Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Learned PDE time-steppers avoid the curse of dimensionality

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 →

For reaction-diffusion, forced parabolic, and viscous conservation law PDEs, neural networks can learn the numerical one-step time map with generalization error that is polynomial in mesh size, not exponential.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Worth a serious referee: the single-input time-stepper results are solid and new, but the multi-input conservation-law theorem is a proof sketch and the global error recursion has a measure-shift gap. the 4 major comments →

arxiv 2509.04256 v1 pith:LQLKOOQO submitted 2025-09-04 math.NA cs.NA

Error analysis for learning the time-stepping operator of evolutionary PDEs

classification math.NA cs.NA MSC 65M1265M1568T01
keywords generalization errorneural operator learningtime-stepping operatorReLU networkscurse of dimensionalityreaction-diffusion equationsviscous conservation lawsimplicit time-stepping
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

This paper attacks a practical bottleneck: implicit time-stepping for evolutionary PDEs is stable but expensive, because each step requires solving a nonlinear system. It asks whether a neural network can learn the one-step map directly, and whether that learning remains tractable as spatial resolution grows. The paper's answer is yes for three broad PDE classes: reaction-diffusion equations, parabolic equations with forcing, and viscous conservation laws. In each case, the time-stepping operator is Lipschitz and can be decomposed into linear maps interleaved with low-dimensional nonlinearities, and this structure yields generalization error that is polynomial in the encoding dimension and decays like n^{-1/2} or n^{-2/3} with the number of training samples. The same guarantee extends to multi-input operators that take both the initial state and a reaction or forcing term, so one trained network can serve a family of PDEs without retraining.

Core claim

The paper's central claim is that implicit time-stepping operators for several important evolutionary PDEs are efficiently learnable by ReLU feedforward networks without the curse of dimensionality. For implicit Euler with Picard iterations and with a quadratically convergent linearized solver, for implicit Euler and Crank-Nicolson for forced parabolic equations, and for implicit Euler with Picard iterations for viscous conservation laws, the paper proves that the target operator is Lipschitz continuous and has a low-complexity structure: it can be written as a composition of linear transformations and nonlinearities that act on at most two coordinates at a time. Feeding these structural fac

What carries the argument

The central mechanism is the low-complexity structure (Definition 2.4): an operator admits a decomposition DY ∘ G_k ∘ ... ∘ G_1 ∘ E_X where each G_i applies linear projections followed by nonlinear functions of at most d_i inputs. This decomposition lets the paper invoke a general theorem converting Lipschitzness and structural width into explicit bounds on generalization error in terms of d_max and ℓ_max. The second load-bearing ingredient is a set of Lipschitz estimates for the numerical time-stepping operators, proved with semigroup non-expansiveness for implicit Euler, maximum-principle arguments for Crank-Nicolson, and an energy estimate for viscous conservation laws under a viscosity c

Load-bearing premise

The multi-input viscous conservation-law result depends on the viscosity coefficient being large enough relative to the Lipschitz constant of the flux, and on the claim that the flux-coupled time-stepping operator is Lipschitz and admits the same low-complexity decomposition as the single-input case; the paper only sketches this Lipschitz estimate by saying it may be obtained by modifying an earlier inequality.

What would settle it

Take a multi-input viscous Burgers-type conservation law with flux u^2/2, encode the flux by p basis coefficients and the initial data on a grid of size d_X1, and train a ReLU network on implicit-Euler/Picard data while roughly doubling d_X1 and d_X2. If the sample size needed to reach a fixed generalization error grows faster than polynomially, or if the empirical error versus n decays with a slope worse than the predicted -1/2, the claimed absence of the curse of dimensionality would be contradicted.

