acceptodds
Under review as a conference paper at ICLR 2027

Principal-Part Decomposition for Neural Operator Learning of Dirichlet-to-Neumann Maps

Abstract

Dirichlet-to-Neumann (DtN) maps encode PDE responses on the boundary by mapping Dirichlet data to Neumann data. Learning these operators across varying geometries is challenging because DtN maps are nonlocal first-order operators with singular high-frequency behavior. For smooth planar boundaries, a boundary integral formulation decomposes the DtN map into a universal Fourier multiplier and a smoother geometry-dependent remainder. We introduce Principal-Part Decomposed Neural Operators (PPDNO), a hybrid analytic-neural framework that computes the principal part exactly via FFT and learns only the remainder using a geometry-conditioned low-rank network that preserves linearity in the boundary data. We prove smoothing and separated-approximation properties of the remainder and derive error bounds for the reconstructed map. Experiments on Laplace and Helmholtz problems show that PPDNO improves accuracy and generalization over standard neural operators and residualized baselines while maintaining high inference efficiency.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operators, operator learning, scientific machine learning, Dirichlet-to-Neumann maps, boundary integral methods

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gDLDOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper considers the operator learning task (Dirichlet-to-Neumann (DtN) map) from a geometry and boundary data to the corresponding boundary response, i.e., boundary gradient along the normal direction, provided that the interior Laplacian problem is known. This problem is a particularly specific operator learning setting (while I have not learned much DtN literature before, I believe that this problem is important). The DtN map can be decomposed into a universal singular part and a smoother geometry-dependent remainder. The singular part could be geometry-independent and thus is handled through explicit calculation. The reminder integral operator is parameterized in a separable form similar to DeepONet, i.e., an MLP processes coordinates and the other encodes geometry to mimic the smooth operator. The overall architecture is physically interpretable and can be applied to other PDEs whose Green's functions-induced BIE hold similar decomposition.

Critical strengths
  1. Propose a physically interpretable operator learning framework for a type of boundary estimation problems.
  2. The presentation is clear and easy to follow.
  3. Experiments show good performances, especially on boundary data and geometry generalization.
Critical weaknesses

The theoretical parts are information-dense and could be better positioned. I was not sure if Theorems 5.1-5.3 provide essentially new theoretical results for the DtN operator and the associated low-rank kernel approximation. Theorems 5.1 is valid under the boundary problem setting, independent of the proposed structure. Proper citations on previous literatures could be appreciated. Theorem 5.2 gives a rank-dependent error estimate for the low-rank kernel operator approximation. However, it is not clear whether this estimate is new or is adopted from existing literature on kernel approximation (e.g., from Mercer's theorem or something like that). Theorem 5.3 gives a generalization error bound. I think it could be a direct combination of the rank-dependent error and standard generalization estimate (e.g., Rademacher or VC dimension), and hence could be shortened.

Questions for authors
  1. Around Eq. (9), while W0{\mathcal W}_0 was defined, I believe that W1{\mathcal W}_1 should also be properly defined nearby.
  2. In Eq (1), the period should be replaced by a comma.
  3. A sensitivity analysis w.r.t. the rank parameter could be important to supply the rank error estimate.
  4. Theorem 5.3 could be simplified and made more rigorous. The parameter space of the DeepONet, Fourier truncation error, and the error terms Equad,EnetE_{quad},E_{net} could be stated explicitly. The theory is itself dense but is conceptually simple, hence could be moved to appendix.
2 comments
AuthorsOpenReview ↗

Response to Reviewer gDLD

We thank the reviewer for the careful reading of the theory. We revised Section 5 to separate what is classical from what is specific to our architecture, shortened Theorem 5.3 and stated its terms explicitly, and fixed the notation points.

W: Novelty and attribution of Theorems 5.1–5.3. The theorems justify the architecture rather than introduce new boundary-integral theory, and the revision now says so for each one.

  • Theorem 5.1 follows from classical mapping properties of boundary integral operators (Kress; McLean; Sauter & Schwab) in our parameterized convention, as now cited directly after Theorem 5.1. In two dimensions, log⁡∣X(s)−X(t)∣\log|\mathbf X(s)-\mathbf X(t)| differs from log⁡∣2sin⁡s−t2∣\log|2\sin\frac{s-t}{2}| by a smooth function for any regular simple closed curve, so subtracting 2W02\mathcal W_0 leaves a smoothing operator. This is the fact the decomposition uses.

