acceptodds
Withdrawn from ICLR 2027

Neural Operators with Exactly Linearizable Structure for Near-Integrable PDEs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learned solution operators accumulate rollout error on systems whose evolution is exactly linear in the right coordinates. For near-integrable PDEs those coordinates exist in closed form: the inverse scattering transform (IST) diagonalizes the integrable flow. The Scattering-Coordinate Neural Operator (SCNO) routes the state through a differentiable nonlinear Fourier transform, applies the exact phase rotation, inverts by machine-precision Darboux dressing, and learns only the integrability-breaking residual. Zero-initialized, the pipeline is the analytic solution operator, verified to round-off; a learned transform cannot satisfy this identity. Soliton perturbation theory predicts that the residual, the learning target, grows as εz; we measure that law directly, and matched-capacity ablations show that the coordinates carry the advantage while three residual architectures inside them are interchangeable. We also measure where structure stops paying: across four fiber-physics perturbation channels the loss channel crosses over at ε* ∈ [0.020, 0.030] over baseline families, selection rules and budgets, beyond which tuned time-domain baselines win, and the same phase diagram re-emerges on damped Korteweg–de Vries through a different spectral problem and a different inverse. Removing the transmitter-side oracle and re-training from blindly recovered spectra changes the diagram by a median of 0%. Off the reflectionless manifold the error is set by the energy the coordinates exclude: a radiative fraction ρ costs ≈√ρ, flat in distance. The closest prior method, an FNO with learned scattering structure, is competent in distribution and diverges at twice its training horizon: learned approximate linearizations compound under extrapolation, exact ones cannot.

Primary area:
learning on time series and dynamical systems
Keywords:
neural operators, dynamical systems, inverse scattering transform, physics-informed learning, integrable systems, nonlinear Schrödinger equation, PDE surrogates, solitons

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ZG6KOpenReview ↗
Rating
1(−1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SCNO (Scattering Coordinates Neural Operator) introduces a framework that integrates exact inverse-scattering evolution with a learned correction for near-integrable partial differential equations. Experiments on perturbed nonlinear Schrodinger dynamics and damped KdV equations evaluate extrapolation capabilities, coordinate selection, and the impact of stronger perturbations.

Critical strengths
  1. Principled architecture: the zero-residual model recovers the exact solution for reflectionless integrable inputs, giving learning a physically meaningful starting point.

  2. Useful empirical exploration: several perturbations and a second PDE are studied, with informative ablations and explicit reporting of limitations.

Critical weaknesses
  1. The novelty, contribution, and potential benefits of the work need stronger justification

    The proposed solution represents a reasonable design. However, integrating physical dynamics with a learned correction is a well-established approach e.g. APHYNITY (Yin et al., 2021) and universal differential equations (Rackauckas et al., 2020). The manuscript should therefore articulate more precisely the novel aspects of this construction. Proposition 1 appears to follow directly from the design. The study is limited in scope, focusing primarily on the integrable soliton context. A clearer explanation of the approach's generalizability beyond the demonstrated cases would strengthen the work's significance.

    The primary concern relates to the comparison of physical models. Table 9 presents first-order loss baselines, and the loss baseline outperforms SCNO across all four strengths. However, first-order perturbation is not applied for TOD and Raman effects. This omission represents a missed opportunity, as evaluating the corresponding approximate drift models could clarify where the learning-based approach captures effects that simpler analytical corrections do not. Furthermore, since the equations are known and models are trained separately for each channel and strength, an accuracy-versus-cost comparison with less expensive classical solver settings is necessary (e.g. Hult, 2007). Comparison solely with the production-accuracy reference solver does not adequately address this issue. Incorporating these comparisons, along with clearer positioning relative to prior methods, would better substantiate the contribution.

  2. Better support and documentation of the theoretical claims and experimental setup are needed

    Proposition 2 offers valuable motivation; however, its assumptions and coordinate selection should be more explicitly linked to the empirical claims. The manuscript differentiates between the path-integrated coordinates used in the bound and the implemented instantaneous coordinates, and it includes a frame ablation. However, the abstract O(ϵz)O(\epsilon z) description should also acknowledge the additional O(ϵz2)O(\epsilon z^2) contribution identified for the implemented coordinates. Clarify which experiments satisfy the separation and adiabatic assumptions, especially in the context of collision-trajectory interpretation. This distinction would help separate conditional theoretical guarantees from experimental motivations that fall outside the theoretical scope.

    The validation and implementation details also need clarification. The loss-decay law is described as exact, while the authors refer to adiabatic theory; the authors should also specify the measured observable and the conditions under which the law is exact or approximate. It is also difficult to identify which configurations generated Figures 5b and 7b. Too many internal log references are distracting and sometimes confusing. The authors should explicitly state the boundary conditions and how they treat solitons reaching the temporal-window boundary. No code is provided to help.

