Pith. sign in

REVIEW 3 major objections 4 minor 62 references

High-dimensional PDEs can be solved by training two networks as a saddle-point game whose error splits cleanly into approximation, sampling, and optimization pieces.

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 ↗

Inf-sup neural networks with dual Lagrange multipliers solve high-dimensional PDEs with proven error decomposition for linear convection-diffusion and empirical success on nonlinear cases.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Clean linear error theory for an InfSupNet extension to parabolic PDEs; the bound is residual-controlled only under linearity, and I_GP is left open. the 3 major comments →

arxiv 2607.11718 v1 pith:GA2CAYBD submitted 2026-07-13 math.NA cs.NAmath.AP

Inf-Sup Neural Networks for High Dimensional PDEs

Ziren Chen , Hailiang Liu This is my paper
classification math.NA cs.NAmath.AP MSC 65K1068Q25
keywords high-dimensional PDEsinf-sup formulationsaddle-point optimizationneural network approximationMonte Carlo samplingerror decompositionconvection-diffusionLagrange multiplier
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.

The reading

High-dimensional PDEs become intractable for grid methods because the number of points grows exponentially with dimension. This paper turns a well-posed PDE into an equivalent constrained optimization problem and then into an unconstrained inf-sup (saddle-point) problem by introducing a Lagrange multiplier that enforces the PDE residual. Two neural networks parameterize the primal solution and the multiplier; they are trained by alternating gradient descent-ascent on a Monte-Carlo approximation of the saddle functional. The authors prove that every classical solution of the original PDE is a saddle point of this functional and conversely, and they show that the total L2 error of the iterates is controlled by three explicit terms: how well the networks can approximate the continuous saddle, how well random samples approximate the continuous integrals, and how close the training algorithm comes to a true saddle of the empirical loss. Concrete bounds are derived for linear convection-diffusion equations under reduced-regularity norms that are easy to discretize. Numerical tests on linear and nonlinear parabolic problems up to dimension 20 show stable training and relative errors on the order of 10^{-3}.

Core claim

For linear well-posed PDEs the continuous inf-sup problem is equivalent to the original PDE, and the neural-network iterates produced by the alternating saddle algorithm satisfy an a-posteriori error bound that decomposes the total approximation error into network approximation error, Monte-Carlo sampling error, and optimization gap, each of which can be estimated quantitatively for linear convection-diffusion under L2-type losses.

What carries the argument

The inf-sup loss L(u,v) = (1/2) dist(Bu,g)^2 + <F[u]-f,v>, whose saddle points are exactly the PDE solutions (Theorems 1-2); the total error of the trained networks is then bounded by the sum of three residual terms I_NN + I_MC + I_GP (Theorems 3-6).

Load-bearing premise

The error analysis needs the PDE operators to be linear and the solution to be controlled by the size of the residual; without that stability estimate the three-term decomposition does not close.

What would settle it

Train the same architecture on a linear convection-diffusion problem whose exact solution is known, drive the network and sampling errors to machine precision, and check whether the observed L2 error continues to track the measured optimization gap I_GP; if the error plateaus far above I_GP, the claimed decomposition fails.

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

If this is right

  • Any linear well-posed PDE that admits a residual-controlled stability estimate can be attacked with the same two-network saddle training and inherits the three-term error bound.
  • Reducing the boundary norm from H^{3/2} to L2 (or fractional L2) yields a practical loss that needs only function values and pairwise differences, making high-dimensional sampling feasible.
  • The dual network is forced toward zero once the primal residual vanishes, giving a built-in diagnostic of constraint satisfaction.
  • The same algorithmic skeleton already produces accurate solutions for several nonlinear reaction-diffusion and Burgers-type problems even though the theory is still linear.

