Pith. sign in

REVIEW 3 major objections 2 minor 21 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · grok-4.3

A hybrid low-regularity integrator augmented by a neural correction trained end-to-end achieves global error C(ε_net + δ) τ^γ ln(1/τ) for nonlinear dispersive PDEs.

2026-07-01 07:42 UTC pith:ZVO3SCT4

load-bearing objection The hybrid low-regularity integrator plus latent neural correction with explicit O(τ) scaling and Bourgain-norm trajectory training is the actual novelty, but the global error bound stays conditional on unshown derivations for stability preservation and training success. the 3 major comments →

arxiv 2605.04853 v3 pith:ZVO3SCT4 submitted 2026-05-06 cs.LG

Hybrid Iterative Neural Low-Regularity Integrator for Nonlinear Dispersive Equations

Zhangyong Liang , Huanhuan Gao This is my paper
classification cs.LG
keywords hybrid numerical methodneural operatorlow-regularity integratornonlinear dispersive PDEsolver-in-the-loop trainingtruncation error correctionGronwall stabilityrough data
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 paper introduces HIN-LRI, which starts from a classical first-order low-regularity integrator for nonlinear dispersive equations and adds a lightweight neural operator that learns the residual truncation defect on a low-dimensional latent space. An explicit scaling of the neural term by the time step keeps its Lipschitz constant O(τ), so the overall stability factor from Gronwall remains bounded uniformly in step size and independent of spatial mesh. Training unrolls the full iteration and penalizes accumulated trajectory error in a Bourgain-type norm rather than single-step targets. Under the stated assumptions the resulting global error bound is C(ε_net + δ) τ^γ ln(1/τ). Numerical tests on three dispersive benchmarks with rough initial data show the scheme outperforming pure analytical integrators, splitting methods, and standalone neural PDE models while remaining stable under spatial refinement.

Core claim

Under the paper's assumptions the global error of the hybrid scheme satisfies C(ε_net + δ) τ^γ ln(1/τ), where ε_net quantifies the network's approximation quality and δ its training shortfall; the explicit time-step scaling of the neural correction guarantees that its contribution to the Lipschitz constant stays O(τ) and therefore does not spoil the uniform Gronwall factor of the underlying low-regularity integrator.

What carries the argument

Solver-in-the-loop training of a latent-space neural correction whose explicit time-step scaling keeps its Lipschitz contribution O(τ) while the base low-regularity integrator supplies the consistent first-order term.

Load-bearing premise

The trained network must keep both its approximation error ε_net and its training shortfall δ small enough that the product with the logarithmic factor still yields the desired accuracy, while the explicit scaling truly confines its Lipschitz contribution to O(τ).

What would settle it

Run the three reported dispersive benchmarks at successively halved time steps and check whether the observed global error in the Bourgain-type norm fails to track the predicted C(ε_net + δ) τ^γ ln(1/τ) scaling once ε_net and δ are measured independently.

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

If this is right

  • Accuracy improves over analytical integrators, splitting methods, and neural PDE surrogates on dispersive problems with rough data.
  • The scheme remains stable under spatial mesh refinement because the Gronwall factor is independent of spatial resolution.
  • Out-of-distribution transfer works because the latent correction is learned on the solver trajectory rather than isolated steps.
  • Online computational overhead stays modest once the network is trained.

Where Pith is reading between the lines

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

  • The same solver-in-the-loop construction could be applied to other families of structure-preserving integrators beyond low-regularity methods.
  • If the latent dimension can be chosen adaptively, the approach might extend to equations whose solution manifolds change dimension over time.
  • The Bourgain-norm training objective suggests a route for incorporating dispersive smoothing estimates directly into the loss without explicit Fourier analysis at inference time.

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 / 2 minor

Summary. The manuscript proposes HIN-LRI, a hybrid method that augments a classical low-regularity integrator for nonlinear dispersive PDEs with a lightweight neural operator acting on a low-dimensional latent manifold to correct structured truncation error. An explicit time-step scaling is applied to the neural correction to keep its Lipschitz factor O(τ), and the network is trained end-to-end via a solver-in-the-loop objective that unrolls the iteration and penalizes trajectory error in a Bourgain-type norm. Under stated assumptions the global error is claimed to satisfy C(ε_net + δ) τ^γ ln(1/τ), where ε_net measures network approximation quality and δ the training shortfall; experiments on three dispersive benchmarks with rough data are said to show gains over analytical integrators, splitting methods, and neural PDE surrogates.

