Source-linked AI summary
Boundary Value Correction for the Boundary Element Method
Erik Burman, Peter Hansbo, Mats G. Larson
TL;DR
High-order boundary element approximation on curved domains requires accurate boundary representation, while curved panels entail singular quadrature. This paper corrects Dirichlet data on flat interior surrogate panels using the existing flux, and derives stability and error bounds under geometric and trace assumptions. With γ ≃ h^-1, the sufficient offset condition yields order h^(k+1/2), supported through cubic Dirichlet approximation for offsets of order h².
Problem
High-order BEM on curved domains requires accurate boundary representation, while curved elements require singular quadrature on curved panels.
Source: §1 · p. 1Method
A first-order Taylor correction imposes Dirichlet data on flat interior surrogate panels using the existing Dirichlet trace and normal flux, without adding an unknown field.
Source: §1 · p. 1, §7 · p. 23Results
With γ ≃ h^-1, δ_h ≲ h^((k+1)/2) yields error order h^(k+1/2) for P_k/P_(k−1) spaces, and O(h²) offsets support the estimate through cubic Dirichlet approximation.
Source: §1 · p. 2Takeaways & Limitations
An O(h²) normal offset supports the predicted rate through cubic Dirichlet approximation.
Source: §1 · p. 2Exterior-surrogate experiments use special data and do not justify exterior surrogates for general Dirichlet problems, whose solutions may lack a sufficiently regular harmonic extension.
Source: §6.6 · p. 21
Abstract
from arXiv · showhide
We develop a boundary element method for the Laplace equation with Dirichlet boundary conditions in a bounded three-dimensional domain. The method uses flat panels of maximum diameter \(h\) on an interior surrogate boundary and corrects the imposed boundary values by a first-order Taylor expansion along the panel normals. The correction uses the normal flux already present in the Calder\'on formulation and introduces no additional unknown field. Under uniform assumptions on the surrogate geometry and boundary operators, the penalised form is coercive when \(\gamma\delta_h\le c_0<4\), where \(\gamma\) is the penalty parameter and \(\delta_h\) is the maximum normal offset. The error estimate separates trace approximation from a geometry term of size \(\gamma^{1/2}\delta_h^2\|u\|_{H^3(\Omega)}\). With \(\gamma\simeq h^{-1}\), the sufficient condition \(\delta_h\lesssim h^{(k+1)/2}\) yields an error of order \(h^{k+1/2}\) for \(P_k/P_{k-1}\) Dirichlet/flux spaces under the stated regularity and uniform trace bounds. Thus an offset of order \(h^2\) supports the estimate through cubic Dirichlet approximation. We analyse two penalty variants and present numerical experiments on a sphere and an ellipsoid to examine convergence and sensitivity to geometry, penalty scaling and quadrature.
1. Introduction.
The paper transfers boundary-value correction to boundary elements, using flat panels and the existing flux unknown to impose corrected Dirichlet data. Under stated uniform assumptions, the penalised formulation is stable and has an error estimate balancing trace approximation with geometric error.
1. Introduction.
- Flat panels avoid curved-panel singular quadrature, while the correction requires evaluating surrogate-to-physical boundary distances at quadrature points.Perturbing those distances allows small geometry variations to be introduced.Source: §1 · p. 1
- The first-order Taylor correction uses the existing Dirichlet trace and normal flux, adding no unknown field.The physical boundary value is approximated by the surrogate trace plus the normal offset times the flux.Source: §1 · p. 1
- For uniformly bounded interior surrogates, stability holds when γδ_h ≤ c_0 < 4.The penalty controls boundary mismatch and the constant mode in the hypersingular operator's kernel.Source: §1 · p. 2
- With γ ≃ h^-1 and δ_h ≲ h^((k+1)/2), the P_k/P_(k-1) spaces attain error order h^(k+1/2).An O(h²) offset supports the estimate through cubic Dirichlet approximation.Source: §1 · p. 2
- The method transfers Taylor correction along discrete normals to boundary elements, using the existing flux unknown; earlier approaches include, [7], and.It incorporates geometric correction into a penalised Calderón structure.Source: §1 · p. 2
2. Boundary Integral Formulation.
This section sets the boundary-integral conventions and trace spaces used by the corrected formulation. Layer potentials yield the Calderón trace identity, while operator bounds and coercivity properties support stability.