Where Pith is reading between the lines

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

  • The three-term split supplies a practical training budget: once I_NN and I_MC are driven below a target, further epochs should be judged solely by the size of I_GP.
  • Because the dual variable is exactly the residual residual map under the Riesz isomorphism, the method can be viewed as a learned residual-corrector that automatically balances interior and boundary residuals.
  • Extending the same saddle construction to operators that are only monotone (not linear) would immediately give a mesh-free solver for a larger class of elliptic and parabolic variational inequalities.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. The paper extends InfSupNet to a broader class of PDEs by reformulating (1.1) as a constrained minimization (1.2) and then as an inf-sup problem (1.3) via a Lagrange multiplier. Theorems 1–2 establish equivalence of the continuous formulations under well-posedness. Two networks parameterize the primal and dual variables and are trained by alternating GDA (Algorithm 1). For linear well-posed problems the residual-controlled stability estimate (4.1)–(4.2) yields the three-term error bound of Theorems 3–4; for linear convection–diffusion under a reduced L^{2} loss the authors further bound I_NN and I_MC explicitly (Theorem 6), leaving I_GP as an optimization residual. Numerical experiments on linear convection–diffusion (d=3,5) and several nonlinear problems (reaction–diffusion up to d=20, Burgers, Allen–Cahn) report relative L^{2} errors of order 10^{-3}.

Significance. The work supplies a clean residual-to-error decomposition for linear PDEs that simultaneously controls the primal solution and the dual multiplier, together with a practical reduced-regularity loss that avoids high-order boundary norms. The continuous equivalence proofs (Theorems 1–2) and the Monte-Carlo / fractional-Sobolev sampling analysis feeding Theorem 6 are carefully tracked. If the optimization gap can be controlled, the framework would give a transparent a-posteriori certificate for neural PDE solvers in high dimensions. The nonlinear experiments, while outside the theory, indicate that the same architecture remains stable on stiff and high-dimensional problems.

major comments (3)
  1. The central quantitative claim (Theorems 3–6) is ∥u_n−u∥^{2}_X + ∥v_n∥^{2}_{Z*} ≤ C(I_NN + I_MC + I_GP). I_GP is defined exactly as the empirical saddle gap at the trained pair (u_n,v_n). Remark 2 states that this gap is uncontrolled for general nonconvex–nonconcave GDA, and the numerical sections never report its magnitude. Consequently the bound does not yet guarantee that the trained network is close to the PDE solution; it only absorbs residual error into an unestimated term. Either a concrete bound (or decay plot) for I_GP under the stated assumptions, or a clear statement that the result is conditional on I_GP o0, is needed for the claim to be load-bearing.
  2. Section 4.1 and the proofs of Theorems 3–4 rely on linearity of F and B so that the residual of e=u_n−u is exactly F[e] and Be. Remark 4 and Subsections 6.2–6.4 present nonlinear experiments (reaction–diffusion, Burgers, Allen–Cahn) without corresponding theory. The abstract and contribution list should explicitly restrict the error analysis to linear problems and label the nonlinear results as empirical only, so that the scope of the theorems is not overstated.
  3. Theorem 6 (and Remark 1) assumes that the parameter spaces of U_θ and V_τ remain uniformly bounded so that the constants C_u, C_v, C_1–C_4 stay finite as n o∞. No mechanism (weight clipping, regularization, or a-posteriori check) is provided to enforce this hypothesis during training. Without it the almost-sure bound (5.4) is not guaranteed for the iterates actually produced by Algorithm 1.
minor comments (4)
  1. Typographical slips: “minimizaiton” (p. 4), “sovles” (p. 11), “hlaved” (p. 21), “inreqased” (p. 26), “Owning to” (p. 29).
  2. Figure 5 captions list “Dimension d=1,…,5” but the corresponding panel is not fully described in the text; a short legend clarifying the plotted quantity would help.
  3. The dual-network L^{2}-norm plot (Figure 6) is shown only for d=3; a brief remark on its behaviour for the higher-dimensional or nonlinear cases would strengthen the saddle-point interpretation.
  4. References [16] and [32] are central; a one-sentence comparison of the present error decomposition with the elliptic analysis of [16] would clarify the incremental contribution.

Circularity Check

1 steps flagged

Minor non-load-bearing self-citation to prior InfSupNet [16]; equivalence proofs and residual-stability error decomposition are self-contained and not forced by definition or fit.

specific steps
  1. self citation load bearing [Section 1 (Related work / Main contributions) and opening of Section 4]
    "Building on this perspective, the present work develops a more general inf–sup neural framework... building on the framework of [16], we establish error bounds for linear well-posed PDEs... The present analysis can be viewed as a natural extension of this framework."