Significance. If the error bound can be established rigorously and the training reliably achieves the required ε_net and δ, the framework would supply a concrete route to improve consistency of low-regularity integrators via learned corrections while preserving uniform-in-τ stability independent of spatial mesh size. The solver-in-the-loop objective and explicit scaling are positive design choices that align learning with multi-step dynamics rather than one-step targets.

major comments (3)
  1. [Abstract / error analysis] Abstract and error-analysis section: the global error bound C(ε_net + δ) τ^γ ln(1/τ) is asserted, yet no derivation is supplied showing that the explicit τ-scaling of the neural correction preserves the consistency order γ of the base low-regularity integrator or keeps the Gronwall multiplier bounded uniformly in τ and independent of spatial resolution. This step is load-bearing for the central theoretical claim.
  2. [Training objective] Training section: the claim that the unrolled solver-in-the-loop loss in the Bourgain-type norm controls the defect δ uniformly for rough initial data (and drives ε_net below the threshold needed for the bound to be informative) is stated without supporting estimates or analysis of how the loss controls the defect across time steps or mesh sizes.
  3. [Experiments] Experiments: the abstract asserts quantitative improvement and stable spatial refinement, but the manuscript supplies neither tabulated error values, baseline comparisons, nor verification that the O(τ) Lipschitz condition and small ε_net, δ are attained on the reported benchmarks.
minor comments (2)
  1. [Preliminaries] Notation for the Bourgain-type norm and the precise definition of the latent manifold dimension should be introduced earlier and used consistently.
  2. [Error bound statement] The dependence of the constant C on the spatial mesh size and on the roughness parameter of the initial data should be stated explicitly.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive report and the identification of points where additional justification and documentation are required. We address each major comment below and will incorporate the necessary revisions.

read point-by-point responses

  1. Referee: [Abstract / error analysis] Abstract and error-analysis section: the global error bound C(ε_net + δ) τ^γ ln(1/τ) is asserted, yet no derivation is supplied showing that the explicit τ-scaling of the neural correction preserves the consistency order γ of the base low-regularity integrator or keeps the Gronwall multiplier bounded uniformly in τ and independent of spatial resolution. This step is load-bearing for the central theoretical claim.

    Authors: We agree that the error-analysis section would benefit from an explicit, self-contained derivation of how the O(τ) Lipschitz scaling of the neural correction preserves the base integrator's consistency order γ and yields a Gronwall factor bounded uniformly in τ and independent of spatial mesh size. The assumptions under which the bound holds are stated, but the intermediate steps were condensed. In the revision we will expand this section with the full argument, including the relevant stability estimates. revision: yes

  2. Referee: [Training objective] Training section: the claim that the unrolled solver-in-the-loop loss in the Bourgain-type norm controls the defect δ uniformly for rough initial data (and drives ε_net below the threshold needed for the bound to be informative) is stated without supporting estimates or analysis of how the loss controls the defect across time steps or mesh sizes.

    Authors: The solver-in-the-loop objective is constructed precisely to align the training with multi-step trajectory error rather than single-step residuals. We acknowledge that explicit estimates demonstrating uniform control of δ for rough data across varying time steps and mesh sizes are not supplied. In the revision we will add a short analysis (or, if the estimates prove technical, a clear statement of the conditions under which the loss is expected to control δ) together with additional numerical diagnostics on the attained ε_net and δ values. revision: partial

  3. Referee: [Experiments] Experiments: the abstract asserts quantitative improvement and stable spatial refinement, but the manuscript supplies neither tabulated error values, baseline comparisons, nor verification that the O(τ) Lipschitz condition and small ε_net, δ are attained on the reported benchmarks.

    Authors: We will augment the experiments section with tabulated L^2 and Bourgain-norm errors for all methods and benchmarks, explicit baseline comparisons, and verification plots or tables confirming that the O(τ) Lipschitz condition holds and that the realized ε_net and δ remain below the thresholds required by the error bound. These additions will make the quantitative claims fully reproducible from the reported data. revision: yes

Circularity Check

1 steps flagged

Global error bound stated directly as C(ε_net + δ) τ^γ ln(1/τ), reducing the main result to the fitted network quality

specific steps
  1. fitted input called prediction [Abstract]
    "Under stated assumptions, the global error satisfies C(ε_net+δ) τ^γ ln(1/τ), where ε_net measures the network approximation quality and δ the training shortfall."

