🔬 condensed matter

Microcanonical Hamiltonian Monte Carlo and the Helmholtz Theorem

This paper provides a thermodynamic analysis of the Microcanonical Hamiltonian Monte Carlo algorithm, demonstrating its adherence to the Helmholtz theorem and its representation of a microcanonical ensemble, while ultimately arguing that canonical Markov Chain Monte Carlo methods are more natural from both thermodynamic and information-theoretic perspectives.

Original authors: Heinrich von Campe, Bjoern Malte Schaefer

Published 2026-09-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Heinrich von Campe, Bjoern Malte Schaefer

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

In the vast landscape of modern science, researchers often face a problem that is less about finding a single answer and more about understanding the shape of all possible answers. When scientists build models to describe everything from the spread of a disease to the expansion of the universe, they must estimate the values of many hidden variables. To do this, they rely on a powerful family of techniques known as Markov Chain Monte Carlo methods. Imagine trying to map the terrain of a foggy mountain range where you cannot see the peaks or valleys, only the ground immediately beneath your feet. These methods act as a systematic way to wander through that fog, taking steps that are guided by probability to eventually create a complete and accurate map of the entire landscape. This process is essential for making sense of complex data, but it requires a delicate balance: the wanderer must explore the terrain thoroughly without getting stuck in one spot or wandering aimlessly forever.

Recently, a new approach to this wandering was proposed, one that borrowed heavily from the laws of thermodynamics, the branch of physics that governs heat and energy. This new method, called Microcanonical Hamiltonian Monte Carlo, was designed to be more efficient than older techniques by keeping the total energy of the system constant, much like a frictionless pendulum swinging back and forth. However, while the algorithm was praised for its speed and performance, no one had stopped to ask if it truly obeyed the fundamental physical laws it was named after. A team of researchers at the University of Heidelberg and the Tübingen AI Center set out to fill this gap. They wanted to know if this digital wanderer was truly behaving like a physical system in a state of constant energy, or if it was merely mimicking the behavior without the underlying reality.

The researchers began by treating the algorithm not just as a piece of code, but as a physical system with its own temperature, pressure, and entropy. In physics, there is a fundamental rule known as the first law of thermodynamics, which describes how energy, heat, and work are related. For systems that keep their energy constant, there is a specific mathematical formulation of this law called the Helmholtz theorem. The team's primary goal was to prove that their algorithm satisfied this theorem. They performed a rigorous analysis, deriving the necessary equations from scratch and then testing them against real-world data. Their findings confirmed that the algorithm does indeed obey the Helmholtz theorem, but only under a very specific condition. The algorithm relies on a mechanism called "random bounces," where the direction of movement is occasionally randomized. The researchers discovered that without these random bounces, the system would fail to explore the entire landscape, violating the physical laws it was supposed to follow. With the bounces in place, the algorithm successfully maintains the delicate balance required by the theorem, proving that it is a genuine realization of a microcanonical ensemble.

To verify this in practice, the team applied the algorithm to two very different problems. First, they tested it on a simple, synthetic mathematical shape known as a Gaussian distribution, which is a standard test case for such methods. They then moved to a much more complex and realistic challenge: analyzing data from 1,590 Type Ia supernovae to understand the expansion of the universe. In both cases, the algorithm behaved exactly as the theory predicted. The researchers calculated the "temperature" of the system based on the data the algorithm produced and compared it to the theoretical value derived from the laws of thermodynamics. The numbers matched perfectly, even when the data was noisy and the mathematical landscape was twisted and difficult to navigate. This provided strong evidence that the algorithm is not just a clever trick, but a system that genuinely respects the deep physical principles of energy conservation.

However, the study also revealed a surprising limitation in the way this new method compares to older, more established techniques. While the algorithm works beautifully in high-dimensional spaces with many variables, the researchers found that its connection to the concept of information is less natural than that of the traditional methods. In the world of statistical inference, there is a specific measure of uncertainty called Shannon entropy, which quantifies how much information is contained in a probability distribution. The researchers demonstrated that for the older, "canonical" methods, the thermodynamic entropy of the system aligns perfectly with this information entropy. For the new microcanonical method, however, the two concepts do not align in the same intuitive way. The thermodynamic entropy of the new method turns out to be a complex mathematical construct that does not directly reflect the information content of the data in the same straightforward manner.