    The method and error-decomposition strategy are presented as an extension of the authors' own prior InfSupNet paper [16]. This is ordinary self-citation of a foundational special case; it is not load-bearing because Theorems 1–2 and the residual-stability argument of Theorem 3 are proved from scratch without relying on unverified claims from [16].

full rationale

The paper's core claims (Theorems 1–2 equivalence of PDE / constrained / inf-sup; Theorems 3–6 residual-controlled error split into I_NN + I_MC + I_GP for linear well-posed problems, with explicit Monte-Carlo and approximation rates for convection–diffusion under reduced L2 loss) are derived from first principles inside the manuscript: well-posedness + Riesz map rewriting of residual norms + best-approximation projections onto network classes + standard MC concentration. The three-term bound is a transparent a-posteriori decomposition, not a tautology that assumes the conclusion; I_GP is left explicitly uncontrolled (Remark 2) rather than claimed small by construction. Self-citation to the authors' earlier InfSupNet [16] appears for the elliptic special case and the overall saddle-point idea, but the new parabolic analysis, reduced-regularity loss (4.20), and proofs do not reduce to that citation. No fitted parameters are re-labeled as predictions, no uniqueness theorem is imported to forbid alternatives, and no ansatz is smuggled. Nonlinear experiments lie outside the linear theory (as the paper itself notes) but do not create circularity. Score 2 reflects only the non-load-bearing self-citation; the derivation chain is otherwise independent.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 1 invented entities

Central claims rest on classical well-posedness/stability of linear PDEs, Riesz duality, Monte Carlo and fractional-Sobolev quadrature theory, universal approximation in Sobolev spaces, and informal assumptions that saddle-point training drives I_GP small and that network parameters stay bounded. Hyperparameters (architecture, samples, step counts, learning rates) are chosen by hand for experiments. No new physical entities are introduced; the dual network is a computational Lagrange multiplier, not a postulated physical object.

free parameters (4)
  • Network depth and width (primal and dual)
    Chosen by hand per experiment (e.g., 4×64 or 8×128); error constants and I_NN depend on these hypothesis spaces.
  • Interior/boundary/initial sample counts (m, m0, m̂, n̂)
    Hand-chosen (e.g., 6400/640/640); directly control I_MC rates in Theorem 6.
  • Ascent/descent steps k,l and learning rates η_θ, η_τ with decay
    Algorithm 1 hyperparameters tuned empirically; affect I_GP and observed convergence.
  • Training epoch budgets
    20k–100k epochs chosen per problem; determine reported final errors without a stopping theorem.
axioms (7)
  • domain assumption The PDE (1.1) is well-posed with unique solution in the chosen Banach/Hilbert spaces.
    Invoked for Theorems 1–2 and throughout Section 4; without uniqueness the constrained and saddle formulations need not recover a single u.
  • domain assumption Linear stability: ∥u∥_X² ≤ C(∥f∥_Z² + ∥g∥_P²) (and the reduced-regularity L2 version for parabolic problems).
    Equation (4.1)/(4.16); converts residual control into solution error in Theorems 3–4 and Lemma 5.
  • domain assumption F and B are linear for the error analysis.
    Stated in Section 4.1; authors note nonlinearity is outside the proved theory.
  • standard math Neural networks in U_θ, V_τ can approximate the required Sobolev/space-time targets (with C^{max(p,s)} activations for the primal).
    Used to argue I_NN is small; cites approximation results [1,8].
  • standard math Monte Carlo and tensor-product fractional-Sobolev quadrature error rates as in [5,6].
    Underpins I_MC bounds leading to Theorem 6.
  • ad hoc to paper Parameter spaces of U_θ and V_τ remain uniformly bounded so error constants stay finite as n→∞.
    Remark 1 after Theorem 6; not enforced by a proved constraint in Algorithm 1.
  • ad hoc to paper Saddle-point training makes the optimization gap I_GP small under suitable structural assumptions.
    Remark 2 cites general GDA literature but does not prove a rate for the nonconvex–nonconcave empirical loss used here.
invented entities (1)
  • InfsupNet / dual Lagrange-multiplier network v_τ no independent evidence
    purpose: Parameterize the dual variable in the saddle-point loss so the PDE residual is enforced adversarially rather than by squared residual minimization.
    Computational construct extending InfSupNet [16]; not a new physical field. independent_evidence is false as a physical entity, but the method is falsifiable via PDE residual and solution error on benchmarks.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Inf-Sup Neural Networks for High Dimensional PDEs." pith.science (2026). https://pith.science/paper/GA2CAYBD