  • Theorem 5.2 is a degenerate-kernel approximation built from a piecewise Taylor expansion of the kernel in ss; the rate r−mr^{-m} for kernels of class CmC^m in ss follows from standard degenerate-kernel estimates (Kress, 1989, Ch. 11), which we now note after Theorem 5.2. We use this construction rather than a Mercer expansion because RX\mathcal R_{\mathbf X} is symmetric but in general indefinite (already for ellipses it has eigenvalues of both signs), and its eigenfunctions in general change with the geometry. Theorem 5.2 is included because its form matches Eq. (17): the trunk functions depend only on the rank budget, the geometry enters only through the coefficients, and the constants are uniform over An\mathcal A_n.

  • Theorem 5.3 is, as the reviewer says, the approximation error combined with the standard uniform-deviation and optimization terms of empirical risk minimization. We shortened the main-text statement and made it explicit (Q4):

    • Θr,J\Theta_{r,J} is the class of networks of the form (17) with rank rr, JJ boundary nodes, fixed depth and width, and weights bounded by BB. Lpop\mathcal L_{\rm pop} and Lemp\mathcal L_{\rm emp} are the mean squared remainder errors under the data distribution μ\mu and over NN i.i.d. samples, and Λ^X=2W0+Nθ^\widehat\Lambda_{\mathbf X}=2\mathcal W_0+\mathcal N_{\widehat\theta}.
    • For bounded data, Rademacher-complexity bounds (Bartlett & Mendelson, 2002) give ΔN=O(N−1/2)\Delta_N=O(N^{-1/2}) with high probability, with a constant that grows with the size of the class.
    • If ∥f∥Cq≤F\lVert f\rVert_{C^q}\le F on the support of μ\mu, then Edet≤2π F(Erank+Egeom+Equad+Enet)E_{\mathrm{det}}\le \sqrt{2\pi}\,F\left(E_{\mathrm{rank}}+E_{\mathrm{geom}}+E_{\mathrm{quad}}+E_{\mathrm{net}}\right), with Erank=O(r−m)E_{\mathrm{rank}}=O(r^{-m}), Egeom=O(J−(n−ν))E_{\mathrm{geom}}=O(J^{-(n-\nu)}), Equad=O(J−q)E_{\mathrm{quad}}=O(J^{-q}) for fixed rr, and EnetE_{\mathrm{net}} linear in the trunk and branch approximation errors up to their product. Theorem C.8 gives EquadE_{\mathrm{quad}} and EnetE_{\mathrm{net}} in closed form.
    • Truncating 2W02\mathcal W_0 to ∣k∣≤K|k|\le K increases the bound by at most K1−σ(E∥f∥Hσ2)1/2K^{1-\sigma}(\mathbb E\|f\|_{H^\sigma}^2)^{1/2} for any σ≥1\sigma\ge1.

    The full statements and proofs are in Appendix C (Theorem C.9 for Theorem 5.3). We keep the short version in the main text because it ties the rank rr and the node count JJ to the error of the reconstructed DtN map.

Q1: Definition of W1\mathcal W_1. Eq. (9) now defines both parts of the splitting, with W1f(s)=12π∂s∫Tlog⁡∣X(s)−X(t)2sin⁡s−t2∣∂tf(t) dt.\mathcal W_1 f(s)=\frac{1}{2\pi}\partial_s\int_{\mathbb T}\log\left|\frac{\mathbf X(s)-\mathbf X(t)}{2\sin\frac{s-t}{2}}\right|\partial_t f(t)\,dt. Its kernel is smooth for smooth X\mathbf X (Lemma C.1).

Q2: Punctuation in Eq. (1). Eq. (1) now ends with a comma.

Q3: Rank sensitivity. Accuracy improves steadily with rank, in line with the role of rr in Theorem 5.2. We added a rank study on ellipses (Section 6.6, Appendix J.2, Table 15) with ranks 4 to 10 at width 144, the same data, and a common 10,000-epoch schedule.

  • From rank 4 to rank 10, the error falls from 0.85% to 0.53% (ID-fixed), from 1.02% to 0.62% (Dirichlet-OOD) and from 3.90% to 2.94% (Joint-OOD).
  • Storage grows linearly, from 1.39 MB to 2.29 MB, and even the rank-4 model, trained with this shorter schedule, is more accurate than every learned baseline in Table 1.

