acceptodds
Under review as a conference paper at ICLR 2027

Improving Hamiltonian Learning through Local Energy Decomposition

Abstract

Hamiltonian neural networks provide a principled framework for learning conservative dynamics by parameterizing a scalar Hamiltonian and deriving the vector field through Hamilton's equations. However, preserving Hamiltonian structure does not guarantee that the energy function itself is represented accurately. When the Hamiltonian is modeled by a conventional multilayer perceptron, it inherits the spectral bias of coordinate-based neural representations, learning smooth components more readily than fine-scale variations that can strongly affect the dynamics after differentiation. We introduce the Local Multiscale Hamiltonian Neural Network (LM-HNN), which restructures the Hamiltonian according to its physical interaction structure. LM-HNN decomposes the energy into kinetic and potential components and represents them as sums of shared local functions over multiple interaction neighborhoods. This replaces a single high-dimensional approximation problem with lower-dimensional, reusable interaction laws, making fine-scale energy structure easier to resolve while preserving Hamiltonian dynamics. Across low-dimensional, interacting, and fast–slow systems, LM-HNN consistently achieves more accurate Hamiltonian derivatives and long-horizon rollouts than competing structure-preserving models, while substantially extending the effective spectral range of the learned energy. The same structured representation also generalizes more reliably to energy levels not observed during training. These results show that accuracy depends on the Hamiltonian representation as well as its structure, and that incorporating locality, decomposition, and shared interaction laws provides a simple and effective route to more accurate and generalizable Hamiltonian learning.

Primary area:
learning on time series and dynamical systems
Keywords:
Hamiltonian neural network, Spectral bias, Multi-scale dynamics

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_7wAMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes LM-HNN, which decomposes the Hamiltonian energy model into shared local functions into multiple interaction neighborhoods. This construction reduces the dimensionality of the functions being learned while retaining dynamics induced via Hamiltonian energy functions. Spectral diagnostics show improved recovery of fine-scale energy variations and their derivatives. Authors tested LM-HNN across six conservative systems, and show a better performance compared with some prior works including HNN, SympNet, FS-HNN, and included improved performance in unseen-energy evaluations.

Critical strengths
  • The idea is relevant in learning a compositional Hamiltonian system.
  • The experiment and ablations are convincing at least to me, for given six dynamical systems and to understand if local interaction and decomposition help in learning a high dimensional energy modeling.
Critical weaknesses
  • The claim that this model is a local multi-scale model is not convincing nor clear to me. In L180-206, authors claim the neighborhood construction and interaction within local variables but I didn't find different scale or rr (I saw R=0,1,2\mathcal{R}=\\{0,1,2\\} but I am unsure if R\mathcal{R} is identical to rr). In appendix it seems three high dimensional cases are only in chain topology. It might be better if there exists a visualization to explain how the multiscale is being properly addressed in LM-HNN.
  • Only conservative systems are considered, where observation noise or dissipative or general systems are ruled out, which is more general and practical if data are given via trajectory observations.
Questions for authors
  • What is the need of multi-scale or neigobourhood setup for single pendulum, double pendulum and fast-slow oscillator? Why it is better than conventional methods even when there is no complex state space topology at all?
  • Could you quantify the overlap between training and test distributions of local strains or neighborhoods? This would distinguish recombination of familiar interactions from extrapolation of the local interaction law itself. Would R\mathcal{R} reported in appendix same as rr defined in Section 4.1 and how does that affect choice of Nr(i)\mathcal{N}_r(i) respectively?
  • Can one generalize the connectivity graph of given systems or use articulated chain models to test more general decompositional energy learning cases?
6 comments

Thank you for the constructive comments and suggestions. In response, we have clarified the mathematical definition and implementation of local multiscale neighborhoods, added experiments on two-dimensional lattices and articulated chains, and investigated robustness to observation noise, energy dissipation, and unseen energy levels. We address each weakness and question below.

W1. Definition, Visualization, and Effectiveness of Local Multiscale Neighborhoods

We clarify the distinction between individual neighborhood radii and the set of radii used by LM-HNN, together with their construction, aggregation, and relationship to physical interaction ranges.

Given a reference index set I\mathcal{I} with distance metric dd, the radius-rr neighborhood is

Nr(i)=j∈I(i,j)≤r.\mathcal{N}_r(i)=\\{j\in\mathcal{I}(i,j)\leq r\\}.

