acceptodds
Under review as a conference paper at ICLR 2027

Gaussian Process Priors for Boundary Value Problems of Linear Partial Differential Equations

Abstract

We propose Boundary Ehrenpreis–Palamodov Gaussian Processes (B-EPGPs), a probabilistic framework for constructing Gaussian Processes (GP) priors for linear constant-coefficient Partial Differential Equations (PDE) with linear boundary conditions that can be conditioned on a finite data set. Starting from the Ehrenpreis–Palamodov representation, we learn the free parameters from data and enforce boundary conditions analytically for piecewise-flat boundaries. This yields priors and posteriors whose sample paths strictly satisfy both PDE and boundary condition by construction. We provide constructive examples, formal correctness proofs, and experiments showing improved accuracy and reduced runtime/memory compared to existing PDE baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Gaussian Processes, Partial Differential Equations

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rRvkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The manuscript proposes an extension of GPs for linear PDEs (without approximations via collocation points) to boundary conditions. Harkonen et al. proposed the original method (EPGPs) but did not extend it to boundary conditions. This paper closes this gap and is a useful follow-up.

Critical strengths

Scope: This paper is a useful extension in the field of GP-based PDE solvers. It fits right into the current scope for GPs. Physics-informed probabilistic ML is a hot topic, and boundary conditions are part of this movement.

Readability: The paper is clearly written and easy to follow. The contributions are easy to identify. I also enjoyed reading about previous work in the field. It's a nice overview.

Methodology: Most importantly, the methodology is a clever extension of Harkonen et al.'s work.

Critical weaknesses

Theorems and proofs: Theorem 1 makes several choices that are difficult to follow or verify. I don't understand why "we suppose" what we suppose there. How are the choices motivated? After Theorem 1, several assertions follow that need support. Why are approximations in the Fréchet topology strong? Why do we expect Theorem 1 to hold for any BVP?

Examples: I find it difficult to follow what I am supposed to learn from the examples; they don't seem to be on the same footing, which makes it hard to compare apples to apples. I would expect 3 or 4 well-defined examples, all applied to the proposed method and all competitors, and a table with results (L1, RMSE for deterministic methods, NLPD, PICP, CRPS for probabilistic methods) Some of this is present in the Appendix but should be moved to the main text.

Questions for authors

How are choices in Theorem 1 motivated?

Why are approximations in the Fréchet topology strong?

Why do we expect Theorem 1 to hold for any BVP?

How does the GP compare against other tools in RMSE, and against other probabilistic tools in CRPS, NLPD, and PICP?

How does the method behave when caustics occur in the wave equations? Given that sources can be diverse, caustics can easily happen.

2 comments

Theorems and proofs: Theorem 1 makes several choices that are difficult to follow or verify. I don't understand why "we suppose" what we suppose there. How are the choices motivated? After Theorem 1, several assertions follow that need support. Why are approximations in the Fréchet topology strong? Why do we expect Theorem 1 to hold for any BVP?

The topology in Theorem 1 is that of C∞C^\infty: the approximation and all its derivatives converge uniformly on compact sets. This is stronger than, and implies, L2L^2 convergence on compact sets, a norm that controls only the function itself. For rough solutions, where the L2L^2 setting is natural, Theorem 1 makes no claim. Theorem 1 is stated for smooth solutions on the half-space with Dirichlet/Neumann conditions. We expect the statement to hold in much greater generality, supported by Section 6 and Appendix A.3, but we have no proof and will state it as a conjecture.

Examples: I find it difficult to follow what I am supposed to learn from the examples; they don't seem to be on the same footing, which makes it hard to compare apples to apples. I would expect 3 or 4 well-defined examples, all applied to the proposed method and all competitors, and a table with results (L1, RMSE for deterministic methods, NLPD, PICP, CRPS for probabilistic methods) Some of this is present in the Appendix but should be moved to the main text.

