acceptodds
Under review as a conference paper at ICLR 2027

Why Projected Hamiltonian Flow Matching Can Learn Zero

Abstract

Generative models have recently incorporated Hamiltonian structure to align the generation process with physical constraints. However, the resulting flows may preserve energy yet fail to learn the required probability transport. We study projected scalar-Hamiltonian velocities and construct smooth positive density paths on compact phase-symmetric shells whose unique population Flow Matching optimum is zero. Orbit averaging identifies the obstruction and characterizes the exact divergence range on round shells. Under a density-ratio assumption, a finite-excess-risk bound gives a quantitative endpoint error floor for approximate fits. We also characterize the closed regression spaces that preserve every teacher's density response: they contain the closure of tangent gradients. A fixed-budget class–path comparison measures how this population capacity translates into neural learning, and a source-filtering diagnostic identifies an exact conditional sampler in the physical-data task. For an explicit two-oscillator construction, the population-optimal flow retains its source law and incurs endpoint total-variation error .

Primary area:
generative models
Keywords:
Flow Matching, Hamiltonian Dynamics, Constrained Generative Modeling, Probability Transport, Expressivity

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Dmc1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper asks whether flow matching over projected scalar-Hamiltonian velocities $ P_H J\nabla h$, which conserve a prescribed invariant HH exactly, can still learn the probability transport needed to change a density on a fixed energy shell. Its contributions are:

  1. Main result. On compact shells H=F(I1,…,In)H=F(I_1,\dots,I_n), n≥2n\ge2, with independent oscillator phase symmetry, every positive action-dependent path ρt=1+tϵa(I)\rho_t=1+t\epsilon a(I) is generated by a phase-invariant tangent gradient that is L2(ρt)L^2(\rho_t)-orthogonal to the whole class. Hence zero is the unique population FM-optimal velocity, and the learned flow keeps its source. On the round sphere the attainable divergences are exactly the functions with zero average along every Hopf circle, {div⁡Sv:v∈VH}=ker⁡A\{ \operatorname{div}_S v: v\in \mathcal{V}_H \}=\ker\mathcal{A}. For two oscillators on S3S^3 with target (z+12)μ(z+\tfrac12)\mu, this gives endpoint error TV=1/8\mathrm{TV}=1/8 (action W1=1/12W_1=1/12).
  2. Response-preserving regression. A closed linear regression space preserves every teacher's density response if and only if it contains the closure of tangent gradients. For a fixed backbone, the minimum-norm correction solves a weighted Poisson equation and costs ∥Rb∥H˙−1(ρ)\|R_b\|_{\dot H^{-1}(\rho)}. This motivates the gradient-completed class PH(J∇h+∇g)P_H (J\nabla h+\nabla g).
  3. Validation. A pre-declared comparison of backbone, gradient, mixed and generic tangent classes at a fixed 1{,}000-update budget uses two paths on S3S^3. On the analytic path, action W1W_1 falls from 0.0840 (backbone) to 0.0079 (gradient). On the geodesic bridge the generic tangent class is best, while the mixed class gains little over the backbone. Further experiments cover a quartic shell, designed physical controls and FPUT-32 trajectory data, where filtering the source is an exact conditional sampler and outperforms all neural models.
Critical strengths
  1. The theory is correct and fully explicit. The counterexample is in closed form, and the stated constants are reproducible: TV=1/8\mathrm{TV} =1/8, action W1=1/12W_1=1/12, the teacher energy, the geodesic-bridge coefficients and the quartic-shell values. It could serve as a clean cautionary example.
  2. Paper gives a checkable design rule and a diagnostic. Prop. 5 gives an if-and-only-if condition. A closed linear regression class preserves every teacher’s density response exactly when it contains the closure of tangent gradients and Prop. 6 gives the minimal correction explicitly, and its cost equals the H˙−1(ρ)\dot H^{-1}(\rho) norm of the missing response.
Critical weaknesses
  1. The motivation and the contribution are not established. The paper does not show that the analyzed class $ P_H J\nabla h$ is used by any existing method; the cited Hamiltonian generative models rely on unprojected or augmented-phase-space fields.
  2. Symmetry. The obstruction requires exact phase symmetry, and no result is given for nearly symmetric settings. This assumption appears strong; however, the paper provides neither a comparison with the assumptions of other approaches nor an analysis of its practical feasibility.
  3. The text is dense and heavily hedged. Terms and notation are used before they are defined, and symbols are reused for different objects. Moreover, the corresponding literature is not discussed enough: equivariant flows/FM, Benamou–Brenier/Otto/Ambrosio–Gigli–Savaré, Riemannian generative models, and Flux Matching (Pao-Huang et al., 2026). Finally, the main text contains more than seven theoretical propositions, yet none of them is singled out as the central or primary one—the very point for which the work was written—supported by clear and convincing motivation.
  4. The experiments do not demonstrate the contribution, and the presentation obscures it. There are no comparisons with competing approaches, such as Riemannian FM on the shell, constrained sampling (PCFM, HardFlow), or likelihood-trained flows. Also the proposed remedy is not supported: the generic tangent class is best or tied wherever it is evaluated, yet it is omitted from the FPUT and physical-control experiments. On the geodesic bridge, the mixed model barely improves on the backbone (action W1W_1 0.08170.0817 vs. 0.08400.0840).
Questions for authors
  1. Which existing generative method uses $ P_H J\nabla h$ on a prescribed shell? If none, why would a practitioner prefer it over projecting a generic velocity field, which also conserves HH?
  2. Is the failure due to the class or to the FM objective? Appendix shows that compositions of backbone fields escape the instantaneous divergence range. Does a backbone flow trained by maximum likelihood still fail to reach the target?
  3. Why does the mixed model, which contains the mean teacher, barely improve over the backbone on the geodesic bridge? Is this an optimization or budget effect, and do learning curves run to convergence change the ranking?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Dmc1

