Research map / Probabilistic and scientific modeling / Variational and generative sampling
Symplectic Normal Coordinates Flow for Generative Modeling of Hamiltonian Systems
Christopher M Pierce
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a symplectic normal-coordinates flow for conditioned generative modeling of Hamiltonian systems, exploiting symplectic relations between densities at different conditioning parameters.
Problem
Conditioned generative models of physical systems must represent densities across conditioning parameter values that are related by symplectic transformations; existing approaches do not exploit this structure.
Approach
Uses symplectic transformations to normal coordinates, where linear stable Hamiltonian dynamics reduce to rotations in each phase-space plane, extending to the nonlinear case away from resonances, so that densities at different conditioning values share a common structured form.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use normal coordinates flow instead of Real NVP | conditioned flow · Duffing chain, beta=0.1 | NLL: 14.1 -> 11.6 | Improved |
| use normal coordinates flow instead of LA-SympNet | conditioned flow · Duffing chain, beta=0.1 | NLL: 15.4 -> 11.6 | Improved |
| use normal coordinates flow instead of Real NVP | conditioned flow · uniform focusing channel | NLL: -11.34 -> -11.49 | Improved |
| use normal coordinates flow instead of Real NVP | conditioned flow · chicane with CSR | NLL: -5.06 -> -5.07 | No clear effect |
| use normal coordinates flow instead of LA-SympNet | conditioned flow · chicane with CSR | NLL: -4.15 -> -5.07 | Improved |
| add amplitude dependence to the rotation layer (ablation) | normal coordinates flow · Duffing chain, beta=0.1 | NLL: 17.0 -> 11.6 | Improved |
| add the rotation layer (ablation) | normal coordinates flow · Duffing chain, beta=0.1 | NLL: 16.7 -> 11.6 | Improved |
| use GradNet (symplectic) instead of MLP in rotation layer (ablation) | normal coordinates flow · Duffing chain, beta=0.1 | NLL: 13.0 -> 11.6 | Improved |
| vary conditioning network width 16-128 (ablation) | normal coordinates flow · Duffing chain, beta=0.1 | NLL: 11.2 to 11.6 | No clear effect |
Closest papers by meaning
- Equivariant Flows: Exact Likelihood Generative Learning for Symmetric Densities (ICML 2020 · similarity 0.88)
- Hamiltonian Generative Networks (ICLR 2020 · similarity 0.88)
- Hamiltonian Score Matching and Generative Flows (NeurIPS 2024 · similarity 0.85)
- Generating Physical Dynamics under Priors (ICLR 2025 · similarity 0.84)
- Physics-Informed Regularization for Domain-Agnostic Dynamical System Modeling (NeurIPS 2024 · similarity 0.84)
- Controlled Generation with Equivariant Variational Flow Matching (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