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

If this is right

  • Learning an implicit time-stepping map with ReLU networks is sample-efficient: error scales polynomially in the encoding dimension and as n^{-1/2} or n^{-2/3}, so refining the spatial mesh does not force an exponential increase in training data.
  • Once trained, the network can march solutions forward without per-step nonlinear solves, and the global error over N steps is bounded by the usual first-order numerical error plus the accumulated per-step learning error.
  • Multi-input operators mapping both initial data and reaction/forcing terms into the next state enjoy the same no-curse-of-dimensionality bounds, enabling one network to predict evolution for a family of PDEs without retraining.
  • The conditions Δt L_f < 1 for Picard iterations and κ ≥ (1+L_f)/2 for conservation laws appear explicitly, and smaller time steps or stronger viscosity lower the generalization error.
  • The same structural analysis covers implicit Euler with Picard and with a quadratically convergent solver, Crank-Nicolson, and first-order IMEX schemes for viscous conservation laws, so the guarantees apply to schemes actually used in practice.

Where Pith is reading between the lines

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

  • A natural extension of the argument would be to higher-order Runge-Kutta or IMEX schemes: if their time-stepping operators also admit low-complexity decompositions with bounded d_max, the same theorem should yield polynomial sample-complexity bounds, though this is not proved in the paper.
  • The paper's non-Lipschitz example for explicit Euler suggests that the stability of implicit methods is doing real theoretical work: a learned explicit stepper would need a different analytical framework, not just the same Lipschitz-plus-structure argument.
  • The theory makes a testable quantitative prediction: for fixed PDE data, log-log error versus sample size should show slopes close to -1/2 or -2/3, and error growth in the encoding dimension should be polynomial; experiments on multi-input conservation laws below the stated viscosity threshold should break the bound, confirming the role of the condition.
  • The explicit network constructions in the proofs could be read as architectural guidance: letting network depth scale with the number of solver iterations and width scale with d_X^2 should realize the time-stepping operator directly.
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

4 major / 6 minor

Summary. The paper studies supervised learning of one-step time-stepping operators for evolutionary PDEs. After recalling a general generalization bound for low-complexity operators (Theorem 2.6, due to the authors' prior work [16]), it derives Lipschitz estimates and low-complexity decompositions for implicit Euler with Picard and Newton iterations for reaction-diffusion equations, implicit Euler and Crank-Nicolson for parabolic equations with forcing, and implicit Euler with Picard iterations for viscous conservation laws. Theorems 3.7, 3.9, 3.20, 4.2, 4.4, 4.5, 5.5, and 5.8 state rates n^{-1/2} or n^{-2/3} with polynomial dependence on the encoding dimension, and Section 6 proposes an accumulated-error bound. The paper claims that the examined time-stepping operators are efficiently learnable without the curse of dimensionality.

Significance. If established, the results would be a valuable contribution to rigorous operator learning for PDEs: they connect classical numerical-analysis estimates for semi-discrete schemes to sample-complexity bounds, and the identification of low-complexity structures in standard implicit time-steppers is conceptually useful. The single-input reaction-diffusion and parabolic sections contain explicit, checkable constructions and are the strongest part; Proposition 5.2's energy estimate for the conservation-law case is also instructive. However, the multi-input conservation-law theorem, the global error propagation argument, and the transfer from the abstract low-complexity definition to concrete ReLU networks are not sufficiently supported. These gaps are load-bearing for the advertised no-Curse-of-Dimensionality claims, so the paper needs substantial revision.

