Pith. sign in

REVIEW 3 major objections 3 minor 131 references

Boundary-adapted PINNs with first-order normalized distance functions guarantee explicit H2(Ω) error bounds for elliptic Dirichlet problems, with rates depending on the choice of activation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 13:16 UTC pith:H4ASKRLM

load-bearing objection The main theorem's hypotheses omit the first-order normalization that the paper's central claim depends on; as stated, the H² bounds fail for ρ=d². Otherwise a genuine and useful contribution. the 3 major comments →

arxiv 2607.19167 v1 pith:H4ASKRLM submitted 2026-07-21 math.NA cs.LGcs.NAmath.STstat.MLstat.TH

Boundary-Adapted PINNs for Elliptic Dirichlet Problems: H²(Ω) A Priori Error Bounds with Application to Mean Escape Time Computation

Nathanael Tepakbong , Jun Fan , Xiang Zhou , Ding-Xuan Zhou This is my paper
classification math.NA cs.LGcs.NAmath.STstat.MLstat.TH MSC 68Q3268T0765N1241A3041A28
keywords physics-informed neural networksmean escape timeelliptic Dirichlet problemsH2 error boundsboundary-adapted PINNsdistance-to-boundary normalizationVC dimensionReQU activation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks when a physics-informed neural network solving an elliptic Dirichlet problem can be guaranteed to converge in the strong H2 norm, not just weakly. The authors prove that exact boundary enforcement by multiplying the network by any smooth distance-to-boundary function is not enough: the multiplier must also be normalized to first order, so its normal derivative equals one on the boundary. Under that condition, they derive explicit non-asymptotic high-probability error bounds for networks with ReQU and tanh activations. The result identifies boundary-adapted PINNs as the right ansatz for stiff problems like mean escape time, where boundary layers are sharp. A sympathetic reader would take away a precise design rule for hard-enforced PINNs and a proof that this rule is essentially necessary.

Core claim

The central claim is that for second-order elliptic Dirichlet problems, an H2(Ω) a priori error bound cannot follow from residual minimization unless the boundary-enforcing multiplier ρ is a smooth distance approximation normalized to first order (∂νρ=1 on ∂Ω). If ρ only vanishes on the boundary but has zero normal derivative there, the quotient τ/ρ becomes unbounded near ∂Ω, destroying any smooth approximation target. With first-order normalization, Proposition 2.8 shows τ/ρ extends smoothly to the closure, and Theorem 3.11 then yields rates: for ReQU, ||u_n − τ||_{H2(Ω)} ≤ c n^{−γ/(2(d+γ))} log(n log n) with γ = min{r−2, (d+1)/2}; for tanh, ≤ c n^{−(r−2)/(2(3d+r−2))} log²(n log n). These a

What carries the argument

The load-bearing object is the smooth normalized distance approximation ρ: a C^{r+2} function that vanishes on ∂Ω, is positive inside, and satisfies ∂νρ=1 on ∂Ω (Definition 2.7). Along with it, the machinery includes Proposition 2.8, which guarantees τ/ρ ∈ C^r(Ω̄), and the boundary-adapted hypothesis space F(ρ,L,m,σ,Q,B,M) of outputs ρ·f for f in a network class with bounded W^{2,∞} norm. New VC-dimension bounds for derivatives of ReQU and tanh networks (Theorem 3.7) feed an oracle inequality (Theorem 3.10) that converts approximation and statistical errors into the final H2 rate.

Load-bearing premise

The training process is assumed to produce a network inside the hypothesis space with a uniform W^{2,∞} bound Q; the paper only encourages this via weight decay and does not prove that empirical risk minimization stays in that ball.

