SSBE-PINN: A Sobolev Boundary Scheme Boosting Stability and Accuracy in Elliptic/Parabolic PDE Learning

1 Introduction

• Even minor boundary-approximation errors can yield O(1) relative H1 errors, motivating an approach that controls boundary regularity.The introduction notes that exact boundary conditions can support H1 approximation properties, while perturbations in boundary loss can cause substantial deviations.Source: §1 · p. 3
• SSBE retains L2 terms for the PDE residual and initial condition while adding an H1(∂Ω) boundary term that constrains boundary values and tangential derivatives.Its boundary-aligned reparameterization uses intrinsic coordinates to evaluate boundary traces and derivatives without requiring an ambient extension of boundary data.Source: §1 · pp. 3–5
• The paper establishes quantitative H1-stability estimates for SSBE that link its population loss to approximation error in the natural energy norm.The introduction describes analyses for both elliptic and parabolic PDEs.Source: §1 · pp. 4–5
• Finite-sample generalization bounds relate population and empirical SSBE losses under finite interior and boundary sampling and yield explicit H1-error estimates.The stated analysis combines Rademacher complexity estimates with Barron-type approximation theory.Source: §1 · pp. 4–5
• SSBE consistently achieves lower relative L2 and H1 errors than conventional PINNs across neural-network architectures and Poisson, heat, nonlinear elliptic, and irregular-domain problems.The introduction also describes theoretical stability and generalization results as supporting these experiments.Source: §1 · p. 5

2 Mathematical Setup and Preliminaries

• The preliminaries organize boundary-space definitions, PDE assumptions, and neural-network function-space notation for the subsequent analysis.The section also presents Figure 1 as a framework of SSBE.Source: §2 · pp. 5–6, Figure 1 · p. 6

2.1 Sobolev Space Restricted on Boundary

• Boundary Sobolev spaces are defined through local boundary coordinates, including L2 and H1 integrability on the boundary.The setup also introduces coordinates that straighten the boundary.Source: §2.1 · pp. 6–7

2.2 PDEs and Assumptions

• The analysis focuses on elliptic and parabolic PDEs on bounded C1,0 domains, with assumptions stated to support solution existence and analysis.The paper also assumes the relevant elliptic and parabolic coefficient conditions.Source: §2.2 · pp. 7–8

2.3 Barron Space

• The neural-network setup introduces feed-forward networks, path norms, and Barron-type function spaces for data and solutions.Barron pairs and triplets are defined for the elliptic and parabolic settings.Source: §2.3 · pp. 8–9

Example Construction

• For the unit-ball Poisson problem with L=−∆, f=4, and zero boundary data, the construction uses harmonic perturbations.Each perturbation is harmonic inside the unit ball.Source: §3 · p. 10

• The sequence nevertheless has O(1) relative H1 error, illustrating failure to achieve H1 convergence despite the small boundary loss.The authors attribute the issue to conventional boundary loss not capturing intrinsic regularity constraints.Source: §3 · pp. 10–11
• Boundary L2 mismatch of order O(ε) can coexist with a PINN objective approaching zero in the constructed sequence.The paper presents this as a counterexample to inferring H1 accuracy from small boundary loss.Source: §3 · pp. 10–11

• SSBE redefines the loss to include an H1 boundary-discrepancy penalty alongside the PDE residual.The boundary weight λ is adjustable; the analysis sets it to |∂Ω| for simplicity.Source: §4 · p. 11
• The algorithm is described for elliptic and time-dependent parabolic problems using the boundary-coordinate definitions introduced earlier.Its stated objective-function regularity requirements differ for time-dependent and elliptic cases.Source: §4 · p. 11
• The authors note that SSBE can extend to hyperbolic problems, but do not elaborate on those cases.Source: §4 · p. 12

5 H1-Stability and Convergence Analysis of SSBE for PDEs

• The theoretical analysis establishes SSBE stability and H1 convergence for both elliptic and parabolic PDEs.The results are presented under the paper’s stated assumptions.Source: §5 · p. 12
• An extension theorem supplies a bounded linear map from boundary H1 data to an interior H1 function on Lipschitz domains.The extension agrees with the boundary data and its norm is bounded by a constant depending on Ω and d.Source: §5 · p. 12

