acceptodds
Under review as a conference paper at ICLR 2027

Learning Contact Maps: Universality and Inversion Through Folded Fronts

Abstract

Symplectic map networks combine exact geometric structure with approximation guarantees. We develop the corresponding theory for contact maps, which also represent dissipative and action-dependent dynamics, and connect it to learned inversion. We propose two architectures: LegendreNet composes strict lifts of Henon-like symplectic maps, while ContactNet alternates these lifts with prolongations of invertible base maps. We prove that LegendreNet is dense in smooth strict contact maps on compact sets. For general contact Hamiltonian flow maps, we establish a modular approximation theorem: any strict family that is dense in , combined with prolongations of any base-map family that is dense in , gives approximation, with nonsingular factors on the relevant compact sets and convergence of local inverses. This turns contact splitting into a criterion for choosing architectural blocks. Smooth neural-flow bases satisfy the theorem's hypotheses, while triangular bases provide algebraic inverses for recovery. We benchmark the architectures against structure-preserving and generic baselines on damped Hamiltonian systems to quantify finite-capacity approximation. Finally, we use the contact structure of the generalized Hamilton-Jacobi equation to enable recovery of initial smooth profiles from terminal folded fronts.

Primary area:
learning on time series and dynamical systems
Keywords:
Structure-Preserving, Machine Learning, Invertible Neural Networks, Contact Geometry, Dynamical Systems

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GBcMOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose two neural architectures, LegendreNet and ContactNet, for learning contact maps that are invertible on regular domains.
LegendreNet composes strict contact blocks, while ContactNet combines them with prolonged base maps to model general contact dynamics.
The paper establishes strict universality and gives sufficient conditions for approximating smooth contact Hamiltonian flow maps on compact sets.
The inverse of the forward-trained map is then used to recover initial profiles from complete ordered one-dimensional folded fronts.

Critical strengths
  1. The modular approximation theorem provides a useful criterion for constructing expressive contact-preserving networks.

  2. The folded-front example provides a concrete illustration of reusing a forward-trained contact map for initial-profile recovery.

Critical weaknesses
  1. For the triangular-base ContactNet used in the inverse experiments, universality for general contact Hamiltonian flow maps relies on the Cr+1C^{r+1}-density of its base-map compositions, which is not established in the paper. Could the authors identify any non-strict target classes for which the current results provide approximation guarantees without this assumption?

  2. Although the paper provides derivative formulas and low-dimensional timing results, it does not assess how derivative-evaluation overhead and the capacity needed to maintain a fixed approximation accuracy change with dimension

Questions for authors
  1. How do computation, memory, and required depth scale with state dimension for the triangular and neural-flow variants?

  2. What geometric assumptions or additional measurements would enable momentum recovery in two or more dimensions?

  3. How does the proposed framework scale to long-horizon prediction in real-world applications, such as damped mechanical systems, in terms of stability and computational cost?

2 comments
AuthorsOpenReview ↗

Response to GBcM

We thank the reviewer for the careful reading and for the precise question about Corollary 4.3(b). It led to a stronger result: the triangular ContactNet is universal with no base-density hypothesis (see the general response). Below we answer that question directly, then the questions on dimension and horizon.

W1 (universality of the triangular ContactNet). The general response gives the stronger answer. LegendreNet atoms composed with positive xx-conditioned action shears are already CrC^r-dense, so Corollary 4.3(b) now holds with no base-density hypothesis.

The reviewer's question also has a direct answer within the submitted paper. For H=H0(x,p)+γ(x)u+β(x)H = H_0(x,p) + \gamma(x)u + \beta(x), which includes our damped oscillator, the flow of γ(x)u+β(x)\gamma(x)u + \beta(x) is an exact action shear defined on all of J1(Rn)J^1(\mathbb{R}^n). A refined Lie–Trotter product of a strict flow (Theorem 4.1) and this shear, with Lemma C.1, gives the guarantee without the new theorem. The new theorem extends this to every smooth contact Hamiltonian. It shows that brackets of strict Hamiltonians with x1ux_1 u generate all of them. The full proof, including the localization to a compact tube and the negative-time substeps, is in the new Appendix C.1.