2. Boundary Integral Formulation.
- On a connected Lipschitz boundary, the Calderón trace identities fix signs in the corrected formulation, and operator bounds underpin its stability proof.The coercivity estimates discussed here concern three dimensions.Source: §2 · p. 2
2.1. Layer Potentials and Boundary Operators.
- The single- and double-layer potentials represent interior harmonic functions, with the source-variable derivative indicated by the subscript y.Their mapping properties are recalled in Appendix A.1.Source: §2.1 · p. 3
- The operators V, K, K′, and D are the single-layer, double-layer, adjoint double-layer, and hypersingular boundary operators.D is the negative weak Neumann trace of the double-layer potential.Source: §2.1 · p. 3
- The pairing ⟨·,·⟩_Γ is duality between H^-1/2(Γ) and H^1/2(Γ), and equals the surface product integral for L2 functions.The operators extend to the stated trace spaces.Source: §2.1 · p. 3
2.2. The Calderón Trace Identity.
- Taking interior traces of the harmonic representation gives the Calderón trace identity for trial pairs ordered as flux then Dirichlet trace.Adjointness cancels off-diagonal operator terms when trial and test pairs coincide.Source: §2.2 · p. 3
2.3. Trace Spaces and Coercivity.
- The natural trace space and norm are specified for the formulation, and weighted flux integrals from the Taylor correction motivate a subspace.The additional condition for single-layer coercivity in two dimensions is discussed in Remark A.3.Source: §2.3 · p. 4
- In three dimensions, V and D satisfy ellipticity estimates, while D is symmetric, nonnegative, and has globally constant functions as its kernel.The Dirichlet penalty controls this constant mode.Source: §2.3 · p. 4
3. Boundary Value Correction.
The method imposes physical Dirichlet data on an interior flat surrogate by correcting the boundary trace along panel normals. Its stability and rate estimates depend on the offset, penalty scaling, and uniform geometric and operator assumptions.
3.1. Surrogate Geometry and the Taylor Correction.
- The normal-ray map sends each surrogate point to the physical boundary with offset 0 ≤ ρ ≤ δ_h, and the method assumes uniformly controlled geometry and segment overlap.The normal-ray projection generally differs from closest-point projection, and each segment must remain in the domain.Source: §3.1 · p. 4
- Neglecting the Taylor remainder, the corrected boundary condition is w + ρλ ≈ g_h, combining the surrogate Dirichlet trace and normal flux.The remainder measures the discrepancy between prescribed data and the corrected trace.Source: §3.1 · p. 5
3.2. The Corrected Discrete Method.
- The penalised Calderón form enforces the Taylor-corrected boundary value when exact traces are inserted, with the Taylor remainder as the only consistency error under exact assembly.The discrete spaces use continuous degree-k Dirichlet traces and discontinuous degree-(k−1) fluxes; analysis assumes exact form and load evaluation.Source: §3.2 · p. 5
- Exterior surrogates fall outside the stability argument because negative offsets can make the stabilising term negative, and general Dirichlet solutions may lack the required harmonic extension.The analysis therefore justifies interior surrogates, not exterior ones.Source: §3.2 · p. 6
- A second penalty variant corrects the test trace; both variants satisfy the error estimate with possibly different constants, though their discrete solutions may differ.The alternative formulation is analysed in Appendix A.4.Source: §3.2 · p. 6