Here, rr denotes a single neighborhood radius, whereas R\mathcal{R} is the set of radii used simultaneously. For Cartesian lattices, we adopt the Chebyshev metric on fixed reference coordinates:

d∞(i,j)=∥xi−xj∥∞=max⁡1≤k≤D∣xi,k−xj,k∣.d_\infty(i,j)=\lVert\mathbf{x}_i-\mathbf{x}_j\rVert_\infty=\max_{1\leq k\leq D}|x_{i,k}-x_{j,k}|.

In one dimension, this gives a window of 2r+12r+1 consecutive indices. In two dimensions, it gives a (2r+1)×(2r+1)(2r+1)\times(2r+1) square patch that includes diagonal neighbors. Near fixed boundaries, out-of-domain entries are zero-padded. These neighborhoods depend on the reference geometry rather than instantaneous displacements or added physical coupling edges.

The extracted variables depend on the energy component. For one-dimensional potentials, neighborhoods are defined over bond strains sb=qb+1−qbs_b=q_{b+1}-q_b, giving sb(r)=(sb−r,…,sb+r)\mathbf{s}_b^{(r)}=(s_{b-r},\ldots,s_{b+r}). Kinetic branches instead use site-position neighborhoods and, when needed, local momenta. Two-dimensional models use the corresponding square patches.

At each radius, a separate network is shared across equivalent neighborhoods, and their contributions are aggregated as

Vθ(q)=∑r∈RπrV∑i∈IVvθ,r(ψi,rV(q)).V_\theta(\mathbf{q})=\sum_{r\in\mathcal{R}}\pi_r^V\sum_{i\in\mathcal{I}_V}v_{\theta,r}\left(\psi_{i,r}^V(\mathbf{q})\right).

Here, πrV\pi_r^V is the learned scale weight and ψi,rV\psi_{i,r}^V extracts local potential variables. The kinetic branches follow Eqs. (10)–(12). For the original chains, R=0,1,2\mathcal{R}=\\{0,1,2\\} corresponds to 1, 3, and 5 consecutive strains, with a shared width-64 network at each scale. Our additional two-dimensional experiments evaluate configurations including 1,2\\{1,2\\} and 1,2,3\\{1,2,3\\}.

Interaction-range experiments. The receptive-field ablations in Appendix B and the new experiments on two-dimensional FPUT lattices demonstrate how the choice of radius relates to the underlying physical interactions.

We consider an L×LL\times L square lattice of unit masses with fixed boundaries, where each site a=(i,j)a=(i,j) has scalar displacement qaq_a and momentum pap_a. The Hamiltonian combines nonlinear nearest-neighbor FPUT bonds with distance-dependent harmonic interactions [1, 2]:

H=∑a12pa2+∑⟨a,b⟩ϕFPUT(qa−qb)+12∑∥a−b∥∞≤rc (a,b) not nearest neighborsg(2dab)αlr(qa−qb)2.H=\sum_a\frac12p_a^2+\sum_{\langle a,b\rangle}\phi_{\mathrm{FPUT}}(q_a-q_b)+\frac12\sum_{\substack{\lVert a-b \rVert_\infty \leq r_c \\\ (a,b)\text{ not nearest neighbors}}}g\left(\frac{\sqrt2}{d_{ab}}\right)^{\alpha_{\mathrm{lr}}}(q_a-q_b)^2.

Here, ϕFPUT\phi_{\mathrm{FPUT}} uses α=0.25\alpha=0.25 and β=0.10\beta=0.10, while dabd_{ab} is the Euclidean distance and rcr_c is the Chebyshev interaction cutoff. We evaluate a short-range setting (rc=1r_c=1, g=0.3g=0.3), with diagonal harmonic coupling, and an extended-range setting (rc=3r_c=3, αlr=1\alpha_{\mathrm{lr}}=1, g=0.3g=0.3), combining anharmonic nearest-neighbor interactions with longer-range harmonic couplings.

The following table reports rollout MSE on 4×44\times4 lattices over three training seeds, with a parameter-matched HNN baseline.

ModelShort-rangeExtended-range
LM-HNN (R=1\mathcal{R}=\\{1\\})3.1e-34.4e-1
LM-HNN (R=2\mathcal{R}=\\{2\\})9.0e-31.3e-2
LM-HNN (R=1,2\mathcal{R}=\\{1,2\\})5.1e-31.3e-2
LM-HNN (R=1,2,3\mathcal{R}=\\{1,2,3\\})5.7e-35.2e-3
HNN5.1e-16.3e-1

LM-HNN substantially outperforms HNN under all tested radius configurations. A smaller radius is most accurate for short-range interactions, whereas broader multiscale neighborhoods improve accuracy when the interaction range increases. A potential term of radius rr couples two sites only if they lie within Chebyshev distance 2r2r, so R=1\mathcal{R}=\\{1\\} misses the couplings at distance 33 in the extended-range setting.