What would settle it

Choose a smooth domain and a function ρ that vanishes on the boundary but has zero normal derivative (e.g., ρ = d² for smooth distance d). If the central claim is right, τ/ρ is unbounded near ∂Ω and the boundary-adapted PINN with this ρ should fail to converge in H2 as the number of collocation points grows; a concrete check is to compute the H2 error for increasing n and observe it not decaying, or to verify numerically that the approximating outputs have unbounded derivatives near the boundary.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim is correct, hard-enforced PINNs should be built only with first-order normalized distance functions when H2 accuracy is the goal; simply multiplying by any zero-on-boundary ρ is insufficient.
  • The explicit rates imply a near-minimax scaling with the number of collocation points: tanh networks approach n^{−1/2} (up to log factors) for smooth solutions, while shallow ReQU networks saturate at a dimension-dependent rate.
  • The theory directly extends to nonhomogeneous Dirichlet conditions by adding a smooth extension G with ρ·f replaced by G+ρ·f, broadening applicability to general elliptic BVPs.
  • Poorly chosen distance functions, such as ρ^{3/2} or ρ^{1/2}, are predicted to fail catastrophically in the H2 norm even though they satisfy the boundary condition exactly; the experiments confirm this.
  • The proof components—derivative VC-dimension bounds and W^{2,∞} shallow ReQU approximation—stand as independent tools for analyzing other PDE learning algorithms.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The necessity result likely extends to mixed or Neumann boundary conditions: any exact-enforcement multiplier must match the leading normal behavior of the solution at the boundary to preserve smoothness of the ratio, so analogous normalization conditions should appear there.
  • For rare-event and narrow-escape problems with strong boundary layers, this analysis suggests that boundary-adapted PINNs with first-order normalization may be the minimal architecture that provably captures boundary-layer structure in H2; testing on such problems would be a natural next step.
  • The dependence of the rate on r and d hints that deeper ReQU networks, while avoiding saturation, would require a separate VC-dimension analysis; the current bounds provide a benchmark for what shallow architectures can guarantee.
  • A testable practical consequence: monitoring τ/ρ near the boundary during training could diagnose whether a chosen ρ preserves smoothness, before committing to expensive H2 evaluation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper develops an a priori error analysis for boundary-enforced PINNs of the form ρ(x)f_θ(x) for elliptic Dirichlet BVPs, motivated by mean escape time computation. The central claim is that exact boundary enforcement alone is insufficient for H²(Ω) error bounds, and that a sufficient and essentially necessary condition is that ρ be a smooth distance approximation normalized to first order (∂_νρ=1 on ∂Ω). The authors prove quotient regularity (τ/ρ∈C^r), new W^{2,∞} approximation bounds for shallow ReQU networks, new VC-dimension bounds for derivatives of ReQU and tanh networks, and combine these with an oracle inequality to obtain non-asymptotic high-probability H²(Ω) rates: essentially n^{-γ/(2(d+γ))} log(n log n) for ReQU and n^{-(r-2)/(2(3d+r-2))} log²(n log n) for tanh. Numerical experiments compare the boundary-adapted ansatz with standard, variational, over-enforced, and under-enforced PINNs.

Significance. If the main sufficiency result is correct, this is a valuable contribution: it identifies a concrete, constructive class of hard-constrained PINN architectures that admit H²(Ω) error bounds, and it provides the first rigorous a priori analysis that explicitly accounts for the boundary-enforcement ansatz. The paper is also genuinely useful for its auxiliary results: the shallow ReQU W^{2,∞} approximation theorem and the VC-dimension bounds for higher-order derivatives of ReQU and tanh networks are of independent interest, and the appendices are detailed and mostly self-contained. However, the main theorem as formally stated is too strong because it omits the first-order normalization hypothesis; the intended theorem is clear and the fix is local, but the error is load-bearing rather than cosmetic.