References

Yin, Y., Le Guen, V., Dona, J., de Bézenac, E., Ayed, I., Thome, N., and Gallinari, P. (2021). Augmenting Physical Models with Deep Networks for Complex Dynamics Forecasting. ICLR. https://arxiv.org/abs/2010.04456

Rackauckas, C., et al. (2020). Universal Differential Equations for Scientific Machine Learning. arXiv:2001.04385. https://arxiv.org/abs/2001.04385

Hult, J. (2007). A Fourth-Order Runge–Kutta in the Interaction Picture Method for Simulating Supercontinuum Generation in Optical Fibers. Journal of Lightwave Technology, 25(12):3770–3775. https://doi.org/10.1109/JLT.2007.909373

Questions for authors
  1. How does SCNO compare with approximate first-order TOD and Raman drift models, and also with classical solvers, in terms of accuracy and cost?
  2. What specific advantage does SCNO offer over existing approaches that combine physical dynamics with learned residuals, and what extends beyond the demonstrated integrable soliton settings?
  3. Which experiments satisfy Proposition 2’s assumptions, and which claims for the implemented coordinates follow once the additional phase contribution is included (see weaknesses above)?
  4. Under what conditions is the loss-decay law exact or approximate?
  5. Which configurations produced Figures 5b and 7b, and what boundary conditions were used?
5 comments
AuthorsOpenReview ↗

Reply to Reviewer ZG6K (Part 1 of 4)

We thank the reviewer for precise questions. They led us to recheck our calculations, and we made two corrections. Proposition 2 now includes an ϵz2\epsilon z^2 term for moving solitons, and the revised Eq. (4), which replaces Eq. (5), states that this term changes the amplitude as well as the phase. After these corrections, we added the requested first-order TOD and Raman models and the classical-solver comparison, and we added APHYNITY. We also retrained SCNO in both coordinates and on top of the first-order rates. The corrected results led us to revise our claims. A cell is one perturbation type at one strength, and one further cell applies loss, TOD and Raman together. With the perturbation known, first-order theory, with zero learned parameters, has lower error than SCNO in most cells. A classical solver at twice SCNO's inference cost is more accurate in all 9 cells with one deterministic perturbation. When the perturbation is unknown and must be learned from data, SCNO has the lowest error of the tested learned methods in the TOD cells at z=8z=8, twice the training horizon. The uploaded manuscript is revised throughout, and we answer each question below.

Experiments added at the reviewer's request.

We ran first-order TOD and Raman drift models, classical solvers (SSFM and RK4IP) for accuracy against cost, a physics-plus-residual model (APHYNITY), and an evaluation of the test fields by soliton geometry. All are reported in the revised manuscript (Sec. 5.2 to 5.4).

Q1. First-order TOD and Raman drift models, and classical solvers, in accuracy and cost.

We added both comparisons. The Raman rate of adiabatic theory, dζ/dz=−6415ϵ3η4d\zeta/dz=-\tfrac{64}{15}\epsilon_3\eta^4, matches single-soliton simulation within 0.1%. We use two methods with zero learned parameters. The first applies the first-order drift rates with the multi-soliton dressing of the positions. The second projects the perturbation onto the NN-soliton manifold. The table gives the median field error at z=8z=8, the median over the 256 test fields of the relative L2L_2 error, for the TOD and Raman cells. For SCNO it is the median over 3 seeds, and Table 2 also gives the four loss cells.

CellSCNOFirst orderProjectionRK4IP, 10 steps/unitRK4IP, 20 steps/unit
TOD 0.0050.0490.0080.0060.7330.024
TOD 0.020.1330.0640.0570.7260.034
TOD 0.050.2590.3710.3370.9490.093
Raman 0.010.4550.0280.0120.9230.115
Raman 0.030.9660.6030.1561.0230.389
Cost per field4.0 ms0.4 to 0.8 ms2400 to 2500 ms3.9 to 5.2 ms8 to 10 ms

SCNO beats first-order theory in one cell, TOD 0.05, and its lead lies in fields with several solitons. There SCNO has median error 0.304 on three-soliton fields, against 0.448 for the projection. On single solitons, the projection is better, 0.185 against 0.237. At Raman ϵ3=0.01\epsilon_3=0.01, collisions break first-order theory, which has error 0.747 on colliding fields. The projection handles them with 0.029, and SCNO does not, with 0.857 (Table 4).

For the classical solvers, we timed SSFM and RK4IP at 10 to 50 steps per unit length on the same machine, one CPU thread, batch 32. SCNO's 4.0 ms assumes the scattering data are given, and computing them from a field adds 238 ms. Training, 340 s per run, is also excluded. At equal inference cost with the scattering data given, RK4IP with 10 steps per unit, SCNO has lower error in 6 of 9 cells. At twice SCNO's cost, RK4IP with 20 steps per unit has lower error than SCNO in all 9 cells. Table 2 adds these methods, Table 9 gives the solvers, and this accuracy-cost comparison replaces the comparison with the production reference (Sec. 5.2).

AuthorsOpenReview ↗

Reply to Reviewer ZG6K (Part 2 of 4)

Q2. Advantage over approaches that combine physics with a learned residual, and reach beyond solitons.

We implemented APHYNITY following Algorithm 1 of Yin et al. (2022). The known physics is the integrable NLSE, plus the loss term in the cell with all three perturbations, with the physical parameters fixed at their known values. A small convolutional network learns the rest. UDE (Rackauckas et al., 2021) has the same known-plus-learned structure in state coordinates, and we use APHYNITY as the representative of this structure. APHYNITY has lower median field error at z=8z=8 in 7 cells, and SCNO in 5.