These results demonstrate that neighborhood selection has a direct physical interpretation and that combining multiple scales allows LM-HNN to capture interactions of different spatial ranges. We have clarified these in revised Section 4.

W2. Robustness to Observation Noise and Dissipative Dynamics

We evaluate LM-HNN on the FPUT chain (N=8N=8) under observation noise and energy dissipation. All experiments use parameter-matched HNN baselines, three independent seeds, and 120-step rollouts from unseen initial conditions.

Observation noise. We add Gaussian noise to the training data, with standard deviations corresponding to 5% and 10% of each coordinate's standard deviation.

Noise levelModelRollout MSEPeak err.
0%LM-HNN4.9e-30.83%
0%HNN2.0e-14.86%
5%LM-HNN1.0e-20.95%
5%HNN2.8e-15.22%
10%LM-HNN6.0e-21.64%
10%HNN6.4e-16.62%

Although both models degrade with increasing noise, LM-HNN maintains substantially lower rollout and peak errors at every noise level, demonstrating robustness to noisy observations.

Dissipative dynamics. We further introduce linear damping −γpi-\gamma p_i with γ=0.1\gamma=0.1, for which the true system retains approximately 54% of its initial energy. When trained on the damped trajectories using their original conservative formulations, neither model can reproduce the energy decay. We therefore introduce the same learnable damping term into both models, following dissipative Hamiltonian formulations [3, 4]:

q˙=∂Hθ∂p,p˙=−∂Hθ∂q−γp.\dot{\mathbf{q}}=\frac{\partial H_\theta}{\partial\mathbf{p}},\qquad \dot{\mathbf{p}}=-\frac{\partial H_\theta}{\partial\mathbf{q}}-\gamma\mathbf{p}.

FormulationModelRollout MSEPeak err.Efinal/E0E_{\mathrm{final}}/E_0Learned γ\gamma
ConservativeLM-HNN1.4e-18.60%0.99—
ConservativeHNN2.6e-19.06%1.00—
With dampingLM-HNN4.1e-30.81%0.540.0999
With dampingHNN1.1e-14.01%0.540.1011

Both dissipative extensions recover the damping coefficient and energy decay, but LM-HNN provides substantially more accurate trajectories. The dissipative extension leaves the local multiscale Hamiltonian representation unchanged and only adds the damping term.

These experiments show that LM-HNN is robust to observation noise and can accommodate dissipative dynamics through an explicit damping term. The results have been included in the revised appendix.

Q1. Role of Energy Decomposition and Multiscale Neighborhoods in Low-Dimensional Systems

The single pendulum, double pendulum, and fast–slow oscillator evaluate a different aspect of LM-HNN from the interacting many-body systems. In these low-dimensional benchmarks, the performance gains primarily arise from the structured decomposition of kinetic and potential energies, rather than spatial locality or multiscale neighborhood aggregation.

Unlike a conventional HNN that learns the entire Hamiltonian through an unconstrained network, LM-HNN separates kinetic and potential energies. For the pendulum systems, it explicitly encodes quadratic momentum dependence:

Tθ(q,p)=12p⊤Mθ−1(q)p.T_\theta(\mathbf{q},\mathbf{p})=\frac12\mathbf{p}^\top M_\theta^{-1}(\mathbf{q})\mathbf{p}.

This reduces the approximation burden by requiring the network to learn configuration-dependent coefficients rather than an arbitrary function of positions and momenta. For the single pendulum, the structure captures quadratic kinetic energy and periodic potential; for the double pendulum, it additionally incorporates the configuration-dependent mass matrix and angular interactions. The fast–slow oscillator tests structured Hamiltonian learning under separated temporal scales, which are distinct from spatial neighborhood radii.

Thus, the low-dimensional benchmarks validate the benefits of kinetic–potential decomposition and structured momentum dependence, while the interacting-chain and lattice experiments demonstrate the additional advantages of locality and multiscale representations. The parameter-matched comparisons in Table 14 further support that the improvements are not simply due to increased model capacity.

Q2(a). Training–Test Overlap of Local Neighborhoods and Unseen-Energy Generalization

We analyze the FPUT chain with N=8N=8 masses to quantify whether unseen-energy generalization arises from recombining familiar local interactions or extrapolating local interaction laws.

