A framework for the use of generative modelling in non-equilibrium statistical mechanics

This paper proposes a framework for modeling non-equilibrium and self-organizing systems using generative models and the variational free energy principle, which offers a tractable, parsimonious explanation of system dynamics as a form of variational inference without requiring the system to literally perform inference.

The Gist

The Big Picture: The "As If" Game

Imagine you are watching a flock of birds swirl in the sky. To a scientist, these birds are just physical objects moving according to wind, gravity, and muscle power. But this paper suggests a different way to look at them.

The authors argue that we can describe these birds as if they were doing math. Specifically, we can pretend they are constantly guessing where they should be and adjusting their flight to make their guesses match reality. They aren't actually sitting in the sky doing calculus; they are just physical objects. But describing them as if they are guessing makes it much easier for us (the scientists) to understand and predict how they move.

The paper calls this the Free Energy Principle (FEP). It's a tool that lets us model complex, messy systems (like cells, brains, or even the weather) by pretending they are trying to minimize "surprise."

The Core Concept: The "Markov Blanket" (The Invisible Wall)

To understand how this works, imagine a house with a very special wall.

• Inside the house: The "Internal States" (the family living there).
• Outside the house: The "External States" (the weather, the neighbors, the world).
• The Wall: The "Markov Blanket." This is the boundary (like sensory organs or skin) that separates the inside from the outside.

The wall has two types of windows:

1. Sensory Windows: You can see out, but you can't touch the outside directly.
2. Active Doors: You can push on the outside (like opening a window to let air in), but you can't see through them.

The paper claims that if a system has this kind of wall, it naturally tends to stay in a state where it isn't "surprised" by what comes through the sensory windows. If a fish is in water, it expects to feel wet. If it suddenly feels dry (high surprise), it's in trouble. The system naturally moves to stay in the "wet" zone.

The Magic Trick: "Surprisal" and "Free Energy"

The authors introduce two key ideas:

1. Surprisal: How shocked a system is by its current situation. If you are a fish and you are in water, your surprisal is low. If you are on land, your surprisal is high.
2. Variational Free Energy: This is a mathematical "upper bound" on surprisal. Think of it as a scorecard.
   • The system doesn't know the exact score of its surprisal (because it can't see the whole world).
   • Instead, it uses a "best guess" model to calculate a scorecard called Free Energy.
   • The paper argues that physical systems naturally drift toward minimizing this scorecard.

The Analogy: Imagine you are playing a video game where you can't see the whole map. You only see a small circle around your character. You want to avoid falling off cliffs (surprise). You don't know exactly where the cliffs are, but you have a "hunch" (a generative model) about where they might be. You move to minimize the risk of falling off a cliff based on your hunch. The paper says physical objects do this automatically, not because they are thinking, but because of how they are built.

The "Map vs. Territory" Distinction

This is the most important philosophical point the paper makes.

• The Territory: The real world (the actual fish, the actual cells, the actual physics).
• The Map: The scientist's mathematical model.

Critics often say: "Wait! You are saying the fish has a map in its head. That's wrong! The fish is just a fish."

The authors say: "No, we aren't saying that."

They argue that the Map (our math) is just a tool we use to describe the Territory.

• We can write a map that says, "The fish behaves as if it is trying to stay in the water."
• This doesn't mean the fish is actually thinking, "I must stay wet."
• It just means that if we describe the fish using this "as if" logic, the math works perfectly.

The paper calls this a "deflationary" view. We aren't giving the fish a brain or a soul; we are just using a clever mathematical trick (variational inference) to describe its movement. The "inference" happens in our model, not necessarily in the fish.

The Examples: How It Works in Practice

The paper tests this idea with two computer simulations:

1. Cellular Morphogenesis (Building a Body):

   • Imagine a group of identical, undifferentiated cells.
   • The scientists give them a "target" (a map of what a head, body, and tail should look like).
   • The cells don't have a blueprint. Instead, they use the "Free Energy" rule. They move and change their chemical signals to match the "surprise" of not being in the right spot.
   • Result: The cells spontaneously organize themselves into a head, body, and tail, just by trying to minimize their "surprise" about where they are.

2. Periodically-Firing Cells (A Rhythm):

   • Imagine a ring of cells that need to fire in a specific rhythm (like a heartbeat).
   • The scientists set up a "target wave" (a sine wave).
   • The cells adjust their firing to match this wave, minimizing the error between what they feel and what they "expect" to feel.
   • Result: The cells lock into a perfect, stable rhythm, behaving as if they are predicting the future beats.

The Conclusion: A Map of Maps

The paper concludes with a clever twist.

If the "Territory" is the real world, and the "Map" is our scientific model...

• The Free Energy Principle is a Map of Maps.

It is a rule that tells us: "Any physical system that exists and stays together must, from the perspective of an observer, look like it is trying to minimize surprise."

It doesn't matter if the system is a rock, a cell, or a human brain. If it has a boundary (a Markov blanket) and stays stable, we can describe it using this "as if" logic. The paper isn't claiming the rock is conscious; it's claiming that our best way to understand the rock is to treat it as if it were a model of its own environment.

Summary in One Sentence

This paper proposes a mathematical framework where we can describe any stable physical system (like a cell or a machine) as if it were constantly guessing and correcting its own state to avoid "surprise," not because the system is actually thinking, but because this "as if" description is the most powerful and accurate way for scientists to model how the world works.