I am satisfied with the response and raised my score to 4. Good luck!

Reviewer_7VJAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on learning Dirichlet-to-Neumann (DtN) maps for 2D PDEs with smooth boundaries. The proposed method (PPDNO) analytically separates the singular principal part of the DtN operator and learns only the smoother, geometry-dependent remainder using a low-rank neural network. However, the experimental setting is highly simplified: it considers only linear, constant-coefficient PDEs, smooth geometries generated from predefined parametric families, analytically constructed boundary conditions, and fixed input discretizations. The main gap is real-world applicability and transferability.

Critical strengths
  1. The mathematical formulation is clear: separates the known singular structure from the part that needs to be learned, rather than approximating the entire operator from end to end.
  2. The method achieves good accuracy and generalization on the controlled 2D PDE benchmarks.
Critical weaknesses

I think the main limitation is that the paper focuses on a rather simplified setting, where the PDEs and their analytical structures are already well understood. I have two main concerns:

  1. Generalization to more complex problems: I'm not sure if the same decomposition would still work for more complex geometries or PDEs, especially with variable coefficients, nonlinearities, or multiphysics coupling. It's also unclear whether the remainder would still be smooth enough and low-rank.

  2. Practical relevance: The experiments are all based on synthetic geometries and analytically generated boundary data. Even the OOD tests are still within predefined function families. I'd like to see how the method performs on more realistic geometries and numerical simulation data, and whether it can actually speed up a practical PDE-solving workflow.

Questions for authors
  1. Would this approach still work for more complex PDEs and geometries? Would the remainder still be smooth and low-rank?

  2. How would the method perform on more realistic problems, with real geometries and numerical simulation data? Would it actually make PDE solving faster?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7VJA

We thank the reviewer for pressing on generality and practical relevance. The revision adds FEM benchmarks with corners and a variable coefficient, and below we explain why the decomposition extends beyond the original setting.

Q1: More complex PDEs and geometries. PPD relies only on the principal part of the DtN map, which remains explicit well beyond constant-coefficient Laplace. It therefore extends naturally to variable-coefficient equations, and the new FEM benchmarks show that it continues to improve accuracy on domains with corners and with spatially varying interior coefficients.

  • Variable coefficients. For −∇⋅(a∇u)=0-\nabla\cdot(a\nabla u)=0 with smooth a>0a>0, the DtN map is a first-order pseudodifferential operator with principal symbol a∣ξ∣a|\xi| on the boundary (Lee & Uhlmann, 1989). When a=1a=1 near the boundary, as in our benchmark, Λa−Λ1\Lambda_a-\Lambda_1 is smoothing, because the difference of the two solutions satisfies an equation whose source lies inside the domain. The remainder after subtracting 2W02\mathcal W_0 is then as smooth as in Theorem 5.1. When aa varies along the boundary, the principal part becomes a∣D∣a|D| and can be subtracted in the same way.
  • Evidence (Section 6.5). On 420 ellipses with a variable coefficient, PPDNO reaches 0.44% error against 17.12% for PPDNO-Direct. On 250 quadrilaterals with corners, it reaches 3.74% against 4.64%. PPD improves all three paired backbones in both problems.
  • Low rank. The remainder has rapid singular-value decay for Laplace, whereas for Helmholtz its numerical rank grows with kLkL (Appendix J.3). PPD itself is unaffected, and a more expressive backbone such as FNO can learn the remainder.
  • Nonlinear and coupled problems. For a nonlinear PDE the DtN map is nonlinear; PPD applies to its linearization, with a nonlinear backbone learning the remainder. Elliptic systems such as linear elasticity have explicit matrix-valued principal symbols, so the same subtraction applies. These are natural next steps.

Q2: Numerical data, realistic geometry and practical speed. The new FEM benchmarks are built entirely from numerical simulation data, and PPD improves every backbone on them. The speed advantage comes from handling a new geometry without assembly or a linear solve, which pays off when a workflow queries the DtN map for many geometries.