Models are trained on energies per mass E/N∈[0.05,2.0]E/N\in[0.05,2.0], where EE is the total energy, and evaluated at unseen values E/N∈2.5,3.0E/N\in\\{2.5,3.0\\}. We compare training–test overlap for individual bond strains, local neighborhoods, and full states.

A test sample counts as covered if it is as close to the training data as training samples are to other training trajectories. Specifically, its Euclidean nearest-neighbor distance to the training set must fall below the 95th percentile of nearest-neighbor distances between different training trajectories, with this threshold set separately for each representation. This feature-space distance measures distribution overlap and is distinct from the Chebyshev metric defining spatial neighborhoods.

RepresentationIn-range testUnseen-energy test
Individual strains within training range100%99.1%
Strain neighborhoods, r=0/1/2r=0/1/297% / 94% / 93%87% / 75% / 72%
Position neighborhoods, r=0/1/2r=0/1/297% / 96% / 94%89% / 80% / 76%
Full configuration q\mathbf{q}93%66%
Full state (q,p)(\mathbf{q},\mathbf{p})94%40%

The individual-strain row reports the fraction of test strains lying between the smallest and largest strains seen in training. All remaining rows use the nearest-neighbor criterion described above.

Despite the shift in global energy, 99.1% of individual strains remain within the training range, and local neighborhoods retain substantially greater coverage than full states. In particular, the full-state coverage drops to 40%, while radius-zero strain and position neighborhoods retain 87% and 89% coverage, respectively. The Toda and Morse chains show similar coverage of single strains and local neighborhoods, as reported in the newly added Table 22 of the updated manuscript.

These findings support the interpretation in Section 4.3 that LM-HNN can reuse familiar local interaction laws in globally unseen configurations. The reduced coverage at larger radii also shows that some local energy functions must extrapolate under more substantial energy shifts. This analysis has been included in Appendix C.10 and connected to the energy-generalization discussion in Section 4.3.

Q2(b). Relationship Between Individual Radius rr and Radius Set R\mathcal{R}

As defined in W1, rr denotes a single neighborhood radius, whereas R\mathcal{R} represents the set of radii combined by LM-HNN. For example, R=0,1,2\mathcal{R}=\\{0,1,2\\} defines three separate neighborhoods N0(i)\mathcal{N}_0(i), N1(i)\mathcal{N}_1(i), and N2(i)\mathcal{N}_2(i), whose learned energy contributions are aggregated through scale weights.

The coverage analysis in Q2(a) also explains the role of different radii. Smaller neighborhoods retain greater training–test overlap under energy shifts, whereas larger neighborhoods capture more extended interactions. Combining multiple radii allows LM-HNN to represent these complementary interaction scales within the same Hamiltonian. The notation and neighborhood selection have been clarified in Section 4.

Reviewer_Fiv2OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents LM-HNN, which represents kinetic and potential energies using shared local functions over multiple neighborhood scales. Experiments on six conservative systems show improved derivative and rollout accuracy, with additional tests on spectral resolution and unseen energy levels.

Critical strengths
  1. I believe the paper provides a practical and useful representation for Hamiltonian learning in interacting physical systems. Representing the total Hamiltonian through shared local energy functions reduces the complexity of the functions each network must learn. The kinetic–potential decomposition further incorporates the structure of the mechanical system, while Hamilton's equations retain the connection between energy and dynamics. Although the methodological novelty is moderate, I consider the resulting representation a meaningful contribution.

  2. The experiments provide encouraging evidence for this practical value. I also appreciate the ablations separating locality, kinetic–potential decomposition, and parameter sharing.

Overall, the paper develops a useful representation and demonstrates its effectiveness on the tested problems. I see potential applications in addressing Hamiltonian learning problems where accurate dynamics of interacting systems are essential, and I am happy to support its acceptance. My main suggestion is to demonstrate that potential on a broader range of multiscale physical systems.

Critical weaknesses

My main reservation is that the examples do not yet fully demonstrate the scope suggested by the multiscale motivation. The pendulum systems and nonlinear chains are useful controlled benchmarks, and the fast–slow oscillator adds a distinct dynamical setting. However, the chain-length study only reaches N = 16, and several examples are variations of the same one-dimensional interaction structure. These tests establish that the representation works on the selected systems, but leave its usefulness for more complex multiscale many-body dynamics less clear.

