🤖 machine learning

Reconciling Causality and Non-Equilibrium Thermodynamics with Hamiltonian Causal Models

This paper introduces Hamiltonian Causal Models (HCMs), a trajectory-level framework that reconciles statistical causal modeling with non-equilibrium thermodynamics by defining causal effects through intervention-controlled Hamiltonian mechanisms and utilizing entropy production to quantify irreversibility and causal influence in non-stationary physical systems.

Original authors: Dario Rancati, Max Welling, Francesco Locatello

Published 2026-06-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dario Rancati, Max Welling, Francesco Locatello

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a movie of a physical system, like a drop of ink spreading in water or a pendulum swinging. Standard ways of studying cause and effect (causality) usually look at the "before" and "after" snapshots. They ask: "If I change the starting point, does the ending point change?"

But the authors of this paper argue that for physical systems, this "snapshot" approach misses the most important part of the story: the journey itself.

Here is a simple breakdown of their new idea, Hamiltonian Causal Models (HCMs), using everyday analogies.

1. The Problem: The "Snapshot" vs. The "Movie"

In standard causal models, we often treat cause and effect like a switch: you flip it, and something happens. But in the real physical world, causes often happen over time.

The Analogy: Imagine you are pushing a child on a swing.
- Standard View: You look at where the child is at the start and where they are at the end. If you push them differently but they end up in the same spot, standard models might say, "No difference!"
- The Reality: The way you pushed them mattered. Maybe you gave one long, smooth push, and another short, jerky push. Even if they end up in the same spot, the path they took was totally different. The child felt different forces, and the air resistance (friction) acted differently during the journey.

The authors say we need a model that watches the whole movie, not just the start and end frames.

2. The Solution: The "Energy Blueprint" (Hamiltonian Models)

The authors introduce a framework called Hamiltonian Causal Models. Think of a "Hamiltonian" as the energy blueprint of a system. It's a mathematical rulebook that says how energy is stored and how it moves between parts of a system.

The Analogy: Imagine a complex Rube Goldberg machine.
- The Blueprint (Hamiltonian) describes exactly how the dominoes, levers, and balls are connected and how much energy is in each part.
- The Rules of Motion: The laws of physics (like gravity) are the "immutable rules" that never change.
- The Intervention: An intervention is like a human hand coming in to tweak the blueprint. You might tighten a spring or change the angle of a ramp. You aren't changing the laws of gravity; you are changing the setup (the mechanism) that the gravity acts upon.

In this model, a "cause" is defined by how changing the blueprint changes the entire path the system takes through time.

3. The Secret Ingredient: The "Thermodynamic Fingerprint"

The most exciting part of the paper is how they detect these causes. They use a concept from physics called Entropy Production.

The Analogy: Think of Entropy as a measure of "messiness" or "irreversibility."
- If you watch a movie of a glass shattering, it looks real. If you play it backward (glass pieces flying up to reassemble), it looks impossible. That feeling of "this can't happen backward" is entropy.
- The Fingerprint: When you intervene in a system (like pushing that swing), you create a specific "thermodynamic footprint." The system generates heat and friction in a way that is unique to how you pushed it.

The authors discovered that Entropy Production acts like a super-sensitive detector for causality.

Why it's special: Standard measurements might say "The swing ended in the same spot, so the push was the same." But the Entropy Production will say, "No! The first push made a lot of heat and noise (high entropy), while the second push was smooth (low entropy). Therefore, the causes were different!"

It can detect causal effects that are invisible to standard methods because it looks at the irreversibility of the path, not just the final result.

4. The "Maxwell's Demon" Lesson

The paper also touches on a famous thought experiment involving a "demon" that sorts particles to reduce entropy. The authors use this to explain that interventions aren't free.

The Analogy: If you want to control a system (like a thermostat), you have to gather information and use energy to do it. The paper argues that to truly understand causality, you have to account for the "cost" of the intervention itself. The "demon" (the controller) is part of the system, and its actions leave a thermodynamic mark that proves a causal link exists.

5. What They Actually Proved (The Results)

The authors didn't just theorize; they tested this with computer simulations:

Detecting Hidden Differences: They created scenarios where two different control methods led to the exact same final result. Standard tools said, "No causal difference." Their new tool (Entropy Production) correctly said, "These are totally different causes!"
Mapping the Connections: In complex, looping systems (where A affects B, and B affects A), they showed that by measuring how much "entropy" a specific part of the system produces when nudged, they could figure out exactly which parts were connected to which. It's like figuring out who is talking to whom in a crowded room just by listening to the noise they make.

Summary

This paper proposes a new way to understand cause and effect in physical systems by:

Looking at the entire journey (trajectory) rather than just the start and end.
Using energy blueprints to separate the unchangeable laws of physics from the things we can control.
Using Entropy Production (the heat and messiness created) as a "fingerprint" to prove that a cause actually happened, even if the final result looks the same.

It bridges the gap between the statistical world of "data" and the physical world of "thermodynamics," offering a tool to see causal connections that were previously invisible.

Technical Summary: Reconciling Causality and Non-Equilibrium Thermodynamics with Hamiltonian Causal Models

Problem Statement

Standard statistical causal models (e.g., Structural Causal Models, potential outcomes) face significant challenges when applied to physical temporal phenomena. These frameworks often struggle to natively accommodate:

Trajectory-level interventions: Interventions that act as time-dependent controls along a system's path rather than static assignments.
Non-stationarity: Induced laws that change over time.
Path-dependence: Causal effects that depend on the specific history of the system, not just endpoint states.
Feedback and Cycles: Cyclic dependencies that unfold dynamically through time.

