pith. machine review for the scientific record. sign in

arxiv: 2605.04853 · v1 · submitted 2026-05-06 · 💻 cs.LG

Recognition: unknown

Hybrid Iterative Neural Low-Regularity Integrator for Nonlinear Dispersive Equations

Zhangyong Liang

Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:01 UTC · model grok-4.3

classification 💻 cs.LG
keywords hybrid neural integratorlow-regularity integratornonlinear dispersive equationssolver-in-the-loopneural operatorerror boundsrough dataGronwall stability
0
0 comments X

The pith

Augmenting low-regularity integrators with scaled neural corrections yields global error C(ε_net + δ) τ^γ ln(1/τ) for nonlinear dispersive equations.

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, a method that combines a classical low-regularity integrator with a neural network to correct its truncation errors for nonlinear dispersive partial differential equations. The network operates on a low-dimensional latent space and is trained by unrolling the solver over multiple steps, penalizing errors in a Bourgain-type norm to match the dynamics of the full simulation. Explicit scaling of the neural correction ensures its Lipschitz constant stays proportional to the time step, which bounds the stability factor uniformly. This setup yields a global error estimate involving the network approximation error and training residual, multiplied by tau to the gamma and a log term. Experiments on three benchmarks with rough initial data demonstrate gains in accuracy and robustness compared to traditional and neural-only approaches.

Core claim

Under stated assumptions the global error of the hybrid integrator satisfies C times (network approximation quality plus training shortfall) times tau to the gamma times the natural log of one over tau. The framework augments a base low-regularity integrator with a lightweight neural operator on a latent manifold, trained end-to-end via unrolled iterations that penalize full-trajectory error. The explicit time-step scaling on the correction term guarantees that its contribution to the Lipschitz constant is order tau, producing a Gronwall factor bounded independently of both time step and spatial resolution.

What carries the argument

The HIN-LRI hybrid framework, consisting of a low-regularity integrator augmented by a neural correction term scaled explicitly by the time step and trained on unrolled solver trajectories.

If this is right

  • The hybrid method yields higher accuracy than analytical integrators, splitting methods, and neural PDE surrogates on dispersive benchmarks with rough data.
  • Accuracy improves stably as spatial resolution increases.
  • Performance transfers well to out-of-distribution initial data without retraining.
  • The online computational cost remains modest while preserving theoretical stability properties.

Where Pith is reading between the lines

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

  • The training strategy of penalizing multi-step trajectory errors could be transferred to other numerical methods to enhance their long-term accuracy.
  • Uniform stability independent of spatial resolution suggests the approach may scale to very fine grids in high-dimensional simulations.
  • If the error bound holds for other dispersive systems, it could guide the design of neural corrections for related wave propagation problems.

Load-bearing premise

The assumption that after scaling the neural correction by the time step, its Lipschitz contribution stays order tau, allowing the Gronwall stability factor to remain bounded independently of spatial resolution even as the network error is small.

What would settle it

Computing the observed global error for decreasing time steps on a dispersive benchmark and checking if it exceeds the predicted C(ε_net + δ) τ^γ ln(1/τ) scaling would falsify the error bound claim.

Figures

Figures reproduced from arXiv: 2605.04853 by 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 ↗
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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 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 correction operating on a low-dimensional latent manifold. An explicit time-step scaling is applied to the neural term to keep its Lipschitz contribution O(τ), which is used to obtain a Gronwall stability factor bounded uniformly in τ and independent of spatial resolution. The network is trained end-to-end by unrolling the solver and minimizing trajectory error in a Bourgain-type norm. Under stated assumptions the global error is bounded by C(ε_net + δ) τ^γ ln(1/τ). Experiments on three dispersive benchmarks with rough initial data report improved accuracy relative to analytical integrators, splitting methods, and neural PDE surrogates, together with stable spatial refinement and out-of-distribution transfer.