Stability Analysis for Elliptic Problem

• For the elliptic problem, Theorem 5.2 states a stability estimate for objectives v∈H2(Ω), with its constant depending on the domain and PDE coefficients.The theorem assumes the elliptic conditions and a C1,0 domain.Source: Stability Analysis for Elliptic Problem · p. 13

Stability Analysis for Parabolic Problem

• For the parabolic problem, Theorem 5.3 gives a stability result for solutions and objectives in the stated space-time Sobolev classes.Its constant depends on Ω, T, d, and the listed PDE coefficients.Source: Stability Analysis for Parabolic Problem · p. 14

6 Error Analysis: A Priori Estimates for SSBE

• The explicit theoretical estimates assume Barron-type forcing, although the authors state that practical SSBE applies beyond this setting.The analysis assumes non-divergence-form operators and uniformly bounded PDE coefficients.Source: §6 · p. 15
• Explicit a posteriori and a priori generalization error bounds are provided for both elliptic and parabolic problems.The section frames these bounds as connecting population loss with empirical loss under finite sampling.Source: §6 · p. 15

Technical Lemma and Sketch of Proof

• For Barron pairs, the hypothesis-class Rademacher complexity is uniformly bounded in terms of parameter norm and data size.The estimates cover the solution class and boundary-related classes used in the analysis.Source: Technical Lemma and Sketch of Proof · p. 18
• For a Barron pair, increasing network width can make the population loss arbitrarily small.The Barron-type forcing assumption is used for tractable theoretical estimates, not required for practical SSBE use.Source: §6 · p. 15, Technical Lemma and Sketch of Proof · p. 18
• The proof strategy combines a Rademacher-complexity bound with approximation estimates to bound population loss through empirical loss.The a priori bound decomposes population loss into approximation error and generalization error.Source: Technical Lemma and Sketch of Proof · pp. 17, 19

7 Numerical Experiments

• Across representative PDE experiments, SSBE improves both L2 and H1 accuracy over standard PINNs, including on irregular domains.The experiments include Poisson, heat, nonlinear Poisson, high-dimensional Poisson, and irregular-domain problems.Source: §7 · p. 19

7.1 Poisson Equation

• SSBE yields lower relative L2 and H1 errors than PINN across tested neural-network architectures for the Poisson equation.The comparison varies the number of hidden layers and neurons per layer.Source: Table 1 · p. 20, §7.1 · p. 20

7.2 Heat Equation

• For the heat equation, SSBE produces lower relative L2 and H1 errors than PINN across tested parameter settings.The reported pointwise errors are uniformly lower for SSBE across the evaluated time steps.Source: §7.2 · p. 21, Table 2 · p. 22

7.3 Nonlinear Elliptic Equation

• For the nonlinear elliptic equation, SSBE achieves lower relative L2 and H1 errors than PINN across tested parameter configurations.Source: Table 3 · p. 23, §7.3 · p. 23

7.4 High Dimension Poisson Equation

• For dimensions d = 3, 10, 20, and 50, SSBE achieves accurate H1 approximations and consistently outperforms PINN.The experiment uses 10,000 interior points and 1,000×d boundary points.Source: Table 4 · p. 24, §7.4 · p. 24

7.5.1 Irregular domain: intrinsic boundary parameterization

• Across the tested irregular shapes, SSBE outperforms standard PINN with L2(∂Ω) enforcement in relative L2(Ω) and H1(Ω) errors.The test uses intrinsic boundary parameterizations and reports results for flower, heart, and lips domains.Source: Table 5 · p. 25, §7.5.1 · p. 25, Figure 8 · p. 25

7.5.2 Irregular domain: porous medium

• Across porous-medium cases, SSBE outperforms standard PINN with an L2(∂Ω) penalty in relative L2(Ω) and H1(Ω) errors.The tested domains remove holes of varying sizes from a square, with later cases more highly perforated.Source: Table 6 · p. 26, §7.5.2 · p. 26