CellAPHYNITYSCNO
Loss 0.010.1120.259
Loss 0.030.0780.276
Loss 0.050.0440.503
Loss 0.10.0171.073
Raman 0.010.2120.455
Raman 0.030.3730.966
All three perturbations0.1210.471
TOD 0.0050.2180.049
TOD 0.020.2380.133
TOD 0.050.5970.259
Noise 0.0010.1520.017
Noise 0.010.2430.171

APHYNITY has lower error in the loss and Raman cells and in the cell with all three perturbations. SCNO has lower error in the three TOD cells and the two noise cells, where the noiseless solution reaches the same floor without learning. Timed in one session on one CPU thread, APHYNITY takes 508 ms per field at inference and SCNO 8.2 ms with the scattering data given. Computing the scattering data from a field adds 238 ms (Sec. 5.7). The 4.0 ms of Q1 was timed on another day, when the same machine ran about twice as fast.

The specific difference from these frameworks is where the residual lives. SCNO places it in scattering coordinates, where the integrable flow is an exact rotation. Our matched-capacity ablation, the same residual model in time-domain coordinates, has 5.8 times higher error at z=4z=4. The corrected results show that this design has lower error than APHYNITY on TOD and at the noise floor, and higher error on loss, Raman and the cell with all three perturbations. Against this class, SCNO's advantages are lower error in the TOD cells and lower inference cost once the scattering data are known. The ablation compares two coordinates inside SCNO's own pipeline, and the APHYNITY numbers above give the comparison with this class. On reach beyond solitons, SCNO needs an explicit inverse scattering transform, with decaying boundary conditions and a computable inverse. The continuous spectrum enters only through the measured error of the discarded energy. The revised manuscript states this scope, cites APHYNITY and UDE, and states the positioning in these terms (Sec. 2, Sec. 5.3, Sec. 6).

Proposition 1 follows directly from the design.

Proposition 1 states a design property of the pipeline. The revised manuscript presents it as such and no longer lists it as a contribution (Sec. 1, Sec. 4).

AuthorsOpenReview ↗

Reply to Reviewer ZG6K (Part 3 of 4)

Q3. Which experiments satisfy Proposition 2's assumptions, and which claims follow once the additional contribution is included?

Proposition 2 is stated for an isolated soliton in the adiabatic regime, and the 76 single-soliton fields satisfy it. Separated fields with several solitons carry dressing terms that drift at O(ϵz)O(\epsilon z). Of the 256 test fields, 110 (43%) keep every pair of solitons at least three widths apart along the unperturbed tracks for z≤8z\le8, and 76 of these are single solitons. The other 146 (57%) come closer, and 93 (36%) collide, with centers less than one width apart. The adiabatic parameter 2ϵ1z2\epsilon_1 z reaches 1.6 at the strongest loss. The revised manuscript reports these fractions and evaluates each group separately (Sec. 5.4, Table 4). The table gives the median field error at z=8z=8 for the projection, the first-order line and SCNO.

GroupLoss 0.01 (projection / first order / SCNO)TOD 0.05Raman 0.01
Single soliton (76)0.006 / 0.007 / 0.2050.185 / 0.185 / 0.2370.006 / 0.005 / 0.303
Separated (110)0.007 / 0.007 / 0.2400.264 / 0.264 / 0.2320.007 / 0.008 / 0.341
Colliding (93)0.015 / 0.067 / 0.2640.469 / 0.487 / 0.3020.029 / 0.747 / 0.857

SCNO's error on colliding fields is 0.95 to 1.6 times that on separated fields on loss and TOD, and 1.4 to 2.5 times on Raman (Table 15 gives every cell). On single solitons SCNO has higher error than the projection in 9 of 9 cells, so the separation assumption does not explain its gap to the projection and first-order theory. The revised Proposition 2 is stated for an isolated soliton, so collision trajectories are outside its scope (Sec. 4, Sec. 5.4).

For the implemented instantaneous coordinate, the corrected calculation gives an O(ϵz)O(\epsilon z) log-amplitude drift on loss and TOD, an O(ϵz2)O(\epsilon z^2) phase drift on loss, and an O(ϵz2)O(\epsilon z^2) log-amplitude drift on Raman. On the single-soliton test fields that stay tracked to z=8z=8, at most 76 per cell and 15 to 29 on Raman, the log-amplitude follows these forms. The Raman log-amplitude exponent is 2.00 against a predicted 2, and the Raman phase exponent is 0.99 to 1.00 against 1 (Table 12). On the first 64 test fields of all geometries, the Raman exponent of the log-amplitude is 1.26 to 1.45, so the isolated-soliton calculation does not extend to fields with several solitons (Table 11). So the claim of an O(ϵz)O(\epsilon z) learning target holds for the log-amplitude on loss and TOD, and it fails for the Raman log-amplitude and the loss phase. The revised abstract states that the drift grows as the square of the distance for the log-amplitude under Raman and for the phase under loss. Fig. 3, which replaces Fig. 4, includes Raman and both coordinates. Retraining SCNO in the path-integrated coordinate raised its median error at z=8z=8 from 0.455 to 0.739 at ϵ3=0.01\epsilon_3=0.01, so removing the first-order Raman term from the target did not help.