    The claimed global error is written as proportional to ε_net and δ, the very quantities obtained from the solver-in-the-loop training. The bound therefore reduces to a scaled version of the training defect by construction, rather than yielding a parameter-free or externally verifiable rate.

full rationale

The paper's central theorem expresses the global error explicitly in terms of ε_net (network approximation quality) and δ (training shortfall). This makes the bound a direct restatement of the training inputs rather than an independent derivation. The O(τ) Lipschitz claim via explicit scaling is asserted to 'ensure' the Gronwall factor but is not shown to preserve the base integrator's order for rough data. No other circular steps found; the method is otherwise self-contained once the assumptions on ε_net and δ are granted.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The central claim rests on the neural network achieving small approximation error after training and on the validity of the O(τ) Lipschitz scaling for uniform Gronwall control.

free parameters (1)
  • neural network parameters
    Learned end-to-end to minimize multi-step trajectory error in the Bourgain norm.
axioms (2)
  • domain assumption The base low-regularity integrator supplies a consistent first-order approximation
    Invoked as the foundation that the neural correction augments.
  • standard math Gronwall inequality yields a uniform stability factor from the O(τ) Lipschitz bound
    Used to obtain the global error estimate independent of spatial resolution.
invented entities (1)
  • neural operator on low-dimensional latent manifold no independent evidence
    purpose: Learns the residual truncation defect
    Core new component introduced to close the analytical gap.

pith-pipeline@v0.9.1-grok · 5755 in / 1362 out tokens · 48739 ms · 2026-07-01T07:42:30.596010+00:00 · methodology

0 comments
read the original abstract

We propose HIN-LRI, a hybrid framework that augments a classical numerical solver with a neural operator trained to correct the solver's structured truncation error. A base low-regularity integrator provides a consistent first-order approximation to nonlinear dispersive PDEs, while a lightweight neural network, operating on a low-dimensional latent manifold, learns the residual defect that analytical methods cannot close. An explicit time-step scaling on the neural correction ensures that its Lipschitz contribution remains $\mathcal{O}(\tau)$, yielding a Gronwall stability factor bounded uniformly in the step size and independent of the spatial resolution. The network is trained end-to-end through a solver-in-the-loop objective that unrolls the full iteration and penalises trajectory error in a Bourgain-type norm, aligning learning with multi-step solver dynamics rather than isolated one-step targets. Under stated assumptions, the global error satisfies $C(\varepsilon_{net}+\delta)\,\tau^\gamma\ln(1/\tau)$, where $\varepsilon_{net}$ measures the network approximation quality and $\delta$ the training shortfall. Experiments on three dispersive benchmarks with rough data show that HIN-LRI improves accuracy over analytical integrators, splitting methods, and neural PDE surrogates, with stable spatial refinement, effective out-of-distribution transfer, and modest online overhead.

Figures

Figures reproduced from arXiv: 2605.04853 by Huanhuan Gao, Zhangyong Liang.