3.3. Parameter Choices and Predicted Rates.
- With γ ≃ h^-1, the estimate combines trace error of order h^(k+1/2) and geometry error of order γ^(1/2)δ_h², giving the sufficient condition δ_h ≲ h^((k+1)/2).For k = 1, choose c_γc_ρ ≤ c_0 < 4; for k > 1, γδ_h ≤ c_0 holds for sufficiently small h.Source: §3.3 · p. 6, Table 3.1 · p. 6
- An affine geometry approximation with δ_h = O(h²) supports predicted rates through k = 3; higher degrees require smaller geometric error, and quadrature must suit the polynomial degree.A higher-order correction would require treatment of derivatives beyond the trace and flux, which is not developed here.Source: §3.3 · p. 6
4. Stability and Error Estimates.
Under uniform geometric and operator assumptions, the corrected formulation is stable and has an a priori error bound that separates best approximation from the Taylor remainder. With additional regularity and uniform trace bounds, the analysis gives rates for Pk/Pk−1 spaces under sufficient geometry conditions.
4.1. Standing Assumptions and the Error Norm.
- The error analysis assumes interior surrogate geometry, conforming shape-regular quasiuniform meshes, and uniform operator and surface-norm bounds.The analysis separately requires overlap control for normal segments and additional trace regularity for the convergence rates.Source: §4.1 · pp. 6–7
4.2. Stability of the Corrected Form.
- The corrected form is coercive, with positive weighted penalty terms, and its continuity constants are independent of h.The proof uses the energy identity, in which the off-diagonal Calderón terms cancel by adjointness.Source: §4.2 · pp. 7–8
4.3. Consistency and the Geometry Error.
- Consistency error is governed by the Taylor remainder, which is controlled using a strip estimate along normal segments.When δh = 0, the remainder rh vanishes.Source: §4.3 · p. 8
4.4. Discrete Well-Posedness and Best Approximation.
- Under the standing assumptions, the discrete problem has a unique solution, and coercivity plus consistency separates best approximation from the Taylor remainder.The unique-solution result is stated for every gh in L2(Γh).Source: §4.4 · p. 8
4.5. Approximation and Convergence Rates.
- The rate statements require bulk regularity and uniform trace bounds; panelwise regularity permits derivative jumps across polyhedral edges.The theorem first gives the a priori estimate under trace regularity, then adds assumptions to express the surface estimate in a bulk norm.Source: §4.5 · pp. 9–10
- For Pk/Pk−1 spaces, the reported rates are h^3/2, h^5/2, and h^7/2 for k = 1, 2, and 3, respectively, under the listed offset conditions.The corresponding conditions are δh ≤ cρh, δh ≤ cρh^3/2, and δh ≤ cρh^2; the bounds assume the bulk regularity and other hypotheses of Theorem 4.4.Source: §4.5 · p. 10
5. Reduced Formulation: Elimination of the Hypersingular Operator.
The reduced formulation eliminates the Dirichlet trace as an independent unknown and solves for the flux, recovering the trace afterward. Under its stated conditions it retains the error rates of the full method, with reduced assembly cost, but its natural formulation is in L2.
5. Reduced Formulation: Elimination of the Hypersingular Operator.
- The Taylor correction eliminates the Dirichlet trace by substituting the corrected datum and solving the flux equation; the trace is then recovered by post-processing.The reduction removes terms involving D and K′ from the formulation.Source: §5 · pp. 10–11
- Because ρλh is not in H1/2(Γh) for elementwise-polynomial λh, the reduced method is posed in L2.The H1/2 mapping property needed for K is unavailable for this argument.Source: §5 · p. 11
- The reduced formulation retains coercivity and consistency under the same smallness condition on γδh, with the consistency error still governed by the Taylor remainder.The passage also states a finite upper bound on γδh suffices for the alternative penalty variant, whereas the first variant uses c0 < 4.Source: §5 · pp. 11–12
- The reduced method reproduces the full method’s rates under the design condition, at a fraction of the assembly cost, but ties Dirichlet-trace accuracy to flux accuracy.The trace is recovered by post-processing from the corrected datum and computed flux.Source: §5 · pp. 11–12
6. Numerical Experiments.
Flat-panel experiments on a sphere and ellipsoid examine convergence and sensitivity to penalty and offset choices, alongside exterior-surrogate and penalty-consistency diagnostics. The reported convergence rates broadly track predictions, while the exterior tests lie outside the convergence theory’s scope.