We agree and will move a unified table for the Wave IBVP (relative L1 error/RMSE for all methods, NLPD, PICP and CRPS for the probabilistic ones) into the main text. We have two kinds of examples: (i) reconstruction from data, plentiful (Section 4) or sparse (Section 6); (ii) well-posed PDE problems (Section 5). We use (i) to demonstrate the generality of our method (see also Appendix A.3) and (ii) for comparison. In Section 5.1 the other methods are prohibitively expensive for the 4D problem, so the comparison with all six other methods is made on the 2D problem (Wave IBVP).

How does the GP compare against other tools in RMSE, and against other probabilistic tools in CRPS, NLPD, and PICP?

We have conducted experiments to evaluate the uncertainty performance. The table below will be added to the manuscript. Relative L1 errors for all methods are given in Table 2. The uncertainty metrics are:

AlgorithmPICP (50%)PICP (90%)PICP (95%)NLPDCRPS
CNON/AN/AN/AN/AN/A
FNON/AN/AN/AN/AN/A
WIGPR44.2%83.9%89.5%-3.720.00142
EPGP46.5%86.1%91.2%-4.180.00095
B-EPGP (ours)51.2%90.8%95.3%-5.120.00041
BCGP48.1%87.4%92.6%-4.450.00078
HC-PINNsN/AN/AN/AN/AN/A

How does the method behave when caustics occur in the wave equations? Given that sources can be diverse, caustics can easily happen.

In our example the ground truth is a focusing spherical pulse. It converges to the point (0,0,1.2)(0,0,1.2) at t=1.2t = 1.2 and diverges afterwards, and it vanishes on the plane z=0z = 0 at all times (the second term is a mirror source). The solution is smooth and finite at the focus. This tests a focusing wave, not a singular caustic. With 4,800 basis elements and 50,000 data points B-EPGP reaches below 1% relative validation RMSE. We ran the B-EPGP algorithm with basis sin⁡(z3z)cos⁡(z0t+z1x+z2y),sin⁡(z3z)sin⁡(z0t+z1x+z2y)\sin(z_3 z)\cos(z_0 t+z_1x+z_2y),\sin(z_3 z)\sin(z_0 t+z_1x+z_2y) where z02=z12+z22+z32z_0^2=z_1^2+z_2^2+z_3^2. This is similar to examples from the paper.

Here is the ground truth we used u(t,x,y,z)=0.1r1[g(τ+r1)−g(τ−r1)]−0.1r2[g(τ+r2)−g(τ−r2)],u(t,x,y,z)=\frac{0.1}{r_1}\Big[g(\tau+r_1)-g(\tau-r_1)\Big]-\frac{0.1}{r_2}\Big[g(\tau+r_2)-g(\tau-r_2)\Big], where τ=t−1.2,r1=x2+y2+(z−1.2)2,r2=x2+y2+(z+1.2)2,\tau=t-1.2,\qquad r_1=\sqrt{x^2+y^2+(z-1.2)^2},\qquad r_2=\sqrt{x^2+y^2+(z+1.2)^2}, g(s)=e−1.778s2sin⁡(5s).g(s)=e^{-1.778 s^2}\sin(5s).

I thank the authors for their thorough reply and will consider adjusting my score accordingly.

Reviewer_ygpyOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new method for constructing GPs when modelling linear PDEs with linear B.Cs. By building kernels from specifically chosen basis functions, they can assure that the resulting GP satisfies the PDE and B.C exactly. The proposed method is shown to achieve low test error across a range of linear constrained physics problems.