  • Numerical data. Dirichlet loads are random Fourier series, and reference fluxes come from quadratic FEM, checked under mesh refinement (Section 6.5, Appendix I). The quadrilateral domains keep their corners.
  • Speed. PPDNO evaluates about 1.5×1061.5\times10^6 samples/s at batch size 2,048 on one GPU, 27–40 times faster than PPD+FNO on the same hardware, with 14–18 times less GPU memory. The Nyström BIE reference runs at 430 samples/s on eight CPU threads, with assembly and a dense solve for every geometry (Table 1).
  • Use cases. Shape optimization, inverse boundary problems and domain decomposition with changing subdomains all evaluate the DtN map on many geometries.
  • Geometry. Because the model requires only boundary node coordinates, contours extracted from CAD or imaging can be used directly.

Input discretization. The output is resolution-independent, and the input grid samples smooth periodic functions that can be resampled from any resolution.

  • Output. The trunk network is continuous in ss. On 100 Functional Dirichlet-OOD ellipses, the error changes from 0.4161% to 0.4164% as the output grid is refined from 256 to 2,048 points (Section 6.4).
  • Input. The network takes geometry and Dirichlet data on 256 nodes; inputs sampled at other resolutions can be resampled to this grid by trigonometric interpolation, which is spectrally accurate for smooth boundaries and data, and the principal part 2W02\mathcal W_0 is applied by FFT at any resolution.

Thanks to the authors for the additional experiments. I think the variable-coefficient elliptic PDE experiment is the most valuable addition. Comparing the same backbones with and without PPD makes the contribution of PPD itself much clearer.

However, the coefficient in this experiment still satisfies a specific condition near the boundary. This is what allows the original decomposition to work and the remainder to stay smooth. So what the experiment shows is that PPD works when the coefficient varies in the interior under this assumption. This is still quite different from solving general variable-coefficient PDEs. The authors suggest that a modified principal part can be subtracted when the coefficient varies along the boundary. But would the remainder still be smooth and low-rank? I don't think this has been answered.

I'm also concerned about the low-rank assumption. The authors mention that the numerical rank grows for Helmholtz problems and suggest using FNO instead. To me, there are two separate questions here. Does PPD help neural operators learn better? And is the low-rank architecture of PPDNO a good choice for more complex problems? The new experiments make me more convinced about the first one, but not the second. In fact, I'm still not sure why we need the specific architecture of PPDNO.

Regarding realistic geometries, the authors say that contours from CAD or imaging can be used directly. But being able to take a CAD geometry as input is one thing. Being able to make accurate predictions on unseen CAD geometries is another. I don't see any experimental evidence for the latter.

The input discretization issue is also still there. The model still takes 256 input nodes. Interpolating data from other resolutions to these 256 nodes is not the same as being discretization-independent. Also, the authors discuss the accuracy of trigonometric interpolation for smooth boundaries and data. What about geometries with corners?

Overall, I think the new FEM experiments are useful and make the paper more complete. However, I'm still not convinced that the remainder will be smooth and low-rank for more general PDEs and geometries. Generalization to realistic geometries and different input discretizations is still unclear. We also haven't seen how much this method would help in an actual PDE-solving workflow.

I appreciate the additional experiments, and they do address some of my original concerns. But I don't see much improvement in the methodological novelty. My main concerns about the generality of the method, the low-rank assumption, and its practical value remain.

AuthorsOpenReview ↗

Response to Reviewer 7VJA (1/2)

Thank you for the follow up and for your positive assessment of the variable coefficient experiment. We address each of your points below.

1. PPD only requires the remainder to be more regular, not smooth and low rank (PPD vs. PPDNO).
Our primary contribution is principal part decomposition (PPD) as a general learning strategy across backbones: evaluate the universal singular part 2W02\mathcal W_0 (or −2W0-2\mathcal W_0 for exterior problems) exactly via FFT and train a neural operator only on the remainder RX\mathcal R_{\boldsymbol X} (Abstract; Section 1, Eq. 1; Section 4.1, Eqs. 12, 20; Section 4.4, Eqs. 21 and 22).