Reviewer_Xd24OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces SCNO, a neural operator that combines an exact inverse scattering solver for the integrable dynamics with a learned correction for the non-integrable residual. The authors show that the method reproduces the analytical solution to machine precision when the residual is absent and that the residual magnitude grows approximately linearly with the perturbation strength and propagation distance. Through ablations, they argue that most of the performance gain comes from the scattering coordinates themselves rather than from the specific residual architecture.

Critical strengths

The topic is of of great interest and I think that some of the ideas may be promising. However the clarity is really a burden. I cannot say more in terms of strength.

Critical weaknesses
  • Limited ML novelty. The proposed approach combines an exact inverse scattering solver for the integrable component with a neural network that learns the residual dynamics. While this is a sensible strategy, much of the reported performance appears to derive from the analytical solver rather than the learned model, and analytical perturbation corrections sometimes outperform the neural residual. As a result, the machine learning contribution seems relatively limited.

  • Highly restrictive scope. The method is confined to reflectionless multi-soliton solutions, a setting where the analytical structure is already well understood. It neglects the continuous spectrum present in more general solutions, and the extension to KdV relies on similar assumptions. This raises concerns about the broader applicability of the approach to general neural operator learning problems.

  • Insufficient support for the broader claims. The theoretical analysis largely builds on established results from inverse scattering and perturbation theory, while the claim that residual dynamics are inherently easier to learn is not rigorously justified. The experimental evidence is also mixed: SCNO is outperformed by a tuned U-Net in some settings, and the strongest extrapolation results are obtained in regimes where learning plays only a minor role. Overall, the evidence does not convincingly establish a general advantage of the proposed framework.

Questions for authors

I suggest that the authors involve some colleagues to have a first review to improve the clarity.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer Xd24 (Part 1 of 2)

We thank the reviewer for a frank review. Its central question, what the learning adds beyond the analytical solver, led us to recheck the mathematics behind our claims, and we made two corrections. Proposition 2 now includes an ϵz2\epsilon z^2 term for moving solitons, and the argument that the residual is easier to learn, which relied on a Barron-type rate that does not apply, is removed. After these corrections, we added first-order models for all three perturbation channels, retrained SCNO and compared it with analytical perturbation corrections, with classical solvers, and with neural baselines from the literature. The corrected results led us to revise our claims. The uploaded manuscript is revised throughout. It reports the new comparisons and states the revised claims, which we summarize below.

Limited ML novelty; the performance comes mainly from the analytical solver.

The revised manuscript separates the evidence for the analytical part from the evidence for the learned part (Sec. 5.1 to 5.3). We now have first-order rates for all three perturbation channels. A cell is one perturbation type at one strength, and one further cell applies loss, TOD and Raman together. With these rates, first-order theory with zero learned parameters has lower error than SCNO at z=8z=8 in 8 of the 9 NLSE cells with one deterministic perturbation and at z=4z=4 in all 5 KdV cells. The learned part adds measurable value when the perturbation is unknown and must be learned from data. There SCNO has the lowest error of the tested learned models in the three third-order dispersion (TOD) cells at z=8z=8, twice the training horizon. At TOD ϵ2=0.05\epsilon_2=0.05 it also has lower error than two methods with zero learned parameters on three-soliton fields. There its median error is 0.304, against 0.481 for first-order theory and 0.448 for a projection of the perturbation onto the NN-soliton manifold.

Highly restrictive scope.

SCNO is exact for reflectionless soliton inputs of the perturbed NLSE and the damped KdV equation. When an input also carries a continuous spectrum, SCNO keeps the discrete part, and Sec. 5.8 measures how the error follows the discarded energy. The KdV extension rests on the same assumptions (Appendix K). The revised abstract and introduction state this scope, and the manuscript describes it in detail in Sec. 3, Sec. 5.8 and Sec. 6.

AuthorsOpenReview ↗

Reply to Reviewer Xd24 (Part 2 of 2)

Insufficient support for the broader claims.

  • The theory builds on established results. Proposition 1 is a design property of the pipeline. The revised Proposition 2 is a formal first-order calculation for an isolated soliton and now includes the ϵz2\epsilon z^2 term (Sec. 4).
  • "Residual dynamics are easier to learn" is not justified. The claim rested on a Barron-type rate. Barron's bound depends on the first moment of the target's Fourier transform, which a small residual does not control, and we never bounded that quantity. The claim is removed (Sec. 1, Sec. 4).
  • The evidence is mixed. The revised manuscript reports the comparison regime by regime (Table 2, Table 3). A tuned U-Net has lower error than SCNO at strong loss, 0.053 against 1.073 at ϵ1=0.1\epsilon_1=0.1, and in the cell with all three perturbations, 0.124 against 0.471.
  • The strongest extrapolation results come from regimes where learning plays a minor role. At zero perturbation and z=40z=40, the rotation-only pipeline, which is SCNO with its learned flow at zero, has error 4.6×10−44.6\times10^{-4}, against 1.62 for the best run of the tuned U-Net, a factor of 3.6×1033.6\times10^3. This separation comes from the exact inverse scattering transform, with no learned component. In the weakly perturbed cells, first-order theory alone has lower error than SCNO. The revised text attributes these results to the analytical part (Sec. 5.1, Sec. 5.2).