@misc{pith2026260711718,
  author       = {Pith},
  title        = {Pith review of: Inf-Sup Neural Networks for High Dimensional PDEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GA2CAYBD}},
  note         = {Machine review of arXiv:2607.11718}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Solving partial differential equations (PDEs) in high dimensions remains challenging due to the curse of dimensionality. We propose a neural-network-based framework that reformulates PDEs as inf--sup optimization problems through the introduction of a Lagrange multiplier. The primal solution and the associated Lagrange multiplier are parameterized by two networks and are computed via an iterative saddle-point optimization procedure. We prove the theoretical equivalence between the proposed optimization formulation and the original PDE problem, and we derive rigorous error estimates that quantify the total approximation error in terms of the network approximation error, statistical (sampling) error, and optimization error. Numerical experiments demonstrate the accuracy, stability, and efficiency of the proposed method for solving high-dimensional PDEs.

Figures

Figures reproduced from arXiv: 2607.11718 by Hailiang Liu, Ziren Chen.

Figure 1
Figure 1. Figure 1: Training dynamics for the convection–diffusion equation in d = 3. network approximation, and the corresponding pointwise absolute error. The learned solution accurately reproduces the spatial structure, with errors that remain small and smoothly distributed. across the domain [PITH_FULL_IMAGE:figures/full_fig_p022_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Solution comparison for the d = 3 convection–diffusion equation at t = 1. From left to right: exact solution u(·, 1), neural network approximation uθ(·, 1), and pointwise absolute error |u−uθ|. The solutions are evaluated on the (x1, x2) slice with x3 = π. Training dynamics and error convergence (d = 5) [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training dynamics for the convection–diffusion equation in d = 5 [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Solution comparison for the d = 5 convection–diffusion equation at t = 1. From left to right: exact solution u(·, 1), neural network approximation uθ(·, 1), and pointwise absolute error |u−uθ|. The solutions are evaluated on the (x1, x2) slice while fixing x3 = x4 = x5 = π. Validation studies. We assess robustness through four validation experiments summarized in [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Validation studies [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Evolution of the adversarial network vτ in d = 3 (L 2 norm over 40,000 epochs) [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training dynamics for the nonlinear reaction–diffusion equation in d = 5. To visualize the learned solution, we evaluate uθ on a two-dimensional slice by varying (x1, x2) ∈ [0, 2π] 2 while fixing x3 = · · · = x5 = 0.5 at the final time t = 1 [PITH_FULL_IMAGE:figures/full_fig_p026_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Solution comparison for the nonlinear reaction–diffusion equation in d = 5 at t = 1. From left to right: exact solution, neural network approximation, and pointwise absolute error. The solution is evaluated on the (x1, x2) slice while fixing x3 = · · · = x5 = 0.5 [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Training dynamics for the nonlinear reaction–diffusion equation in d = 20 [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Solution comparison for the nonlinear reaction–diffusion equation in d = 20 at t = 1. From left to right: exact solution, neural network approximation, and pointwise absolute error. The solution is evaluated on the (x1, x2) slice with x3 = · · · = x20 = 0.5. 6.3. One-Dimensional Viscous Burgers’ and Allen–Cahn Equations. In this sub￾section, we assess the performance of the proposed inf–sup neural network… view at source ↗
Figure 11
Figure 11. Figure 11: Training dynamics for the one-dimensional viscous Burgers’ equa￾tion. To further evaluate the approximation accuracy, [PITH_FULL_IMAGE:figures/full_fig_p028_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Exact solution and neural network approximation of the viscous Burgers’ equation at t = 1. Owning to the stiff cubic reaction term, a longer training horizon is required. Neverthe￾less, the training process remains stable, without noticeable oscillations or divergence in either the primal or dual variables. After 100,000 epochs, the relative L 2 error reaches approximately 8.62 × 10−3 , with an absolute L… view at source ↗
Figure 13
Figure 13. Figure 13: Training dynamics for the one-dimensional Allen–Cahn equation [PITH_FULL_IMAGE:figures/full_fig_p029_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Exact solution and neural network approximation of the Allen–Cahn equation at t = 1. These one-dimensional experiments demonstrate that the inf–sup neural network effec￾tively resolves nonlinear convection, diffusion, and stiff reaction mechanisms. Even in the presence of ramp–cliff structures and sharp transition layers, the method achieves stable training, consistent loss decay, and accurate solution ap… 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

