations.The paper notes that such extensions are discussed elsewhere and identifies efficiency, robustness, and automatic adaptivity as open questions.
- Implementation issues: The paper does not fully address computational cost; a straightforward implementation has asymptotic complexity O(NNt).An example applies a single-layer operator at p = 5 to 7680 source nodes and 1280 targets in 0.6 seconds using 16 cores.
6. Conclusions
The conclusions characterize QBX as an easily implemented, high-order method for singular and weakly singular layer-potential evaluation. They frame it as a basis for robust software tools serving large-scale physics and engineering simulations.
- Conclusions: QBX evaluates layer potentials rapidly and accurately by exploiting smoothness of the induced potential in the interior or exterior domain.The method is described as having a complete convergence theory and requiring only minor modifications for arbitrarily close off-surface evaluation.
- Conclusions: QBX could support robust software tools for integral operators with singular or weakly singular kernels in large-scale physics and engineering simulations.