Clarity.

We thank the reviewer for this suggestion. We have reorganized the language of the whole paper.

Reviewer_UgewOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This work targets accurate long-horizon prediction of near-integrable wave PDEs by exploiting known analytical dynamics. SCNO transforms wavefields into scattering coordinates, evolves the integrable dyanmics analytically, and learns only the remaining correction. On nonlinear Schrodinger and damped kdv benchmakrs it reports 5.8x lower error than a matched capacity time domain residual and 1.29-2.57x improvements over 1st order kdv, although analytical baselines already match or outperform it for NLSE and noise.

Critical strengths
  • Ablation studies are informative. Matched capacity comparisons help isolate the benefit of scattering coordinates
  • Exploiting known analytical dynamics and learning the residual can be very effective inductive bias for the targeted types of PDEs.
  • Computational efficiency seems promising for inference, and the #params comparison is good too, although the end-to-end training time is not shown.
Critical weaknesses

Empirical Weakness

  • I feel only 2 PDE benchmarks are too slim. I don't think they are representative enough of the near-integrable wave PDEs the authors advertise in their motivation.
  • 2 PDEs are both 1D and too simple with training distributions only one to three solitons. I am nowhere convinced that SCNO is truly impactful and practical for the targeted class of PDEs in reality.
  • I feel authors should include a few more baselines, prior works like PDE Refiner (NeuRIPS 2023) and RecurrFNO (ICLR 2026) that worked on exact same problem and KdV rollouts.
  • Also, I am not convinced that the experimental results of SCNO is anywhere impactful. There are no order improvement and gains around x1.3~5, I don't think that is impressive especially for very simple benchmarks.

Theoretical Weakness

I find substantial faults in Proposition 2. Here are the list of my concerns.

  • page 4, lines 209-215 Eq (4), page 20 lines 1076-1079, page 21 lines 1109-1113: I think Proposition 2 is missing the growing amplitude drift in its corrected coordinates here. Authors should check that an ez^2 term is missing in eq (4). For a moving soliton under loss, attenuation interacts with its growing displacement and should produce ez^2 in the corrected norming contant's log-amplitude even after the path-integrated correction. So for z ~ e^-1/2, ez^2 produce order-one change and the bound doesn't tend to zero. So I think the bound does not hold uniformly for the general moving-soliton class described. Please check this issue.
  • page 5, lines 221-231 Eq(5), page 21 lines 1115-1119: Eq(5) here is an amplitude and phase correction, not a pure phase correction. Because lambda is complex, the quantity delta_phi is generally complex too, so its exponential should change magnitude as well as phase. The manuscript's phase-only interpretation works it special cases like stationary solitons, but does not justify the manuscript's general separation between amplitude and phase drift.
  • page 17, lines 914-917, page 7, lines 364-370: Here, the loss-induced eigenvalue decay is not exact. The exponential decay law is exact for the total wave power, while applying it to each soliton eigenvalue is an adiabatic approximation. Loss also transfers power into radiation, so the individual eigenvalues cannot generally obey the asserted exact exponential law. So the "nothing is left to learn" is not justified mathematically.
  • page 23, lines 1195-1198, page 4 lines 188-190: Random forcing does not imply that the expected training gradient vanishes. Zero-mean noise can change the average solution after nonlinear evolution, and the manuscript's relative error loss additionally weights noisy targets according to their random norms. So the randomness alone does not justify a zero expected gradient nor establish that learning no correction is optimal.
  • page 1 lines 29-31, page 2 lines 88-93, page 3 lines 120-124: Approximate learned linearization does not necessarily cause growing rollout error. An approximate invertible change of coordinates surrounding a stable evolution can remain uniformly close to the true dynamics at every prediction time, like a slightly inaccurate coordinate transformation around an exact rotation. The tested IS-FNO failures therefore do not establish a general mathematical obstruction to stable learned linearizations.
  • page 21 lines 1125-1128: Having a Lax pair is insufficient for the claimed general extension. A Lax commutator relation can preserve eigenvalues without supplying complete, invertible scattering coordinates with explicitly solvable evolution. So extending the construction to "any equation admitting a Lax pair" requires additional assumptions.
  • page 8 Fig 5 caption line 391: Relative error of one does not mean that a prediction contains no information.

Also, there are some unproved theoretical implications I notice:

  • page 5, lines 237-240: Small residual size does not estabiish the claimed learning-efficiency advantage. A bound on accumulated drift does not itself guarantee the bounds on the vector field's magnitude or Lipschitz constant. A small amplitude function can still vary rapidly and Barron's approximation bounds require control of a stronger function norm that the manuscript never discuss. So the claimed factor-e reduction in learning budget does not follow from Proposition 2
  • page 20, lines 1076-1079, page 21 lines 1109-1113: The crucial long-time remainder estimate is assumed rather than proved. O(e^2z^2) remainder control need for the long-time argument is assumed without establishing the required uniform estimates.
  • page 9, lines 454-463, page 2 lines 95-99: The trace formula does not establish the claimed general waveform error law. The trace formula divides total energy between discrete and continuous scattering data, but removing the continuous spectrum is not generally an orthogonal projection in waveform space.
Questions for authors
  • is SCNO easy to train? That is, how is the optimization difficulty? Also, could the authors provide end-to-end training time comparisons with the baselines?
  • Lei et al., “Long-Time Integration of Nonlinear Wave Equations with Neural Operators”, it seems this work is highly related to this manuscript, could the authors compare SCNO with this work? And if viable, add this work into baselines?
  • I find that the maths around Proposition 2 is very sloppy here and there. Please check the list above and provide explanations on my concerns, I may be wrong or misinterpreted the manuscript.
  • I also generally feel that the boasted "near-integrable wave PDE" target is not well represented by the 2 PDEs benchmarks in this manuscript. They are too simple and only 1D. I am nowhere convinced SCNO can be effective on actual application of practical interest. And even for those simple benchmarks SCNO do not really demonstrate meaningful improvements. If authors disagree, please explain why the 2 PDEs are representative enough.
6 comments
AuthorsOpenReview ↗

Reply to Reviewer Ugew (Part 1 of 5)

We thank the reviewer for a careful and detailed review. The questions on Proposition 2 led us to recheck our mathematics, and we found oversights in our calculations. We corrected them and retrained SCNO. We then compared it with first-order perturbation theory, with classical solvers, and with PDE-Refiner and the method of Lei et al. The corrected results led us to revise our claims. Two of the revised claims bear on this review. For reflectionless soliton inputs, the analytic pipeline is exact in the integrable limit. When the perturbation is unknown and must be learned from data, SCNO has the lowest error of the tested learned methods in the TOD cells at z=8z=8, twice the training horizon. The uploaded manuscript is revised throughout, and we answer each point below.

Experiments added at the reviewer's request.

We added PDE-Refiner and RecurrFNO as baselines, compared SCNO with the method of Lei et al., and measured end-to-end training time and optimization difficulty. The revised manuscript reports the baselines and the training times (Sec. 5.3, Table 3, Appendix B), and Part 4 of this reply gives the optimization result.

Only two PDE benchmarks, not representative of the class.

The study covers the perturbed nonlinear Schrödinger equation (NLSE) and the damped KdV equation. The inverse scattering transform is explicit for both, and for reflectionless soliton inputs, which carry no continuous spectrum, the transform in SCNO is exact. The revised abstract and introduction state the claims for these two equations with soliton inputs, and the manuscript describes this scope in detail (Sec. 1, Sec. 3, Sec. 6).

One-dimensional problems with one to three solitons.

The revised manuscript states this setting in its Limitations section (Sec. 6) and adds two methods with zero learned parameters. Here a cell is one perturbation type at one strength, and one further cell applies loss, TOD and Raman together. The NLSE results below cover the 9 cells with one deterministic perturbation. First-order perturbation theory has lower error than SCNO at z=8z=8 in 8 of the 9 NLSE cells and at z=4z=4 in all 5 KdV cells. The second method projects the perturbation onto the NN-soliton manifold. Of all methods except the classical solvers, learned methods included, it has the lowest error in 6 of the 9 NLSE cells (Tables 2 and 3). Table 2 reports both methods next to SCNO, and Table 20 reports first-order theory on KdV.

Baselines such as PDE-Refiner and RecurrFNO.

We implemented PDE-Refiner with the training recipe of Table 3 in Lippe et al. (2023) and trained it on our data. Its median field error at z=8z=8, the median over the 256 test fields of the relative L2L_2 error, is as follows, with SCNO in parentheses.

  • 0.208 at loss ϵ1=0.01\epsilon_1=0.01 (SCNO 0.259);
  • 0.157 at ϵ1=0.03\epsilon_1=0.03 (SCNO 0.276);
  • 0.292 at Raman ϵ3=0.01\epsilon_3=0.01 (SCNO 0.455);
  • 0.366 at third-order dispersion (TOD) ϵ2=0.005\epsilon_2=0.005 (SCNO 0.049).

PDE-Refiner thus has the lower median in three of the four cells. At ϵ1=0.01\epsilon_1=0.01, 13 to 21 of its 256 test fields exceed 100% error, against none for SCNO. We also implemented RecurrFNO, the FNO instance of the recurrent-depth simulator of Abdul Majid et al. (2026), with the configuration of their Section 4.3, Appendix C and Table 6. Its median field error at z=8z=8 is 0.405 at ϵ1=0.01\epsilon_1=0.01, 0.204 at ϵ1=0.03\epsilon_1=0.03, 0.491 at ϵ2=0.005\epsilon_2=0.005 and 0.508 at ϵ3=0.01\epsilon_3=0.01, so it is below SCNO in one of the four cells. Both baselines use three seeds at ϵ1=0.01\epsilon_1=0.01 and one seed in the other cells. These results are in Table 3, and Appendix B gives the configurations and the differences from the original papers. We did not run these baselines on KdV, where Table 20 compares SCNO with FNO and first-order theory.

Gains of 1.3 to 5 times are not impressive.

