Stochastic dynamics from maximum entropy in action space

This paper establishes a unified, covariant, and information-theoretic framework for stochastic dynamics by maximizing Shannon entropy over a joint distribution of actions and endpoints, thereby deriving a Boltzmann-like action-space distribution that reproduces standard Brownian motion, extends naturally to relativistic regimes, and bridges the principle of least action with statistical inference without relying on functional path integration.

The Gist

Imagine you are trying to predict where a drunk person will end up after walking for a while. In the old way of thinking (the "path-based" approach), you would try to map out every single possible wobbly step they could take. You'd imagine them stepping left, then right, then tripping, then recovering. You'd have to calculate the probability of every single specific route they might take. It's like trying to count every single grain of sand on a beach to predict the tide. It's messy, complicated, and if you try to do this while moving at the speed of light (relativity), the math breaks down because "steps" don't make sense when time and space are flexible.

This paper proposes a much smarter, simpler way to look at the problem. Instead of counting every single path, the authors say: "Let's just count the total 'effort' or 'cost' of the journey."

Here is the breakdown of their idea using everyday analogies:

1. The New Way of Counting: "The Cost of the Trip"

Imagine you are a travel agent.

• The Old Way: You list every possible route a tourist could take from New York to London. Route A goes through Paris, Route B goes through Tokyo, Route C goes through a black hole. You assign a probability to each specific route.
• The New Way (This Paper): You stop caring about the specific cities they visit. You only care about the total price of the ticket.
  • Some routes cost $100.
  • Some cost $1,000.
  • Some cost $1,000,000.

The authors argue that instead of tracking the tourist's specific path, we should track the probability of the price. They call this "Action Space." In physics, "Action" is a measure of the "cost" or "effort" a particle exerts to get from point A to point B.

2. The Two Competing Forces: "The Price Tag vs. The Crowd"

The paper uses a concept called Maximum Entropy (which is just a fancy way of saying "be as uncertain as possible until you have to be specific"). They balance two things:

1. The "Least Effort" Rule: Nature generally likes to take the easiest, cheapest path. In our travel analogy, everyone wants the $100 ticket. This is the Principle of Least Action.
2. The "Crowd" Rule (Entropy): Sometimes, there are so many different ways to get a \1,000 ticket that it becomes statistically more likely to see someone with that ticket. Maybe there is only one \100 route, but there are a million different ways to spend $1,000.

The paper shows that the most likely outcome is a compromise between these two.

• If the "cheap" path is unique, the particle takes it.
• If the "expensive" path has a massive "crowd" of different routes leading to it, the particle might take the expensive path because there are simply more ways to get there.

They call this balance an "Action Free Energy." It's like a traveler deciding: "Is the extra cost of the expensive ticket worth the variety of routes available?"

3. Why This is a Big Deal for Relativity (The "Speed of Light" Problem)

The old method (counting specific steps) has a fatal flaw when dealing with Einstein's theory of relativity.

• The Problem: In the old method, you have to slice time into tiny steps (Step 1, Step 2, Step 3). But in relativity, "now" is different for everyone. If you slice time for one person, it looks messy for someone moving fast. The math breaks, and you can't predict things correctly at high speeds.
• The Solution: The "Total Cost" (Action) is a Lorentz Scalar. In plain English, this means the "price tag" of the trip looks the same to everyone, whether they are standing still or zooming past at the speed of light.
  • Because the authors are counting "prices" instead of "steps," their math works perfectly for slow particles (like a ball rolling) AND fast particles (like light or high-speed electrons). They don't have to force the math to work; it just works naturally.

4. The "Gaussian" Hill (The Shape of the Crowd)

The authors did the math to see what the "crowd" of routes looks like. They found that for a simple particle (like a speck of dust in water), the "crowd" of routes forms a bell curve (a Gaussian shape).

• The peak of the bell curve is the "cheapest" path (the straight line).
• The sides of the bell curve represent paths that are slightly more expensive but still very common.
• The further you go out, the fewer paths there are.

This allows them to use a mathematical shortcut (the "saddle-point approximation"). It's like saying, "The crowd is so huge right at the cheapest price that we can basically ignore the expensive paths for most calculations." This makes the math incredibly fast and easy compared to the old method.

5. The Result: A Unified Theory

By switching from "counting paths" to "counting costs," the authors achieved three things:

1. Simplicity: They replaced a nightmare of infinite-dimensional math (counting every path) with a simple one-dimensional integral (counting costs).
2. Covariance: Their theory works for both slow and fast particles without breaking.
3. Clarity: It clearly shows how the "laws of physics" (taking the easiest path) and "statistics" (the sheer number of options) fight and cooperate to determine where a particle ends up.

In summary: The paper suggests that to understand how particles move randomly, we shouldn't obsess over the specific wiggles and turns they take. Instead, we should look at the "total cost" of their journey. By doing this, we can easily predict their behavior whether they are moving slowly in a jar of water or racing through space at near-light speeds, all while using a single, elegant mathematical framework.

