arXiv eess.SY: Puiggros T. and Phani propose learning momenta

arXiv eess.SY: Puiggros T. and Phani propose learning momenta

Ignacio Puiggros T. and A. Srikantha Phani propose extending Hamiltonian-based architectures to learn a system’s generalized momenta, Hamiltonian structure and dynamics from data. Hamiltonian neural networks (HNNs) encode Hamiltonian mechanics in their structure and loss function, but the authors identify a limitation: these networks rely on data containing the target system’s generalized momenta. That can hinder their use when the momenta’s analytical form is unknown. The proposed architecture is intended for that situation: it learns generalized momenta from data on the state and its derivatives, which may be non-sequential. The authors consider a dissipative mass-spring system, a double pendulum and a rolling disk under nonholonomic constraints as examples for assessing performance. The motivating idea is to retain a Hamiltonian-based approach without requiring generalized momenta as known input. The abstract describes the problem, proposed method and examples, but does not give performance results. How well the approach works beyond the stated assessments is also not established by the abstract.

Publication: c448a6fe-7f78-447d-8bd7-59274dd10a7f

attribution — pass: The proposal and the limitation are correctly attributed to authors Ignacio Puiggros T. and A. Srikantha Phani. Their method is presented as a proposal, not an established result. Labeling the source arXiv eess.SY matches the publisher field, though the paper's primary listing is cs.LG.

Authors:Ignacio Puiggros T., A. Srikantha Phani

summary — uncertain: The central terms are never explained in plain language: generalized momenta, Hamiltonian mechanics, dissipative and nonholonomic constraints. The draft also never says HNNs build physics knowledge into neural networks to improve modeling, so non-specialists cannot see why the limitation matters. It weakens the authors' force: 'severely hampers' becomes 'can hinder', and it drops their claim that unknown momenta is frequently the case in real-world applications. Smallest repair: restore the frequency claim with attribution, add the physics-priors purpose, and add tooltips for generalized momenta and nonholonomic.

This severely hampers their applicability when the analytical form of the generalized momenta is unknown, which frequently is the case in real-world applications.

date — pass: The originalPublished date of 2026-10-31 matches the v1 submission date. The later feed date was correctly not used.

[Submitted on 31 Oct 2026]

support — pass: All four claims are directly stated in the abstract. The limitations and the question correctly note that no performance results appear in the evidence. The draft adds no benefits, costs or forecasts beyond the source.

simultaneously learning the system's generalized momenta, Hamiltonian structure, and dynamics from data

media — pass: The selected media is the verified direct link to the original arXiv abstract page, which is the source itself.