  • PPD does not require RX\mathcal R_{\boldsymbol X} to be C∞C^\infty or low rank. It only requires RX\mathcal R_{\boldsymbol X} to be more regular than the full DtN map of order one, so that high frequency amplification is removed from the learning target. In every paired comparison in the paper (DeepONet, FNO, Operator Transformer (OT), and our low rank network), adding PPD reduces test error in every benchmark and regime:
    • Interior Laplace on ellipses and Fourier boundaries (Sections 6.1 and 6.2, Tables 1 and 2, Figure 3; Appendix H, Table 10);
    • Exterior Helmholtz at k=1k=1 and k=10k=10 (Section 6.3, Table 3, Figure 3; Appendix H, Table 11);
    • Quadrilaterals with corners and elliptic PDEs with variable coefficients (Section 6.5, Figure 5);
    • Single mode input frequencies m=1,…,6m=1,\dots,6 (Section 6.6; Appendix J.1, Table 14). In particular, PPD consistently improves all paired backbones on quadrilaterals with corners, where Theorem 5.1 does not apply (DeepONet: 17.25%→12.84%17.25\% \to 12.84\%, FNO: 8.04%→6.03%8.04\% \to 6.03\%, low rank: 4.64%→3.74%4.64\% \to 3.74\%; Section 6.5, Figure 5), and on Helmholtz at k=10k=10, where the singular values of the remainder decay more slowly (FNO: 6.07%→4.51%6.07\% \to 4.51\%, low rank: 5.80%→5.10%5.80\% \to 5.10\% on Joint-OOD; Section 6.3, Table 3).
  • Why we introduce PPDNO. When the remainder is smooth and low rank (Theorems 5.1 and 5.2; Appendices C.1 to C.3), a heavy nonlinear operator network is unnecessary. PPDNO (Section 4.2, Eqs. 13 to 17, Figure 2) is designed for this regime: it preserves exact linearity in the Dirichlet data ff, achieves the lowest error on Laplace, k=1k=1 Helmholtz, and both FEM benchmarks (Tables 1 to 3, Figure 5), and runs 26.626.6 to 40.140.1 times faster than PPD+FNO while using 1414 to 1818 times less GPU memory (Sections 6.1 to 6.3, Tables 1 to 3; Appendix G.4). When the numerical rank of the remainder grows with kk (Section 6.6; Appendix J.3, Table 16), PPD can be paired with a more expressive backbone such as PPD+FNO (Section 6.3, Table 3).

2. Uniformity of the coefficient near the boundary.
A coefficient that is uniform near the boundary is standard in physical applications of DtN maps, such as inverse inclusion problems (for example in electrical impedance tomography), where one probes unknown interior inclusions from measurements on a known background boundary.

  • Our FEM benchmark with a variable coefficient directly models this setting: we solve −∇⋅(a(x)∇u)=0-\nabla\cdot(a(x)\nabla u)=0 on 420 ellipses with two interior inclusions supported in ∣x∣≤0.24|x|\le 0.24, while every ellipse contains ∣x∣≤0.5|x|\le 0.5 so that a(x)=1a(x)=1 near the boundary (Section 6.5, Figure 5; Appendix I.1). The interior inclusions strongly alter the boundary response (the FFT-only error is 26.93%26.93\%), and PPD reduces error across all backbones: DeepONet from 24.55%24.55\% to 5.03%5.03\%, FNO from 7.66%7.66\% to 4.73%4.73\%, and PPDNO-Direct from 17.12%17.12\% to 0.44%0.44\% for PPDNO (Section 6.5, Figure 5).
  • Even if a(x)a(x) varies smoothly along the boundary, the leading singular part remains explicit (a1/2∣X′∣−1(2W0)a1/2−12∂naa^{1/2}|X'|^{-1}(2\mathcal W_0)a^{1/2} - \tfrac12\partial_n a, requiring one FFT and pointwise scalings), and subtracting it still leaves a remainder of order −1-1 that is two orders smoother than the DtN map.
AuthorsOpenReview ↗

Response to Reviewer 7VJA (2/2)