We thank the reviewer for the detailed review, and for confirming that the theory is correct and that Props. 5–6 give a checkable design rule and diagnostic.

W1 / Q1: Motivation; why not project a generic field? We agree that the motivation needed to be clearer. The revised introduction gives FINDE as the origin of the form and explains why it is a natural candidate for conservation-preserving generation (see our response to 3M58, Q1). We also agree with the implied recommendation: a projected generic field (V, the RFM tangent parameterization) or a gradient branch added to the backbone (M) satisfies Prop. 5 and is the right design. The paper explains why the projected-Hamiltonian class fails where generic projection succeeds, and gives a criterion for auditing any closed linear class.

W2: Exact phase symmetry. The revision adds Theorem 7 on the ellipsoid ΛδS3\Lambda_\delta S^3 (General Response §4): the optimum is of order ∣δ∣|\delta| with a positive endpoint lower bound for small ∣δ∣|\delta|. A single Limitations paragraph (Section 5) states the boundary. Our strict-zero construction uses phase symmetry; such obstructions are non-generic, but they arise naturally when generic states have a nontrivial stabilizer under the symmetries of (H,ρ)(H,\rho), as for orbit orientation in central-force motion, where we prove a response obstruction for smooth fields (General Response §5). On generic asymmetric shells, including FPUT-32 and Hénon–Heiles, B's deficit is quantitative rather than a zero, which is where the empirical witness is useful.

W3: Presentation and literature. We accept this criticism. In the revision we: (i) name Theorem 1 and Prop. 5 as the central results in the introduction, keeping the numbering of the original main-text results; (ii) keep the notation paragraph of the Appendix guide, which separates DρD_\rho from the radial derivative D=x⋅∇D=x\cdot\nabla and the two uses of bb, and add the new symbols there (the 10-page limit leaves no room for a table); new symbols are renamed to avoid clashes; (iii) collect hedges into one Limitations paragraph and remove several from the main text; (iv) add the literature: symmetry breaking in equivariant networks and equivariant flows (Smidt et al., 2021; Kaba & Ravanbakhsh, 2023; Lawrence et al., 2025; Köhler et al., 2020), Benamou–Brenier / Otto / AGS (identifying Gρ\mathcal G_\rho as the Wasserstein tangent space), FINDE, Riemannian FM, and Flux Matching (Pao-Huang et al., 2026), which changes the objective whereas our obstruction is at the level of the velocity class.