Critical strengths
  1. The paper is well written, clear and self-contained.
  2. The mathematics appears correct and well justified, including the universal approximation result (Theorem 1).
  3. The literature review is broad and comprehensive, apart from the key omissions noted below.
Critical weaknesses
  1. Limited novelty. The core idea of restricting the EPGP feature space so each basis function satisfies the boundary condition (essentially via reflection) feels too incremental for a top-tier ML venue.
  2. Missing key related work. Two highly relevant references are not discussed: Solin & Kok (AISTATS 2019, https://proceedings.mlr.press/v89/solin19a/solin19a.pdf) and Solin & Särkkä (Statistics and Computing 2020, https://link.springer.com/article/10.1007/s11222-019-09886-w). Both build boundary-respecting GP priors via Hilbert-space reduced-rank basis expansions, closely resembling the rectangular and heat experiments, but they handle arbitrarily shaped domains rather than only polyhedral ones. The authors should discuss and ideally compare against them.
  3. No uncertainty quantification. Despite the GP framing, only predictive error is reported. There is no evaluation of calibration, predictive log-likelihood or posterior uncertainty, which is the main motivation for using a GP.
  4. Overfitting risk and unclear hyperparameters. Optimising many basis-function frequencies by maximum marginal likelihood can overfit, and it is not stated clearly which hyperparameters (frequencies, weight variances, noise) are optimised in each experiment. Weak experimental evidence. The test problems are relatively easy, and key baselines are missing. EPGP is absent from Table 1, even though it is the most informative comparison given how close the two methods are. For Helmholtz, error does not decrease with more data (0.211% at 10 points versus 0.288% at 50), no error bars or repeated runs are reported, and the substantial existing literature on Moreover, Helmholtz GPs is not compared against.
  5. Unclear baseline setup. CNO and FNO are operator-learning methods normally trained on many input–output function pairs. More detail is needed on how they were applied to single-solution regression and whether the comparison is fair.
Questions for authors
  1. Please can you elaborate + test against the work of Solin described above, as well as something from the existing Helmholtz GP literature?
  2. Please can you explore the uncertainty quantification side of this work. In my mind (and according to your into), this is the primary benefit of using GPs?
  3. Please could you elaborate exactly on how/which kernel parameters were estimated for each experiments and how you performed CNO and FNO?
3 comments

Limited novelty. The core idea of restricting the EPGP feature space so each basis function satisfies the boundary condition (essentially via reflection) feels too incremental for a top-tier ML venue.

We respect the Referee's opinion. We agree that, for scalar operators of the form A0(∂′)+∂n2A_0(\partial') + \partial_n^2 with Dirichlet or Neumann conditions, the B-EPGP basis is an even or odd reflection in the normal variable (s↦−ss \mapsto -s in Appendix C.2), with tangential frequencies that are learned. In this case our basis is the method of images, built into a GP prior. Robin conditions, as in Section 6, give the non-reflection combination with coefficients β−αs\beta - \alpha s and −(β+αs)-(\beta + \alpha s). Our contribution is the construction of GP priors whose samples satisfy PDE and boundary condition exactly, with trained frequencies, a closed-form posterior, and uncertainty quantification. For systems, the compatible combinations are not reflections. In Appendix A.3 (Schrödinger equation, curl system with tangential conditions, reduced Maxwell system with perfect electric conductor conditions) we compute them on a half-space, and Appendix C.2 shows how this is a syzygy computation over the boundary characteristic variety.

Missing key related work. Two highly relevant references are not discussed: Solin & Kok (AISTATS 2019, https://proceedings.mlr.press/v89/solin19a/solin19a.pdf) and Solin & Särkkä (Statistics and Computing 2020, https://link.springer.com/article/10.1007/s11222-019-09886-w). Both build boundary-respecting GP priors via Hilbert-space reduced-rank basis expansions, closely resembling the rectangular and heat experiments, but they handle arbitrarily shaped domains rather than only polyhedral ones. The authors should discuss and ideally compare against them.

We thank the reviewer for these two references and agree that they are related. We will discuss them in the camera ready version and add the comparison below to the paper. Both works handle general domains, while we currently treat piecewise-flat boundaries. We state this limitation clearly. The two approaches differ in several points: (i) The technical constructions are different. (ii) Both papers start from stationary covariance functions. Ehrenpreis–Palamodov features, e.g. exp⁡(iωx−ω2t)\exp(i \omega x− \omega^2t) for the heat equation, are non-stationary already before boundary conditions are imposed. (iii) Neither paper constructs a prior that satisfies a PDE; a PDE can only enter as a soft constraint. (iv) Our features form a basis of the solution set of PDE and boundary conditions, whereas the two papers approximate a basis of a different solution set (boundary conditions only) with high accuracy.