3. Domains with corners, input discretization (N=256N=256), and boundary representations.

  • Geometries with corners: We already evaluate on polygonal domains with corners in Section 6.5 (Figure 5; Appendices I.1 to I.3, Tables 12 and 13), using 250 random convex quadrilaterals with sharp corners (no corner rounding or smoothing) solved by quadratic Lagrange FEM on unstructured triangular meshes. Both the weak flux reference labels and the network inputs use 64 midpoint nodes per straight edge (tj=(j+12)2π/256t_j = (j+\tfrac12)2\pi/256), avoiding any undefined pointwise normal at the corners (Appendices I.1 and I.2).
  • N=256N=256 is not a hard constraint: In Eqs. (14) to (17), N=256N=256 is simply the quadrature grid size used in our runs; the formulation accepts any node count JJ or quadrature rule (such as periodic trapezoidal quadrature on smooth curves in Section 6 or composite midpoint quadrature on polygon edges in Appendix I.1), and our error bound explicitly tracks the node count JJ (Theorem 5.3, Eq. 27; Assumptions C.6 and C.7; Theorem C.8, Eq. 120). On the output side, the trunk network MLPs(s)\mathrm{MLP}_s(s) (Eq. 17) is continuous in ss and the FFT principal part applies at any grid size: when refining the output grid from Nout=256N_{\mathrm{out}}=256 to 512,1024,2048512, 1024, 2048 without retraining, PPDNO's error remains virtually unchanged at 0.4161%0.4161\% to 0.4164%0.4164\% (Section 6.4; Appendix G.5, Table 9).
  • Unseen geometries and boundary representations (CAD shapes): All reported errors are evaluated on unseen test geometries (10,00010{,}000 unseen geometries per profile in Sections 6.1 to 6.3, with strictly disjoint parameter ranges in Joint-OOD, Appendices G.1 and G.2; and an 80%/20% split by geometry in the FEM benchmarks, Section 6.5, Appendix I.2). Across the paper, we already test five distinct boundary representations:
    1. Analytic ellipses (Section 6.1, Table 1; Appendix G.2);
    2. Trigonometric polynomial (Fourier) curves (Section 6.2, Eq. 28, Table 2; Appendix G.2, Figure 6);
    3. Freeform periodic Gaussian RBF spline curves interpolated through 5 to 8 random control points with random scaling, rotation, and translation (Section 6.3, Table 3; Appendix G.2, Eqs. 169 to 172, Figure 7);
    4. Polygons with corners on FEM meshes (Section 6.5, Figure 5; Appendix I.1);
    5. Elliptical boundaries enclosing interior inclusions with variable coefficients (Section 6.5, Figure 5; Appendix I.1). While applying PPD to CAD datasets for specific applications is an interesting future direction, introducing a separate CAD preprocessing pipeline here could obscure the controlled comparisons across these five boundary representations.

We thank the reviewer again for the constructive discussion.

Reviewer_xSBTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work aims to solve the problem of learning Dirichlet-to-Neumann (DtN) operators across different geometric structures. The proposed Principal-Part Decomposed Neural Operator (PPDNO) decomposes the DtN into a geometric-independent, accurately computable singular principal part, and a geometric-dependent smooth remainder based on the boundary integral theory. For the remainder, a geometry-conditioned low-rank neural operator is designed, which reduces the learning complexity while retaining the linearity of the DtN with respect to Dirichlet boundary data. Additionally, the paper provides theoretical analyses of the smoothing property of the remainder, low-rank approximation, and reconstruction operator error. PPDNO outperforms the compared neural operators and their principal-part variants in multiple experimental settings. Ablation experiments verify the effectiveness of the principal-part decomposition. However, the validation of PPDNO still has some key issues.

Critical strengths
  1. Motivated by the known singular structure of the DtN operator, this work separates the analytically computable principal part from the smooth remainder that needs to be learned. This structured learning paradigm has a clear mathematical basis. Theorems 5.1–5.3 establish the smoothness, low-rank approximability, and conditional error bounds of the remainder.

  2. The current manuscript compares DeepONet, FNO, and Operator Transformer, and also introduces principal-part residualized variants of the same backbone network, PPDNO-Direct, and FFT-only. In most Laplace and Helmholtz tests, PPDNO exhibits lower prediction errors, better out-of-distribution generalization, and advantages in efficiency.

Critical weaknesses
  1. The experiments lack a systematic analysis of the Dirichlet input frequency. It is not clear to what extent the performance improvement of principal-part decomposition is due to the alleviation of the difficulty in modeling the high-frequency response. For the low-rank approximation of smooth remainder, the paper should analyze the impact of different ranks on accuracy and efficiency.