W2 and Q1 (cost, memory and depth as nn grows). The triangular blocks were designed to avoid n×nn \times n solves. Each needs only scalar fields and their gradients on Rn\mathbb{R}^n, and its inverse is in closed form at the same cost. The uu-conditioned spatial block adds a rank-one correction instead of a linear solve. Only the neural-flow base of ContactNet-ODE needs an (n+1)×(n+1)(n+1) \times (n+1) Jacobian, with an O(n3)O(n^3) solve. We timed each block at width 96 for n=1n = 1 to 6464 (Appendix J.3, Table 25). Time and memory grow near-linearly in nn for the triangular blocks, and a full stage goes from 2.0 to 46 µs per state. Inverses cost the same as forward passes, and the neural-flow base costs 15 to 40 times a triangular stage.

A bound on depth at fixed accuracy remains open. The theorem is qualitative, like the universality results for coupling flows. We did measure depth on the one-step map of H=∣p∣2/2+u2/4H = |p|^2/2 + u^2/4 for n=1,2,4n = 1, 2, 4 and 88. At three stages, the error barely moves across widths 8 to 128, by less than a factor of two for n≥2n \geq 2. It does grow with nn, from between 2.3×10−72.3 \times 10^{-7} and 1.05×10−61.05 \times 10^{-6} at n=1n = 1, depending on width, to about 1.4×10−51.4 \times 10^{-5} at n=8n = 8. Doubling the number of stages cuts the error 6.6 times at n=4n = 4 and 2.5 times at n=8n = 8. Four times the training cuts it only 1.3 to 1.4 times. So at these sizes, the number of composed stages, not the width of each field, has to grow with nn. This rests on one doubling at one width with two seeds, so we report it as an observation, not a rate (Table 26).

Q2 (momentum recovery in two or more dimensions). The relation DsU=P DsXD_s U = P\, D_s X fixes PP wherever DsXD_s X is invertible, with no extra assumption. At a fold point, DsXD_s X has corank one with kernel vv. Differentiating along vv gives P⋅D2X(v,v)=D2U(v,v)P \cdot D^2X(v,v) = D^2U(v,v), which supplies the missing component when D2X(v,v)D^2X(v,v) is not in the range of DsXD_s X. Where that fails, as at the isolated cusps of a generic 2D front, PP follows by continuity, provided the singular set has empty interior. This gives an n-dimensional version of Proposition 6.1 (new Proposition H.2, Appendix H.3). In practice, the 2D inverse can now run from the projected front alone. It stays finite on all 11 test profiles, with median initial-front error 4.5×10−34.5 \times 10^{-3} against 2.3×10−52.3 \times 10^{-5} when momentum is supplied, limited by the momentum lift near caustics.

Q3 (long horizons on damped mechanical systems). Cost grows linearly with the horizon. Each step is one map evaluation with constant memory at inference. Training uses one-step pairs, so its cost does not depend on the horizon. Every step is an exact contactomorphism on its regular domain, so what accumulates over a rollout is fit error, not loss of structure. The conformal factor is a product over steps, so long-horizon dissipation accuracy is set by the per-step factor error, which we now report (Appendix J.1).

We ran the saved oscillator models to ten periods (800 steps, Appendix J.4), and every model stays finite on the test sets. Damped state MSE at ten periods is 2.3×10−52.3 \times 10^{-5} for GCF, 2.8×10−42.8 \times 10^{-4} for the MLP, 7.4×10−47.4 \times 10^{-4} for ContactNet, 1.4×10−31.4 \times 10^{-3} for RealNVP and 0.97 for LegendreNet. In the conservative case, MSE is 0.083 for GCF, 0.36 for ContactNet, 0.80 for LegendreNet, 3.5 for the MLP and 182 for RealNVP. On the wide-range test set, GCF diverges in every seed, between steps 58 and 92. A ContactNet step costs 0.11 ms at batch 32, against 0.8 to 1.1 ms for GCF.

Thank you for your thoughtful and thorough rebuttal. I sincerely appreciate the significant effort you have put into addressing my concerns, especially the additional theoretical results, experiments, and detailed analyses.