62 extracted references · 4 linked inside Pith

  1. [1]

    3, 499–541

    Ahmed Abdeljawad and Philipp Grohs,Approximations with deep neural networks in Sobolev time- space, Analysis and Applications20(2022), no. 3, 499–541

    2022

  2. [2]

    David Alvarez-Melis, Yair Schiff, and Youssef Mroueh,Optimizing functionals on the space of prob- abilities with input convex neural networks, Transactions on Machine Learning Research (2022)

    2022

  3. [3]

    6, C688–C715

    Silvia Bertoluzza, Erik Burman, and Cuiyu He,WAN discretization of PDEs: Best approximation, stabilization, and essential boundary conditions, SIAM Journal on Scientific Computing46(2024), no. 6, C688–C715

    2024

  4. [4]

    Boffi and Eric Vanden-Eijnden,Probability flow solution of the Fokker–Planck equation, Machine Learning: Science and Technology4(2023), no

    Nicholas M. Boffi and Eric Vanden-Eijnden,Probability flow solution of the Fokker–Planck equation, Machine Learning: Science and Technology4(2023), no. 3, 035012

    2023

  5. [5]

    Siegel,Convergence and error control of consistent PINNs for elliptic PDEs, IMA Journal of Numerical Analysis46(2025), no

    Andrea Bonito, Ronald DeVore, Guergana Petrova, and Jonathan W. Siegel,Convergence and error control of consistent PINNs for elliptic PDEs, IMA Journal of Numerical Analysis46(2025), no. 1, 90–148

    2025

  6. [6]

    Caflisch,Monte Carlo and quasi-Monte Carlo methods, Acta Numerica7(1998), 1–49

    Russel E. Caflisch,Monte Carlo and quasi-Monte Carlo methods, Acta Numerica7(1998), 1–49

    1998

  7. [7]

    4, 303–314

    George Cybenko,Approximation by superpositions of a sigmoidal function, Mathematics of Control, Signals and Systems2(1989), no. 4, 303–314

    1989

  8. [8]

    Tim De Ryck, Samuel Lanthaler, and Siddhartha Mishra,On the approximation of functions by tanh neural networks, Neural Networks143(2021), 732–750

    2021

  9. [9]

    Tim De Ryck and Siddhartha Mishra,Error analysis for physics-informed neural networks (PINNs) approximating kolmogorov PDEs, Advances in Computational Mathematics48(2022), no. 6, 79

    2022

  10. [10]

    2, 811–841

    Tim De Ryck, Siddhartha Mishra, and Roberto Molinaro,wPINNs: Weak physics-informed neural networks for approximating entropy solutions of hyperbolic conservation laws, SIAM Journal on Numerical Analysis62(2024), no. 2, 811–841

    2024

  11. [11]

    3, 195– 201

    Gamini Dissanayake and Nhan Phan-Thien,Neural-network-based approximations for solving partial differential equations, Communications in Numerical Methods in Engineering10(1994), no. 3, 195– 201

    1994

  12. [12]

    Weinan E and Bing Yu,The deep Ritz method: A deep learning-based numerical algorithm for solving variational problems, Communications in Mathematics and Statistics6(2018), no. 1, 1–12

    2018

  13. [13]

    6185–6215

    Jiaojiao Fan, Qinsheng Zhang, Amirhossein Taghvaei, and Yongxin Chen,Variational Wasserstein gradient flow, Proceedings of the 39th International Conference on Machine Learning, PMLR, 2022, pp. 6185–6215

    2022

  14. [14]

    32 ZIREN CHEN AND HAILIANG LIU

    Tianhao Hu, Bangti Jin, and Fengru Wang,An iterative deep Ritz method for monotone elliptic problems, Journal of Computational Physics527(2025), 113791. 32 ZIREN CHEN AND HAILIANG LIU