This distinction led the authors to a broader conclusion about the nature of these computational tools. They argued that while the new microcanonical method is a valid and powerful tool, the older canonical methods are a more "natural" fit for the problems of statistical inference. The canonical approach, which allows the system's energy to fluctuate, mirrors the way information is naturally processed in these problems, creating a seamless link between the physics of the simulation and the logic of the inference. The microcanonical approach, by forcing the energy to remain constant, requires extra steps and transformations to work correctly, making it feel somewhat forced in comparison. The researchers did not declare the new method useless; in fact, they showed it works well and even proposed a new variation that could handle simpler, lower-dimensional problems more effectively. But they made it clear that for the vast majority of complex inference tasks, the older, canonical way of thinking remains the most coherent and natural description of the process.

Ultimately, this work serves as a bridge between the abstract world of computer algorithms and the concrete laws of physics. By proving that the new algorithm obeys the Helmholtz theorem, the researchers have validated its physical consistency, giving users confidence in its results. At the same time, by highlighting the mismatch between its thermodynamic properties and information theory, they have provided a deeper understanding of why the older methods have remained the standard for so long. The study does not suggest that one method is a magic solution to all problems, nor does it dismiss the new approach. Instead, it offers a clear, physics-based perspective on how these tools work, showing that while the new method is a valid realization of a specific physical state, the older methods offer a more intuitive and natural connection to the information they are trying to uncover. This clarity helps scientists choose the right tool for the job, ensuring that their digital wanderers are not just moving fast, but moving in a way that is fundamentally sound.

Technical Summary: Microcanonical Hamiltonian Monte Carlo and the Helmholtz Theorem

Problem Statement
The Microcanonical Hamiltonian Monte Carlo (MCHMC) algorithm, recently proposed by Robnik et al. (2023), offers a sampling method based on the microcanonical ensemble of statistical physics. While its technical and algorithmic performance has been studied, its thermodynamic properties remain unexplored. Specifically, it is unclear whether MCHMC formally satisfies the fundamental laws of thermodynamics, such as the first law, when applied to Bayesian inference problems. Furthermore, the relationship between the algorithm's design (e.g., "random bounces") and thermodynamic prerequisites like ergodicity and monocyclicity requires clarification. This work aims to fill this gap by deriving thermodynamic state variables and potentials for MCHMC and verifying if the construction formally represents a microcanonical thermodynamic ensemble.

Methodology
The authors approach the problem by constructing a thermodynamic framework analogous to the canonical partition function used in standard Hamiltonian Monte Carlo (HMC).

Theoretical Derivation: The paper defines microcanonical partition functions ( $\Omega_0$ and $\Omega_2$ ) corresponding to two variants of the MCHMC Hamiltonian ( $H_0$ and $H_2$ ). Using the Dirac delta distribution and integrating out angular momentum degrees of freedom, the authors derive analytical expressions for the surface area ( $\Omega$ ) and volume ( $\phi$ ) of the phase space.
Thermodynamic Verification: The authors test the Helmholtz theorem, an alternative formulation of the first law of thermodynamics for microcanonical systems ( $dS = \frac{1}{T}dE - \frac{1}{T}\Lambda \cdot dJ$ ). They analytically compute the temperature ( $T$ ) and the conjugate variable to the external source ( $\Lambda$ ) using statistical mechanical definitions and compare them against the partial derivatives of the volume entropy ( $S = \log \phi$ ).
Numerical Experiments: The theoretical findings are validated through numerical experiments on three levels of complexity:
- Analytical: Verification for general likelihoods under $H_0$ and $H_2$ .
- Toy Problem: A 3-dimensional Gaussian likelihood where temperature estimates from MCHMC samples are compared to analytical predictions and finite-difference calculations of entropy derivatives.
- Real-world Application: A cosmological inference problem using the Pantheon+SH0ES Type Ia supernova dataset. Due to the high conditioning number of the likelihood and the requirement to evaluate the likelihood $L$ directly (rather than $\log L$ ) in $H_2$ , the authors employ data whitening and coordinate transformations to ensure the sampler functions correctly.
Comparative Analysis: The paper contrasts the thermodynamic nature of MCHMC (microcanonical) with standard MCMC methods like HMC and Rosenbluth-Metropolis-Hastings (canonical), specifically analyzing how thermodynamic entropy relates to information-theoretic (Shannon) entropy in each framework.