After carefully reading your responses, I realize that I may have misunderstood certain aspects of your proposed methodology. Your explanations and additional results have helped clarify these points and given me a better understanding of the contributions. I have decided to increase my score by one point.

Reviewer_ET7ROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LegendreNet and ContactNet, two structure-preserving architectures for learning invertible contact maps. It establishes universality for strict contact maps and a modular approximation theorem for general contact Hamiltonian flows, then investigates forward prediction and inverse recovery of folded Hamilton–Jacobi fronts.

Critical strengths
  • The proposed architectures preserve contact geometry by construction while providing tractable inverses, offering a principled extension of symplectic networks to dissipative and action-dependent dynamics.

  • The theoretical analysis provides a rigorous modular approximation framework, clarifying the smoothness requirements for contact-map approximation and local inverse convergence.

  • The experiments examine geometric consistency, forward prediction, and inverse recovery across several dynamical systems, providing useful evidence of the benefits and limitations of contact-preserving architectures.

Critical weaknesses
  • Corollary 4.3(b) leaves the required Cr+1C^{r+1} density of triangular base maps unproven, so the ContactNet variant used in the inverse experiments lacks a universality guarantee. Moreover, Equations (4) and (31) impose regularity conditions that the sampled training penalties do not guarantee on unseen trajectories
  • The inverse recovery results rely on restrictive observations and remain sensitive to noise. Proposition 6.1 assumes complete ordered fronts, while Table 20 shows recovery MSE increasing from 6.29×10−56.29 \times 10^{-5} to 3.54×10−33.54 \times 10^{-3} under noise of standard deviation 10−310^{-3}.
Questions for authors

see wekaness.

2 comments
AuthorsOpenReview ↗

Response to ET7R

We thank the reviewer for the supportive assessment and for pinpointing where our guarantees were conditional. Both points now have firmer answers. The general response removes the density hypothesis of Corollary 4.3(b). Below we show that the universal family it identifies needs neither the denominator condition of Eq. (4) nor the contraction of Eq. (31). For the optional blocks that do need them, we report the measured margins. We also show that the noise sensitivity in Table 20 stays at the scale of the momentum-lift error.

W1, density. The general response gives the revised theorem. One consequence matters for the reviewer's next point. The universal subfamily, LegendreNet atoms with positive xx-conditioned action shears, consists of maps defined on all of J1(Rn)J^1(\mathbb{R}^n) with closed-form inverses and positive conformal factors. It has no prolongation denominator and no fixed-point inverse. So universality no longer rests on the blocks whose regularity is enforced only on samples.

W1, regularity off the samples. We agree that sampled penalties certify nothing off the samples, and the paper says so in Section 8 and Appendix D. Beyond the point above, the two conditions differ in kind.

The contraction in Eq. (31) can be imposed by construction. Scale FF so that ∣ρ∣|\rho| times its Lipschitz constant in uu is at most λ<1\lambda < 1, as in invertible residual networks (Behrmann et al., 2019). Then the fixed-point iteration is a global contraction, and the factor 1+ρFu1 + \rho F_u stays at least 1−λ1 - \lambda everywhere. The revision adds this to Appendix D.4.

The denominator in Eq. (4) is intrinsic. By the matrix determinant lemma, for the n-dimensional block it equals a+p⋅wa + p \cdot w, with w=a′(u)x+b′(u)w = a'(u)x + b'(u). It therefore vanishes on a hyperplane of momenta unless the spatial map is independent of uu. A base map that mixes uu into xx cannot be prolonged over all momenta, however it is trained. This is an intrinsic property for contact systems. It can be kept regular on a stated range ∣p∣≤pˉ|p| \leq \bar{p} by bounding its uu-dependence (Appendix D.3).