Figure 1
Figure 1. Figure 1: Schematic of the HIN-LRI alternating iteration. The LRI branch (orange) produces the predictor v (m) and exposes the resonance defect Edefect. The neural branch (blue) maps the latent residual to a time-scaled correction τ Hneural. The hybrid update (green) combines both terms as u (m) = v (m) + τ e. The bottom equation shows the corresponding predictor-correction decomposition. From the perspective of mac… view at source ↗
Figure 2
Figure 2. Figure 2 view at source ↗
Figure 5
Figure 5. Figure 5: fig. 5. The Fourier amplitude of the one-step error for ELRI2 grows as view at source ↗
Figure 3
Figure 3. Figure 3: An illustration of the HIN-LRI framework. Blue modules denote latent Neural Operator components (SN: Scaling Net Escale; TB: Trunk Basis Φ), while brown modules represent exact physical dispersion and base LRI discretization. The diamond branch activates the neural correction cycle only when m ≡ 0 (mod κ) (χm=1); otherwise the iteration reduces to a purely spectral update (χm=0). The dashed loop arrow indi… view at source ↗
Figure 4
Figure 4. Figure 4: Taylor expansion collapse on rough data (KdV & cubic NLS). Top: KdV (ETD1, Lawson1, RES1); bottom: cubic NLS (Lie, Strang, RES1, BS22); each row contrasts smooth (γ = 3.0) and rough (γ = 0.5) data. while HIN-LRI remains stable across all tested grids up to N = 4096 view at source ↗
Figure 5
Figure 5. Figure 5: KdV LRI schemes across regularity levels. (a)–(b) Convergence at γ ∈ {0.30, 1.50}; (c) empirical order across γ (shaded: γ < 1); (d) Fourier error spectrum at γ = 0.40, τ = 0.05, with ELRI2 amplitude ∼ |k| 3−γ−1/2 view at source ↗
Figure 6
Figure 6. Figure 6: Algebraic rigidity and combinatorial explosion. (a) Factorization residual under perturbation ε ∈ {0, 10−3 , 10−2 , 10−1}; (b)–(c) FFT and brute-force convo￾lution scaling, with per-step conv1 counts and Catalan extrapolation (p = 3, 4, 5); (d) per-step wall-clock across N for all five methods. the filter cap, a 69% degradation that is more severe than Strang’s 61% drop (from 2.18 to 0.85). HIN-LRI maintai… view at source ↗
Figure 7
Figure 7. Figure 7: ULRI logarithmic/CFL defects on rough H0.5 KdV data. (a) τ - convergence with the τ γ ln(1/τ ) envelope; (b) L 2 error across N at τ = 10−3 ; (c) pseudodifferential amplitudes and spurious low-frequency energy injection; (d) FFT transform counts and wall-clock timing view at source ↗
Figure 8
Figure 8. Figure 8: Filter-induced convergence cap in NLS-RES1 and BS22. (a)–(b) RES1 convergence curves for γ ∈ {0.25, 0.5, 1.0, 2.0} and spectral attenuation |φ1(−2hk2 )|. (c) Empirical order across γ for RES1 and BS22 with min(1, γ) cap curves. (d) Convergence comparison of RES1 and BS22 at γ = 0.5 view at source ↗
Figure 9
Figure 9. Figure 9: Implicit curse: structure drift of explicit LRI methods (cubic NLS, smooth datum). (a)–(b) Normalized mass M(t)/M(0) and Hamiltonian H(t)/|H(0)| over T = 20 at τ = 0.05 (Lie, Strang, RES1, BS22). (c)–(d) Log-log drift of M and H across τ at T = 5. 5.6 Comparison with Neural PDE Solvers We compare HIN-LRI against three representative purely data-driven neural PDE solvers on the cubic NLS benchmark (γ = 0.5,… view at source ↗
Figure 10
Figure 10. Figure 10: KdV: HIN-LRI against RES1 on rough H0.5 data (N = 1024, T = 1.0). (a) τ -convergence with the τ γ ln(1/τ ) reference envelope; (b) L 2 error across N at τ = 10−3 . (c)–(d) Spatial absolute-error profile at τ = 2−8 and SITL training loss across epochs. 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 Sobolev regularity 0.0 0.5 1.0 1.5 2.0 Observed convergence order (a) first-order ceiling min(1, ) second-order ceil… view at source ↗
Figure 11
Figure 11. Figure 11: KdV: HIN-LRI against ELRI1/ELRI2 on rough H0.5 data (N = 1024, T = 1.0). (a)–(b) τ -convergence and empirical order across regularity γ. (c) L 2 error across N at τ = 2−8 ; (d) per-step wall-clock time across N. 5.7 Ablation Study We systematically ablate the key components of HIN-LRI on the KdV equation (γ = 0.5, τ = 2−8 , N = 1024). The ablated variants are: (A) base RES1 without any neural cor￾rection;… view at source ↗
Figure 12
Figure 12. Figure 12: Cubic NLS: HIN-LRI against operator-splitting and BS22 on rough H0.5 data. (a)–(b) τ -convergence at N = 1024 and L 2 error across N at τ = 2−8 . (c)–(d) Empirical order across γ and L 2 error across the number of time steps. 10 2 Time step 10 5 10 4 10 3 10 2 10 1 R elativ e L 2 error (a) OS1 BS22 HIN-LRI = 0.5 = 1.0 = 2.0 OS1 BS22 HIN-LRI 10 8 10 7 10 6 10 5 10 4 10 3 10 2 10 1 10 0 Magnitude (log scale… view at source ↗
Figure 13
Figure 13. Figure 13: Cubic NLS: structure preservation over T = 100 on rough H0.5 data. (a) Normalized mass M(t)/M(0); (b) normalized Hamiltonian H(t)/|H(0)| (Strang, HIN-LRI, RES1, implicit LRI). (c)–(d) Spatiotemporal evolution |u(x, t)| and discrete L∞ solution norm over time view at source ↗
Figure 14
Figure 14. Figure 14: Quadratic NLS: joint (τ, N) convergence landscape, γ = 0.5. Cell colour encodes L 2 error (log scale): the filtered integrator (top) saturates at small τ ; ULRI (middle) diverges beyond the CFL diagonal τ = O(N −2 ); HIN-LRI (bottom) stays uniformly low over the entire grid. contributes 1.8×; SITL re-optimisation (D vs. E) contributes 2.4×. The full HIN-LRI achieves 62× lower error than the base RES1 view at source ↗
Figure 15
Figure 15. Figure 15: HIN-LRI against neural PDE solvers: L 2 error across τ (cubic NLS, γ = 0.5, N = 1024). Neural solvers (FNO, PINN, DeepONet) yield flat error curves independent of τ ; HIN-LRI tracks O(τ ) across all tested step sizes view at source ↗
Figure 16
Figure 16. Figure 16: OOD transfer: ULRI failure against HIN-LRI stability (KdV, Rie￾mann step datum, τ = 2−8 , N = 512). (a)–(c) Solution profiles at T = 1.0 for ULRI, HIN-LRI (zero-shot), and HIN-LRI (mini-retrained, 10 SITL steps). (d) Fourier amplitude of the absolute error for all three methods. 5.9 Long-time Invariant Preservation and Total Computational Time We integrate the rough KdV and cubic NLS wave profiles up to T… view at source ↗
Figure 17
Figure 17. Figure 17: Long-time stability (T = 100, rough H0.5 data, τ = 2−8 , N = 1024). (a)– (b) Spatiotemporal heatmaps |u(x, t)| for HIN-LRI and the unfiltered integrator. (c) Relative Hamiltonian drift ∆H(t) for HIN-LRI, ULRI, and the fully implicit LRI. (d) Relative mass drift ∆M(t) for HIN-LRI, Strang, ULRI, and RES1 view at source ↗