major comments (3)
  1. [Corollary 3.5 / Theorem 3.11] The main results are stated for an '(r+2)-smooth distance approximation' in the sense of Definition 2.5, which does not include the first-order normalization ∂_νρ=1 on ∂Ω. The proof of Corollary 3.5 invokes Lemma 2.11, whose hypothesis is supplied by Proposition 2.8 and therefore requires Definition 2.7. This is not cosmetic: take ρ=d², the squared Euclidean distance to ∂Ω. Then ρ satisfies Definition 2.5, ∂_νρ=0 on ∂Ω, and τ/ρ behaves like 1/d near the boundary, so no W^{r,∞}(R^d) extension u of τ/ρ exists and the bound (3.5) cannot hold. Moreover, every ansatz ρf has zero normal derivative on ∂Ω, while τ has non-zero normal derivative, so H²(Ω) convergence is impossible for such ρ. The statements should require ρ to be normalized to first order (Definition 2.7).
  2. [Abstract / §1.3 / footnote 2] The abstract and §1.3 claim that first-order normalization is 'essentially necessary' for H²(Ω) error bounds. The only argument supplied is footnote 2: if ∂_νρ=0 at a boundary point, then τ/ρ is unbounded. This shows that the quotient-based extension argument fails, but it is not a lower bound on inf_{f} ∥ρf−τ∥_{H²(Ω)} over admissible neural networks. A rigorous necessity claim would require a lower bound over all such f, or the claim should be softened to 'necessary for the present proof strategy'.
  3. [Definition 2.6 / Remark 2.4 / Theorem 3.11] The error analysis applies to the empirical risk minimizer within the hypothesis class F = {ρf : f∈Σ, max_{|α|≤2}∥∂^α f∥_{L∞}≤Q}. Theorem 3.11 therefore does not, by itself, control the function produced by gradient-based training unless the training trajectory stays inside this Q-ball. Remark 2.4 suggests weight decay as a heuristic but supplies no proof. This is acknowledged in the introduction, but the abstract and conclusion describe 'a complete a priori error analysis' for PINN solutions; the Q-ball assumption should be stated as a standing hypothesis of the main theorem, and the gap between the theorem and the trained networks in Section 4 should be made explicit.
minor comments (3)
  1. [Theorem 3.11, proof] In the ReQU case the proof writes '∥m∥_∞ = r d n^a' with a=1/(2(d+γ(r))). To match the theorem statement and the subsequent VC-dimension bound O(n^{2ad}), the width should be r d n^{ad} (equivalently, set N=n^a and width=N^d). As written, the displayed width and the parameter count are inconsistent.
  2. [Section 4] The over-enforced and under-enforced variants ρ^{3/2} and ρ^{1/2} do not satisfy Definition 2.5 or Definition 2.7 (ρ^{1/2} has unbounded normal derivative at ∂Ω, and ρ^{3/2} is not C² up to the boundary). The experiments are therefore outside the theory; this is fine as a numerical illustration, but it should be stated explicitly.
  3. [Throughout] There are several typographical issues: the title uses 'BOUNDAR Y'; Definition 2.10 introduces 'boundary-adapted FNNs' but Theorems 3.11 and Corollary 3.5 do not refer to this definition; and the constants M_ReQU/M_tanh are introduced with inconsistent fonts. These should be cleaned up.

Circularity Check

0 steps flagged

The H^2 error bounds are derived from external approximation and statistical-learning results; no prediction in the paper reduces to a fitted value or to an equivalent restatement of the ansatz.

full rationale

The central derivation chain is not circular. The paper proves Proposition 2.8 (sufficiency of first-order normalization for tau/rho in C^r) by a local-coordinate/FTC argument in Appendix A, then combines external approximation theorems (De Ryck et al., 2021; Yang and Zhou, 2025; Siegel, 2025) with new VC-dimension bounds proved in Appendix C and an oracle inequality adapted from Lei et al. (2025) to obtain Theorem 3.11. The rates n^{-gamma/(2(d+gamma))} and n^{-(r-2)/(2(3d+r-2))} come from balancing approximation and statistical terms, not from matching data. Self-citations appear (Lei et al., 2025; Tepakbong et al., 2026; Yang and Zhou, 2025) but are used as independent published building blocks, not as the conclusion being derived. The numerical experiments compare against exact or FEM reference solutions rather than recycling the model output as ground truth. The main manuscript-level concern is a hypothesis gap, not circularity: Corollary 3.5 and Theorem 3.11 state only that rho is an '(r+2)-smooth distance approximation' (Definition 2.5), while the proof invokes Lemma 2.11 and Proposition 2.8, which need the first-order normalization of Definition 2.7. Repairing this by adding the normalization hypothesis would strengthen the assumptions; it is a correctness issue, not a case of an output being equivalent to an input by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 0 invented entities

The central claim rests on classical elliptic regularity, the ADN estimate, and prior learning-theory bounds. No physical entities are invented; the only ad hoc element is the assumed uniform W^{2,∞} bound Q in the hypothesis space, which is not exactly enforced in training.