For further improvement, I would suggest a larger many-body system with interactions across different scales. A representative physical application would help readers understand how the choice of neighborhoods relates to the underlying interactions and whether the learned representation remains accurate as the system becomes more complex. The fixed parameter count in the chain-length sweep is encouraging, but a larger example would give stronger evidence of practical applicability than parameter sharing alone (I am not an expert in the full range of multiscale Hamiltonian systems and would leave the choice of a suitable example to the authors).

Questions for authors

None that would change my current acceptance recommendation.

1 comment

Thank you for this suggestion, we agree that a larger many-body system further illustrates the benefits of repeated multiscale aggregation. We now extend LM-HNN to two-dimensional FPUT lattices [1], testing both different interaction ranges and scalability. All experiments follow the original training protocol, using three training seeds and parameter-matched HNN baselines.

We consider an L×LL\times L square lattice of unit masses with fixed boundaries, where each site a=(i,j)a=(i,j) has scalar displacement qaq_a and momentum pap_a. The Hamiltonian combines nonlinear nearest-neighbor FPUT bonds with distance-dependent harmonic interactions [2]: H=∑a12pa2+∑⟨a,b⟩ϕFPUT(qa−qb)+12∑∣a−b∣∞≤rc(a,b) not nearest neighborsg(2dab)αlr(qa−qb)2.H=\sum_a\frac12p_a^2+\sum_{\langle a,b\rangle}\phi_{\mathrm{FPUT}}(q_a-q_b)+\frac12\sum_{\substack{\\|a-b\\|_\infty\leq r_c\\\\(a,b)\text{ not nearest neighbors}}}g\left(\frac{\sqrt2}{d_{ab}}\right)^{\alpha_{\mathrm{lr}}}(q_a-q_b)^2. Here, ϕFPUT\phi_{\mathrm{FPUT}} uses α=0.25\alpha=0.25 and β=0.10\beta=0.10, while dabd_{ab} is the Euclidean distance and rcr_c is the Chebyshev interaction cutoff.

We evaluate a short-range setting (rc=1r_c=1, g=0.3g=0.3), with diagonal harmonic coupling, and an extended-range setting (rc=3r_c=3, αlr=1\alpha_{\mathrm{lr}}=1, g=0.3g=0.3), combining anharmonic nearest-neighbor interactions with longer-range harmonic couplings. The following table reports rollout MSE on 4×44\times4 lattices for LM-HNN with different sets of neighborhood radii R\mathcal R.

ModelShort-rangeExtended-range
LM-HNN (R=1\mathcal R=\\{1\\})3.1e-34.4e-1
LM-HNN (R=2\mathcal R=\\{2\\})9.0e-31.3e-2
LM-HNN (R=1,2\mathcal R=\\{1,2\\})5.1e-31.3e-2
LM-HNN (R=1,2,3\mathcal R=\\{1,2,3\\})5.7e-35.2e-3
HNN5.1e-16.3e-1

LM-HNN outperforms HNN in both settings under all radius sets, and substantially so whenever the radii span the interaction range. A smaller radius is most accurate for short-range interactions, whereas broader multiscale neighborhoods improve accuracy when the interaction range increases. A potential term of radius rr couples two sites only if they lie within Chebyshev distance 2r2r, so R=1\mathcal R=\\{1\\} misses the couplings at distance 33 in the extended-range setting.

We also test scalability by increasing the short-range lattice from 4×44\times4 to 8×88\times8 (16–64 masses, phase-space dimensions 32–128), so that the largest lattice has four times as many masses as the longest chain (N=16N=16) in the original sweep. LM-HNN uses the same architecture with R=1,2\mathcal R=\\{1,2\\} and 21,384 parameters throughout, while HNN is parameter-matched at each size.

Lattice sizeLM-HNN rollout MSEHNN rollout MSE
4×44\times4 (16 masses)5.1e-35.1e-1
6×66\times6 (36 masses)4.0e-31.5
8×88\times8 (64 masses)5.0e-33.3

LM-HNN maintains nearly constant prediction accuracy as the lattice grows, whereas HNN errors increase considerably. These results demonstrate LM-HNN's applicability to two-dimensional multiscale interactions (W1) and its scalability to larger many-body systems (W2). We have included the experimental details and results in the revised manuscript (Table 3, Appendices A.2 and C.6).

