Variational Principles and Least Action for Stochastic Dynamics

This direction is split into other directions: Euler-Lagrange equation for most probable paths. Its last version stays readable and citable here.

Version 1 of 152 papers67 conceptsfrozen 2026-10-04

In classical physics, the "Principle of Least Action" tells us that nature is efficient. A ball thrown through the air doesn't wander aimlessly; it follows the single path that minimizes a physical quantity called "action." However, when we look at the microscopic world—the jittery molecules inside a cell or the fluctuating weights of a neural network—this deterministic elegance seems to vanish into a cloud of noise.

Recent research across biophysics, fluid dynamics, and machine learning has revealed that this disappearance is an illusion. By applying variational principles to stochastic systems, researchers have shown that even in the presence of chaos, there is a "most probable" path that behaves remarkably like a classical trajectory. This framework allows us to predict rare "black swan" events, understand how generative AI creates images, and explain how biology builds stable switches from noisy parts.

The Cost of a Path

To understand why a noisy system follows a specific route, we first need a way to "score" different trajectories. In a deterministic system, paths that don't obey the laws of physics are impossible. In a stochastic system, every path is possible, but some are vastly more likely than others.

The mathematical tool used to score these paths is the Onsager-Machlup functional. Think of this as a "cost function" for noise. If a system is pushed by random fluctuations, the Onsager-Machlup functional measures how much "effort" the noise must exert to force the system along a specific trajectory.

As Zachary Selk and Harsha Honnappa demonstrate in A Feynman-Kac Type Theorem for ODEs: Solutions of Second Order ODEs as Modes of Diffusions, the most likely path of a diffusion process is actually the solution to a deterministic second-order differential equation. By minimizing this action, we find the "mode" of the stochastic process—the path where the probability density is highest. This same principle is used by John J. Vastola in Dynamical symmetries in the fluctuation-driven regime to apply Noether’s theorem to noisy systems, deriving conservation laws for the "most likely" trajectories of neural networks.

Rare Events and the Minimum Action Path

One of the most powerful applications of this theory is predicting rare events—those "once in a lifetime" fluctuations that cause a system to flip its state. William Bialek explored this in Stability and Noise in Biochemical Switches, asking how biological memories can last for years despite being made of jittery molecules. He found that by identifying the Minimum Action Path (MAP), one can calculate the exact sequence of fluctuations required to flip a genetic switch.

This "instanton" approach—finding the most probable trajectory to an unlikely state—is a recurring theme. Whether it is Jean‐Philippe Bouchaud analyzing extreme velocity fluctuations in Burgers turbulence or Peijia Yu and Qing Nie modeling stem cell differentiation, the logic remains the same: to understand a rare catastrophe or a rare transition, don't look at the average behavior. Instead, solve a deterministic optimization problem to find the "cheapest" way the noise can break the system.

In systems with "heavy tails" (where extreme events are more common), this path often takes the form of a "single big jump." Bert Zwart and Roger J. A. Laeven show in Large deviations asymptotics for unbounded additive functionals of diffusion processes that for certain processes, a rare extreme average isn't caused by a steady accumulation of small errors, but by one massive excursion that outweighs everything else.

Hamiltonian Flows on the Density Manifold

While the Onsager-Machlup approach focuses on individual paths, another facet of this research looks at the evolution of entire populations or probability distributions. This is the world of Wasserstein Hamiltonian Flows.

Instead of thinking about a single particle moving through space, researchers like Wuchen Li and Shui-Nee Chow in Wasserstein Hamiltonian flows treat the space of all possible probability distributions as a geometric landscape (the Wasserstein manifold). In this view, the evolution of a density is a "flow" that follows a principle of least action.

Jianbo Cui and Shu Liu extended this to Stochastic Wasserstein Hamiltonian Flows, showing that even when a population is buffeted by "common noise," its macro-scale evolution remains a structured geometric flow. This provides a unified language for problems as diverse as the Schrödinger bridge (finding the most likely way a cloud of particles moved from A to B) and mean-field games.

From Physics to Generative AI

This deep connection between action minimization and probability is now transforming machine learning. Generative models, particularly diffusion models, work by slowly reversing a noise process to create data.

In Understanding Diffusion Models by Feynman's Path Integral, Yuji Hirono and colleagues use path integrals to calculate the exact likelihood of these models, treating noise as analogous to quantum fluctuations. Similarly, Sanjeev Raja and Michael Psenka demonstrate in Action-Minimization Meets Generative Modeling that we can use the learned "score" of a diffusion model to find high-probability transition paths in molecular systems, bypassing the need for expensive simulations.

Even the behavior of Stochastic Gradient Descent (SGD) can be understood through this lens. Waïss Azizian and Panayotis Mertikopoulos show in The Global Convergence Time of Stochastic Gradient Descent that the time it takes for an optimizer to escape a local minimum is determined by the "action" required to climb over the energy barrier—a direct application of Large Deviations Theory to the training of AI.

By treating noise not as an obstacle, but as a medium that follows a principle of least action, this research direction provides a rigorous bridge between the deterministic laws of the macro-world and the stochastic reality of the micro-world.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_variational_principles_and_least_action__09d55814,
  title = {Variational Principles and Least Action for Stochastic Dynamics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/variational-principles-and-least-action-for-stochastic-dynamics/rdr_09d5581469a24c84a2f282773fdcfc76},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_variational_principles_and_least_action__09d55814_v1,
  title = {Variational Principles and Least Action for Stochastic Dynamics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/variational-principles-and-least-action-for-stochastic-dynamics/rdr_09d5581469a24c84a2f282773fdcfc76/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}