axioms (6)
  • domain assumption Elliptic regularity: under assumptions A.1–A.3, the BVP (2.3) has a unique solution τ ∈ C^{r+2,κ}(Ω̄) (Gilbarg–Trudinger, Theorems 6.14 and 6.19).
    Invoked at the start of Section 2; the entire error analysis needs τ/ρ ∈ C^r and H² regularity of τ.
  • standard math Agmon–Douglis–Nirenberg estimate (2.8): for zero-trace H² functions, ∥u−τ∥_{H²} ≲ ∥L(u−τ)∥_{L²}.
    This converts residual control into H² error and is the key inequality in Theorem 3.11; it requires the boundary to be C² and the operator to satisfy the complementing condition.
  • domain assumption Existence of a C^{r+2} defining function φ with ∇φ ≠ 0 on ∂Ω, and a tubular neighborhood with nearest-point projection (Algorithm 2.1).
    Algorithm 2.1 and Definition 2.7 require this; for many domains (balls, ellipses) it is explicit, but for arbitrary smooth domains it is a nontrivial geometric input.
  • standard math The quotient τ/ρ extends to a W^{r,∞}(R^d) function with controlled norm (Lemma 2.11, via Gilbarg–Trudinger Theorem 7.25).
    Extension lemma invoked in Corollary 3.5 to apply Euclidean approximation theorems (Theorems 3.3, 3.4).
  • domain assumption Oracle inequality of Lei et al. 2025 (Theorem 3.10 here).
    The paper asserts the argument is agnostic to the network class and only needs pseudo-dimension and bounded loss; this is a check on a prior result rather than a proof in this paper.
  • ad hoc to paper The empirical risk minimizer u_n lies in F with uniform W^{2,∞} bound Q (Definition 2.6, Theorem 3.11).
    Remark 2.4 proposes weight decay to encourage this, but the paper does not prove that the trained network satisfies the Q-bound; this is a gap between theory and practice.

pith-pipeline@v1.3.0-alltime-deepseek · 52257 in / 27458 out tokens · 274322 ms · 2026-08-01T13:16:25.899923+00:00 · methodology

0 comments
read the original abstract

Motivated by the numerical computation of the Mean Escape Time (MET) $\tau:\Omega\to\mathbb{R}$ of a stochastic process from a bounded domain $\Omega\subseteq\mathbb{R}^d$, we study elliptic Dirichlet boundary value problems (BVPs) using boundary-enforced Physics-Informed Neural Networks (PINNs), in which the Dirichlet condition is imposed exactly by multiplying the network output with a predefined distance-to-boundary approximation $\rho$. Combining approximation-theoretic and statistical-learning arguments for Rectified Quadratic Unit (ReQU) and hyperbolic tangent (tanh) networks, we derive a priori error bounds that make explicit the dependence on $\rho$. In particular, we show that exact boundary enforcement alone is not enough for $H^2(\Omega)$ error bounds, and that a sufficient and essentially necessary condition is for $\rho$ to be a smooth distance approximation $\textit{normalized to first order}$, of the kind constructed in arXiv:2104.08426 [math.NA]. We thereby identify this subclass of $\textit{boundary-adapted}$ PINNs as the appropriate neural network ansatz for solving Dirichlet BVPs. Numerical experiments support the theory, showing that appropriate choices of $\rho$ improve accuracy and convergence, while poorly chosen distance functions can substantially degrade the solution. Our proof also yields new VC-dimension bounds for hypothesis spaces of higher-order derivatives of ReQU and tanh networks, together with new approximation bounds for shallow ReQU networks in higher-order Sobolev norms, all of which are of important independent interest.

Figures

Figures reproduced from arXiv: 2607.19167 by Ding-Xuan Zhou, Jun Fan, Nathanael Tepakbong, Xiang Zhou.