major comments (4)
  1. [Section 5, Theorem 5.8] The multi-input conservation-law bound is not proved; the paragraph after the theorem is only a sketch. The multi-input energy equation contains the extra term -Δt∫ ∂x w^m [f(v^{m-1}) - g(v^{m-1})] dx, which cannot be bounded by the single-input argument in Proposition 5.2. One must derive a control of ||f(v)-g(v)||_{L2} in terms of ||f-g||_{C^1} and then obtain the stated viscosity condition κ ≥ (1+L_f)/2 by a Young-inequality estimate. In addition, the claimed low-complexity realization with ℓmax = dX1 dX2 + dX1 + dX2 is not established: computing ∂x f(u) from the coefficients of f requires assembling f'(u) from basis terms and then multiplying by Tu, so the two-layer block of Proposition 5.4 does not carry over verbatim. Without these two steps, the theorem's conclusion does not follow.
  2. [Section 5, Theorem 5.5] There is a norm mismatch. The theorem states the target ΦP : L2(D) → L2(D), but Proposition 5.2 proves an H1-Lipschitz estimate, and Proposition 5.4's construction is H1-based. Theorem 2.6 requires the operator to be Lipschitz in the norms of the chosen spaces X and Y. Thus the proof as written is invalid. The authors should either prove an L2-Lipschitz estimate (which the energy inequality (5.9) can support when κ ≥ Lf/2, since the H1 term is nonnegative) or restate the theorem with H1 spaces and the corresponding encoder/decoder assumptions.
  3. [Section 6, Proposition 6.2] The global accumulation bound does not follow from the definition of Egen as an expectation over the fixed initial measure γ. In the recursive decomposition of the second term in (6.5), one encounters expressions of the form E_{u0∼γ}[||Φ^{N-1}_{Δt}(Φ_NN(u0)) - Φ^{N-1}_{NN}(Φ_NN(u0))||], where the inner argument Φ_NN(u0) is not distributed according to γ. The one-step generalization error Egen cannot be inserted at that point without a uniform (worst-case) one-step error or a quantitative control of the distribution shift. As stated, inequality (6.2) is unproved.
  4. [Section 2.3, Theorem 2.6 and Definition 2.4] The paper imports Theorem 2.6 from [16] but does not state the assumptions on the component nonlinearities g_j^i in Definition 2.4. In the applications these functions include monomials (x,y) → y x^k, division for the LU solver, and the flux f and its derivative; none of these are ReLU networks. Theorem 2.6 as written has no term for approximating the g_j^i, so it cannot apply to arbitrary low-complexity structures. The manuscript should either quote the precise conditions from [16] (e.g., Barron regularity or explicit ReLU implementation of each g with controlled size) and verify them for the specific nonlinearities, or include the resulting approximation terms in the error bounds. This is a load-bearing issue for every theorem that invokes Theorem 2.6.
minor comments (6)
  1. [Throughout] The symbol L_p is used both for the network depth parameter in Theorem 2.6 and for the Lipschitz constant in Theorems 3.7, 3.9, and 5.5. Please disambiguate.
  2. [Theorem 5.8] The statement 'Assumptions 2.1, 5.1–5.7' is not meaningful because 5.4 is a proposition and 5.6 is a remark. The intended assumptions are 5.1, 5.3, and 5.7.
  3. [Theorems 3.20, 4.2, 4.4] The M-bound sometimes uses LEX where the framework requires LEY; please check the constants and make them consistent.
  4. [Theorem 3.9] The proof says 'follows the same idea as Theorem 4.2', but Theorem 4.2 appears later. It likely means Theorem 3.7.
  5. [Section 3, Proposition 3.6] The proof references 'Figure 1', but no figure is actually included in the manuscript text.
  6. [Section 3.1] The number of Taylor basis functions p is overloaded with the network width p in FNN(d, L, p, M). This is a potential source of confusion.

Circularity Check

0 steps flagged

No circular reduction found; central bounds quote a general prior theorem by the same authors, but the PDE-specific assumption checks are independent and no prediction is fitted.

full rationale

Section 2.3 imports Theorem 2.6 from the authors' own prior work [16]: 'The following theorem from [16] bounds this error under the stated assumptions.' This theorem is logically load-bearing for every subsequent generalization bound. I do not count it as circular, however, because it is a parameter-free general operator-learning bound whose stated Assumptions 2.1-2.5 (compact support, fixed Lipschitz encoders/decoders, Lipschitz operator, low-complexity structure) are not the PDE-specific conclusions of this paper. Sections 3-5 independently verify these assumptions for reaction-diffusion (Propositions 3.2, 3.6, 3.15, 3.18-3.19), parabolic (Theorems 4.2, 4.4, 4.5), and conservation-law operators (Propositions 5.2, 5.4). Each final bound is an instance of Theorem 2.6 after those verifications; no parameter is fitted to a subset of the predicted quantity and no result is defined in terms of its own conclusion. The only notable weakness is Theorem 5.8, whose proof is a sketch: 'A Lipschitz estimate on the operator may be obtained by modifying (5.7)... The low complexity structure follows as in Proposition 3.7.' This is an omitted/terse argument (and 'Proposition 3.7' appears to be a typo for Proposition 3.6), but an incomplete proof is a rigor gap, not a circular reduction; the asserted bound does not become an input by construction. The paper also explicitly discloses the inviscid case as out of scope in Section 7. Hence no fully circular step is exhibited; the score of 2 reflects only the load-bearing self-citation, which under the stated rules is independent support rather than circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 8 axioms · 0 invented entities