We implemented a baseline built on these papers (boundary-exact Hilbert-space GP with the PDE residual as noise-free observations at collocation points) for the Heat IBVP of Section 5.2 on [0,1]2[0,1]^2. Its spatial basis consists of the Laplace eigenfunctions (sines at integer multiples of π), with the same number of spatial frequencies as B-EPGP. In addition it needs temporal frequencies; B-EPGP does not, because its temporal frequency is determined by the spatial ones. With 10410^4 and 10510^5 collocation points the error stagnates at about 0.38%. B-EPGP reaches 0.25% without any collocation point. The baseline needs a nonlinear hyperparameter optimisation and is therefore slower by roughly three orders of magnitude (16–31 s against 0.013 s).

MethodColloc. PtsRel. RMSEPDE Res. RMSETraining time
baseline1,0003.2548%1.5139e-0128.83 s
baseline10,0000.3788%1.5086e-0315.93 s
baseline100,0000.3792%7.8553e-0430.57 s
B-EPGP (ours)00.2502%00.0128 s

No uncertainty quantification. Despite the GP framing, only predictive error is reported. There is no evaluation of calibration, predictive log-likelihood or posterior uncertainty, which is the main motivation for using a GP.

We have conducted uncertainty quantification experiments (Wave IBVP) and will add them to the manuscript. PICP is taken at nominal levels 50/90/95%; lower is better for NLPD and CRPS. CNO, FNO and HC-PINNs are deterministic and therefore have no entries.

AlgorithmPICP (50%)PICP (90%)PICP (95%)NLPDCRPS
CNON/AN/AN/AN/AN/A
FNON/AN/AN/AN/AN/A
WIGPR44.2%83.9%89.5%-3.720.00142
EPGP46.5%86.1%91.2%-4.180.00095
B-EPGP (ours)51.2%90.8%95.3%-5.120.00041
BCGP48.1%87.4%92.6%-4.450.00078
HC-PINNsN/AN/AN/AN/AN/A

Overfitting risk and unclear hyperparameters. Optimising many basis-function frequencies by maximum marginal likelihood can overfit, and it is not stated clearly which hyperparameters (frequencies, weight variances, noise) are optimised in each experiment. Weak experimental evidence. The test problems are relatively easy, and key baselines are missing. EPGP is absent from Table 1, even though it is the most informative comparison given how close the two methods are. For Helmholtz, error does not decrease with more data (0.211% at 10 points versus 0.288% at 50), no error bars or repeated runs are reported, and the substantial existing literature on Moreover, Helmholtz GPs is not compared against.

We will tackle the questions one by one:

  • While we observe overfitting at N=3 data points (Appendix Table 5), we did not observe overfitting in more reasonable experiments. The above table about uncertainty quantification supports this point on held-out data.
  • In almost all experiments we optimize frequencies, weight variances and the noise variance, as is clearly stated in the formula for our loss function LL, just under (B-EPGP) formula. The details are outlined in Appendix C.3. The only exceptions are the box and triangular domains (Section 5.2 and Appendix A.1.1-2), where the frequencies are discrete, so we fix them and fit only the weights by least squares.
  • We agree that about half the experiments are "relatively easy", as indeed the polynomial solutions we fit are meant to demonstrate that the method works really well for very different equations and boundary conditions. The other half of the experiments involve fitting Gaussian pulses using sums of sines and cosines. This is computationally challenging as the Gaussian pulses have numerically negligible values in large parts of the computational domain.
  • Below is a comparison for Table 1 with EPGP, which we will include in the camera-ready version. We benchmarked all 3 Dirichlet BC ground truths in the following manner: we use the same data and record the time until the prediction reaches 0.5% relative RMSE at interior points and 1e-5 absolute RMSE on the boundary.
Algorithm∂1−∂22−∂32\partial_1-\partial_2^2-\partial_3^2∂12−∂22−∂32\partial_1^2-\partial_2^2-\partial_3^2∂12+∂22+∂32\partial_1^2+\partial_2^2+\partial_3^2
EPGP28.870s23.707s49.895s
B-EPGP (ours)2.243s1.014s5.539s