Figure 1
Figure 1. Figure 1: Projection on the {(x, y) : x = 0} axis of the MET for the SDE (2.5) in the 2D unit ball, with θ = 1 and varying ε > 0. The closed-form expression for this MET has been derived in (Kersting et al., 2023) [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Reference solution τ for the 2D Ornstein–Uhlenbeck problem. Method L 2 rel(∥τ∥L2 = 1.155 × 101 ) H1 rel(∥τ∥H1 = 2.266 × 101 ) H2 rel(∥τ∥H2 = 1.841 × 102 ) Standard 4.887 × 10−2 ± 3.872 × 10−2 4.994 × 10−2 ± 3.926 × 10−2 5.041 × 10−2 ± 3.964 × 10−2 Variational 1.169 × 10−1 ± 5.697 × 10−2 6.491 × 10−1 ± 6.106 × 10−1 3.269 ± 4.211 Over-enforced 1.013 ± 5.774 × 10−4 1.006 ± 0.000 1.001 ± 5.774 × 10−4 Under-enf… view at source ↗
Figure 3
Figure 3. Figure 3: Predicted solutions (left column) and pointwise error fields (right column) for [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Predicted solutions (left column) and pointwise error fields (right column) for [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FEM solution τref for the 2D double-well MET problem. Method L 2 rel(∥τ∥L2 = 9.538 × 10−1 ) H1 rel(∥τ∥H1 = 1.834) H2 rel(∥τ∥H2 = 8.577) Standard 2.953 × 10−1 ± 1.200 × 10−2 4.499 × 10−1 ± 1.393 × 10−2 6.530 × 10−1 ± 2.366 × 10−2 Variational 5.639 × 10−1 ± 1.100 × 10−2 1.173 ± 9.637 × 10−2 1.197 × 101 ± 1.703 Over-enforced 4.857 × 10−1 ± 3.981 × 10−1 5.865 × 10−1 ± 4.187 × 10−1 8.943 × 10−1 ± 2.767 × 10−1 U… view at source ↗
Figure 6
Figure 6. Figure 6: Predicted solutions (left column) and pointwise error fields (right column) [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Predicted solutions (left column) and pointwise error fields (right column) [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

131 extracted references · 2 linked inside Pith

  1. [1]

    N. G. van Kampen , Edition =. Stochastic Processes in Physics and Chemistry , Year =

  2. [2]

    2014 , publisher=

    Stochastic processes in cell biology , author=. 2014 , publisher=

    2014

  3. [3]

    Analysis and approximation of rare events

    Budhiraja, Amarjit and Dupuis, Paul. Analysis and approximation of rare events

  4. [4]

    Escape from an attractor: importance sampling and rest points:

    Paul Dupuis and Konstantinos Spiliopoulos and Xiang Zhou , Journal =. Escape from an attractor: importance sampling and rest points:

  5. [5]

    Enforcing

    Berrone, Stefano and Canuto, Claudio and Pintore, Moreno and Sukumar, Natarajan , journal=. Enforcing. 2023 , publisher=

    2023

  6. [6]

    A comparison study of deep

    Chen, Jingrun and Du, Rui and Wu, Keke , journal =. A comparison study of deep

  7. [7]

    International Conference on Learning Representation , year=

    An Operator Preconditionning Perspective on Training in Physics-Informed Machine Learning , author=. International Conference on Learning Representation , year=

  8. [8]

    Mean first exit times of

    Kersting, Hans and Orvieto, Antonio and Proske, Frank and Lucchi, Aurelien , journal=. Mean first exit times of. 2023 , publisher=

    2023

  9. [9]

    2012 , publisher=

    Random Perturbations of Dynamical Systems , author=. 2012 , publisher=

    2012

  10. [10]

    Challenges in training

    Rathore, Pratik and Lei, Weimu and Frangella, Zachary and Lu, Lu and Udell, Madeleine , booktitle=. Challenges in training

  11. [11]

    Handbook of Numerical Analysis , volume=

    Two-layer neural networks for partial differential equations: Optimization and generalization theory , author=. Handbook of Numerical Analysis , volume=. 2024 , publisher=

    2024

  12. [12]

    On the Convergence of Physics Informed Neural Networks for Linear Second-Order Elliptic and Parabolic Type

    Shin, Yeonjong and Darbon, J. On the Convergence of Physics Informed Neural Networks for Linear Second-Order Elliptic and Parabolic Type. Communications in Computational Physics , volume=. 2020 , publisher=

    2020

  13. [13]

    Communications in Computational Physics , volume=

    Convergence of Physics-Informed Neural Networks Applied to Linear Second-Order Elliptic Interface Problems , author=. Communications in Computational Physics , volume=

  14. [14]