W4: Experiments and comparisons.

  • Riemannian FM: on the geodesic bridge, V trained with geodesic labels is Riemannian conditional FM, and the revision now says so.
  • Constrained sampling (PCFM, HardFlow): we added ambient FM with final or step-wise state projection, which trails V on S3S^3 and FPUT (General Response §6); it is not PCFM or HardFlow, and we make no claim against them.
  • V in FPUT and all five physical controls: on independent-trajectory FPUT, V has the lowest full-state test SW1 (0.1101), while filtering keeps the highest hit rate and the lowest mode W1W_1. In the physical controls, V is close to the oracle; its large fput16 advantage over M (0.056 vs. 1.059) is on the target observable only, since full-state SW1 is nearly equal.

Q2: Does a maximum-likelihood (CNF) backbone reach the target? With exact surface-divergence likelihood training (500 updates, 5 seeds), B stays at W1=0.0840±0.0024W_1=0.0840\pm0.0024 (final loss −0.0004-0.0004) and V reaches 0.0083±0.00240.0083\pm0.0024 (−0.0502-0.0502). Because compositions of backbone fields can leave the instantaneous range (App. A.5), we report this as an observation for this protocol.

Q3: Training up to 30k updates. See General Response §3. At 3k updates the gap largely closes (G 0.0086±0.0032, M 0.0077±0.0023, V 0.0083±0.0024). At 30k, M reaches 0.0034±0.0003 (mean-teacher NMSE 0.010), G 0.0037±0.0006 (0.009), and V 0.0047±0.0014 (0.018), while B stays at 0.0840 (NMSE 1.000) at every budget. This is consistent with the 1k deficit of M being a label-noise and budget effect, whereas B's failure is structural.

Thank you for your response, it addressed some of my questions. Unfortunately, I could not discern any clear practical advantage - based on your answers - of the constructed counterexample and the proposed method over other approaches to Hamiltonian or conservative-field flow matchings

Reviewer_3M58OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors demonstrate that when flow matching is performed using the velocity field v=PHJ∇hv = P_HJ\nabla h, which structurally ensures energy conservation, the population-level optimum is v≡0v \equiv 0 even on phase-symmetric energy shells where the density varies. The authors also show that adding the gradient field restores the density response at the population level. These results have been verified in small-scale experiments.

Critical strengths
  • The counterexample to the closed-form solution and the precise characterization of the divergence range on the sphere are significant contributions.
  • The authors use both theory and comparative experiments to distinguish that the cause of failure lies in the velocity class rather than the objective function.
  • They explicitly state the limitations of the method’s scope.
Critical weaknesses
  • No specific applications requiring the targeted conservative generation flow have been presented, so its significance is unclear.
  • No existing methods using the PHJ∇hP_HJ\nabla h under analysis have been presented. The cited methods, such as HGN, are techniques that learn HH and thus have a different setup.
Questions for authors
  • This paper is theoretically sound, and its theoretical results are interesting; it serves as a useful guideline for embedding Hamiltonian structures into machine learning models. However, it is unclear which existing methods actually experience the collapse described in this paper. Are HGNs and Hamiltonian generative flows included in this setting? If not, please clarify the scope of the claim.
  • Furthermore, can the authors identify at least one specific application scenario? Are there settings where simple filtering or posterior projection cannot be used, necessitating the proposed countermeasures?

If the usefulness and significance of these theoretical results are made clear, I am prepared to swing my vote toward acceptance.

1 comment
AuthorsOpenReview ↗

Response to Reviewer 3M58

We thank the reviewer for recognizing the counterexample and the divergence-range characterization, and for stating clearly what would change the assessment.