References

[1] G. Benettin, "Time scale for energy equipartition in a two-dimensional FPU model," Chaos, 15, 015108 (2005). https://doi.org/10.1063/1.1854278

[2] H. Christodoulidi, C. Tsallis, and T. Bountis, "Fermi-Pasta-Ulam model with long-range interactions: Dynamics and thermostatistics," EPL, 108, 40006 (2014). https://doi.org/10.1209/0295-5075/108/40006

Reviewer_2ffiOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The goal of this paper is to propose a main method for capturing physical dynamics as different modes throgh a learnable and local multiscale NN architecture. This architecture is meant to avoid spectral bias, which the authors discuss as a problem in the NN architectures needed for, or used in, existing HNN modeling. According to the authors, the novelty lies in reformulating the high-dimensional parameterization problem discussed in prior methods as low-dimensional sub-problems, and they claim to have shown so called a “reusable” interaction law.

Critical strengths

The core value of this paper comes from Section 4.2. What I value highly in this section is the careful modeling of local interactions, which existing HNNs have not paid much attention to, and especially the user-specific hand design of the potential and kinetic energy terms of the Hamiltonian. The attempt to describe physical phenomena as a sum of multiple sub-components is in line with similar approaches in applied science, and I think the intention to point out the problem and solve it from that angle is very good. In the experiments, the method shows very good results on some systems compared with existing global and blind NN modeling methods, and at least the results are consistent with what the authors intended, which empirically validates their argument.

Critical weaknesses

While I highly value the strengths described above, due to several technical issues, I am considering not recommending acceptance of this paper in my initial review. The followings are main concerns listed:

First, there is no theoretical investigation at all of the advantage of the multiscale nature of the sub-components described above. In which systems is the R\mathcal{R}-order complexity described by the authors helpful for improving performance? An understanding of, for example, generalization at the level of path measures over dynamical trajectories, or at least an approximation error bound, is necessary.

Second, while the paper aims to improve performance through sub-components, there is no background or narrative explaining why this local interaction is needed at all. In some physical systems, such interactions are entirely absent. The authors should describe the difference between the physical systems they aim to solve and the systems their claims are based on, and should connect this explanation to the modeling in (8) and (10).

Third, and related to the above: since the paper clearly deals with a variant of HNNs that describes physical systems, mathematical rigor is essential. I ask the authors to provide a clear definition of the “interaction neighborhood of degree of freedom i at radius r”. How is this defined mathematically? Even though this is the most important object and forms the core contribution, the paper only describes it briefly in a very vague sentence. The complexity of the system would be defined according to the degree of this “local interaction”. Please make this clear.

Questions for authors

Please, see the above "Critical Weaknesses" section for the questions.

3 comments

Thank you for emphasizing the need for a more rigorous mathematical characterization of LM-HNN. Below, we clarify the motivation for local multiscale decomposition, the physical systems for which it is appropriate, and how its approximation properties relate to dynamical prediction.

Physical motivation and applicability (W2). LM-HNN is motivated by mechanical systems whose Hamiltonians contain repeated finite-range interactions. For such systems, the total energy can be expressed as a sum of lower-dimensional interaction functions rather than approximated through a single network over the full 2N2N-dimensional state. Sharing these functions across equivalent interactions allows repeated instances to contribute learning signals to the same parameters. This representation can also help resolve fine-scale energy variations that are difficult for conventional HNNs to learn because of spectral bias (Eqs. (3)–(5)).

The benefits of locality depend on whether the underlying Hamiltonian admits a suitable finite-range decomposition. As shown in Tab. 7, local representations are effective for finite-range momentum coupling, whereas globally coupled systems may require dense parameterizations. These results illustrate the importance of matching the energy parameterization to the underlying physical interactions.

Definition of interaction neighborhoods. For an index set I\mathcal{I} with distance metric dd, the radius-rr neighborhood is

Nr(i)=j∈I:d(i,j)≤r.\mathcal{N}_r(i)=\\{j\in\mathcal{I}:d(i,j)\leq r\\}.

Here, rr denotes a single neighborhood radius, and R\mathcal{R} is the set of radii whose contributions LM-HNN combines, such as R=0,1,2\mathcal{R}=\\{0,1,2\\}. For regular Cartesian lattices, we use the Chebyshev metric on fixed reference coordinates, given by

d∞(i,j)=∥xi−xj∥∞=max⁡1≤k≤D∣xi,k−xj,k∣.d_\infty(i,j)=\lVert\mathbf{x}_i-\mathbf{x}_j\rVert_\infty=\max_{1\leq k\leq D}|x_{i,k}-x_{j,k}|.