Measured margins. Across every evaluated state of the 16 ContactNets behind the 1D and oscillator results, inverse recovery included, the smallest prolongation or inverse denominator is 0.263. That is above the training margin of 0.2. The largest measured vertical-block slope is 0.0215, and the fixed-point iteration ends with residual at most 9×10−169 \times 10^{-16}. The 2D block denominators stay at or above 0.766. Two cases leave the regular domain, and we report both. With noisy lifts up to σ=10−3\sigma = 10^{-3}, the smallest denominator stays at or above 0.44. With raw lifts at σ=10−2\sigma = 10^{-2}, the states leave the regular domain: prolongation denominators fall to 4×10−44 \times 10^{-4} and inverse denominators change sign, although every state stays finite. On a grid 10% beyond the test box, some seeds cross zero, and so do the oscillator models on the wide-range set (Appendix G.3).

W2 (observation assumptions and noise). Completeness and dimension are weaker assumptions than they read, ordering remains an assumption for folded fronts, and the noise sensitivity stays at the scale of the momentum-lift error.

Completeness. Proposition 6.1 is local, and the inverse acts point by point. From a contiguous quarter, half and three quarters of each folded front, the median recovery errors are 1.3×10−41.3 \times 10^{-4}, 8.5×10−58.5 \times 10^{-5} and 7.1×10−57.1 \times 10^{-5}, against 6.7×10−56.7 \times 10^{-5} for the full front, and every case stays finite.

Ordering. A nearest-neighbour reordering recovers the order of the unfolded fronts exactly but fails on all twelve folded ones, because the chain jumps branches at the cusp. For folded fronts the order must come from the measurement, such as rays labelled by launch angle or tracked markers.

Dimension. A new n-dimensional identifiability result (Proposition H.2, Appendix H.3) lets the 2D inverse run from the projected front alone. It stays finite on all 11 test profiles, with median initial-front error 4.5×10−34.5 \times 10^{-3} against 2.3×10−52.3 \times 10^{-5} when momentum is supplied, limited by the momentum lift near caustics.

Noise. In every row of Table 20, the recovery error stays between 0.75 and 1.26 times the momentum-lift error, so it stays at the scale of the lift error. The growth the reviewer notes is consistent with the ill-posedness of numerical differentiation (Proposition H.1), which any method that needs terminal momentum shares. A smoothing-spline lift, with its smoothing chosen by generalized cross-validation, cuts the recovery error at σ=10−2\sigma = 10^{-2} from 0.264 (raw lift) to 0.016.

Thank you for your quick reply and all my concerns have been addressed. And I keey my positive score.

Reviewer_QLmgOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LegendreNet and ContactNet, two neural architectures for learning invertible contact maps. The goal is to extend structure-preserving symplectic networks to dissipative and action-dependent dynamics. LegendreNet uses strict lifts of Henon-like symplectic maps, while ContactNet combines these lifts with prolongations of invertible base maps. The authors establish approximation guarantees and apply the learned inverse to recover initial profiles from folded Hamilton-Jacobi fronts. Experiments show that the proposed architectures preserve contact geometry and support inverse recovery.

Critical strengths
  • The proposed architectures preserve contact geometry by construction.
  • The theoretical analysis provides approximation guarantees with clear assumptions.
  • The folded-front recovery task is an interesting application, supported by geometric and noise-sensitivity analyses.
Critical weaknesses
  • The universality result for general contact flows applies to the neural-flow base, but the triangular base used in the main inverse experiments only has a conditional guarantee (Corollary 4.3). Algebraic inversion is one of the main motivations for the proposed architecture, yet inverse recovery is not evaluated for the theoretically supported neural-flow variant.
  • The empirical results do not show a substantial advantage over existing methods. GCF performs better on both oscillator benchmarks, and RealNVP achieves lower state error on the two action-dependent 1D front benchmarks. The strong 2D result is promising, but it uses only one training seed and selects checkpoints using the reported test profiles. Unequal training budgets in some comparisons also make the results not convincing enough.
  • The inverse recovery experiments rely on restrictive observation assumptions. The 1D setting requires complete and correctly ordered terminal fronts, and the 2D experiment provides the ground-truth terminal momentum. The noise experiments also show substantial error increases even after smoothing. This limits the demonstrated applicability of the recovery pipeline beyond clean synthetic settings.
Questions for authors
  • Could the authors compare the learned contact conformal factor with the analytical factor on the damped oscillator? This would help assess whether the model learns the correct dissipation, rather than simply satisfying the contact constraint.
  • Could the authors report the minimum absolute prolongation denominator along the evaluated trajectories, especially during inverse recovery?
3 comments
AuthorsOpenReview ↗

Response to QLmg (1/2)

We thank the reviewer for a careful reading and for two concrete diagnostics we should have reported. The general response above shows that the triangular ContactNet used in all our inverse experiments is universal without the density hypothesis. That closes the gap between the variant we prove results for and the one we evaluate. Below we answer each point, and we state the empirical scope more plainly: ContactNet is not the most accurate model everywhere, and we say where it leads and why.