Q1: Which existing methods experience the collapse? Are HGNs and Hamiltonian generative flows included? No. HGN (including its Hamiltonian normalizing flow), Hamiltonian generative flows (Holderrieth et al., 2024), symplectic generative networks, and HNN-type models are not in the analyzed class: they learn HH to model dynamics, or use unprojected volume-preserving fields on an augmented phase space and marginalize. The revised introduction says so and states the scope:

The zero-optimum results (Thms. 1–3) concern path-regression training of velocity classes that conserve a prescribed invariant HH exactly per sample and are parameterized as PHJ∇hP_HJ\nabla h for a learned scalar hh. They do not apply to augmented-phase-space models. The response criterion (Prop. 5) applies to any closed linear velocity class.

The form arises when Hamiltonian-network dynamics are projected onto a first integral (FINDE, Matsubara & Yaguchi, 2023); on a prescribed shell it gives a Hamiltonian velocity exact conservation without augmentation, which makes it a natural candidate for conservation-preserving generation. The revised introduction motivates the class this way. Used directly as an FM velocity, J∇hJ\nabla h is divergence-free and faces the equimeasurability obstruction (Draxler et al., 2024); projection onto a fixed shell makes the divergence nonzero and appears to remove it. We show it removes it only partially: responses stay in ker⁡A\ker\mathcal A, and regression can discard the whole teacher. The mechanism is not tied to the designed example: General Response §5 gives a response obstruction for orbit orientation in 3D central-force motion.

This is related to, but distinct from, Curie's principle for equivariant networks (Smidt et al., 2021; Kaba & Ravanbakhsh, 2023; Lawrence et al., 2025). Here hh need not be equivariant; the symmetry enters through the class's density response, so network-side symmetry breaking (pose conditioning, symmetry-breaking encodings, noise) does not remove it. These references are now cited.

Physical intuition. A single sample moved by PHJ∇hP_HJ\nabla h can change its energy share; what is restricted is the distribution. At phase-invariant densities the backbone has zero weighted response to every action-dependent test, so it cannot change the law of the energy share at fixed total energy. In central-force motion, likewise, smooth backbone fields have no instantaneous response in the orbit-orientation law at rotation-invariant densities. Adding PH∇gP_H\nabla g supplies the missing channel; it satisfies the degeneracy condition of a metric bracket (PH∇H=0P_H\nabla H=0), so M is analogous to a symplectic + metric split, although B is not the conservative branch of a metriplectic system. Prop. 6 gives the minimal completion.

Q2: A concrete scenario where filtering or posterior projection is not adequate.

  1. Uniform-proposal rejection becomes inefficient. Its acceptance rate scales with 1/sup⁡(ρ1/ρ0)1/\sup(\rho_1/\rho_0). On the 16D tilted target (table in the response to eeT5), acceptance with 16,384 candidates drops from 17.8% (β=2\beta=2) to 0.17% (β=8\beta=8, W1W_1 0.0472 vs. 0.0219 for M) and to 2.3 samples at β=12\beta=12, where we report failure; M and V move every sample. Rejection is most accurate at β=2\beta=2, and for this particular target a one-dimensional inverse-CDF sampler would also apply.
  2. Projection after ambient FM distorts the law when the flow drifts off-shell (raw energy error 0.0429 on S3S^3, 0.2810 on FPUT): W1W_1 0.0325–0.0366 on S3S^3 vs. 0.0107 for V at the same budget, and test SW1 0.1502 on FPUT vs. 0.1101 for V. Projection also does not by itself preserve several coupled invariants, such as HH and LL in central.
Reviewer_eeT5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates the expressive limitations of projected scalar-Hamiltonian velocity fields in constrained Flow Matching. The authors construct smooth, energy-conserving probability paths for which the population-optimal Flow Matching solution within the restricted Hamiltonian class is the zero velocity, despite the target distribution changing over time.

The paper provides a theoretical characterization of the attainable density responses through orbit averaging, derives a quantitative endpoint error bound under additional density-control assumptions, and studies the conditions under which regression preserves the target density response. To address the identified limitation, the authors consider augmenting the Hamiltonian backbone with tangent-gradient fields.

The theoretical findings are supported by experiments on analytic and geodesic probability paths, a nonquadratic energy shell, and several physical-system benchmarks, including FPUT systems.

Critical strengths
  1. Valuable theoretical insights.

