Reproduce “A Call to Lagrangian Action: Learning Population Mechanics from Temporal Snapshots” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Vincent Guan, Lazar Atanackovic, Kirill Neklyudov

The authors introduce WLM, a damped second-order model that learns population-level potential energy and friction from observed marginals, updating particle positions and velocities using initial velocities. On simulated Boids, trained on 50 snapshots through time 24.5, it forecasts the next 50 with mean Wasserstein-1 distance 1.309 ± 0.034, versus 3.078 ± 0.199 for NN-APPEX. This benchmark does not establish identifiable individual paths or universal performance.

Key insights

  • WLM provides a unified theoretical framework that encompasses classical mechanics, quantum mechanics, and Wasserstein gradient flows, demonstrating that gradient flows are specific (often overdamped or unstable conservative) cases.
  • The introduced damping parameter in the Wasserstein Lagrangian enables WLM to effectively model dissipative dynamics, while setting damping to zero allows for conservative systems, broadening its expressivity beyond traditional gradient flows.
  • The functional derivative of the potential energy can be approximated from individual particle gradients, allowing deep neural networks with attention mechanisms to parameterize and learn complex interaction kernels directly from population data.
  • WLM can infer complex, emergent dynamics (e.g., Boids flocking) without explicit rules, by learning the underlying population-level potential from snapshots.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.