  2. The current manuscript only cover two-dimensional smooth boundaries and artificially constructed analytic function families. Besides, at Helmholtz (k=10), the advantage of PPDNO weakens, especially the Joint-OOD error is 5.095%, higher than 4.514% of PPD+FNO. This indicates that the current low-rank architecture may face limitations in expression ability or generalization. The current manuscript lacks more systematic wave number expansion experiments to characterize this performance boundary. The main experiments still focus on general Neural Operator baselines. The lack of direct comparison with more targeted boundary integral learning methods makes the superiority of PPDNO relative to related domain methods less clear.

Questions for authors
  1. Have the authors analyzed the prediction errors under different Dirichlet input Fourier frequencies?
  2. In Appendix E, different tasks used different ranks. How were these ranks determined?
  3. When the Helmholtz wavenumber increased from (k=1) to (k=10), why did the relative advantage of PPDNO significantly weaken?
3 comments
AuthorsOpenReview ↗

Response to Reviewer xSBT (1/2)

We thank the reviewer for the detailed reading and for recognizing the decomposition and the paired comparisons.

W1/Q1: Dirichlet input frequency. The gain from PPD depends on frequency. At high frequencies the DtN response is almost entirely the principal part, which PPD computes exactly instead of learning; at low frequencies, where the remainder is significant, the gain comes from learning that smooth remainder.

We added Appendix J.1 (Table 14, summarized in Section 6.6) to show this. All methods are evaluated on 128 ID ellipses with single-mode inputs 2sin⁡(mt)\sqrt2\sin(mt) and 2cos⁡(mt)\sqrt2\cos(mt), m=1,…,6m=1,\ldots,6, against exact references. PPDNO, PPDNO-Direct and FFT-only are linear in ff, so their relative errors do not depend on amplitude and vary only with frequency across m=1,…,6m=1,\ldots,6.

  • The FFT-only error equals the relative size of the remainder: 37.5% at m=1m=1 and 0.18% at m=6m=6.
  • Without subtraction, the same low-rank network (PPDNO-Direct) degrades from 0.74% to 14.2%. PPDNO stays between 0.20% and 1.19%, 25 times lower at m=6m=6.
  • At m=1m=1, the learned remainder lowers the error from 37.5% (FFT-only) to 0.20%.
  • For m≥4m\ge4, PPDNO is slightly above FFT-only (by at most 0.38 percentage points), likely because these modes are weakly represented in the training traces.

W1/Q2: Rank sensitivity and selection. Accuracy improves steadily with rank at a modest, linear cost in storage, and PPDNO's advantage does not hinge on this choice: even at rank 4 it outperforms all learned baselines. The original ranks were set by a storage budget: each rank, together with the width, keeps PPDNO no larger than the FNO baseline for that task.

We added a rank study on ellipses (Appendix J.2, Table 15) with ranks 4, 6, 8 and 10, width 144, the same data, and a common 10,000-epoch schedule.

  • From rank 4 to rank 10, the error falls from 0.85% to 0.53% (ID-fixed), from 1.02% to 0.62% (Dirichlet-OOD) and from 3.90% to 2.94% (Joint-OOD).
  • Each extra rank adds one output block to the geometry network (about 0.15 MB), so storage grows from 1.39 MB to 2.29 MB.
  • The rank-4 model is about half the size of the original model, yet even with this shorter schedule it is more accurate than every learned baseline in Table 1 on all three test sets.

Appendix E now states how the ranks were chosen: in every task, PPDNO's stored size is 84–99% of FNO's, counting complex spectral weights at their actual storage size.

W2: Smooth geometries and analytic data. The original benchmarks use smooth boundaries and analytic solution families. The revision adds two FEM benchmarks that drop both assumptions, and subtracting the principal part still improves every backbone on them (Section 6.5, Appendix I).

  • Laplace on 250 random quadrilaterals, whose corners make the boundary non-smooth, and −∇⋅(a∇u)=0-\nabla\cdot(a\nabla u)=0 on 420 ellipses with a spatially varying coefficient a(x)a(x).
  • The Dirichlet data are random Fourier series rather than traces of known solutions, and the reference fluxes come from quadratic FEM, checked under mesh refinement.
  • PPD improves all three paired backbones in both problems, and PPDNO is the most accurate method: 3.74% vs 4.64% for PPDNO-Direct on quadrilaterals, and 0.44% vs 17.12% with the variable coefficient (Figure 5, evaluated on validation loads scaled by four; Appendix I.3).