No fitted parameters: the bounds involve theoretical constants (Lipschitz constants, m, ∆t) that are inputs to the theorem, not fitted values. The main unproved inputs are the cited theorem [16] and the ad hoc structural constructions for the multi-input conservation law.

axioms (8)
  • standard math Theorem 2.6 from [16] (Chen-Wang-Yang) provides the generalization bound for operators with low-complexity structure.
    Invoked in Section 2.3 and used for every subsequent error bound; treated as a black box.
  • standard math Heat semigroup representation and maximum principle imply non-expansiveness of [I - ∆t∆]^{-1} (Lemma 3.1).
    Used in Lemma 3.1 and throughout Sections 3-4.
  • standard math Banach fixed-point theorem guarantees a unique solution for the Picard map when ∆t L_p < 1 (Remark 3.3, Proposition 3.2).
    Basis for the Lipschitz estimate of the Picard solver.
  • standard math Kantorovich theorem ensures convergence of Newton's method under Assumption 3.14 (Section 3.2.2).
    Needed for Proposition 3.15 to hold.
  • domain assumption Reaction function f' ≤ 0 (Assumption 3.11) gives invertibility of DΦ and the bound ||[DΦ]^{-1}|| ≤ 2.
    Restricts to monotone decreasing reactions; used in Lemma 3.12.
  • domain assumption Viscosity κ ≥ L_f/2 (or (1+L_f)/2 for multi-input) makes the energy estimate in Proposition 5.2 non-degenerate.
    If the diffusion term is smaller, the recursive inequality loses the positive H1 term and the Lipschitz proof fails.
  • domain assumption Compact support of the data measure and Lipschitz encoder/decoder pairs (Assumptions 2.1-2.2).
    Standard in operator learning frameworks; controls projection error and boundedness.
  • ad hoc to paper Low-complexity structure of each scheme can be realized with the stated dmax, ℓmax (Definition 2.4, Propositions 3.6, 3.19, 5.4).
    Verified by constructing layers for each algorithm; for the multi-input conservation law this construction is only sketched.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Error analysis for learning the time-stepping operator of evolutionary PDEs." pith.science (2026). https://pith.science/paper/LQLKOOQO

@misc{pith2026250904256,
  author       = {Pith},
  title        = {Pith review of: Error analysis for learning the time-stepping operator of evolutionary PDEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQLKOOQO}},
  note         = {Machine review of arXiv:2509.04256}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deep neural networks (DNNs) have recently emerged as effective tools for approximating solution operators of partial differential equations (PDEs) including evolutionary problems. Classical numerical solvers for such PDEs often face challenges of balancing stability constraints and the high computational cost of iterative solvers. In contrast, DNNs offer a data-driven alternative through direct learning of time-stepping operators to achieve this balancing goal. In this work, we provide a rigorous theoretical framework for analyzing the approximation of these operators using feedforward neural networks (FNNs). We derive explicit error estimates that characterize the dependence of the approximation error on the network architecture -- namely its width and depth -- as well as the number of training samples. Furthermore, we establish Lipschitz continuity properties of time-stepping operators associated with classical numerical schemes and identify low-complexity structures inherent in these operators for several classes of PDEs, including reaction-diffusion equations, parabolic equations with external forcing, and scalar conservation laws. Leveraging these structural insights, we obtain generalization bounds that demonstrate efficient learnability without incurring the curse of dimensionality. Finally, we extend our analysis from single-input operator learning to a general multi-input setting, thereby broadening the applicability of our results.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. From Spectral Methods to Sample Complexity Bounds for Fourier Neural Operators

    stat.ML 2026-07 unverdicted novelty 6.0

    FNOs achieve polynomial sample complexity for learning time-T solution operators of dissipative evolution equations when those operators admit stable spectral discretizations, with rates depending on smoothness, dimen...