To reach this boundary accuracy EPGP needs 9 to 23 times more time than B-EPGP (28.9 s against 2.2 s, 23.7 s against 1.0 s, 49.9 s against 5.5 s).

  • Here is a table including error bars for 10 repeated runs for the Helmholtz example in Section 6:
Data pointsMean time/run (s)Mean rel. RMSEStd. dev.
1015.142.4668e-036.0625e-04
2015.012.3146e-032.4356e-04
3014.972.1694e-032.9521e-04
4015.101.6549e-032.4918e-04
5015.161.5756e-032.2655e-04

Mean over 10 runs, 1000 optimisation steps, best validation error over all steps. The error decreases from 0.25% at 10 points to 0.16% at 50 points.

  • A comparison with PDE solvers or PINNs is not meaningful here, since the problem is underdetermined (only one face carries a Robin condition) and such methods need collocation points in the 3D domain. The closest comparison is EPGP, which we ran on the same 50 data points with the boundary condition supplied as data: relative interior RMSE 1.48% and boundary RMSE 6.6e-2. With the B-EPGP basis we obtain 0.19% (single run; mean over 10 runs 0.16%) and exactly 0 on the boundary, i.e. about eight times lower interior error.

We will include all these results in the camera-ready version.

Unclear baseline setup. CNO and FNO are operator-learning methods normally trained on many input–output function pairs. More detail is needed on how they were applied to single-solution regression and whether the comparison is fair.

We agree that we need to make more clarification to justify that we are not comparing apples to oranges. For CNO, we use the default settings from the code in the repository provided by the corresponding paper, which is, learning_rate = 0.001, epochs = 50, step_size = 15, gamma = 0.5, N_layers = 4 (number of up/downsampling blocks), N_res = 4 (number of residual blocks in the middle networks), N_res_neck = 4 (number of residual blocks in the bottleneck), channel_multiplier = 16 (regulates the number of channels in the network). For FNO, we apply the codes provided in the same repository (which was also used for their comparison). We will include this in the camera-ready version.

Thank you for your response. I believe the paper has improved due to this rebuttal, but the limited scope of the contribution remains and so I will not be raising my score.

Reviewer_L2asOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses a limitation of Ehrenpreis–Palamodov Gaussian Processes (EPGPs), whose realizations satisfy linear constant-coefficient PDEs exactly but do not generally enforce prescribed boundary conditions. The proposed B-EPGP framework restricts the characteristic exponential representation to boundary-compatible subspaces. GP priors and posteriors are then defined on this basis, with weights marginalized analytically and frequencies and variance hyperparameters estimated by marginal likelihood. The paper gives explicit constructions for elliptic, parabolic, and hyperbolic PDEs, including Dirichlet, Neumann, and Robin conditions, and compares B-EPGP with neural operators, GP-based methods, hard-constrained PINNs, and a finite-difference solver, including sparse-data and higher-dimensional settings. It also proves a density result for a class of half-space Dirichlet/Neumann problems under a structural assumption on the differential operator.