6.1. Computational Setting.
- Flat-panel experiments cover sphere and ellipsoid geometries, with errors measured against exact surrogate traces and an independent coarse-mesh library comparison.The offset condition was checked for the listed baseline meshes, but uniformity over the refinement family remains an assumption.Source: §6 · p. 12, §6.1 · pp. 12–14
6.2. Baseline Convergence on a Sphere.
- Sphere Erep slopes were 1.45, 2.74, and 3.40 versus predicted 1.5, 2.5, and 3.5; the cubic last-refinement rate was 3.11, insufficient to establish asymptotic convergence.GMRES required 11–19 iterations without direct fallback.Source: §6.2 · p. 14, Figure 6.1 · p. 15, Table 6.1 · p. 15
6.3. Convergence on an Ellipsoid.
- Ellipsoid Erep fitted slopes were 1.45, 2.88, and 3.92, with 15–19 GMRES iterations; refinement-family uniformity remains a hypothesis.The geometric checks verify the offset condition for the tested meshes, not uniformly across the refinement family.Source: §6.3 · p. 16, Figure 6.3 · p. 17, Table 6.3 · p. 17
6.4. Sweep in the Penalty Parameter.
- Across a sixteen-fold cγ change, error varied by at most a factor of 1.6 and GMRES took 10–23 iterations; larger penalties generally reduced iterations, with a small k = 1 reversal.Part of the slight Erep increase comes from its weighted Dirichlet contribution; unweighted error was essentially flat for k = 2, 3 and varied by 15% for k = 1.Source: §6.4 · p. 17, Figure 6.4 · p. 19
6.5. Sweep in the Offset.
- At fixed panel count, k = 1 total error was almost flat, while k = 2, 3 errors continued decreasing beyond the design exponent; sampled potential error was more offset-sensitive.The design exponent is a sufficient asymptotic condition, not necessarily an error minimizer at fixed resolution.Source: §6.5 · p. 20, Table 6.5 · p. 20, Figure 6.5 · p. 21
6.6. Exterior Surrogate Diagnostic.
- Exterior tests are outside the convergence theory: general Dirichlet solutions need not admit a sufficiently regular harmonic extension, and negative ρ invalidates the nonnegative-penalty argument.The experiment’s manufactured solution supplies a harmonic extension only where the enlarged domain avoids its singularity.Source: §6.6 · p. 21
- Exterior k = 1 errors were nonmonotone; k = 2 fell from 2.13 to 0.0616 and then 0.00811, while three configurations hit 5000 iterations and required direct solves.Those tabulated errors are from direct solutions after iterative solves failed.Source: §6.6 · p. 21, Table 6.6 · p. 22
6.7. Penalty Consistency and Quadrature Sensitivity.
- The intentionally inconsistent half-coefficient form had larger errors on every listed mesh and decreased more slowly than the corrected form; its fitted slopes were 0.52 and 0.92.The full mixed penalty coefficient cancels the first-order boundary displacement in the exact-solution residual.Source: §6.7 · p. 22, Table 6.7 · p. 22
7. Conclusions.
The corrected formulation is stable under the stated assumptions and separates trace approximation error from a geometry contribution quadratic in the normal offset. With the prescribed offset scaling, it supports the predicted convergence rate through cubic Dirichlet approximation.
7. Conclusions.
- With γ ≃ h^-1 and δh ≲ h^((k+1)/2), the error is of order h^(k+1/2) for Pk/Pk-1 spaces.The Dirichlet trace and flux use degrees k and k-1, respectively.Source: §7 · p. 23
- Under Theorem 4.4's assumptions, including γδh ≤ c0 < 4, the corrected formulation is stable and its geometry contribution is quadratic in the normal offset.The correction uses the existing Dirichlet and flux unknowns.Source: §7 · p. 23
- An offset of order h^2 supports the estimate through cubic Dirichlet approximation.Source: §7 · p. 23
- Numerical checks support implementation on tested meshes, but do not establish all uniform assumptions on mesh families.The theorem concerns interior surrogates and exact assembly; exterior geometries, unrestricted cut meshes, and a priori quadrature-error control require further analysis.Source: §7 · p. 23
A.1. Additional Operator Bounds.