Key Contributions

Formal Thermodynamic Characterization: The paper demonstrates that MCHMC formally represents a microcanonical thermodynamic ensemble. It derives the necessary state variables and potentials, confirming that the algorithm is grounded in statistical physics.
Verification of the Helmholtz Theorem: The authors prove analytically and numerically that MCHMC satisfies the Helmholtz theorem. This verification holds for general likelihoods and specific inference problems, confirming the algorithm's adherence to the first law of thermodynamics.
Role of Random Bounces: The study clarifies that the "random bounces" (resampling of momentum direction) introduced in MCHMC are not merely heuristic but are essential for satisfying the condition of monocyclicity. Monocyclicity is shown to be a stronger requirement than ergodicity; without it, the sampler fails to explore the full constant-energy surface, and the Helmholtz theorem is violated.
New Sampling Algorithm (VMCHMC): By utilizing the volume entropy ( $\phi$ ) rather than the surface entropy ( $\Omega$ ), the authors propose a new volume-based microcanonical sampling algorithm (VMCHMC). This algorithm is shown to be applicable to lower-dimensional inference problems where standard MCHMC struggles, while converging to MCHMC behavior in high-dimensional settings.
Canonical vs. Microcanonical Nature: The paper argues that canonical MCMC algorithms are more "natural" from a thermodynamic and information-theoretic perspective. It demonstrates that for canonical algorithms, the thermodynamic entropy at unit temperature coincides exactly with the Shannon entropy of the posterior distribution. In contrast, for MCHMC, the thermodynamic entropy at zero energy does not correspond to a meaningful information entropy measure for the inference problem.

Results

Analytical Consistency: For both $H_0$ and $H_2$ , the computed temperature and pressure-like variables satisfy the Helmholtz theorem relations ( $T = (\partial S / \partial E)^{-1}$ and $\Lambda = T \partial S / \partial J$ ).
Numerical Agreement: In the 3D Gaussian toy example and the supernova cosmology application, the temperature estimated from MCHMC samples matches the theoretical prediction ( $T = 2E/n$ for the Gaussian case) and the finite-difference derivatives of the volume entropy.
Limitations of $H_2$ : While $H_2$ satisfies the thermodynamic laws, the authors note practical difficulties in real-world applications (like the supernova dataset) because $H_2$ requires evaluating the likelihood $L$ directly, which can lead to numerical underflow or require specific data transformations (whitening). The authors note that Robnik et al. (2023) previously reported $H_0$ as the better-performing sampling algorithm, though $H_0$ yields trivial energy dynamics in this thermodynamic analysis.
Entropy Mismatch: The analysis confirms that the thermodynamic entropy of the microcanonical ensemble does not map to the Shannon entropy of the posterior, whereas the canonical ensemble does.

Significance and Claims
The paper claims to provide a rigorous thermodynamic foundation for MCHMC, verifying that it is a valid realization of a microcanonical ensemble that obeys the first law of thermodynamics. A central insight is the identification of "monocyclicity" as the critical physical condition ensuring the algorithm's validity, which is enforced by the random bounce mechanism.

However, the authors modestly conclude that despite these thermodynamic successes, canonical MCMC algorithms (like HMC) offer a more natural description for Bayesian inference. This is primarily because the canonical framework naturally unifies thermodynamic entropy with information-theoretic entropy (Shannon entropy), a property that the microcanonical framework lacks. The paper suggests that while MCHMC is a valid microcanonical realization, the canonical approach remains superior for connecting statistical physics with the theory of statistical inference. The work also opens the door for future research into volume-based sampling (VMCHMC) and the application of thermodynamic tools to other sampling algorithms like Langevin Monte Carlo.

Technical Summary: Microcanonical Hamiltonian Monte Carlo and the Helmholtz Theorem

More like this