    A Rate of Convergence of Physics Informed Neural Networks for the Linear Second Order Elliptic

    Jiao, Yuling and Lai, Yanming and Li, Dingwei and Lu, Xiliang and Wang, Fengru and Wang, Yang and Yang, Jerry Zhijian , journal=. A Rate of Convergence of Physics Informed Neural Networks for the Linear Second Order Elliptic

  15. [15]

    Generalization of

    Jiang, Xuelian and Wang, Ziming and Bao, Wei and Xu, Yingxiang , journal=. Generalization of. 2024 , publisher=

    2024

  16. [16]

    Error estimates of residual minimization using neural networks for linear

    Shin, Yeonjong and Zhang, Zhongqiang and Karniadakis, George Em , journal=. Error estimates of residual minimization using neural networks for linear. 2023 , publisher=

    2023

  17. [17]

    Jiao, Yuling and Li, Ruoxuan and Wu, Peiying and Yang, Jerry Zhijian and Zhang, Pingwen , journal=

  18. [18]

    Physical Review Research , volume=

    First passage time distribution of active thermal particles in potentials , author=. Physical Review Research , volume=. 2021 , publisher=

    2021

  19. [19]

    Physical Review E—Statistical, Nonlinear, and Soft Matter Physics , volume=

    Extinction of metastable stochastic populations , author=. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics , volume=. 2010 , publisher=

    2010

  20. [20]

    Physical Review , volume=

    On the first passage time probability problem , author=. Physical Review , volume=. 1951 , publisher=

    1951

  21. [21]

    Applied Mathematics and Mechanics , volume=

    A neural network solution of first-passage problems , author=. Applied Mathematics and Mechanics , volume=. 2024 , publisher=

    2024

  22. [22]

    Stochastic Analysis and Applications , volume=

    Adaptive multilevel splitting for rare event analysis , author=. Stochastic Analysis and Applications , volume=. 2007 , publisher=

    2007

  23. [23]

    2019 Winter Simulation Conference (WSC) , pages=

    Efficient estimation of the mean hitting time to a set of a regenerative system , author=. 2019 Winter Simulation Conference (WSC) , pages=. 2019 , organization=

    2019

  24. [24]

    Performance Evaluation , volume=

    Accelerating mean time to failure computations , author=. Performance Evaluation , volume=. 1996 , publisher=

    1996

  25. [25]

    Journal of Statistical Physics , volume=

    Flows in complex networks: theory, algorithms, and application to Lennard--Jones cluster rearrangement , author=. Journal of Statistical Physics , volume=. 2014 , publisher=

    2014

  26. [26]

    Journal of Scientific Computing , volume=

    Scientific machine learning through physics--informed neural networks: Where we are and what’s next , author=. Journal of Scientific Computing , volume=. 2022 , publisher=

    2022

  27. [27]

    2023 , publisher=

    Overcoming the timescale barrier in molecular dynamics: Transfer operators, variational principles and machine learning , author=. 2023 , publisher=

    2023

  28. [28]

    Verhulst, Ferdinand , year=

  29. [29]

    Convergence and error control of consistent

    Bonito, Andrea and DeVore, Ronald and Petrova, Guergana and Siegel, Jonathan W , journal=. Convergence and error control of consistent. 2026 , publisher=

    2026

  30. [30]

    , title =

    Krishnapriyan, Aditi and Gholami, Amir and Zhe, Shandian and Kirby, Robert and Mahoney, Michael W. , title =. Advances in Neural Information Processing Systems , volume =. 2021 , pages =

    2021

  31. [31]

    , title =

    Escapil-Inchauspé, Paul and Ruz, Gonzalo A. , title =. Neurocomputing , volume =

  32. [32]

    arXiv preprint arXiv:2504.09804 , year =

    Zhang, Rui and Li, Liang and Lanteri, Stéphane and Kang, Hao and Li, Jiaqi , title =. arXiv preprint arXiv:2504.09804 , year =

    Pith/arXiv arXiv

  33. [33]

    Bernoulli , volume =

    Doumèche, Nathan and Biau, Gérard and Boyer, Claire , title =. Bernoulli , volume =

  34. [34]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Multi-task learning using uncertainty to weigh losses for scene geometry and semantics , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