Furthermore, there is a conceptual gap between statistical causality and physical causation. In physics, causal direction is tied to the thermodynamic arrow of time (irreversibility), whereas standard causal models often treat time-reversal symmetric laws as compatible with causal asymmetry. Existing frameworks require ad-hoc constructions to handle these physical realities, failing to distinguish between immutable equations of motion and intervenable mechanisms.

Methodology: Hamiltonian Causal Models (HCMs)

The authors introduce Hamiltonian Causal Models (HCMs), a trajectory-level framework where observed variables interact with local, unobserved thermal environments (baths).

Core Definitions

System Structure: An HCM is defined on a symplectic manifold $\Gamma = \Gamma_X \times \Gamma_E$ , where $X$ represents observed variables and $E$ represents unobserved baths.
Hamiltonian Decomposition: The system evolves via a Hamiltonian $H(z, \lambda)$ $H (z, λ)$ that decomposes into:
- Intrinsic potentials ( $H_{self}$ ).
- Couplings to causal parents ( $H_{mech}$ ).
- Couplings to local baths ( $H_{bath}$ ).
- Free environment dynamics ( $H_E$ ).
  This decomposition plays the role of the Causal Markov factorization in classical Structural Causal Models (SCMs).
Interventions: Interventions are modeled as time-dependent policies $\lambda(t)$ that modify the Hamiltonian function (specifically the mechanisms $H_{mech}$ and $H_{bath}$ ) but do not alter the underlying equations of motion (the symplectic structure). This separates intervenable mechanisms from immutable physical laws.
Causal Effect: Defined as a discrepancy between the path laws of a variable under different intervention policies. Unlike standard Average Treatment Effects (ATE) which often focus on single-time points or cumulative sums, HCMs define causal effects over the entire trajectory space.

Thermodynamic Integration

The framework leverages non-equilibrium thermodynamics to quantify causal influence:

Work: The energy performed by an intervention. The authors show that work rates can identify which variables participate in a local mechanism (graphical parenthood), provided the Hamiltonian is known.
Entropy Production ( $\Sigma$ ): A path-space quantity measuring the irreversibility of a process by comparing the forward path law to its time-reversed counterpart.
- Projected Local Entropy Production: A quantity estimable from the marginal trajectory of a single node, serving as a witness for causal effects.
- Local Entropy Production Rate: A refined metric for controlled Langevin diffusions that, when subjected to infinitesimal "right-local" interventions, can distinguish direct parents from indirect ancestors.

Key Contributions

Formal Framework: The introduction of HCMs, which unifies controlled Hamiltonian systems with causal modeling, allowing for the native description of time-dependent, adaptive, and cyclic interventions.
Thermodynamic Causal Observables:
- Demonstration that work rates identify graphical parents when the Hamiltonian is known.
- Proof that projected entropy production acts as a witness for causal effects, detecting path-level discrepancies invisible to endpoint or cumulative ATEs.
- Derivation of a local entropy production response criterion (Theorem 3.9) that identifies direct parent structures in cyclic, non-stationary Langevin systems via infinitesimal interventions.
Theoretical Unification: Showing that HCMs recover classical SCMs and cyclic generalizations as stationary special cases (via the limit of large bath size and small inertial parameters), while extending naturally to non-equilibrium regimes.

Experimental Results

The authors validate the theory through two sets of experiments:

Validation of Path-Level Causal Detection (Proposition 3.7):
- Setting A (Same Endpoints): A linear chain of oscillators driven by two protocols reaching the same final state via different paths. Standard endpoint ATE is zero by design.
- Setting B (Reversed Time): A system with a hidden mediator receiving pulses in reverse order. Cumulative ATE is zero due to time-translation symmetry.
- Result: In both cases, standard estimands failed to detect the causal difference. However, the projected entropy production successfully distinguished the protocols with high statistical significance, confirming it detects path-level effects invisible to standard ATEs.
Parent Discovery in Cyclic Systems (Theorem 3.9):
- Setting: Random cyclic Erdős–Rényi graphs ( $n=15$ ) instantiated with three families of potentials (Bilinear, Sigmoid, Cubic).
- Method: Estimating the local entropy production response $R_{i|j}$ to infinitesimal interventions.
- Result: The method achieved near-perfect recall and strong precision in recovering direct parent structures in these challenging non-stationary, cyclic settings, outperforming standard approaches that struggle with cycles.

Significance and Claims

The paper claims to offer a "new language" for statistical causality that aligns naturally with how causation emerges in physical systems.

Reconciliation: It bridges the gap between statistical causal models and non-equilibrium thermodynamics, treating the thermodynamic arrow of time as the source of causal asymmetry.
Novelty in Causal Discovery: It provides tools to describe causality in systems where interventions are dynamic and feedback loops exist, areas where standard SCMs often require ad-hoc fixes.
Physical Interpretability: By grounding causal effects in entropy production, the framework offers a physically meaningful, data-estimable metric for causality that captures irreversibility.

The authors remain modest regarding future directions, noting that while HCMs generalize SCMs, they do not immediately describe general temporal SCMs with non-gradient drifts. They identify the extension to such frameworks and the development of causal discovery algorithms based on Hamiltonian learning as prominent next steps.