8 Conclusion

• The work targets conventional PINNs’ difficulty achieving accurate H1 convergence when boundary loss terms are inadequate, particularly with minor boundary inaccuracies.The conclusion points to Section 3 for an explicit illustration of this limitation.Source: §8 · pp. 26–27
• SSBE incorporates boundary regularity into training by redefining the boundary loss in a Sobolev space, with stability analysis bounding H1 error by the proposed loss.The conclusion links this approach to mitigating numerical instabilities and improving solution accuracy and smoothness.Source: §8 · p. 27
• The generalization estimates quantify the relationship between population and empirical losses and indicate robust convergence under limited sampling.The conclusion characterizes these estimates as a finite-sample reliability result.Source: §8 · p. 27
• SSBE consistently achieves lower relative L2 and H1 errors than conventional PINNs across architectures, PDE types, and irregular domains.Figure 9 compares pointwise errors for standard PINN and SSBE-PINN across three porous-medium cases.Source: §8 · p. 27, Figure 9 · p. 27
• Future work includes extending SSBE to more complex PDE systems and hyperbolic equations, optimizing architectures, and adapting boundary segmentation for gradient computations.These directions are proposed to improve computational efficiency, accuracy, and gradient-computation precision.Source: §8 · pp. 27–28

A Embedding Theorem: Wk,p(∂Ω)→W1,q(Ω)

• For boundary data g in W1,p(∂Ω), the appendix states a bounded linear extension operator into W1,q(Ω) whose trace equals g.The passage defines q in terms of p and d.Source: A Embedding Theorem: Wk,p(∂Ω)→W1,q(Ω) · p. 30
• The proof handles non-flat boundary neighborhoods by straightening the boundary with a C1 mapping before constructing local extensions.It then combines local constructions using smooth cutoffs and a cover of the boundary.Source: A Embedding Theorem: Wk,p(∂Ω)→W1,q(Ω) · pp. 31–32
• The extension satisfies a W1,q(Ω) norm bound by the W1,p(∂Ω) norm and has trace equal to the given boundary data.The proof states this bound as the concluding estimate.Source: A Embedding Theorem: Wk,p(∂Ω)→W1,q(Ω) · p. 32
• For Ω in the stated C0,1 class, H1 boundary data also admit an extension into H1(Ω).The appendix obtains this consequence by taking p=2 and invoking the stated Sobolev-space inclusion.Source: A Embedding Theorem: Wk,p(∂Ω)→W1,q(Ω) · p. 32

B Rademahcer Complexity and Contraction Lemma

• Rademacher complexity is defined for a real-valued function class on a sample, with the appendix specifying independent random signs taking values ±1.These definitions set up the complexity results that follow.Source: B Rademahcer Complexity and… · p. 32
• The appendix states a contraction lemma for coordinatewise functions with Lipschitz constant ρ.It identifies the result as rephrased from Lemma 26.9 in.Source: B Rademahcer Complexity and… · p. 33
• A separate result gives a Rademacher-complexity bound for linear predictors on a sample.The appendix identifies it as rephrased from Lemma 26.20 in.Source: B Rademahcer Complexity and… · p. 33
• For uniformly bounded functions, the generalization-gap theorem gives a bound holding with probability at least 1−δ over an i.i.d. sample.The stated conditions include 0≤f(z)≤B for every function and input.Source: B Rademahcer Complexity and… · p. 33

C Technical Details for Generalization Error Estimate

• The proof estimates Rademacher complexities for boundary-data classes and their coordinate derivatives.The passage describes separate estimates for rGQ and Dxrα rGQ.Source: C Technical Details for Generalization… · p. 34
• The derivation defines function classes indexed by boundary patches and coordinates for use in the generalization estimate.The appendix also applies contraction and linear-predictor complexity lemmas to functions involving σ′′, with Lipschitz constant M.Source: C Technical Details for Generalization… · pp. 34, 37
• The derivation combines intermediate estimates to obtain an equation holding with probability at least 1−δ over the sampled set.The stated probability follows after the preceding bounds are combined.Source: C Technical Details for Generalization… · pp. 38–39