Reference graph

Works this paper leans on

80 extracted references · 69 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    Abadias and E

    L. Abadias and E. Alvarez. Asymptotic behavior for the discrete in time heat equation. Math- ematics, 10(17):3128, 2022

  2. [2]

    S. M. Allen and J. W. Cahn. Ground state structures in ordered binary alloys with second neighbor interactions. Acta Metallurgica, 20(3):423–433, 1972

  3. [3]

    Andrychowicz, M

    M. Andrychowicz, M. Denil, S. Gomez, M. W. Hoffman, D. Pfau, T. Schaul, B. Shillingford, and N. De Freitas. Learning to learn by gradient descent by gradient descent. Advances in neural information processing systems , 29, 2016

  4. [4]

    U. M. Ascher, S. J. Ruuth, and R. J. Spiteri. Implicit-explicit runge-kutta methods for time- dependent partial differential equations. Applied Numerical Mathematics, 25(2-3):151–167, 1997

  5. [5]

    K. A. Atkinson. An Introduction to Numerical Analysis . John Wiley & Sons, New York, 2nd edition, 1989

  6. [6]

    F. Bach. Breaking the curse of dimensionality with convex neural networks. Journal of Machine Learning Research, 18(19):1–53, 2017

  7. [7]

    A. Barron. Universal approximation bounds for superposition of a sigmoidal function. IEEE Transaction on Information Theory , 19:930–944, 1991

  8. [8]

    P. W. Battaglia, J. B. Hamrick, V. Bapst, A. Sanchez-Gonzalez, V. Zambaldi, M. Malinowski, A. Tacchetti, D. Raposo, A. Santoro, R. Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261 , 2018

  9. [9]

    Bauer and M

    B. Bauer and M. Kohler. On deep learning as a remedy for the curse of dimensionality in nonparametric regression. projecteuclid.org, 2019

  10. [10]

    Berestycki, L

    H. Berestycki, L. Rossi, et al. Reaction-diffusion equations for population dynamics with forced speed i-the case of the whole space.Discrete and Continuous Dynamical Systems, 21(1):41– 67, 2008

  11. [11]

    Berman and B

    J. Berman and B. Peherstorfer. Randomized sparse neural galerkin schemes for solving evolution equations with deep networks. Advances in Neural Information Processing Systems , 36, 2024

  12. [12]

    Bhattacharya, B

    K. Bhattacharya, B. Hosseini, N. B. Kovachki, and A. M. Stuart. Model reduction and neural networks for parametric pdes. The SMAI journal of computational mathematics , 7:121– 157, 2021

  13. [13]

    A. Bressan. Hyperbolic systems of conservation laws: the one-dimensional Cauchy problem , volume 20. OUP Oxford, 2000

  14. [14]

    J. C. Butcher. Implicit runge-kutta processes. Mathematics of computation , 18(85):50–64, This manuscript is for review purposes only. 24 M. KRISHNAN, K. CHEN, AND H. YANG 1964

  15. [15]

    H. Chen, R. Wu, E. Grinspun, C. Zheng, and P. Y. Chen. Implicit neural spatial representations for time-dependent pdes. In International Conference on Machine Learning , pages 5162–

  16. [16]

    K. Chen, C. Wang, and H. Yang. Deep operator learning lessens the curse of dimensionality for pdes. arXiv preprint arXiv:2301.12227 , 2023

  17. [17]

    M. Chen, H. Jiang, W. Liao, and T. Zhao. Efficient approximation of deep relu networks for functions on low dimensional manifolds. Advances in neural information processing systems, 32, 2019

  18. [18]

    R. T. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems , 31, 2018

  19. [19]