The paper identifies an interesting limitation of Projected Hamiltonian Flow Matching, demonstrating that energy conservation does not necessarily guarantee correct probability transport. The theoretical analysis is rigorous and provides useful insights into constrained generative modeling.

  1. Systematic theoretical analysis.

The authors develop a relatively complete theoretical framework and provide experiments across different probability paths and physical systems to support their findings.

Critical weaknesses
  1. Limited practical significance and insufficient validation on realistic generative tasks.

My main concern is the gap between the theoretical findings and their practical implications. Most experiments primarily validate the theoretical observations rather than demonstrate clear benefits in practical generative modeling. For example, in the FPUT-32 experiment, the target conditional distributions can already be obtained through source filtering, which substantially outperforms the learned models.

I believe the paper would benefit from more challenging physical conditional generation tasks where simple source filtering is insufficient. Comparisons with established approaches, such as Physics-Constrained Flow Matching (NeurIPS 2025) and Riemannian Flow Matching (ICLR 2024), under comparable settings would also help demonstrate the practical significance of the proposed approach.

  1. Insufficient experimental analysis and presentation.

Although the paper contains several experiments, some results and visualizations provide limited additional insight. For example, Figure 2 mainly presents changes in Mean Action Share, offering relatively little information beyond the quantitative comparisons.

More importantly, the substantial performance gap between the Mixed and Tangent models on the geodesic path remains insufficiently explained, despite their theoretical capacity to represent the required density response. A deeper investigation into the effects of optimization, model parameterization, and probability path selection would strengthen the experimental conclusions.

Questions for authors

Questions For Authors

1. Practical applicability.

Could the authors clarify how the theoretical findings can guide the design of practical generative models? Beyond demonstrating the identified limitations, what concrete benefits does the proposed approach offer over existing methods in realistic applications?

2. Gap between theory and experiments.

In Table 1, the Mixed model significantly underperforms the Tangent model on the geodesic path, despite having sufficient theoretical expressivity. Could the authors explain the main reasons for this discrepancy? Does it indicate optimization difficulties or limitations of the proposed parameterization?

1 comment
AuthorsOpenReview ↗

Response to Reviewer eeT5

We thank the reviewer for finding the analysis rigorous and the framework systematic.

W1 / Q1: Practical significance and design guidance. We agree this was under-stated. Section 5 now gives three rules:

  1. Audit the velocity class as well as the loss. A class can conserve HH exactly and still learn zero transport. An empirical estimate of the witness Eq. (9), computed from conditional labels and model outputs, flags a lost direction.
  2. For closed linear classes, include the tangent gradients (Prop. 5). This guarantees the response for every teacher. G, M (the backbone plus a gradient branch), and V (a projected generic field) all qualify, though a given task may need fewer directions. In our runs V learned fastest at short budgets (e.g., the fput16 observable at 300 updates), and M was within one SD of V from 3k updates on S3S^3.
  3. A projected-Hamiltonian branch is a prior, not a requirement. On hh_dynamics, whose flow preserves the source law, B, M, and V coincide; on transport teachers B fails. With a gradient branch added, M keeps the Hamiltonian prior and reaches V's endpoint accuracy with enough updates.

W1: FPUT and source filtering. We agree. In the original task the condition is a function of the source state and the source is an equal mixture of the three pools, so filtering is exact. The revision adds two tasks without this feature:

  • Independent trajectories (216 trajectories in 3 datasets, trajectory-level splits): full-state test SW1 is V 0.1101, M 0.1135, filtering 0.1268, B 0.1408; filtering keeps the highest hit rate and the lowest mode W1W_1.
  • Tilted target on S15S^{15} (∝eβzμ\propto e^{\beta z}\mu, 3 seeds, 16,384 uniform candidates):
β\betaPriorRejection: accepted, W1W_1MV
20.02922918, 0.00300.02910.0061
40.0672533, 0.00860.06660.0069
80.178627.3, 0.04720.02190.0227
120.31902.3, fail0.05020.0561

B equals the prior at every β\beta. Rejection is best at mild tilt (β=2\beta=2) and close to V at β=4\beta=4, but its acceptance collapses as the tilt grows. M stays near the prior at β≤4\beta\le4, which we have not yet explained. Since this target's action law is one-dimensional, an inverse-CDF sampler would also work here; the point is the cost of uniform-proposal rejection under a fixed candidate budget (training cost not included).