Review history (3 revisions) →

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

21 extracted references · 21 canonical work pages · 3 internal anchors

  1. [1]

    ISSN 0377-0427. B. Azulay and M. Treister. A multigrid approach for training deep neural networks.arXiv preprint arXiv:2201.07470,

    work page arXiv

  2. [2]

    Yvonne Alama Bronsard, Yvain Bruned, Georg Maierhofer, and Katharina Schratz

    doi: 10.1093/imanum/drad093. Yvonne Alama Bronsard, Yvain Bruned, Georg Maierhofer, and Katharina Schratz. Sym- metric resonance based integrators and forest formulae.Foundations of Computational Mathematics,

    work page doi:10.1093/imanum/drad093

  3. [3]

    Published online

    doi: 10.1007/s10208-026-09742-0. Published online

    work page doi:10.1007/s10208-026-09742-0

  4. [4]

    Jiachuan Cao, Buyang Li, and Yanping Lin

    doi: 10.1017/fmp.2021.13. Jiachuan Cao, Buyang Li, and Yanping Lin. A new second-order low-regularity integrator for the cubic nonlinear schr¨ odinger equation.IMA Journal of Numerical Analysis, 44(3): 1313–1345,

    work page doi:10.1017/fmp.2021.13 2021

  5. [5]

    SH Chan, X Wang, and OA Elgendy

    doi: 10.1093/imanum/drad017. SH Chan, X Wang, and OA Elgendy. Plug-and-play admm for image restoration.IEEE Transactions on Computational Imaging,

    work page doi:10.1093/imanum/drad017

  6. [6]

    45 Zhangyong Liang Y

    doi: 10.1109/72.392253. 45 Zhangyong Liang Y. Cui, S. Wang, and L. Zheng. A coarse-grid correction method built on fourier neural network for solving partial differential equations.Journal of Computational Mathematics, 1,

    work page doi:10.1109/72.392253

  7. [7]

    Yue Feng, Georg Maierhofer, and Chushan Wang

    doi: 10.1090/mcom/3922. Yue Feng, Georg Maierhofer, and Chushan Wang. Explicit symmetric low-regularity inte- grators for the nonlinear schr¨ odinger equation.arXiv preprint arXiv:2411.07720,

    work page doi:10.1090/mcom/3922

  8. [8]

    Rectangles in latin squares

    Jun-Ting Hsieh, Shengjia Zhao, Stephan Eickenberg, Arthur Balvert, and Qiaoning Liao. Learning core design for multigrid methods.arXiv preprint arXiv:1902.05656,

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  9. [9]

    A fourier integrator for the cubic nonlinear schr¨ odinger equation with rough initial data.SIAM Journal on Numerical Analysis, 57(4):1967–1986,