    2025

  15. [15]

    Karniadakis, and Kenji Kawaguchi,Score-based physics- informed neural networks for high-dimensional Fokker–Planck equations, SIAM Journal on Scientific Computing47(2025), no

    Zheyuan Hu, Zhongqiang Zhang, George E. Karniadakis, and Kenji Kawaguchi,Score-based physics- informed neural networks for high-dimensional Fokker–Planck equations, SIAM Journal on Scientific Computing47(2025), no. 3, C680–C705

    2025

  16. [16]

    Xiaokai Huo and Hailiang Liu,Inf-sup neural networks for high-dimensional elliptic PDE problems, Journal of Computational Physics514(2024), 113188

    2024

  17. [17]

    Jagtap and George E

    Ameya D. Jagtap and George E. Karniadakis,Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition framework for nonlinear partial differential equa- tions, Communications in Computational Physics28(2020), no. 5, 2002–2041

    2020

  18. [18]

    Yuling Jiao, Yanming Lai, Yang Wang, Haizhao Yang, and Yunfei Yang,Convergence analysis of the deep Galerkin method for weak solutions, From Classical Analysis to Analysis on Fractals: A Tribute to Robert Strichartz, Volume 1, Springer International Publishing, Cham, 2023, pp. 53–82

    2023

  19. [19]

    4880–4889

    Chi Jin, Praneeth Netrapalli, and Michael Jordan,What is local optimality in nonconvex-nonconcave minimax optimization?, Proceedings of the 37th International Conference on Machine Learning, PMLR, 2020, pp. 4880–4889

    2020

  20. [20]

    Richard Jordan, David Kinderlehrer, and Felix Otto,The variational formulation of the Fokker– Planck equation, SIAM Journal on Mathematical Analysis29(1998), no. 1, 1–17

    1998

  21. [21]

    Karniadakis,Variational physics-informed neu- ral networks for solving partial differential equations, 2019

    Ehsan Kharazmi, Zhongqiang Zhang, and George E. Karniadakis,Variational physics-informed neu- ral networks for solving partial differential equations, 2019

    2019

  22. [22]

    Mahoney,Charac- terizing possible failure modes in physics-informed neural networks, Advances in Neural Information Processing Systems, vol

    Aditi Krishnapriyan, Amir Gholami, Shandian Zhe, Robert Kirby, and Michael W. Mahoney,Charac- terizing possible failure modes in physics-informed neural networks, Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 26548–26560

    2021

  23. [23]

    Lagaris, Aristidis Likas, and Dimitrios I

    Isaac E. Lagaris, Aristidis Likas, and Dimitrios I. Fotiadis,Artificial neural networks for solving ordinary and partial differential equations, IEEE Transactions on Neural Networks9(1998), no. 5, 987–1000

    1998

  24. [24]

    Lingxiao Li, Samuel Hurault, and Justin Solomon,Self-consistent velocity matching of probability flows, Advances in Neural Information Processing Systems36(2023), 57038–57057

    2023

  25. [25]

    Shaofan Li and Wing Kam Liu,Meshfree and particle methods and their applications, Applied Me- chanics Reviews55(2002), no. 1, 1–34

    2002

  26. [26]

    Wuchen Li, Shu Liu, Hongyuan Zha, and Haomin Zhou,Parametric Fokker–Planck equation, Geo- metric Science of Information: 4th International Conference, GSI 2019 (Cham), Springer, 2019, pp. 715–724

    2019

  27. [27]

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, and Andrew Stuart,Fourier neural operator for parametric partial differential equations, arXiv preprint arXiv:2010.08895 (2021)

    Pith/arXiv arXiv 2010

  28. [28]

    6083–6093

    Tianyi Lin, Chi Jin, and Michael Jordan,On gradient descent ascent for nonconvex-concave minimax problems, Proceedings of the 37th International Conference on Machine Learning, PMLR, 2020, pp. 6083–6093

    2020

  29. [29]

    Jacques-Louis Lions and Enrico Magenes,Non-homogeneous boundary value problems and applica- tions, Springer, 1972

    1972

  30. [30]

    Min Liu, Zhiqiang Cai, and Karthik Ramani,Deep Ritz method with adaptive quadrature for linear elasticity, Computer Methods in Applied Mechanics and Engineering415(2023), 116229