Critical strengths
  1. The problem formulation and motivation are clear, and the proposed construction is mathematically clean and meaningful. B-EPGP incorporates both the governing PDE and homogeneous linear boundary conditions directly into the prior, yielding zero PDE and boundary residuals by construction. This is important because most of the GP, neural-operator, and PINN baselines considered in the paper enforce at most one of these constraints exactly, and the experiments suggest that simultaneous exact enforcement can improve predictive accuracy and data efficiency.
  2. The empirical evaluation is reasonably comprehensive and includes several challenging settings. The sparse-data 3D Helmholtz reconstruction and the four-dimensional wave example are valuable demonstrations of the method beyond simple low-dimensional test problems.
Critical weaknesses
  1. The advantages reported in the sparse-data experiments and in the comparison against the numerical solver are somewhat overstated and lack in-depth analysis.
  2. The theoretical guarantee is established under relatively structured settings, and its extension to the broader constructions demonstrated is not yet guaranteed by the present analysis
Questions for authors
  1. How sensitive is B-EPGP to the finite number of basis functions and to the optimization of the free frequencies? Since the universal-approximation result is asymptotic in representation capacity, it would be useful to understand whether the practical performance reported in the paper is robust to these implementation choices.
  2. For general polyhedral domains whose faces are not aligned with the coordinate axes, how are B-EPGP basis functions constructed systematically, and how does the computational cost scale with the number and arrangement of boundary faces?
  3. EPGP itself does not enforce the boundary condition analytically. In the Wave IBVP comparison, how exactly was boundary information supplied to the EPGP baseline, and was this configuration matched to the information available to B-EPGP? Clarifying this would help assess the fairness of the comparison.
2 comments

The advantages reported in the sparse-data experiments and in the comparison against the numerical solver are somewhat overstated and lack in-depth analysis.

We agree that the comparison with deterministic solvers needs care, but we do not see it as a critical issue: our main comparisons are with probabilistic methods, and deterministic methods appear in only a few experiments. Benchmarking probabilistic against deterministic methods is difficult in any case. Section 6 addresses a different task from FEM or PINNs. The problem stated there is underdetermined: only one face carries a Robin condition, so no deterministic solver can be applied, and the observations select a solution. Collocation-based methods such as PINNs would need many residual points for the PDE and the boundary condition on top of the observations; B-EPGP needs none, since both enter the prior. For the well-posed Wave IBVP of Section 5, where HC-PINNs can be applied, their error decreases with the number of PDE soft constraints but stays above that of B-EPGP, which needs none:

MethodPDE Soft ConstraintsRelative L1L_1 Error (%)
HC-PINN108.20 ± 1.47%
HC-PINN1001.12 ± 0.18%
HC-PINN1,0000.43 ± 0.07%
HC-PINN10,0000.18 ± 0.02%
B-EPGP (Ours)N/A0.06 ± 0.01%

We therefore present Section 6 as a proof of concept, not as a benchmark, and we will tone down the wording and complete Appendix B.4. The method is not guaranteed to work for arbitrary few data: with N = 3 training fails (Table 5).

The theoretical guarantee is established under relatively structured settings, and its extension to the broader constructions demonstrated is not yet guaranteed by the present analysis

We agree that Theorem 1 covers a restricted class of problems. We chose it because it contains most examples of the paper and has a short, self-contained proof. The construction of the B-EPGP basis does not depend on the theorem. It is explicit for linear constant-coefficient PDE systems and boundary conditions on a half-space (Appendix C.2), and for boxes by successive restriction (Appendix A.1.1). The experiments in Section 6 (Robin condition) and Appendices A.3.2 and A.3.3 (systems) test it beyond the setting of the theorem. For the general case we will state density as a conjecture, supported by these experiments, and say so clearly in the paper.

How sensitive is B-EPGP to the finite number of basis functions and to the optimization of the free frequencies? Since the universal-approximation result is asymptotic in representation capacity, it would be useful to understand whether the practical performance reported in the paper is robust to these implementation choices.

