When observing a population of cells differentiating, a fluid swirling, or a cloud of particles dispersing, we rarely get to track individual trajectories over time. Instead, we typically collect "snapshots"—static cross-sections of the population distribution at discrete time points. To reconstruct the continuous motion between these snapshots, standard machine learning frameworks often assume the system is governed by a gradient flow.
In a gradient flow, the population moves down the steepest descent of an energy landscape. This assumption is mathematically elegant and underpins frameworks like Neural Langevin Dynamics, where a neural network learns a potential energy function to model long-term behavior, as discussed by Koop et al..
However, this elegance comes with a severe physical cost: gradient fields are fundamentally conservative. They assume that the system's velocity field is the gradient of a scalar potential ($\mathbf{v} = -\nabla \Psi$). Because the curl of any gradient field is identically zero ($\nabla \times \nabla \Psi = 0$), gradient flows are mathematically incapable of representing rotational, cyclic, or periodic behaviors. As Hashimoto et al. point out, this constraint completely breaks down when individuals exhibit periodic behavior or are driven by non-conservative forces.
Why Optimal Transport Erases Rotation
Modern generative approaches, such as Optimal Transport (OT) and standard Flow Matching, exacerbate this issue. These frameworks are explicitly designed to minimize kinetic energy, finding the "shortest path" to morph one distribution into another.
Imagine a whirlpool. If you take a snapshot of particles before they enter the whirlpool and another snapshot after they have rotated 180 degrees, an OT-based model will not reconstruct the swirling trajectory. Instead, because it seeks to minimize the total distance traveled, it will draw straight lines directly across the center of the vortex.
As Petrović et al. and Schwerdtner et al. demonstrate, this energy-minimizing bias actively suppresses the "curl" and circulation inherent in fluid dynamics, plasma physics, and biological processes like the cell cycle. The models favor dissipative, aperiodic motion, collapsing complex physical structures into oversimplified straight-line approximations, as noted by Berman et al. and Guan et al.. Models strictly based on gradient flow assumptions are simply unable to represent solenoidal or rotational components that do not alter the system's energy (Terpin et al.).
Gradient Flow (OT) Assumption:
[Initial State] --------(Straight Line / Least Energy)--------> [Final State]
(Erases loops, cycles, and vortices)
Non-Gradient Flow Reality:
[Initial State] ~~~~~~~~(Rotational / Curly Path)~~~~~~~~> [Final State]
(Preserves physical curl, cycles, and momentum)
Breaking the Constraint: Mechanisms for Non-Gradient Inference
To recover these hidden rotations without direct trajectory data, recent research focuses on redesigning generative frameworks to accommodate non-gradient fields. The core intuition is to allow the model to learn velocity fields that contain both a conservative gradient component (which changes density) and a solenoidal, divergence-free component (which handles rotation without altering the density profile).
Several distinct mechanisms have emerged to achieve this:
1. Non-Zero Drift Reference Processes
Standard Schrödinger bridges find a path between distributions by regularizing against a simple Brownian motion prior, which has zero drift. To capture cyclic behaviors, Petrović et al. introduce Curly Flow Matching, which utilizes a non-zero drift reference process within the Schrödinger bridge framework. Similarly, Zhang et al. leverage multivariate Ornstein-Uhlenbeck (mvOU) reference dynamics to model non-equilibrium diffusions. By embedding rotational or directional priors directly into the reference process, the model no longer defaults to straight-line paths and can successfully reconstruct underlying rotational dynamics.
2. Exploiting Gauge Freedom
The continuity equation dictates how a velocity field moves a probability density over time. However, this equation suffers from "gauge freedom"—meaning infinitely many different velocity fields can yield the exact same density evolution. Traditional OT resolves this ambiguity by picking the unique gradient flow that minimizes kinetic energy.
Instead of forcing a gradient structure, Berman et al. introduce Non-Gradient Inference Flows (NGIF), which exploit this gauge freedom by applying divergence-based penalties. This allows the model to incorporate physical priors like incompressibility or low curl, selecting fields that preserve swirling and turbulent trajectories rather than erasing them.
3. Second-Order Lagrangian Mechanics
First-order systems react instantly to forces, meaning they cannot model inertia, acceleration, or momentum. To capture these effects, Guan et al. propose Wasserstein Lagrangian Mechanics (WLM). By shifting the problem from first-order gradient flows to second-order mechanics, the framework learns acceleration and potential energy from temporal snapshots. This allows the system to naturally capture momentum, oscillations, and emergent group behaviors like flocking or ocean vortices.
Identifiability and Broader Applications
If multiple non-gradient fields can explain the same snapshot data, how do we know we are recovering the true underlying physics?
A key mathematical breakthrough by Guan et al. proves that as long as a population is actively changing (in a transient state), gradient-flow SDEs possess unique transient dynamics that allow the simultaneous inference of both drift and diffusion. However, when non-conservative forces, feedback loops, or material heterogeneities are present, moving beyond pure gradient constraints becomes mandatory to avoid model misspecification.
Expanding these frameworks to non-gradient dynamics unlocks solutions to long-standing problems across disciplines:
- Physical Systems: It enables the sampling of conditioned paths in complex mechanical systems with inertia and friction, moving past simpler elliptic systems (Hairer et al.).
- Biological Oscillators: It allows geometric stochastic control to map biological oscillators and limit cycles where systems do not follow simple gradient descent (Maoutsa).
- Rare Event Simulation: It empowers data-driven frameworks to simulate rare events in non-gradient systems exhibiting oscillatory behavior or non-normal transient growth, where traditional energy-landscape techniques fail (Zhang et al.).
Key Foundational Work
Learning Population-Level Diffusions with Generative RNNs (2016) — contributed 1 concept elements to this direction.
Sampling conditioned hypoelliptic diffusions (2011) — contributed 1 concept elements to this direction.