Significance. If the claimed error bound and its uniformity with respect to spatial resolution can be rigorously established, the work would offer a concrete route to higher-order accuracy for low-regularity solutions of dispersive equations while retaining the consistency properties of the underlying integrator. The end-to-end training objective that aligns with multi-step dynamics and the modest online overhead are practical strengths. The experimental gains on rough-data benchmarks suggest potential utility beyond purely analytical or purely data-driven approaches, provided the Lipschitz-control assumption holds uniformly.

major comments (2)
  1. [Abstract / stability analysis] Abstract and stability analysis: the central error bound C(ε_net + δ) τ^γ ln(1/τ) rests on the claim that explicit time-step scaling renders the neural correction's Lipschitz constant exactly O(τ) uniformly in the spatial discretization parameter and in the Sobolev regularity of the data. The manuscript provides no explicit derivative bounds on the trained network after scaling, nor a quantitative verification that this O(τ) property is preserved under the Bourgain-norm training and across the range of spatial resolutions tested. Without such control the Gronwall factor may acquire a hidden dependence on the number of modes, undermining the stated uniformity.
  2. [Abstract] The abstract states that the global error satisfies the displayed bound under 'stated assumptions,' yet the provided text supplies neither the precise list of assumptions nor a derivation sketch showing how the O(τ) Lipschitz condition is verified and how it interacts with the low-regularity integrator. This gap is load-bearing for the theoretical contribution.
minor comments (1)
  1. [Abstract] The abstract refers to 'three dispersive benchmarks' and 'stable spatial refinement' without specifying the exact equations, the range of spatial resolutions, or quantitative error tables; adding these details would improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough review and for identifying key points that require clarification in the stability analysis and theoretical presentation. We address each major comment below and will incorporate the necessary additions and expansions in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract / stability analysis] Abstract and stability analysis: the central error bound C(ε_net + δ) τ^γ ln(1/τ) rests on the claim that explicit time-step scaling renders the neural correction's Lipschitz constant exactly O(τ) uniformly in the spatial discretization parameter and in the Sobolev regularity of the data. The manuscript provides no explicit derivative bounds on the trained network after scaling, nor a quantitative verification that this O(τ) property is preserved under the Bourgain-norm training and across the range of spatial resolutions tested. Without such control the Gronwall factor may acquire a hidden dependence on the number of modes, undermining the stated uniformity.

    Authors: We agree that the current version lacks explicit derivative bounds on the scaled network and quantitative verification of the O(τ) Lipschitz property. In the revision we will add a new lemma in the stability section that derives the uniform O(τ) bound from the network architecture and the explicit scaling factor, with estimates independent of the spatial discretization parameter and Sobolev regularity. We will also include a short numerical study verifying that the empirical Lipschitz constant remains O(τ) across the tested resolutions and data regularities, confirming that the Gronwall factor stays uniform. revision: yes

  2. Referee: [Abstract] The abstract states that the global error satisfies the displayed bound under 'stated assumptions,' yet the provided text supplies neither the precise list of assumptions nor a derivation sketch showing how the O(τ) Lipschitz condition is verified and how it interacts with the low-regularity integrator. This gap is load-bearing for the theoretical contribution.

    Authors: We acknowledge that the manuscript refers to 'stated assumptions' without enumerating them or providing a derivation sketch. In the revised version we will (i) expand the abstract to list the principal assumptions in one sentence and (ii) insert a dedicated paragraph immediately after the statement of the main theorem that enumerates all assumptions and gives a concise derivation outline: the explicit scaling ensures the neural term contributes an O(τ) perturbation to the Lipschitz constant of the integrator, which is absorbed into the Gronwall factor without introducing dependence on the number of modes, yielding the claimed global error bound. revision: yes

Circularity Check

0 steps flagged

No circularity: error bound follows from explicit scaling and standard Gronwall

full rationale

The paper's central derivation applies an explicit time-step scaling to the neural correction term so that its Lipschitz contribution is O(τ), then invokes the Gronwall inequality to obtain the global error bound C(ε_net + δ) τ^γ ln(1/τ) under stated assumptions. This is a conventional stability argument and does not reduce by construction to the network weights, the training data, or the Bourgain-norm trajectory loss; the scaling is a deliberate design choice stated separately from the learned residual. The training objective is defined externally via unrolled solver-in-the-loop trajectory error rather than being tautological with the error bound itself. No self-citations, uniqueness theorems, or ansatzes imported from prior author work appear in the provided text, and the low-regularity integrator is taken as a classical base method. The result is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only access yields no concrete free parameters, axioms, or invented entities beyond the general trained neural weights and the unstated assumptions on network approximation quality.

pith-pipeline@v0.9.0 · 5519 in / 1195 out tokens · 42765 ms · 2026-05-08T17:01:33.311741+00:00 · methodology

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 · 18 canonical work pages · 2 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]

    Mathematics of Computation , year =

    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]

    Learning core design for multigrid methods.arXiv preprint arXiv:1902.05656,

    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 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,

    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 arXiv 2003

  12. [12]

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

    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,

    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