    2023

  31. [31]

    3, 1385–1449

    Shu Liu, Wuchen Li, Hongyuan Zha, and Haomin Zhou,Neural parametric Fokker–Planck equation, SIAM Journal on Numerical Analysis60(2022), no. 3, 1385–1449

    2022

  32. [32]

    Shu Liu, Stanley Osher, and Wuchen Li,A natural primal-dual hybrid gradient method for adversarial neural network training on solving partial differential equations, 2024, arXiv:2411.06278

    Pith/arXiv arXiv 2024

  33. [33]

    Jianfeng Lu, Yue Wu, and Yang Xiang,Score-based transport modeling for mean-field Fokker–Planck equations, Journal of Computational Physics503(2024), 112859

    2024

  34. [34]

    Karniadakis,Learning non- linear operators via DeepONet based on the universal approximation theorem of operators, Nature Machine Intelligence3(2021), no

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George E. Karniadakis,Learning non- linear operators via DeepONet based on the universal approximation theorem of operators, Nature Machine Intelligence3(2021), no. 3, 218–229

    2021

  35. [35]

    3196–3241

    Yulong Lu, Jianfeng Lu, and Min Wang,A priori generalization analysis of the deep Ritz method for solving high dimensional elliptic partial differential equations, Proceedings of the Thirty Fourth Conference on Learning Theory, 2021, pp. 3196–3241. INF–SUP 33

    2021

  36. [36]

    McClenny and Ulisses M

    Levi D. McClenny and Ulisses M. Braga-Neto,Self-adaptive physics-informed neural networks, Jour- nal of Computational Physics474(2023), 111722

    2023

  37. [37]

    Ensela Mema, Ting Wang, and Jaroslaw Knap,Deep Ritz method with Fourier feature mapping: A deep learning approach for solving variational models of microstructure, Journal of Computational Science91(2025), 102631

    2025

  38. [38]

    Piotr Minakowski and Thomas Richter,A priori and a posteriori error estimates for the deep Ritz method applied to the Laplace and Stokes problem, Journal of Computational and Applied Mathe- matics421(2023), 114845

    2023

  39. [39]

    Shiv Mishra and Arbaz Khan,A priori error analysis of consistent PINNs for parabolic PDEs, 2025, arXiv:2506.17614

    Pith/arXiv arXiv 2025

  40. [40]

    Siddhartha Mishra and Roberto Molinaro,Estimates on the generalization error of physics-informed neural networks for approximating PDEs, IMA Journal of Numerical Analysis43(2023), no. 1, 1–43

    2023

  41. [41]

    34, 2021, pp

    Petr Mokrov, Alexander Korotin, Lingxiao Li, Aude Genevay, Justin Solomon, and Evgeny Burnaev, Large-scale Wasserstein gradient flows, Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 15243–15256

    2021

  42. [42]

    1, 229–251

    Arkadi Nemirovski,Prox-method with rate of convergenceO(1/t)for variational inequalities with lipschitz continuous monotone operators and smooth convex-concave saddle point problems, SIAM Journal on Optimization15(2004), no. 1, 229–251

    2004

  43. [43]

    Maher Nouiehed, Maziar Sanjabi, Tianjian Huang, Jason D. Lee, and Meisam Razaviyayn,Solv- ing a class of non-convex min-max games using iterative first order methods, Advances in Neural Information Processing Systems, Curran Associates, Inc., 2019, pp. 14934–14942

    2019

  44. [44]

    Maziar Raissi, Paris Perdikaris, and George E. Karniadakis,Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differ- ential equations, Journal of Computational Physics378(2019), 686–707

    2019

  45. [45]

    5, 55–66

    Meisam Razaviyayn, Tianjian Huang, Songtao Lu, Maher Nouiehed, Maziar Sanjabi, and Mingyi Hong,Nonconvex min-max optimization: Applications, challenges, and recent theoretical advances, IEEE Signal Processing Magazine37(2020), no. 5, 55–66

    2020

  46. [46]

    Martin Schechter,OnL p estimates and regularity II, Mathematica Scandinavica13(1963), 47–69

    1963

  47. [47]