In one dimension, this gives a window of 2r+12r+1 consecutive indices. In two dimensions, it gives a (2r+1)×(2r+1)(2r+1)\times(2r+1) square patch that includes diagonal neighbors. Near fixed boundaries, out-of-domain entries are zero-padded. These neighborhoods depend on the reference geometry rather than instantaneous displacements or added physical coupling edges. For one-dimensional chain potentials, they operate on bond strains sb=qb+1−qbs_b=q_{b+1}-q_b, with inputs ψb,rV(q)=(sb−r,…,sb+r)\psi^V_{b,r}(\mathbf{q})=(s_{b-r},\ldots,s_{b+r}), while kinetic neighborhoods use site positions, and also site momenta under momentum coupling. These constructions are illustrated in Fig. 1 and formalized in App. D.1.

Connection to the energy parameterization. Following Eqs. (6)–(12), LM-HNN decomposes the Hamiltonian as Hθ=Tθ+VθH_\theta=T_\theta+V_\theta. This separation allows the kinetic and potential energies to be modeled independently with parameterizations tailored to their respective physical structures, improving parameter efficiency and enabling component-specific multiscale representations. The potential energy is then expressed as

Vθ(q)=∑r∈RπrV∑i∈IVvθ,r(ψi,rV(q)).V_\theta(\mathbf{q})=\sum_{r\in\mathcal{R}}\pi_r^V\sum_{i\in\mathcal{I}_V}v_{\theta,r}\left(\psi_{i,r}^V(\mathbf{q})\right).

Here, ψi,rV\psi_{i,r}^V extracts local variables and πrV\pi_r^V is the learned scale weight. Each radius uses a separate network shared across equivalent neighborhoods. The kinetic representation likewise follows the physical coupling: Eq. (10) encodes diagonal quadratic momentum dependence, while Eq. (12) handles finite-range momentum coupling. The assembled Hamiltonian generates dynamics through Hamilton's equations, while its constituent networks approximate lower-dimensional functions. Because these networks are shared, the parameter count for fixed radii does not grow with the number of repeated interactions.

Interaction range and radius selection. We use the FPUT chain, with and without an added second-neighbor coupling, to demonstrate why different radii are useful. For nearest-neighbor interactions, V(q)=∑bϕ(sb)V(\mathbf{q})=\sum_b\phi(s_b), so a radius-zero potential neighborhood is sufficient. Adding the second-neighbor coupling of Eq. (30) introduces

Vextra(q)=g2∑i(qi+2−qi)2=g2∑i(si+si+1)2.V_{\mathrm{extra}}(\mathbf{q})=\frac{g}{2}\sum_i(q_{i+2}-q_i)^2=\frac{g}{2}\sum_i(s_i+s_{i+1})^2.

For interior second-neighbor pairs, this interaction has the mixed derivative

∂2Vextra∂qi∂qi+2=−g,\frac{\partial^2 V_{\mathrm{extra}}}{\partial q_i\partial q_{i+2}}=-g,

whereas this derivative vanishes for any sum of single-strain functions. A radius-zero neighborhood therefore cannot represent this interaction. A radius-one strain neighborhood includes both strains and can represent it. Thus, the appropriate radius depends on interaction support. A single radius-one neighborhood can also represent both chains, since its three-strain window contains the nearest-neighbor and second-neighbor terms together. Multiscale aggregation allows each interaction to be represented at the smallest radius that contains it, without selecting a radius in advance.

Local-to-global force error. We characterize how local approximation errors affect the learned dynamics. Suppose the true potential admits the decomposition

V(q)=∑r∈R∑i∈Irvr∗(ψi,rV(q))+Vmiss(q),V(\mathbf{q})=\sum_{r\in\mathcal{R}}\sum_{i\in\mathcal{I}_r}v_r^*\left(\psi_{i,r}^V(\mathbf{q})\right)+V_{\mathrm{miss}}(\mathbf{q}),

where VmissV_{\mathrm{miss}} contains interactions not represented by the selected neighborhoods. Let ei,re_{i,r} denote the error of each learned local contribution, including its scale weight. Suppose ∣∂ei,r/∂qj∣≤ϵr|\partial e_{i,r}/\partial q_j|\leq\epsilon_r for every physical coordinate on which it depends, and let DrD_r bound the number of radius-rr contributions involving any coordinate.

Prop. 1 in App. D.3 establishes

∥∇q(Vθ−V)∥2N≤∑r∈RDrϵr+∥∇qVmiss∥2N.\frac{\lVert\nabla_{\mathbf{q}}(V_\theta-V)\rVert_2}{\sqrt N}\leq\sum_{r\in\mathcal{R}}D_r\epsilon_r+\frac{\lVert\nabla_{\mathbf{q}}V_{\mathrm{miss}}\rVert_2}{\sqrt N}.