W1 (universality and the inverse experiments). The reviewer rightly points out a gap. Our unconditional guarantee covered the neural-flow variant, while the inverse experiments used the triangular one. Both sides of this gap are now closed.

First, the new theorem (general response) makes the triangular ContactNet universal. Its universal subfamily, LegendreNet atoms with positive xx-conditioned action shears, is invertible in closed form everywhere. On the free fronts of Table 11 this subfamily is comparable to the full ContactNet: its forward error is lower (3.92×10−13.92 \times 10^{-1} against 5.03×10−15.03 \times 10^{-1} at step 12) and its inverse error higher (3.24×10−13.24 \times 10^{-1} against 2.19×10−12.19 \times 10^{-1}). We have now also trained it on the quadratic fronts. There it inverts exactly (cycle error 1.2×10−121.2 \times 10^{-12}) and recovers jets as Legendrian as ContactNet's. Its recovery error is higher (1.57×10−21.57 \times 10^{-2} against 2.74×10−42.74 \times 10^{-4}), because its forward fit is worse on this action-dependent target.

Second, we now evaluate the ContactNet-ODE inverse: a backward solve followed by the prolongation inverse. Over 80 steps it returns to its input with mean squared error below 10−1610^{-16}, or about 10−2910^{-29} with a Newton step on the discrete base map. Its recovery error tracks its forward fit. It is two to five times ContactNet's on the oscillator and level with ContactNet's on transport. The results are in Appendix J.2 of the revision.

W2 (the empirical case). We agree that no architecture dominates, and the revision makes this the point of the comparison. As the general response explains, the rankings reflect inductive bias at finite depth. Below, each expectation comes from a method's generators first, with the measured result after.

GCF. Its generators M(p)p2/2M(p)p^2/2, V(x)V(x) and F(x)uF(x)u contain the oscillator Hamiltonian term by term. So it should lead on the oscillators, and it does (Section 5.2). A u2u^2 term costs GCF two nested brackets, while ContactNet's uu-dependent blocks represent it directly. On the quadratic fronts, ContactNet's forward error is about 55 times lower than GCF's, and its recovery error is 22 to 29 times lower. In the sweep over H=p2/2+x2/2+γuk/kH = p^2/2 + x^2/2 + \gamma u^k/k (Appendix F.3), ContactNet leads for k=2k = 2 and 44 and GCF for k=1k = 1 and 33. The account does not explain k=3k = 3, and we note that in the text.

RealNVP. On the action-dependent fronts, RealNVP has lower state error (5.5×10−55.5 \times 10^{-5} against 1.1×10−41.1 \times 10^{-4} quadratic, 6.8×10−56.8 \times 10^{-5} against 8.0×10−48.0 \times 10^{-4} intensity). That comes with Legendrian residuals 24 to 42 times larger (Table 16), so its predicted momentum is not the slope of its predicted front. Its recovered initial jets are 39 to 50 times less Legendrian than ContactNet's. In the intensity regime its lead also depends on a longer L-BFGS schedule. At the matched budget of Table 17, the MLP is first.

2D experiments. We agree that one seed was not enough. We retrained with checkpoints selected on fresh validation profiles, three seeds per architecture. The MLP-to-ContactNet error ratios are 103.6, 20.8 and 141.3. We now report a gap of 20 to 140 times, about 80 on average, in place of the single figure of 151. ContactNet's forward state MSE is stable across seeds (6.07×10−5±1.0×10−56.07 \times 10^{-5} \pm 1.0 \times 10^{-5}), while the MLP's is not (4.95×10−3±3.1×10−34.95 \times 10^{-3} \pm 3.1 \times 10^{-3}).