Technical Summary

Technical Summary: Stochastic Dynamics from Maximum Entropy in Action Space

Problem Statement
Standard formulations of stochastic dynamics, particularly for Brownian motion, often rely on the Wiener measure, which constructs probability distributions over ensembles of individual spacetime trajectories. While successful in non-relativistic regimes, this path-based approach encounters fundamental difficulties when extended to relativistic systems. The standard Wiener construction requires a preferred time parameter and a foliation of spacetime into spacelike hypersurfaces, which breaks Lorentz covariance. Consequently, relativistic stochastic processes derived via path integrals often fail to preserve causality or require ad hoc phenomenological modifications to restore covariance. Furthermore, the computational complexity of functional integration over infinite-dimensional path spaces poses significant challenges.

Methodology
The authors propose an information-theoretic reformulation of stochastic dynamics where the fundamental stochastic variable is not the individual path, but the total action ($A$) connecting two spacetime points, $a$ and $b$.

1. Maximum Entropy in Action Space: Instead of maximizing Shannon entropy over a distribution of paths $p_k(b|a)$, the authors maximize the relative entropy over a joint distribution of actions and endpoints, $p(b, A|a)$. The constraints are normalization and a fixed mean action $\langle A \rangle$.
2. Density of States: A central component of the framework is the introduction of a density of states, $g(A, b)$, which quantifies the measure (or degeneracy) of trajectories arriving at endpoint $b$ with a total action $A$. This function acts as the Jacobian when transforming from path space to action space.
3. Variational Principle: By maximizing the relative entropy $S[p\|g] = -\int p \ln(p/g)$ subject to constraints, the authors derive a Boltzmann-like distribution in action space:
   $$p(b, A|a) = \frac{g(A, b)e^{-\eta A}}{Z(\eta)}$$
   where $\eta$ is a Lagrange multiplier associated with the mean action constraint, acting as an inverse "informational temperature."
4. Derivation of $g(A, b)$: For free Brownian motion, the authors explicitly derive $g(A, b)$ using large deviation theory and microscopic collision models. They demonstrate that in the diffusive regime ($\Delta t \gg \tau_{relax}$), the density of states is Gaussian, centered at the minimal classical action $A_{min}$ with a variance scaling as $\sigma_A^2 \sim m k_B T D \Delta t$.
5. Saddle-Point Approximation: In the diffusive limit, the exponential factor $e^{-\eta A}$ varies much more rapidly than the Gaussian density of states. This justifies a saddle-point (Laplace) approximation, reducing the integration over action to a simple evaluation at $A_{min}$.

Key Results

• Non-Relativistic Limit: Applying the formalism to a free particle recovers the standard classical Brownian propagator. The Lagrange multiplier is identified as $\eta = 1/(2mD)$, linking the information-theoretic parameter to the diffusion coefficient. The resulting propagator satisfies the Chapman-Kolmogorov equation, confirming that the Markovian property emerges naturally from the Gaussian structure of the kernel rather than being imposed via path discretization.
• Relativistic Regime: The framework naturally extends to relativistic Brownian motion. Because the action is a Lorentz scalar, the probability distribution defined over action values is manifestly Lorentz covariant. The causal constraint $|\Delta x| \leq c \Delta t$ is naturally incorporated as the upper bound of the action integral ($A_{max} = 0$ for lightlike paths). The resulting propagator matches the covariant diffusion kernel previously derived by Dunkel and Hänggi, but here it emerges from a first-principles variational argument without ad hoc substitutions.
• Thermodynamic Analogy: The authors establish a structural analogy between equilibrium thermodynamics and action-space dynamics. The "informational temperature" $T_{info} = 1/\eta$ plays the role of temperature, and the system minimizes an effective potential $\Phi(A, b) = A - T_{info} \ln[g(A, b)/g_0]$. This potential balances the dynamical cost (minimizing action) against entropic degeneracy (maximizing the density of states).
• Fluctuation Theorems: The formalism allows for the derivation of action-space analogs to the Jarzynski equality and Crooks fluctuation theorem. By defining "action-work" ($W_A$) as the difference in action between forward and reverse protocols, the authors show that $\langle e^{-\eta W_A} \rangle = e^{-\eta \Delta F}$, where $\Delta F$ is the free action difference.

Significance and Claims
The paper claims to establish a unified, covariant, and information-based foundation for both classical and relativistic stochastic processes. Its primary contributions are:

1. Manifest Lorentz Covariance: By shifting the inference problem from path space to action space, the formulation avoids the non-covariant issues inherent in the Wiener measure, providing a rigorous derivation of relativistic diffusion kernels.
2. Computational and Conceptual Simplification: The approach bypasses the need for functional integration over paths. By integrating over scalar action values, it replaces infinite-dimensional integrals with ordinary integrals, simplifying calculations and making the role of entropic degeneracy explicit through $g(A, b)$.
3. Connection Between Variational Principles and Inference: The work clarifies the relationship between the principle of least action and statistical inference. It demonstrates that the principle of least action is the deterministic limit ($\eta \to \infty$) of a maximum entropy principle, while finite $\eta$ introduces a competition between action minimization and entropic effects.
4. Self-Contained Framework: The derivation of the density of states $g(A, b)$ from first principles (microscopic collisions and large deviation theory) renders the framework self-contained, avoiding the need to assume specific stochastic equations (like Langevin) at the variational level.

The authors conclude that this reformulation provides a robust alternative to path-based constructions, particularly for systems where path discretization is problematic or where Lorentz covariance is essential, while remaining fully consistent with established results in the non-relativistic limit.