    36, 2023, pp

    Zebang Shen and Zhenfu Wang,Entropy-dissipation informed neural network for McKean–Vlasov type PDEs, Advances in Neural Information Processing Systems, vol. 36, 2023, pp. 59227–59238

    2023

  48. [48]

    Zebang Shen, Zhenfu Wang, Satyen Kale, Alejandro Ribeiro, Amin Karbasi, and Hamed Hassani, Self-consistency of the Fokker–Planck equation, Proceedings of the Thirty Fifth Conference on Learn- ing Theory, PMLR, 2022, pp. 817–841

    2022

  49. [49]

    Justin Sirignano and Konstantinos Spiliopoulos,DGM: A deep learning algorithm for solving partial differential equations, Journal of Computational Physics375(2018), 1339–1364

    2018

  50. [50]

    4, 507–554

    Eitan Tadmor,A review of numerical methods for nonlinear partial differential equations, Bulletin of the American Mathematical Society49(2012), no. 4, 507–554

    2012

  51. [51]

    Tijmen Tieleman and Geoffrey Hinton,Lecture 6.5—RMSProp: Divide the gradient by a running average of its recent magnitude, COURSERA: Neural Networks for Machine Learning, 2012

    2012

  52. [52]

    Sifan Wang and Paris Perdikaris,Long-time integration of parametric evolution equations with physics-informed DeepONets, Journal of Computational Physics475(2023), 111855

    2023

  53. [53]

    5, A3055–A3081

    Sifan Wang, Yujun Teng, and Paris Perdikaris,Understanding and mitigating gradient flow patholo- gies in physics-informed neural networks, SIAM Journal on Scientific Computing43(2021), no. 5, A3055–A3081

    2021

  54. [54]

    40, eabi8605

    Sifan Wang, Hanwen Wang, and Paris Perdikaris,Learning the solution operator of parametric partial differential equations with physics-informed DeepONets, Science Advances7(2021), no. 40, eabi8605

    2021

  55. [55]

    Sifan Wang, Xinling Yu, and Paris Perdikaris,When and why PINNs fail to train: A neural tangent kernel perspective, Journal of Computational Physics449(2022), 110768

    2022

  56. [56]

    1, 360–395

    Hao Wu, Shu Liu, Xiaojing Ye, and Haomin Zhou,Parameterized Wasserstein Hamiltonian flow, SIAM Journal on Numerical Analysis63(2025), no. 1, 360–395

    2025

  57. [57]

    9, 10629–10636

    Wuzhe Xu, Yulong Lu, and Li Wang,Transfer learning enhanced DeepONet for long-time prediction of evolution equations, Proceedings of the AAAI Conference on Artificial Intelligence37(2023), no. 9, 10629–10636. 34 ZIREN CHEN AND HAILIANG LIU

    2023

  58. [58]

    Karniadakis,Gradient-enhanced physics-informed neural networks for forward and inverse PDE problems, Computer Methods in Applied Mechanics and Engineering393(2022), 114823

    Jeremy Yu, Lu Lu, Xuhui Meng, and George E. Karniadakis,Gradient-enhanced physics-informed neural networks for forward and inverse PDE problems, Computer Methods in Applied Mechanics and Engineering393(2022), 114823

    2022

  59. [59]

    Yaohua Zang, Gang Bao, Xiaojing Ye, and Haomin Zhou,Weak adversarial networks for high- dimensional partial differential equations, Journal of Computational Physics411(2020), 109409

    2020

  60. [60]

    Mo Zhou, Stanley Osher, and Wuchen Li,Score-based neural ordinary differential equations for computing mean field control problems, Journal of Computational Physics542(2025), 114369

    2025

  61. [61]

    ,Simulating Fokker–Planck equations via mean field control of score-based normalizing flows, 2025, arXiv:2506.05723

    Pith/arXiv arXiv 2025

  62. [62]

    Qijia Zhou, Yiyang Wang, Shengyuan Deng, and Chenliang Li,Deep Ritz physics-informed neural network method for solving the variational inequality problems, IAENG International Journal of Applied Mathematics55(2025), no. 12. AppendixA.Proof of Lemma 5 Proof.We employ the method of transposition. Letψ∈L 2(ΩT ) be an arbitrary test function. Consider the bac...

    2025

This paper was first reviewed by grok-4.5 on July 14, 2026.