Learning Non-Gradient Dynamics from Population Snapshots

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 213 papers22 conceptsfrozen 2026-10-04

In fields ranging from developmental biology to fluid mechanics, researchers often face a "snapshot problem." We can observe the state of a population at specific moments—such as the gene expression of thousands of individual cells or the distribution of particles in a turbulent flow—but the act of observation often destroys the subjects or loses their identity. We see the crowd, but we cannot follow the individual.

To reconstruct the "movie" of how these systems evolve, traditional machine learning models have leaned heavily on the principle of least action. However, a growing body of research suggests that these energy-minimizing assumptions are fundamentally blind to the "curly" reality of nature: rotations, cycles, and non-conservative forces.

The Gradient Flow Assumption

Most existing frameworks for inferring dynamics from snapshots, such as Optimal Transport (OT) or standard Flow Matching, assume that the underlying velocity field is a gradient flow. Mathematically, this means the velocity of a particle is the gradient of some potential function ($v = -\nabla \Psi$).

As Paul Schwerdtner and Benjamin Peherstorfer point out in Two-Parameter Flows for Learning Population Dynamics of Physical Systems, this assumption is computationally convenient because it forces the model to find the most "efficient" path between two distributions. However, this efficiency is also a straightjacket. Gradient flows are conservative; they cannot represent systems with "curl," such as the swirling vortices of a fluid or the periodic oscillations of a cell cycle.

The Blind Spot of Static Density

The fundamental challenge was articulated by Caleb Weinreb and Allon M Klein in Fundamental limits on dynamic inference from single cell snapshots. They demonstrated that static density maps are inherently insensitive to rotational dynamics. If a population of cells is moving in a perfect circle, the overall "cloud" of cells doesn't change shape. To a gradient-based model, this looks like the cells are standing still.

This creates a "gauge freedom" problem. As Jules Berman explains in Leveraging Gauge Freedom for Learning Non-Gradient Population Dynamics of Stochastic Systems, there are infinitely many velocity fields that can result in the same observed change in population density. Standard methods pick the simplest one (the gradient), which often results in physically unrealistic, straight-line trajectories that ignore the complex rotational structures inherent in the data.

Beyond the Potential Landscape

To capture these missing dynamics, researchers are moving toward models that explicitly account for non-conservative forces. This shift involves three primary strategies:

1. Non-Zero Drift Priors Instead of starting with a "blank slate" (Brownian motion) that assumes no inherent direction, new methods use reference processes with built-in "drift." In Curly Flow Matching for Learning Non-Gradient Field Dynamics, Katarina Petrović and Michael M. Bronstein introduce CURLY-FM. By solving a Schrödinger bridge problem with a non-zero drift reference, they can recover "curly" trajectories like ocean currents that standard flow matching misses.

2. Decomposing the Vector Field Another approach is to break the velocity field into two parts: a gradient component (which changes the density) and a rotational/solenoidal component (which moves particles without changing the overall density). Simon Koop and colleagues propose Neural Langevin Dynamics (NLD) in Neural Langevin Dynamics: towards interpretable Neural Stochastic Differential Equations, which augments energy landscapes with rotation fields to model periodic orbits.

3. Non-Equilibrium Schrödinger Bridges In biological contexts, the cell cycle is the ultimate non-equilibrium process. Stephen Y. Zhang and Michael Stumpf argue in Learning non-equilibrium diffusions with Schrödinger bridges: from exactly solvable to simulation-free that using multivariate Ornstein-Uhlenbeck (mvOU) processes as priors allows for the reconstruction of directional, cyclic velocities that are invisible to traditional Brownian-based bridges.

Why It Works: Recovering the "Hidden" Physics

The intuition behind these new methods is that while a single snapshot cannot reveal rotation, the transition between snapshots—combined with physical constraints like incompressibility or low curl—can.

For example, the NGIF framework by Berman et al. uses divergence-based penalties to ensure that the learned flow respects the physical properties of the medium. Similarly, Vincent Guan and Geoffrey Schiebinger show in Gradient-flow SDEs have unique transient population dynamics that if a population is in a "transient" state (actively changing), it is mathematically possible to disentangle the driving forces (drift) from the random noise (diffusion), even in complex non-conservative systems.

By moving away from the "shortest path" bias of Optimal Transport, this research direction allows us to see the vortices in the fluid and the cycles in the cell, turning static snapshots into dynamic, physically grounded narratives.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 12 papers
  2. Version 1 · 2026-10-04 · 13 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_non_conservative_dynamics_modeling_a61138e3,
  title = {non-conservative dynamics modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-conservative-dynamics-modeling/rdr_a61138e36a494325b3f42e493581ec55},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_learning_non_gradient_dynamics_from_popu_a61138e3_v1,
  title = {Learning Non-Gradient Dynamics from Population Snapshots},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/learning-non-gradient-dynamics-from-population-snapshots/rdr_a61138e36a494325b3f42e493581ec55/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}