The relative L2L_2 error of a predicted field is the error vector magnitude (EVM) of communication systems, and a ratio rr in it is 20log⁡10r20\log_{10}r dB in normalized mean-square error. A factor of 1.3 to 5 is therefore 2.3 to 14 dB. For scale, LTE requires an EVM below 17.5% for QPSK and 8% for 64-QAM (3GPP TS 36.104), so a few dB decide which format a link supports. This is a comparison of scale, because a receiver removes phase and timing offsets that our field error counts. At z=8z=8 (Table 3), SCNO reaches 0.049 at TOD 0.005, where APHYNITY, the best of the other learned methods, reaches 0.218 and FNO 0.624. These gaps are 13 dB and 22 dB, and the second exceeds one order of magnitude. At TOD 0.02 and 0.05 the gap to the next-best learned method is 5.1 dB and 1.6 dB. Under loss and Raman APHYNITY has lower error than SCNO. With the perturbation known, first-order theory has lower error than SCNO in most cells, and RK4IP at twice SCNO's inference cost, with SCNO given the scattering data, has lower error in all 9 NLSE cells, by factors of 2 to 133 (Table 2). The revised abstract states these results.

AuthorsOpenReview ↗

Reply to Reviewer Ugew (Part 2 of 5): Proposition 2, Eq. (5), the loss law and random forcing

Submitted Eq. (4) misses an ϵz2\epsilon z^2 term for moving solitons.

The log-modulus of the norming constant is ln⁡∣Qk∣=2ηkτ0,k\ln|Q_k|=2\eta_k\tau_{0,k}. Our calculation omitted the product term that appears when the amplitude ηk\eta_k decays while the position τ0,k\tau_{0,k} moves. The appendix example used a stationary soliton (ζk=0\zeta_k=0), where this term vanishes. The revised Proposition 2 therefore states no uniform bound. It is a formal first-order calculation for an isolated soliton (see the remainder point in Part 3). Under loss, the corrected first-order drift of the path-integrated coordinate is

Δln⁡∣P~k∣=2(ηk(z)−ηk(0))τ0,k(0)+8ζk(∫0zηk ds−z ηk(z))≈2(ηk(z)−ηk(0))τ0,k(0)+8ζkηk(0) ϵ1z2,\Delta\ln|\tilde P_k| = 2\big(\eta_k(z)-\eta_k(0)\big)\tau_{0,k}(0) + 8\zeta_k\Big(\int_0^z\eta_k\,ds - z\,\eta_k(z)\Big) \approx 2\big(\eta_k(z)-\eta_k(0)\big)\tau_{0,k}(0) + 8\zeta_k\eta_k(0)\,\epsilon_1 z^2 ,

which matches split-step simulation within 0.4%. The revised Proposition 2 gives the drift of each component under each deterministic channel, loss, TOD and Raman (Sec. 4, Table 1). It does so for both coordinates, the instantaneous coordinate PkP_k that SCNO uses and the path-integrated coordinate P~k\tilde P_k. In the instantaneous coordinate, the log-amplitude drifts as O(ϵz)O(\epsilon z) on loss and TOD and as O(ϵz2)O(\epsilon z^2) on Raman. On the first 64 test fields at ϵ1≤0.03\epsilon_1\le0.03 and on Raman, the difference between the two coordinates grows as zz to the power 1.83 to 2.07, near the predicted 2. We also fitted the growth exponent of the instantaneous log-amplitude on the single-soliton test fields that stay tracked to z=8z=8, at most 76 per cell and 15 to 29 on Raman (Table 12). It is 1.00 on TOD against a predicted 1, 2.00 on Raman against 2, and 0.71 to 0.96 on loss, where the closed form gives 0.70 to 0.97. After the correction we retrained SCNO in the path-integrated coordinate, where the first-order Raman amplitude drift vanishes. Its median error at z=8z=8 rose from 0.455 to 0.739 at ϵ3=0.01\epsilon_3=0.01 and from 0.966 to 1.226 at ϵ3=0.03\epsilon_3=0.03. Retraining in the corrected coordinate therefore did not improve the model (Sec. 5.5, Table 17).

Submitted Eq. (5) changes amplitude as well as phase.

We evaluated this step with λk=jηk\lambda_k=j\eta_k, that is ζk=0\zeta_k=0, which makes λk2\lambda_k^2 real. With λk=ζk+jηk\lambda_k=\zeta_k+j\eta_k, the difference Δϕk=4(λk(z)2z−∫0zλk2 ds)\Delta\phi_k=4\big(\lambda_k(z)^2z-\int_0^z\lambda_k^2\,ds\big) has, under loss,

Im Δϕk=8ζk(ηk(z) z−∫0zηk ds)≈−8ζkηk(0) ϵ1z2.\mathrm{Im}\,\Delta\phi_k = 8\zeta_k\Big(\eta_k(z)\,z-\int_0^z\eta_k\,ds\Big)\approx -8\zeta_k\eta_k(0)\,\epsilon_1 z^2 .

Since Pk=P~ke−jΔϕkP_k=\tilde P_k e^{-j\Delta\phi_k}, this imaginary part enters ln⁡∣Pk∣\ln|P_k|. It is the z2z^2 term of the previous point with the opposite sign. The two therefore cancel in ln⁡∣Pk∣\ln|P_k| and remain in ln⁡∣P~k∣\ln|\tilde P_k|. The revised Eq. (4) replaces the submitted Eq. (5). Together with Proposition 2, it states that the imaginary part enters the log-amplitude and the real part enters the phase (Sec. 4, Table 1).