  35. [35]

    1991 , publisher=

    Fewnomials , author=. 1991 , publisher=

    1991

  36. [36]

    Complexity of computations with

    Gabrielov, Andrei and Vorobjov, Nicolai , journal=. Complexity of computations with. 2004 , publisher=

    2004

  37. [37]

    Polynomial bounds for

    Karpinski, Marek and Macintyre, Angus , journal=. Polynomial bounds for. 1997 , publisher=

    1997

  38. [38]

    Scarselli, Franco and Tsoi, Ah Chung and Hagenbuchner, Markus , journal=. The. 2018 , publisher=

    2018

  39. [39]

    2025 , publisher=

    D’Inverno, Giuseppe Alessio and Bianchini, Monica and Scarselli, Franco , journal=. 2025 , publisher=

    2025

  40. [40]

    1999 , publisher=

    Neural Network Learning: Theoretical Foundations , author=. 1999 , publisher=

    1999

  41. [41]

    2013 , publisher=

    The nature of statistical learning theory , author=. 2013 , publisher=

    2013

  42. [42]

    High dimensional probability II , pages=

    Rademacher processes and bounding the risk of function learning , author=. High dimensional probability II , pages=. 2000 , publisher=

    2000

  43. [43]

    The Annals of Statistics , pages=

    Local Rademacher Complexities and Oracle Inequalities in Risk Minimization , author=. The Annals of Statistics , pages=. 2006 , publisher=

    2006

  44. [44]

    The Annals of Statistics , volume=

    Local Rademacher Complexities , author=. The Annals of Statistics , volume=

  45. [45]

    Nearly optimal

    Yang, Yahong and Yang, Haizhao and Xiang, Yang , journal=. Nearly optimal

  46. [46]

    Nearly-tight

    Bartlett, Peter L and Harvey, Nick and Liaw, Christopher and Mehrabian, Abbas , journal=. Nearly-tight

  47. [47]

    2013 , month =

    Boucheron, Stéphane and Lugosi, Gábor and Massart, Pascal , title =. 2013 , month =. doi:10.1093/acprof:oso/9780199535255.001.0001 , url =

    arXiv 2013

  48. [48]

    Proceedings of the National Academy of Sciences , volume=

    The narrow escape problem for diffusion in cellular microdomains , author=. Proceedings of the National Academy of Sciences , volume=. 2007 , publisher=

    2007

  49. [49]

    Analysis and Applications

    Stochastic narrow escape in molecular and cellular biology , author=. Analysis and Applications. Springer, New York , volume=. 2015 , publisher=

    2015

  50. [50]

    2009 , publisher=

    Theory and applications of stochastic processes: an analytical approach , author=. 2009 , publisher=

    2009

  51. [51]

    Journal of the Royal Statistical Society: Series A (General) , volume=

    A stochastic model for the duration of a strike , author=. Journal of the Royal Statistical Society: Series A (General) , volume=. 1972 , publisher=

    1972

  52. [52]

    Redner, Sidney , year=

  53. [53]

    2014 , publisher=

    Stochastic processes and applications , author=. 2014 , publisher=

    2014

  54. [54]

    Comptes Rendus

    Analytical and numerical results for first escape time in 2D , author=. Comptes Rendus. Math

  55. [55]

    Fast and accurate

    Drugowitsch, Jan , journal=. Fast and accurate. 2016 , publisher=

    2016

  56. [56]

    The estimates of the mean first exit time of a bistable system excited by

    Xu, Yong and Li, Hua and Wang, Haiyan and Jia, Wantao and Yue, Xiaole and Kurths, J. The estimates of the mean first exit time of a bistable system excited by. Journal of Applied Mechanics , volume=. 2017 , publisher=

    2017

  57. [57]

    Journal of Mathematical Psychology , volume=

    A fast numerical algorithm for the estimation of diffusion model parameters , author=. Journal of Mathematical Psychology , volume=. 2008 , publisher=

    2008

  58. [58]

    Efficient numerical approximation of a non-regular

    Boehm, Udo and Cox, Sonja and Gantner, Gregor and Stevenson, Rob , journal=. Efficient numerical approximation of a non-regular. 2022 , publisher=

    2022

  59. [59]