    Chen and H

    T. Chen and H. Chen. Approximations of continuous functionals by neural networks with application to dynamic systems. IEEE Transactions on Neural networks , 4(6):910–918, 1993

  20. [20]

    Chen and D

    Y. Chen and D. Xiu. Learning stochastic dynamical system via flow map operator. Journal of Computational Physics , 508:112984, 2024

  21. [21]

    Z. Chen, J. Lu, and Y. Lu. On the representation of solutions to elliptic pdes in barron spaces. Advances in neural information processing systems , 34:6454–6465, 2021

  22. [22]

    Z. Chen, J. McCarran, E. Vizcaino, M. Soljaˇ ci´ c, and D. Luo. Teng: Time-evolving natural gradient for solving pdes with deep neural nets toward machine precision. arXiv preprint arXiv:2404.10771, 2024

  23. [23]

    Cockburn, F

    B. Cockburn, F. Coquel, and P. LeFloch. An error estimate for finite volume methods for multidimensional conservation laws. mathematics of computation , 63(207):77–103, 1994

  24. [24]

    R. M. Colombo. Hyperbolic phase transitions in traffic flow. SIAM Journal on Applied Math- ematics, 63(2):708–721, 2003

  25. [25]

    Darbon, G

    J. Darbon, G. P. Langlois, and T. Meng. Overcoming the curse of dimensionality for some hamilton–jacobi partial differential equations via neural network architectures. Research in the Mathematical Sciences , 7(3):20, 2020

  26. [26]

    D. R. Durran. Numerical methods for wave equations in geophysical fluid dynamics, volume 32. Springer Science & Business Media, 2013

  27. [27]

    Engel, R

    K.-J. Engel, R. Nagel, and S. Brendle. One-parameter semigroups for linear evolution equa- tions, volume 194. Springer, 2000

  28. [28]

    Y. Geng, Y. Teng, Z. Wang, and L. Ju. A deep learning method for the dynamics of clas- sic and conservative allen-cahn equations based on fully-discrete operators. Journal of Computational Physics , 496:112589, 2024

  29. [29]

    J. Han, A. Jentzen, and W. E. Solving high-dimensional partial differential equations using deep learning. Proceedings of the National Academy of Sciences, 115(34):8505–8510, 2018

  30. [30]

    E. Hopf. The partial differential equation ut + uux − νuxx. Communications on Pure and Applied Mathematics, 3:201–230, 1950

  31. [31]

    Hu and R

    J. Hu and R. Shu. On the uniform accuracy of implicit-explicit backward differentiation formu- las (imex-bdf) for stiff hyperbolic relaxation systems and kinetic equations. Mathematics of Computation , 90(328):641–670, 2021

  32. [32]

    Hutzenthaler, A

    M. Hutzenthaler, A. Jentzen, T. Kruse, and T. A. Nguyen. A proof that rectified deep neural networks overcome the curse of dimensionality in the numerical approximation of semilinear heat equations. SN partial differential equations and applications , 1(2):10, 2020

  33. [33]

    Jiang, M

    Z. Jiang, M. Zhu, and L. Lu. Fourier-mionet: Fourier-enhanced multiple-input neural operators for multiphase modeling of geological carbon sequestration. Reliability Engineering & System Safety , 251:110392, 2024

  34. [34]

    P. Jin, S. Meng, and L. Lu. Mionet: Learning multiple-input operators via tensor product. SIAM Journal on Scientific Computing , 44(6):A3490–A3514, 2022

  35. [35]

    Kolmogorov, I

    A. Kolmogorov, I. Petrovskii, and N. Piskunov. A study of the diffusion equation with in- crease in the amount of substance, and its application to a biological problem. In V. M. Tikhomirov, editor, Selected Works of A. N. Kolmogorov I , pages 248–270. Kluwer, 1991. Translated from Bull. Moscow Univ., Math. Mech. 1, 1–25, 1937

  36. [36]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, and A. Anandku- mar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research , 24(89):1–97, 2023

  37. [37]