The appendix records operator and trace properties used for the Calderón form and its uniform analysis on surrogate surfaces. For exact harmonic traces, adjointness cancels off-diagonal terms and yields a positive bound.
A.1. Additional Operator Bounds.
- On a fixed bounded Lipschitz domain, both layer potentials are harmonic in the interior, with cited trace results in [26, Lemmas 6.6 and 6.10].Source: §A.1 · p. 23
- The boundary-operator estimates hold on a fixed surface; uniform bounds for surrogate-surface families are assumed separately.The boundary operators are cited to [26, Chapter 6].Source: §A.1 · p. 23
- For exact harmonic traces λ = γNu and w = γDu, operator bounds, trace duality, and adjointness yield a positive bound for some α > 0.Source: §A.1 · p. 23
A.2. Continuity and Coercivity of the Corrected Form.
The appendix establishes continuity and coercivity for the corrected form under the standing mesh and penalty assumptions. Coercivity requires γδh ≤ c0 < 4, while the penalty also controls the constant mode in the hypersingular operator's kernel.
A.2. Continuity and Coercivity of the Corrected Form.
- The offset may vary over Γh and vanish on subsets, within the standing assumptions.Source: §A.2 · p. 24
- The continuity bound for Ah is uniform in h and follows from uniform operator bounds, trace duality, and estimates for weighted and mixed terms.Source: §A.2 · p. 24
- If γδh ≤ c0 < 4, the corrected form is coercive with a constant αB independent of h.Source: §A.2 · p. 24
- The penalty controls the boundary mismatch and the constant mode in the hypersingular operator's kernel.The proof uses mean-zero coercivity because the hypersingular operator annihilates globally constant functions.Source: §A.2 · p. 24
- The theorem does not cover arbitrary planar domains; a two-dimensional extension needs normalization ensuring single-layer coercivity.The mean-zero restriction for the hypersingular operator does not provide that coercivity.Source: §A.2 · p. 24
A.3. A Surface Inverse Estimate.
The appendix proves a discrete flux inverse estimate using a conforming panelwise bubble test function. The resulting negative-norm bound scales with h^1/2.
A.3. A Surface Inverse Estimate.
- The discrete flux inverse estimate holds under the stated mesh and penalty assumptions.Source: §A.3 · p. 25
- A conforming test function equal to hbT μh on each panel yields h^1/2∥μh∥Γh ≤ C∥μh∥H−1/2(Γh).The proof adapts the localization and duality argument of Nédélec [19, Lemma 4, pp. 68–69].Source: §A.3 · p. 25
A.4. Alternative Dirichlet Penalty.
The alternative Dirichlet penalty uses the corrected trace and is continuous and coercive under the standing assumptions. Its residual bound lets the convergence proof apply as well, with possibly different constants.
A.4. Alternative Dirichlet Penalty.
- The alternative penalty uses the corrected trace for testing; both variants replace u composed with p by w + ρλ but differ in the test trace.Its energy identity includes the single-layer and hypersingular terms plus a weighted penalty norm.Source: §A.4 · pp. 25–26
- The alternative form is continuous and coercive in the stated energy norm, and its residuals satisfy the same dual-norm bound.The convergence theorem therefore applies to this variant, with possibly different constants.Source: §A.4 · pp. 25–26
- For the alternative form, a finite upper bound on γδ_h suffices for coercivity, while the first variant uses the sufficient condition c0 < 4.The penalty terms control the separate weighted norms, and the constant mode is controlled as in Lemma A.2.Source: §A.4 · p. 26
Use of artificial intelligence and computational tools.
- The authors report using ChatGPT and Codex for language editing, presentation, drafting, mathematical checks, and numerical-code assistance and verification.Python with NumPy, SciPy, and Matplotlib was used for computations, independent verification, and figure preparation; the authors retain responsibility for the manuscript.Source: Use of artificial intelligence and… · p. 26