W1: PCFM and Riemannian FM. On the geodesic bridge, V trained with geodesic labels is Riemannian conditional FM. We also added ambient FM with final or step-wise state projection: W1W_1 0.0325 / 0.0366 on S3S^3, against 0.0107 for V at the same budget, and test SW1 0.1502 on FPUT (General Response §6). This baseline is not PCFM or HardFlow, so we make no claim against those methods.

W2: Figure 2. Replaced by three panels: (a) 30k-update endpoint histograms against the exact target, (b) error against training budget (1k–30k), and (c) the label ablation.

W3 / Q2: Mixed vs. Tangent on the geodesic bridge. Mean-teacher labels bring M from 0.0817 to 0.0073 at 1k updates, and conditional-label training to 30k brings M to 0.0034±0.0003 (V 0.0047±0.0014). The 1k gap is therefore a label-noise and budget effect, not a capacity limit (General Response §3). We do not claim M is generally better than V; any class satisfying Prop. 5 preserves the response.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers (Part 1 of 2)

We thank all reviewers for careful reviews; all three found the theory correct and explicit. Shared concerns were (i) motivation and scope, (ii) practical significance and comparisons, (iii) the Mixed vs. Tangent gap on the geodesic bridge, and (iv) presentation. A revised PDF is uploaded; new text is in green. The numbering of the original main-text results and equations is retained; new results are added as Theorem 7 and Proposition 8, and appendix proposition numbering has been updated.

1. Reframed contribution. The revised abstract and introduction lead with the general question: when FM regresses a teacher onto a structured, conservation-preserving velocity class, which density changes survive? Prop. 5 answers it for closed linear classes: population regression preserves every teacher's density response iff the class contains the closed tangent gradients, i.e., the Wasserstein tangent space at ρ\rho (Benamou–Brenier, Otto). The projected scalar-Hamiltonian class (B) is the sharpest instance: on phase-symmetric shells its population optimum for a changing path is zero. The witness Eq. (9) gives a population lower bound; its sample estimate is an empirical diagnostic. The introduction now names Theorem 1 and Prop. 5 as the main results.

2. Class V and Riemannian FM. PHwP_Hw is the tangent parameterization used by RFM (Chen & Lipman, 2024), and on the geodesic bridge, V trained with geodesic labels is Riemannian conditional FM. Any class satisfying Prop. 5 (V, or the backbone plus a gradient branch, M) preserves the response.

3. The 1k-update geodesic gap is a label-noise and budget effect. Replacing only the conditional labels by the exact mean teacher at 1,000 updates (40 paired runs) removes the gap of G and M, while B stays at the zero-flow error (1/12≈0.0831/12\approx0.083):

Action W1W_1, 1k, 5 seedsBGMV (RFM)
conditional labels0.0840±0.00240.0652±0.01030.0817±0.00260.0107±0.0013
mean-teacher labels0.0840±0.00240.0078±0.00240.0073±0.00240.0076±0.0019

With conditional labels and more training (each budget trained from scratch with the same seeds), G and M close most of the gap by 3k updates and are below V at 10k and 30k, while B stays at the zero-flow error (NMSE 1.000):

UpdatesBGMV (RFM)
3,0000.0840±0.00250.0086±0.00320.0077±0.00230.0083±0.0024
10,0000.0840±0.00240.0063±0.00270.0054±0.00170.0072±0.0019
30,0000.0840±0.00240.0037±0.00060.0034±0.00030.0047±0.0014

4. Near-symmetric shells (new Thm. 7, §3.3; proof in App. J). On the ellipsoid ΛδS3\Lambda_\delta S^3, Λδ=diag(1+δ,1,1,1)\Lambda_\delta=\mathrm{diag}(1+\delta,1,1,1), the full-class optimum is nonzero but of order ∣δ∣|\delta|: for 0<∣δ∣≤10−30<|\delta|\le10^{-3}, ∣δ∣/64≤∥vδ⋆∥≤100∣δ∣|\delta|/64\le\|v^\star_\delta\|\le100|\delta|, and the endpoint action-W1W_1 error is at least (1/24−120∣δ∣)+(1/24-120|\delta|)_+, hence at least 89/300089/3000 for ∣δ∣≤10−4|\delta|\le10^{-4}.