    N. B. Kovachki, S. Lanthaler, and H. Mhaskar. Data complexity estimates for operator learning, 2024

  38. [38]

    Krushkov

    S. Krushkov. First order quasilinear equations with several space variables. Math. USSR. Sb , 10:217–243, 1970

  39. [39]

    Kuznetsov

    N. Kuznetsov. Accuracy of some approximate methods for computing the weak solutions of This manuscript is for review purposes only. LEARNING TIME-STEPPING ALGORITHMS FOR PDES 25 a first-order quasi-linear equation. USSR Computational Mathematics and Mathematical Physics, 16(6):105–119, 1976

  40. [40]

    I. E. Lagaris, A. Likas, and D. I. Fotiadis. Artificial neural networks for solving ordinary and partial differential equations. IEEE transactions on neural networks , 9(5):987–1000, 1998

  41. [41]

    Lanthaler

    S. Lanthaler. Operator learning with pca-net: upper and lower complexity bounds. Journal of Machine Learning Research, 24(318):1–67, 2023

  42. [42]

    R. J. LeVeque. Finite volume methods for hyperbolic problems, volume 31. Cambridge university press, 2002

  43. [43]

    R. J. LeVeque and R. J. Leveque. Numerical methods for conservation laws , volume 214. Springer, 1992

  44. [44]

    Li and J

    K. Li and J. Malik. Learning to optimize. arXiv preprint arXiv:1606.01885 , 2016

  45. [45]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandku- mar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895, 2020

  46. [46]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anand- kumar. Neural operator: Graph kernel network for partial differential equations. arXiv preprint arXiv:2003.03485, 2020

  47. [47]

    Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, and A. Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. J. Mach. Learn. Res , 24(89):1–97, 2023

  48. [48]

    Z. Li, H. Zheng, N. Kovachki, D. Jin, H. Chen, B. Liu, K. Azizzadenesheli, and A. Anandkumar. Physics-informed neural operator for learning partial differential equations. ACM/JMS Journal of Data Science , 1(3):1–27, 2024

  49. [49]

    Liniger and R

    W. Liniger and R. A. Willoughby. Efficient integration methods for stiff systems of ordinary differential equations. SIAM Journal on Numerical Analysis , 7(1):47–66, 1970

  50. [50]

    H. Liu, H. Yang, M. Chen, T. Zhao, and W. Liao. Deep nonparametric estimation of operators between infinite dimensional spaces. Journal of Machine Learning Research , 25(24):1–67, 2024

  51. [51]

    L. Lu, P. Jin, and G. E. Karniadakis. Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193, 2019

  52. [52]

    L. Lu, P. Jin, G. Pang, Z. Zhang, and G. E. Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218–229, 2021

  53. [53]

    Mattey and S

    R. Mattey and S. Ghosh. A novel sequential method to train physics informed neural networks for allen cahn and cahn hilliard equations. Computer Methods in Applied Mechanics and Engineering, 390:114474, 2022

  54. [54]

    H. Mhaskar. Local approximation of operators, 2022

  55. [55]

    P. Monk. Finite element methods for Maxwell’s equations . Oxford university press, 2003

  56. [56]

    Nakada and M

    R. Nakada and M. Imaizumi. Adaptive approximation and generalization of deep neural network with intrinsic dimensionality. Journal of Machine Learning Research , 21(174):1–38, 2020

  57. [57]

    A. C. Newell and J. A. Whitehead. Finite bandwidth, finite amplitude convection. Journal of Fluid Mechanics, 38(2):279–303, sep 1969

  58. [58]

    Y. Z. Ong, Z. Shen, and H. Yang. Iae-net: Integral autoencoders for discretization-invariant learning. arXiv preprint arXiv:2203.05142 , 2022

  59. [59]

    Perthame

    B. Perthame. Kinetic formulation of conservation laws , volume 21. Oxford University Press, 2002

  60. [60]

    Poggio, H

    T. Poggio, H. Mhaskar, L. Rosasco, B. Miranda, and Q. Liao. Why and when can deep-but not shallow-networks avoid the curse of dimensionality: a review. International Journal of Automation and Computing , 14(5):503–519, 2017