Budgets. Table 17 already compares all models at a matched 150 Adam epochs and 300 L-BFGS steps, which changes one ranking (intensity). The MLP's threefold parameter count remains, as Appendix E.2 states.

AuthorsOpenReview ↗

Response to QLmg (2/2)

W3 (observation assumptions). We agree these assumptions matter. The new experiments show which of them can be relaxed and which cannot.

Completeness. Proposition 6.1 is local. An observed arc fixes pp on its interior, and the inverse acts point by point. We recovered from a contiguous quarter, half and three quarters of each folded front. The median recovery errors are 1.3×10−41.3 \times 10^{-4}, 8.5×10−58.5 \times 10^{-5} and 7.1×10−57.1 \times 10^{-5}, against 6.7×10−56.7 \times 10^{-5} for the full front, and every case stays finite.

Ordering. This stays an assumption for folded fronts. A nearest-neighbour reordering recovers the order of the unfolded fronts exactly. It fails on all twelve folded ones, because the chain jumps branches at the cusp. In practice the order comes from rays labelled by launch angle, tracked markers, or a time series of fronts.

Dimension. A new n-dimensional identifiability result (Proposition H.2, Appendix H.3) fixes the momentum wherever the front is regular and at fold points, and by continuity elsewhere. With it, the 2D inverse runs from the projected front alone and stays finite on all 11 test profiles. Its median initial-front error (4.5×10−34.5 \times 10^{-3}, against 2.3×10−52.3 \times 10^{-5} with supplied momentum) is limited by the momentum lift near caustics on our coarse 20×2020 \times 20 mesh.

Noise. The noise sensitivity stays at the scale of the momentum-lift error. In every row of Table 20, the recovery error stays between 0.75 and 1.26 times the lift error. A smoothing-spline lift, with its smoothing chosen by generalized cross-validation, cuts the recovery error at σ=10−2\sigma = 10^{-2} from 0.264 (raw lift) to 0.016. Fed the same noisy lifts, RealNVP recovers equally well. ContactNet's recovered jets are 94 times more Legendrian than RealNVP's on clean lifts and 25 times with the smoothed lift at σ=10−3\sigma = 10^{-3}. With raw noisy lifts the lift error dominates, and the advantage shrinks to 1.5 to 3.5 times.

Q1 (learned versus analytic conformal factor). The conformal factor can be read off every trained contact model, and we should have reported it in the original submission. For H=(p2+x2)/2+0.15uH = (p^2 + x^2)/2 + 0.15u, the exact time-hh factor is the constant e−0.15h=0.98829e^{-0.15h} = 0.98829. LegendreNet, whose factor is one, is off by 1.17×10−21.17 \times 10^{-2} by construction. Along the test trajectories, the mean per-step error is 3.2×10−33.2 \times 10^{-3} for ContactNet, about a quarter of the per-step dissipation. It is 5.6×10−45.6 \times 10^{-4} for GCF, whose damping generator matches 0.15u0.15u exactly, and 2.0×10−32.0 \times 10^{-3} for ContactNet-ODE. Over one period, ContactNet's accumulated factor is right on average (0.385 against 0.390), but it varies across trajectories: relative error 0.22 on average and 1.41 at worst, against 0.026 and 0.063 for GCF. The learned factors stay positive along these trajectories, and their accuracy follows the fit. The revision reports the factor for every contact model (Appendix J.1, Table 23).

Q2 (smallest prolongation denominator). We now report it along every evaluated trajectory, inverse recovery included (Appendix G.3). On clean data, across the 16 ContactNets behind the 1D and oscillator results, the smallest prolongation or inverse denominator is 0.263, above the training margin of 0.2. The 2D block denominators stay at or above 0.766. We note that regularity is local. With noisy lifts up to σ=10−3\sigma = 10^{-3}, the smallest denominator stays at or above 0.44. With raw lifts at σ=10−2\sigma = 10^{-2}, the states leave the regular domain: prolongation denominators fall to 4×10−44 \times 10^{-4} and inverse denominators change sign, although every state stays finite. On a grid 10% beyond the test box, some seeds cross zero, and so do the oscillator models on the wide-range set. This is why the universal subfamily we propose matters: it has no prolongation denominator at all (see general response).