This separates local approximation error from unrepresented interactions. For fixed finite-range neighborhoods on regular lattices, DrD_r is independent of system size. Thus, the normalized force error can remain bounded independently of system size when local derivative errors and omitted-interaction contributions remain controlled.

For canonical Hamiltonian dynamics, J⊤J=IJ^\top J=I implies ∥fθ−f∥2=∥∇(Hθ−H)∥2\lVert\mathbf{f}_\theta-\mathbf{f}\rVert_2=\lVert\nabla(H_\theta-H)\rVert_2. Combining this identity with Prop. 1 gives Corollary 1

∥fθ−f∥22N≤∥∇(Tθ−T)∥22N+12(∑r∈RDrϵr+∥∇qVmiss∥2N).\frac{\lVert\mathbf{f}_\theta-\mathbf{f}\rVert_2}{\sqrt{2N}}\leq\frac{\lVert\nabla(T_\theta-T)\rVert_2}{\sqrt{2N}}+\frac{1}{\sqrt 2}\left(\sum_{r\in\mathcal{R}}D_r\epsilon_r+\frac{\lVert\nabla_{\mathbf{q}}V_{\mathrm{miss}}\rVert_2}{\sqrt N}\right).

The kinetic term can similarly be controlled by errors in the inverse-mass coefficients and their derivatives under bounded-momentum and locality assumptions. If the normalized vector-field error is uniformly bounded by δ\delta along the relevant trajectories and the true vector field is LL-Lipschitz, Grönwall's inequality yields

sup⁡0≤t≤T∥zθ(t)−z(t)∥22N≤δL(eLT−1).\sup_{0\leq t\leq T}\frac{\lVert\mathbf{z}_\theta(t)-\mathbf{z}(t)\rVert_2}{\sqrt{2N}}\leq\frac{\delta}{L}(e^{LT}-1).

For L=0L=0, the bound becomes δT\delta T. Coupling trajectories with the same initial conditions also yields a Wasserstein-1 path-measure bound under the corresponding normalized supremum metric. These bounds establish how local approximation accuracy translates into global dynamical accuracy under the stated assumptions. While they do not establish that local networks necessarily achieve smaller approximation errors than global HNNs, the following experiments provide evidence of the advantages of this representation.

Empirical support. Figure 2 supports the motivation above, showing that LM-HNN recovers finer Hamiltonian variations and their derivatives than a parameter-matched HNN, and Table 1 shows that generic spectral interventions do not reproduce this accuracy. Table 6 confirms the radius dependence of the FPUT example, with R=0\mathcal{R}=\\{0\\} best on the nearest-neighbor chain, R=1\mathcal{R}=\\{1\\} the best single radius with second-neighbor coupling, and the multiscale set competitive on both and lowest in derivative error on the two-scale chain. Tabs. 8 and 18 further show that LM-HNN maintains a fixed parameter count and nearly constant rollout error as chain and lattice sizes increase, consistent with the benefits of shared finite-range interaction representations.

Together, the theoretical analysis and experiments clarify the conditions under which local multiscale decomposition is beneficial and how its approximation accuracy affects the resulting dynamics. We have incorporated the formal neighborhood definitions and applicability assumptions into Sec. 4, added the visualization in Fig. 1, and provided the interaction-range analysis and error bounds in App. D.

Thank you for your careful rebuttal.

Since the proposed method aims to describe real physical systems, the philosophy of how to decompose Hamiltonian is a critical issue. Neural dynamical systems have become the backbone of many modern AI4Science methods for molecular dynamics, open quantum systems, neural quantum fields and so on. In such systems, scale and locality depend on the specific physical nature of each system, and so do many properties such as stability and exactness.

In my view, the authors (and "recent" prior works as well) should first categorize physical systems by their characteristics, and then conduct controlled experiments under more precise assumptions. I hope the authors will not regard this as a future work.

From my personal perspective, the novelty of deriving more accurate model in terms of MSE would have been enough until about three years ago, but the new types of neural dynamical systems need a new type of novelty that goes beyond it. That is why I gave weak reject. To be clear, the other performance aspects are comparable or better than prior works, so my final score is not a devaluation of the performance itself.

Comments on the paper

1 comment

Dear Reviewers,

The authors have posted their rebuttals. As the author–reviewer discussion period is coming to a close soon, please read their responses and engage in the discussion. Thank you for your time and effort.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.