    Physical Review E , volume=

    Studying first passage problems using neural networks: A case study in the slit-well microfluidic device , author=. Physical Review E , volume=. 2022 , publisher=

    2022

  60. [60]

    New Journal of Physics , volume=

    Mean exit time for diffusion on irregular domains , author=. New Journal of Physics , volume=. 2021 , publisher=

    2021

  61. [61]

    Communications in Nonlinear Science and Numerical Simulation , volume=

    Approximate analytical solution for transient heat and mass transfer across an irregular interface , author=. Communications in Nonlinear Science and Numerical Simulation , volume=. 2022 , publisher=

    2022

  62. [62]

    Journal of Physics A: Mathematical and Theoretical , volume=

    Mean exit time in irregularly-shaped annular and composite disc domains , author=. Journal of Physics A: Mathematical and Theoretical , volume=. 2022 , publisher=

    2022

  63. [63]

    2022 , url =

    Holmes-Cerfon, Miranda , title =. 2022 , url =

    2022

  64. [64]

    2022 , publisher=

    Evans, Lawrence C , volume=. 2022 , publisher=

    2022

  65. [65]

    Journal of Computational physics , volume=

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , author=. Journal of Computational physics , volume=. 2019 , publisher=

    2019

  66. [66]

    Estimates on the generalization error of physics-informed neural networks for approximating

    Mishra, Siddhartha and Molinaro, Roberto , journal=. Estimates on the generalization error of physics-informed neural networks for approximating. 2023 , publisher=

    2023

  67. [67]

    E, Weinan and Yu, Bing , journal=. The. 2018 , publisher=

    2018

  68. [68]

    A priori generalization analysis of the deep

    Lu, Yulong and Lu, Jianfeng and Wang, Min , booktitle=. A priori generalization analysis of the deep. 2021 , organization=

    2021

  69. [69]

    International Conference on Learning Representations , year=

    Fourier Neural Operator for Parametric Partial Differential Equations , author=. International Conference on Learning Representations , year=

  70. [70]

    Scientific reports , volume=

    DeepGreen: deep learning of Green’s functions for nonlinear boundary value problems , author=. Scientific reports , volume=. 2021 , publisher=

    2021

  71. [71]

    Learning nonlinear operators via

    Lu, Lu and Jin, Pengzhan and Pang, Guofei and Zhang, Zhongqiang and Karniadakis, George Em , journal=. Learning nonlinear operators via. 2021 , publisher=

    2021

  72. [72]

    Handbook of financial time series , pages=

    Maller, Ross A and M. Handbook of financial time series , pages=. 2009 , publisher=

    2009

  73. [73]

    2003 , publisher=

    Stochastic differential equations , author=. 2003 , publisher=

    2003

  74. [74]

    Machine Learning For Elliptic

    Yiping Lu and Haoxuan Chen and Jianfeng Lu and Lexing Ying and Jose Blanchet , booktitle=. Machine Learning For Elliptic. 2022 , url=

    2022

  75. [75]

    Lei, Guanhang and Lei, Zhen and Shi, Lei and Zeng, Chenyu and Zhou, Ding-Xuan , journal=. Solving. 2025 , publisher=

    2025

  76. [76]

    Computer-Aided Civil and Infrastructure Engineering , volume=

    Efficient training of physics-informed neural networks via importance sampling , author=. Computer-Aided Civil and Infrastructure Engineering , volume=. 2021 , publisher=

    2021

  77. [77]

    Computer Methods in Applied Mechanics and Engineering , volume=

    A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks , author=. Computer Methods in Applied Mechanics and Engineering , volume=. 2023 , publisher=

    2023

  78. [78]

    Failure-informed adaptive sampling for

    Gao, Zhiwei and Yan, Liang and Zhou, Tao , journal=. Failure-informed adaptive sampling for. 2023 , publisher=

    2023

  79. [79]

    Journal of Machine Learning Research , volume=

    Deep nonparametric estimation of operators between infinite dimensional spaces , author=. Journal of Machine Learning Research , volume=

  80. [80]

    Neural Networks , volume=

    Basis operator network: A neural network-based model for learning nonlinear operators via neural basis , author=. Neural Networks , volume=. 2023 , publisher=

    2023

Showing first 80 references.