    Marvin Kn¨ oller, Alexander Ostermann, and Katharina Schratz. A fourier integrator for the cubic nonlinear schr¨ odinger equation with rough initial data.SIAM Journal on Numerical Analysis, 57(4):1967–1986,

    work page 1967

  10. [10]

    A Kopanicakova and GE Karniadakis

    doi: 10.1137/18M1198375. A Kopanicakova and GE Karniadakis. Deeponet-based hybrid preconditioners.Computer Methods in Applied Mechanics and Engineering,

    work page doi:10.1137/18m1198375

  11. [11]

    Neural Operator: Graph Kernel Network for Partial Differential Equations

    doi: 10.1007/ s10208-025-09702-0. Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial differential equations.arXiv preprint arXiv:2003.03485, 2020a. Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattachary...

    work page internal anchor Pith review Pith/arXiv arXiv 2003

  12. [12]

    I. Luz, M. Galun, H. Maron, R. Basri, and I. Yavneh. Learning algebraic multigrid using graph neural networks.ICML 2020,

    work page 2020

  13. [13]

    Nils Margenberg, Jonas K”ohler, Rick Boelens, and Tijn Oosterlee

    doi: 10.1090/mcom/4105. Nils Margenberg, Jonas K”ohler, Rick Boelens, and Tijn Oosterlee. Neural networks as structural priors for deep learning-based image reconstruction.arXiv preprint arXiv:2201.07722,

    work page doi:10.1090/mcom/4105

  14. [14]

    Low regularity exponential-type integrators for semilinear schr¨ odinger equations.Foundations of Computational Mathematics, 18(3): 731–755, 2018b

    Alexander Ostermann and Katharina Schratz. Low regularity exponential-type integrators for semilinear schr¨ odinger equations.Foundations of Computational Mathematics, 18(3): 731–755, 2018b. doi: 10.1007/s10208-017-9352-1. Alexander Ostermann, Fr´ ed´ eric Rousset, and Katharina Schratz. Error estimates of a fourier integrator for the cubic schr¨ odinger ...

    work page doi:10.1007/s10208-017-9352-1

  15. [15]

    Alexander Ostermann, Yifei Wu, and Fangyan Yao

    doi: 10.1007/s10208-020-09468-7. Alexander Ostermann, Yifei Wu, and Fangyan Yao. A second-order low-regularity integrator for the nonlinear schr¨ odinger equation.Advances in Continuous and Discrete Models, 2022(23),

    work page doi:10.1007/s10208-020-09468-7 2022

  16. [16]

    Alexander Ostermann, Fr´ ed´ eric Rousset, and Katharina Schratz

    doi: 10.1186/s13662-022-03695-8. Alexander Ostermann, Fr´ ed´ eric Rousset, and Katharina Schratz. Fourier integrator for periodic NLS: Low regularity estimates via discrete bourgain spaces.Journal of the European Mathematical Society, 25(10):3913–3952,

    work page doi:10.1186/s13662-022-03695-8

  17. [17]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis

    doi: 10.4171/JEMS/1275. Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural net- works: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics, 378:686–707,

    work page doi:10.4171/jems/1275

  18. [18]

    Improved error estimates for low-regularity integrators using space-time bounds

    doi: 10.1137/20M1371506. Fr´ ed´ eric Rousset and Katharina Schratz. Convergence error estimates at low regularity for time discretizations of KdV.Pure and Applied Analysis, 4(1):127–152, 2022b. doi: 10.2140/paa.2022.4.127. Maximilian Ruff. Improved error estimates for low-regularity integrators using space-time bounds.arXiv preprint arXiv:2503.22621,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1137/20m1371506 2022

  19. [19]

    K. Um, R. Brand, Y. R. Fei, P. Holl, and N. Thuerey. Solver-in-the-loop: Learning from differentiable physics to interact with iterative pde-solvers.NeurIPS 2020,

    work page 2020

  20. [20]

    Wu and X

    Y. Wu and X. Zhao. Embedded exponential-type low-regularity integrators for kdv equation under rough data. doi:10.1007/s10543-021-00895-8.BIT,

    work page doi:10.1007/s10543-021-00895-8.bit

  21. [21]

    Fangyan Yao

    doi: 10.1007/s10543-021-00895-8. Fangyan Yao. A second-order embedded low-regularity integrator for the quadratic non- linear schr¨ odinger equation on torus.International Journal of Numerical Analysis and Modeling, 19(5):656–668,

    work page doi:10.1007/s10543-021-00895-8