Thank you for the detailed rebuttal and the additional experiments. I appreciate the authors' effort in addressing my concerns.

My concerns in W1, W3, Q1, and Q2 are largely resolved. For W2, the results remain somewhat mixed, but I view this more as a limitation on the strength of the empirical claim, rather than a fundamental issue with the paper's core contribution. The main theoretical contribution is better supported in the revision, and the new experiments make the empirical picture clearer and more convincing.

Overall, I think the authors have made a substantial effort in the rebuttal, and the paper has improved. I will therefore raise my rating from 2 to 3.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank the reviewers for their careful reading. All three asked whether the triangular ContactNet is universal without the base-density hypothesis of Corollary 4.3(b). Motivated by this question, we can now prove that it is.

The proof is Lie-algebraic. Strict Hamiltonians together with the single non-strict Hamiltonian x1ux_1u generate every contact Hamiltonian under brackets. So any family that approximates their flows approximates every contact flow map. Triangular ContactNets meet this with their simplest blocks, LegendreNet atoms and positive x-conditioned action shears, which are defined on all of J1(Rn)J^1(\mathbb{R}^n) and have explicit inverses. GCF meets it at the level of its ideal generators. The modular theorem (Theorem 4.2) remains a valid route for base families already known to be dense. The full statement and proof are in the new Appendix C.1 (Theorem C.9).

This changes how the experiments should be read. Every contact method we compare that has a non-strict generator is universal in the contact class. That covers ContactNet, ContactNet-ODE, and GCF at the generator level. The MLP and RealNVP are universal outside that class, without preserving its geometric structure, and LegendreNet is universal only for strict maps. The comparisons therefore measure inductive bias at finite depth rather than expressivity. Methods tend to be most accurate where the target's generators lie in their own vocabulary. That is why GCF leads under linear damping and ContactNet leads among the contact models on the quadratic-action fronts, though this reading does not predict every ranking. The introduction and Section 5 of the revision now note this directly.

Beyond the theorem, the revision adds new experiments in three groups:

  • The first reports diagnostics of the experiments already performed, such as the learned conformal factor and the regularity margins along every evaluated trajectory.
  • The second tests the method's limits: cost and depth as the dimension grows, long horizons, more seeds in 2D, and weaker assumptions on the observed fronts.
  • The third checks the subfamily singled out by the new theorem, and the inverse of the ContactNet-ODE variant. We discuss numerical details in our individualized responses below, and all of the numerical results appear in the revised appendices.
AuthorsOpenReview ↗

Closing Note

We thank all three reviewers and the area chair for their time and engagement during this discussion period. We are grateful that all three reviewers found that the responses and the revision addressed their concerns, and for the updated assessments.

For convenience, we summarize the additions in the revision below. Changes are marked in blue in the annotated PDF.

  • Universality of the triangular ContactNet (Appendix C.1, Theorem C.9). Strict Hamiltonians, together with the non-strict Hamiltonian x1ux_1 u, generate every contact Hamiltonian under brackets, so LegendreNet atoms with positive xx-conditioned action shears are CrC^r-dense. Corollary 4.3(b) now holds with no base-density hypothesis, through a subfamily that is defined on all of J1(Rn)J^1(\mathbb{R}^n) with closed-form inverses.
  • Reading of the experiments. Since every contact method we compare with a non-strict generator is universal by the new theorem (GCF at the level of its ideal generators), the comparisons measure inductive bias at finite depth rather than expressivity. The introduction and Section 5 now say this directly.
  • Momentum recovery in nn dimensions (Proposition H.2, Appendix H.3). The 2D inverse now also runs from the projected front alone.
  • New diagnostics and tests (Appendices G.3 and J). The learned conformal factor of every contact model, regularity margins along every evaluated trajectory, the ContactNet-ODE inverse, cost and depth as the dimension grows, ten-period rollouts, three seeds in 2D, and recovery from partial and noisy fronts.

The reviewers' questions led directly to the main new result, and the paper is stronger for them. We thank everyone again for a constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.