We have conducted experiments in 1-d wave to compare the relationship of basis number and predict performance between B-EPGP and EPGP, shown in the table below, we will include this in an updated manuscript:

Number of Basis ElementsAbsolute L1 Error (B-EPGP)Absolute L1 Error (EPGP)
1005.88 × 10010^06.09 × 10010^{0}
2004.24 × 10010^06.78 × 10010^{0}
3006.31 × 10010^08.37 × 10010^{0}
4003.68 × 10−410^{-4}4.35 × 10010^{0}
5003.11 × 10−410^{-4}1.12 × 10−310^{-3}
6002.09 × 10−410^{-4}3.02 × 10−310^{-3}
7001.13 × 10−410^{-4}1.80 × 10−310^{-3}
8006.55 × 10−510^{-5}5.27 × 10−410^{-4}
9003.54 × 10−510^{-5}5.54 × 10−410^{-4}
10007.67 × 10−510^{-5}2.04 × 10−410^{-4}

B-EPGP fits from 400 elements, EPGP from 500. From there EPGP’s error is about 3 to 16 times larger.

For general polyhedral domains whose faces are not aligned with the coordinate axes, how are B-EPGP basis functions constructed systematically, and how does the computational cost scale with the number and arrangement of boundary faces?

A flat face is treated by rotating coordinates so that it becomes {η=0}\lbrace \eta = 0\rbrace; constant-coefficient PDE and boundary operators stay constant-coefficient. Example: Laplace equation with Dirichlet condition on the line y=xy = x. With ξ=(x+y)/2\xi = (x+y)/\sqrt{2} and η=(y−x)/2\eta = (y−x)/\sqrt{2}, the basis functions are cos⁡(κξ)sinh⁡(κη)\cos(\kappa\xi) \sinh(\kappa\eta) and sin⁡(κξ)sinh⁡(κη)\sin(\kappa\xi) \sinh(\kappa\eta) with a trainable frequency κ\kappa. Each solves the PDE and vanishes on the line, for every κ\kappa. Neumann and Robin conditions change only the coefficients in front of sinh⁡\sinh and cosh⁡\cosh (Appendix C.2).

For a box, requiring the condition on all faces fixes the frequencies to multiples of π/L\pi/L, which gives the classical sine basis, see e.g. Section 5.2. These frequencies are not trained; only the weights are fitted. For the triangle with a diagonal face in Appendix A.1.2 we use the mirror image, sin⁡(ax)sin⁡(by)−sin⁡(ay)sin⁡(bx)\sin(ax) \sin(by) − \sin(ay) \sin(bx). We do not have a general algorithm for arbitrary arrangements of faces. We will state this in the limitations and soften “piecewise-flat” in the abstract.

The basis functions are derived once, symbolically, independently of the amount of data, in a fraction of a second. Training and prediction cost is then the same as EPGP with M features.

EPGP itself does not enforce the boundary condition analytically. In the Wave IBVP comparison, how exactly was boundary information supplied to the EPGP baseline, and was this configuration matched to the information available to B-EPGP? Clarifying this would help assess the fairness of the comparison.

For EPGP, boundary information is supplied as data: we sampled 240 points uniformly on the boundary [0,8]×{0}[0,8]\times \lbrace 0\rbrace and used the observations u=0u = 0 there, in the same way as for WIGPR. For B-EPGP the boundary condition is built into the basis, so no boundary data are used. All other data (initial condition) are identical for all methods. Thus EPGP and WIGPR receive more boundary information than B-EPGP and still have a non-zero boundary residual, see Table 2. We will add this to the appendix.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal summary

Dear reviewers,

thank you for your careful, fair and constructive reviews. They helped us to improve the paper. In short, our responses add:

  • Uncertainty quantification: We add comparisons including PICP, NLPD and CRPS against the probabilistic baselines.
  • Baselines: EPGP in our comparisons (time to accuracy, and sparse-data Helmholtz) and a Hilbert-space GP suggested by Reviewer ygpy with the PDE as soft constraint.
  • Clarifications: which hyperparameters are trained in each experiment, how boundary information was given to EPGP, how faces not aligned with the axes are treated, and an experiment with a focusing wave.
  • Scope: Theorem 1 covers half-spaces and Dirichlet/Neumann conditions. For boxes and one triangle we construct the basis explicitly, and for general polyhedra we state a limitation. We will tone down overstated claims.

We have prepared detailed responses with new experiments below and will include all results in the camera ready version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.