(Continued in Part 2.)

AuthorsOpenReview ↗

General Response to All Reviewers (Part 2 of 2)

5. A natural response obstruction: orbit orientation in 3D central-force motion (new Prop. 8, §3.3; proof App. K; experiment §4.4, Table 4). For a smooth central force H=∥p∥2/2+ϕ(∥q∥2)H=\|p\|^2/2+\phi(\|q\|^2), every function gg of the orbit normal L^=L/∥L∥\hat L=L/\|L\| is a first integral with ∇H⋅∇g=0\nabla H\cdot\nabla g=0 on the shell. Prop. 8 shows that, at rotation-invariant densities, no smooth backbone field has an instantaneous response in the law of L^\hat L, whatever the network hh. Along an orientation-tilted path the same holds for the inclination marginal L^z\hat L_z; the full orientation law can respond there (App. K gives an example). Because ∇g\nabla g is not square-integrable at L=0L=0, this is a response obstruction for smooth fields, not a zero population optimum as on S3S^3. In a task that concentrates orbit normals toward ±ez\pm e_z (EP2(L^z)\mathbb E P_2(\hat L_z): 0→0.300\to0.30; width-128 MLPs, 5k updates, uniform time sampling, 5 seeds):

BGMV
EP2(L^z)\mathbb E P_2(\hat L_z)0.003±0.0110.165±0.0150.220±0.0190.239±0.012
L^z\hat L_z-marginal W1W_1 (source 0.184)0.187±0.0060.090±0.0090.058±0.0110.047±0.006
Teacher NMSE0.838±0.0410.521±0.0190.345±0.0430.285±0.009

B's marginal W1W_1 equals the source's in every seed to within 0.005, although it fits other velocity components.

6. New experiments (App. L gives protocols and full tables with SDs).

  • Ambient FM with final or step-wise state projection (S3S^3, 5 seeds): raw energy drift 0.0429; action W1W_1 0.0325±0.0013 (final) and 0.0366±0.0017 (step-wise), about 3× the error of V trained with comparable labels and budget (0.0107). The pipelines also differ in path and labels, and this baseline is not PCFM or HardFlow, so we make no claim against those methods. On independent-trajectory FPUT, test SW1 after retraction is 0.1502 (V 0.1101, M 0.1135).
  • Likelihood training (CNF, exact surface divergence, 500 updates, 5 seeds): B stays at W1=0.0840W_1=0.0840; V reaches 0.0083. Compositions of backbone fields can reach further (App. A.5), so we do not generalize this to all likelihood training.
  • V in all five physical controls (300 updates, 5 seeds; same width, so V has 2–17% more parameters than B): observable W1W_1/SD 0.046, 0.049, 0.034, 0.062, 0.056. On fput16, V's 0.056 vs. M's 1.059 is specific to the target observable; full-state SW1 is nearly equal (M 0.0578, V 0.0554, oracle 0.0550).
  • Independent-trajectory FPUT (216 trajectories in 3 datasets): V and M have the lowest full-state test SW1; filtering keeps the highest condition hit rate and the lowest mode W1W_1.
  • Tilted target on S15S^{15} (3 seeds): rejection is the most accurate sampler at β=2\beta=2, but with 16,384 uniform candidates it accepts only 2.3 samples on average at β=12\beta=12, while M and V transport every sample (W1W_1 0.0502, 0.0561) and B stays at the prior.
  • New Figure 2: 30k-update histograms, error against training budget, and the label ablation.

7. Presentation. The revision adds a Discussion section with design guidance and one Limitations paragraph, defines all new symbols, and adds the missing literature (see the response to Dmc1, W3).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.