Technical Summary

Technical Summary: A Framework for the Use of Generative Modelling in Non-Equilibrium Statistical Mechanics

Problem Statement
The paper addresses the challenge of mathematically modelling open, non-equilibrium systems composed of coupled objects (e.g., particles, cells, or populations). Traditional approaches often rely on explicitly defining random dynamical systems, which can be intractable and difficult to interpret regarding why a joint system evolves in a specific manner. The authors seek a framework that offers a parsimonious explanation for the dynamics of interacting systems based on the properties of their coupling, without necessarily presupposing that the physical objects themselves perform literal inference. A central philosophical and technical hurdle addressed is the potential "map-territory fallacy"—the accusation that Free Energy Principle (FEP) modelling conflates the scientific model (the map) with the physical system itself (the territory), erroneously reifying inferential processes as literal features of the object.

Methodology
The authors propose a framework grounded in the Variational Free Energy Principle (FEP), utilizing generative models to represent dependence relationships between the states (or trajectories) of coupled systems. The methodology proceeds through the following steps:

1. Markov Blanket Partitioning: The system is partitioned into internal states ($\mu$), external states ($\eta$), and a Markov blanket ($b$) consisting of sensory and active states. This partition establishes conditional independence between internal and external states given the blanket.
2. Generative Modelling: A joint probability distribution $p(\eta, b, \mu)$ is defined as a generative model. This model encodes the statistical dependencies of the system.
3. Variational Inference as Dynamics: The authors demonstrate that if a system possesses a non-equilibrium steady state (NESS) density, its dynamics can be expressed as a gradient flow on variational free energy ($F$).
   • They utilize the decomposition of stochastic differential equations (SDEs) into a solenoidal (skew-symmetric) component and a dissipative (positive semi-definite) component.
   • Under the assumption that the system minimizes surprisal (negative log-probability) on average, and that the Kullback-Leibler (KL) divergence between a variational density $q$ and the true posterior vanishes, the dynamics of the internal states are shown to follow a gradient flow on $F$.
   • Crucially, $F$ serves as a Lyapunov function for the system's dynamics, guaranteeing stability and providing a tractable upper bound on surprisal.
4. Distinction of Models: The framework rigorously distinguishes between:
   • The Generative Model: The scientific model constructed by the observer (the "map").
   • The Variational Density: The probability distribution over external states parameterized by the internal states of the system.
   • The Physical System: The actual object (the "territory").

Key Contributions and Results

• Formal Tractability: The paper establishes that whenever a system admits a Markov blanket and a decomposition of its dynamics into solenoidal and dissipative flows (as per Eq. 2.1.1), its behavior can be mathematically recast as minimizing variational free energy. This allows complex, non-linear coupling dynamics to be described via the more tractable mathematics of statistical inference and gradient flows.
• Empirical Illustrations (Toy Models): The authors present two simulations to demonstrate the constructive application of the FEP:
  1. Cellular Morphogenesis: A simulation of eight undifferentiated cells migrating and differentiating to form a target morphology (head, body, tail). By equipping cells with a generative model predicting target locations and signal profiles, the system self-organizes. The cells' active states (position and signaling) minimize free energy, effectively "inferring" their unique identity and position within the ensemble.
  2. Periodically-Firing Cells: A ring of gap-junction-coupled excitable cells with a generative model encoding a periodic target waveform. The internal states infer the phase of the cycle, while active states (ion-channel gating) minimize prediction errors. The system converges to a limit cycle (a stable circulating orbit) rather than a fixed point, demonstrating that FEP dynamics can describe oscillatory self-organization.
• Resolution of the Map-Territory Fallacy: The paper argues that the FEP does not commit the map-territory fallacy. The claim that a system "performs inference" is an 'as if' description provided by the observer's generative model. The system itself does not need to literally calculate gradients or perform Bayesian updates; rather, its physical dynamics are mathematically equivalent to such a process. The "inference" is a property of the model of the system, not necessarily a literal mechanism within the system.

Significance and Claims
The authors position this work as a return to the foundational principles of the FEP, emphasizing its role as a constructive framework rather than a universal empirical law.

• Modest Scope: The paper explicitly states that the provided examples are "worked examples" of a constructive use of the FEP, not independent empirical validations of its universality. The examples were engineered to possess the necessary decomposition (Markov blanket, specific $Q$ and $\Gamma$ fields) to illustrate how the formalism works when the required conditions are met.
• A "Map of Maps": The central philosophical claim is that the FEP provides a "map of that part of the territory that behaves as if it were a map." It offers a set of ultimate explanatory constraints on what constitutes a model of a physical process. By treating the generative model as a scientific tool to describe how subsystems track one another, the FEP allows for the construction of nested models that respect known relations without reifying the inferential aspects of the model as literal features of the physical object.
• Philosophical Alignment: The framework aligns with the sentiments of Wittgenstein and Jaynes, suggesting that statistical mechanics and modelling are processes of making sense of sensory streams and microstates given macroscopic constraints. The FEP is presented as a principled approach to formalizing the relationship between a physical system and the models used to describe it, clarifying that the "inference" resides in the mathematical description (the map) used to explain the stability and self-organization of the physical system (the territory).

In summary, the paper argues that using generative models to represent relations between subsystems leads to a statistical theory where the dynamics of interacting systems can be read as variational inference. This provides a powerful, tractable tool for modelling non-equilibrium phenomena, provided one maintains a clear distinction between the scientific model (the generative model) and the physical system it describes.