The quadrilateral benchmark shows that PPD remains effective on domains with corners, where the remainder is no longer smooth. In three dimensions the principal symbol is still ∣ξ∣|\xi|, so the same decomposition applies, with a surface spectral method in place of the FFT.

AuthorsOpenReview ↗

Response to Reviewer xSBT (2/2)

W2/Q3: Helmholtz wavenumber. The k=10k=10 Joint-OOD result (PPD+FNO 4.51%, PPDNO 5.10%) compares two models that both use PPD, so the difference comes from the remainder network rather than the decomposition itself. As kk grows, the singular values of the remainder decay more slowly, so the more expressive FNO backbone fits it better than our rank-32 network. In both cases PPD improves its backbone: FNO drops from 6.07% to 4.51%, and PPDNO-Direct from 5.80% to 5.10%. This gain across backbones is the main contribution of our work, and the abstract, introduction, Section 6.3 and conclusion now state it this way.

We added a wavenumber study and a singular-value analysis in Appendix J.3.

  • PPDNO with rank 32 is trained at k=5,10,15,20k=5,10,15,20, with grid resolution checked at each kk (Table 16). The ID-fixed error is 1.5%, 2.0%, 7.1% and 19.7%, and the Joint-OOD error is 4.5%, 4.9%, 14.1% and 29.5%.
  • On the unit circle, where the exterior DtN map is known exactly, the remainder has norm close to kk and is concentrated on the roughly 2k+12k+1 propagating modes. Its best rank-32 approximation leaves a relative error of 0.39, 0.80 and 0.96 at k=10,15,20k=10,15,20. Our boundaries have radius close to one, so a rank-32 model should lose accuracy between k=10k=10 and k=15k=15, which is where Table 16 degrades.

PPD removes the singularity exactly at every kk. What must scale with kLkL is the capacity of the remainder model, and PPD leaves the choice of that model open.

W2: Boundary-integral learning methods. PPD can be paired with any model that predicts the Neumann trace: the principal part 2W02\mathcal W_0 is evaluated exactly, and the network learns only the remainder RX\mathcal R_{\mathbf X}. We therefore tested it across backbones. In every paired comparison in Tables 1–3 and Figure 5, across DeepONet, FNO, the operator Transformer and our low-rank network, adding PPD lowers the mean error. Table 1 also reports a classical Nyström boundary integral solver.

The closest learned methods are:

  • MAD-BNO (Wu et al., 2026), which learns Dirichlet–Neumann boundary maps from fundamental-solution data;
  • M-PCNO (Han et al., 2026), which learns singular kernel integrals on varying geometries, including an exterior Neumann-to-Dirichlet map;
  • BI-DeepONet and BI-TDONet (Meng et al., 2024), which learn boundary-integral densities.

M-PCNO builds a near/far-field split into each trained layer, whereas PPD removes the principal part from the target before training. A boundary-map learner such as MAD-BNO could be trained on RX\mathcal R_{\mathbf X} in the same way. We now cite these works in the Related Work section.

Thank you for the response, which has addressed most of my concerns. I will consider these improvements and determine my final score during the subsequent discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their careful and constructive reviews. The main changes are:

  • Positioning. We now present principal-part decomposition (PPD) as the main contribution: a backbone-independent strategy that evaluates the singular part of the DtN map exactly and learns only the smooth remainder. PPDNO is its linear low-rank realization. In every paired comparison (DeepONet, FNO, operator Transformer and our low-rank network, on Laplace, Helmholtz and FEM benchmarks), adding PPD lowers the error.
  • New FEM benchmarks (Section 6.5, Appendix I). We add interior Laplace on random quadrilaterals with corners and a variable-coefficient elliptic problem, both using FEM reference fluxes and random Fourier Dirichlet data.
  • Sensitivity studies (Section 6.6, Appendix J). We study sensitivity to input frequency, rank, and Helmholtz wavenumber up to k=20k=20, with a singular-value analysis showing how the numerical rank of the remainder grows with kk.
  • Theory (Section 5). We clarify the attribution to classical results, define W1\mathcal W_1 explicitly in Eq. (9), and make all terms in Theorem 5.3 explicit.
  • Related work. We discuss recent learned boundary-integral and boundary-to-boundary methods (Meng et al., 2024; Wu et al., 2026; Han et al., 2026).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.