  61. [61]

    T. Qin, K. Wu, and D. Xiu. Data driven governing equations approximation using deep neural networks. Journal of Computational Physics , 395:620–635, 2019

  62. [62]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics , 378:686–707, 2019

  63. [63]

    B. L. Rozhdestvenski and N. N. I Anenko. Systems of quasilinear equations and their appli- cations to gas dynamics , volume 55. American Mathematical Soc., 1983

  64. [64]

    R. Sanders. On convergence of monotone finite difference schemes with variable spatial differ- encing. mathematics of computation , 40(161):91–106, 1983

  65. [65]

    Scarselli, M

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini. The graph neural network model. IEEE transactions on neural networks , 20(1):61–80, 2008

  66. [66]

    M. H. Schultz. l∞-multivariate approximation theory. SIAM Journal on Numerical Analysis , This manuscript is for review purposes only. 26 M. KRISHNAN, K. CHEN, AND H. YANG 6(2):161–183, 1969

  67. [67]

    Z. Shen, H. Yang, and S. Zhang. Neural network approximation: Three hidden layers are enough. Neural Networks, 141:160–173, 2021

  68. [68]

    Z. Shen, H. Yang, and S. Zhang. Optimal approximation rate of relu networks in terms of width and depth. Journal de Math´ ematiques Pures et Appliqu´ ees, 157:101–135, 2022

  69. [69]

    Sirignano and K

    J. Sirignano and K. Spiliopoulos. Dgm: A deep learning algorithm for solving partial differential equations. Journal of computational physics , 375:1339–1364, 2018

  70. [70]

    G. A. Sod. A survey of several finite difference methods for systems of nonlinear hyperbolic conservation laws. Journal of computational physics , 27(1):1–31, 1978

  71. [71]

    R. Temam. Navier–Stokes equations: theory and numerical analysis , volume 343. American Mathematical Society, 2024

  72. [72]

    J.-P. Vila. Convergence and error estimates in finite volume schemes for general multidi- mensional scalar conservation laws. i. explicite monotone schemes. ESAIM: Mathematical Modelling and Numerical Analysis , 28(3):267–295, 1994

  73. [73]

    S. Wang, H. Wang, and P. Perdikaris. Learning the solution operator of parametric partial differential equations with physics-informed deeponets. Science advances, 7(40):eabi8605, 2021

  74. [74]

    Wojtowytsch and E

    S. Wojtowytsch and E. Weinan. Can shallow neural networks beat the curse of dimensionality? a mean field training perspective. IEEE Transactions on Artificial Intelligence , 1(2):121– 129, 2020

  75. [75]

    Yang and S

    L. Yang and S. J. Osher. Pde generalization of in-context operator networks: A study on 1d scalar nonlinear conservation laws. arXiv preprint arXiv:2401.07364 , 2024

  76. [76]

    Approximation of Functionals by Neural Network without Curse of Dimensionality

    Y. Yang and Y. Xiang. Approximation of functionals by neural network without curse of dimensionality. arXiv preprint arXiv:2205.14421 , 2022

  77. [77]

    Yarotsky

    D. Yarotsky. Optimal approximation of continuous functions by very deep relu networks. In Conference on learning theory , pages 639–649. PMLR, 2018

  78. [78]

    W. Zhai, D. Tao, and Y. Bao. Parameter estimation and modeling of nonlinear dynamical systems based on runge–kutta physics-informed neural network. Nonlinear Dynamics , 111(22):21117–21130, 2023

  79. [79]

    Zhang, Z

    S. Zhang, Z. Shen, and H. Yang. Deep network approximation: Achieving arbitrary accuracy with fixed number of neurons. Journal of Machine Learning Research, 23(276):1–60, 2022

  80. [80]

    M. Zhu, B. Chang, and C. Fu. Convolutional neural networks combined with runge–kutta methods. Neural Computing and Applications , 35(2):1629–1643, 2023. This manuscript is for review purposes only

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.