The loss-induced eigenvalue decay is not exact; "nothing is left to learn".

The exponential law is exact for the total power, E(z)=E(0)e−2ϵ1zE(z)=E(0)e^{-2\epsilon_1 z}. We applied it to individual eigenvalues, where it is the first-order adiabatic approximation, and our check measured total power only. A single-soliton test shows that the eigenvalue departs from the law by 2×10−52\times10^{-5} at ϵ1=0.01\epsilon_1=0.01, 7×10−47\times10^{-4} at 0.03, 6×10−36\times10^{-3} at 0.05, and 15% at 0.1, all at z=8z=8. We drew "nothing is left to learn" from first-order theory beating SCNO in the four loss cells, without accounting for the error of first-order theory itself. The revised text calls the law a first-order adiabatic approximation, reports the deviation, and removes the sentence (Sec. 4, Table 10).

Random forcing does not imply a zero expected gradient.

We estimated the conditional mean by Monte Carlo, with 32 inputs and 64 noise realizations for each. At z=8z=8 the conditional mean differs from the noiseless solution by 0.0023 at σ=0.001\sigma=0.001 and by 0.0375 at σ=0.01\sigma=0.01. Measured against the noisy realizations, the conditional mean and the noiseless solution have the same median error, 0.0208 at σ=0.001\sigma=0.001 and 0.2075 at σ=0.01\sigma=0.01. Over seeds, SCNO has 0.0209 to 0.0211 and 0.2080 to 0.2117. The noise cells therefore show that SCNO sits at the noise floor, which the noiseless solution reaches without any learning. SCNO is also farther from the conditional mean than the noiseless solution, 0.0034 to 0.0047 against 0.0023 at σ=0.001\sigma=0.001 and 0.0452 to 0.0590 against 0.0375 at σ=0.01\sigma=0.01. The relative loss weights each target by its random norm, so its minimizer need not be the conditional mean. The revised text replaces those sentences with this measurement (Appendix D, Table 13).

AuthorsOpenReview ↗

Reply to Reviewer Ugew (Part 3 of 5): linearization, Lax pair, relative error, Barron, remainder, trace formula

Approximate learned linearizations need not accumulate error.

In the reviewer's example, a fixed coordinate error around an exact rotation, the error does not accumulate. We have removed the general statement from the abstract and introduction. We also measured why IS-FNO diverges. Its learned latent operator has a spectral radius per step of 1.21 to 1.68, outside the unit circle. Our earlier IS-FNO runs rolled out in latent space only, without the re-lifting step of Yu et al. (their Eq. 32), which every few steps decodes the latent state to the field and encodes that field again. With re-lifting, IS-FNO has 0.8% error at step 1000 on KdV. On the unperturbed NLSE its median error at z=8z=8 is 0.211 on the training frames the other methods use, and 0.074 on frames 12.5 times denser. The revised manuscript replaces the earlier numbers with these (Sec. 5.1, Fig. 2, Table 18).

A Lax pair is insufficient for the general extension.

A Lax pair gives isospectral evolution. The construction also needs decaying boundary conditions, complete scattering data, an explicit evolution of those data, and a computable inverse transform. The revised sentence lists these conditions (Sec. 6, Appendix K).

A relative error of one can still carry information.

A field multiplied by ejπ/2e^{j\pi/2} has a relative L2L_2 error of 2\sqrt2 and an exact intensity. We added an intensity error, the relative L2L_2 error of ∣q∣2|q|^2. At ϵ1=0.01\epsilon_1=0.01 and z=8z=8, the rotation-only prediction, which evolves the input with the integrable equation alone, has relative L2L_2 error 1.381 and intensity error 0.296. Of its fields with L2L_2 error above 1, 84% have intensity error below 0.5. The revised text no longer uses "no information". Where the point matters it reads "no better than the zero prediction in relative L2L_2", and it reports the intensity error alongside (Fig. 4, Appendix D, Table 14).

Small residual size does not establish a learning-efficiency advantage.

A bound on the accumulated drift gives no bound on the vector field or on its Lipschitz constant, and we measured neither. Barron's bound depends on the first moment of the target's Fourier transform, which a small amplitude does not control. We have removed the factor-ϵ\epsilon argument and the O(ϵ)O(\epsilon) and Lipschitz statements (Sec. 4).

The long-time remainder is assumed.

The revised Proposition 2 is a formal first-order calculation for an isolated soliton, stated without a uniform remainder estimate. The long-time conclusions that relied on the remainder are removed (Sec. 4, Appendix J).

The trace formula does not give a waveform error law.

Removing the continuous spectrum is not an orthogonal projection in waveform space. At z=0z=0, on seven test inputs, the discrete-spectrum part and the remainder have a cosine of −0.02-0.02 to −0.07-0.07, and the waveform error is 1.18 to 1.30 times ρ\sqrt\rho. The revised text presents ρ\sqrt\rho as an empirical scaling on the tested input families, reports these deviations, and drops "law" and "controls" (Sec. 5.8).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.