Abstract Markov Random Fields

Markov random fields are known to be fully characterized by properties of their information diagrams, or I-diagrams. In particular, for Markov random fields, regions in the I-diagram corresponding to disconnected vertex sets in the graph vanish. Recently, I-diagrams have been generalized to F-diagrams, for a larger class of functions F satisfying the chain rule beyond Shannon entropy, such as Kullback-Leibler divergence and cross-entropy. In this work, we generalize the notion and characterization of Markov random fields to this larger class of functions F and investigate preliminary applications. We define F-independences, F-mutual independences, and F-Markov random fields and characterize them by their F-diagram. In the process, we also define F-dual total correlation and prove that its vanishing is equivalent to F-mutual independence. We then apply our results to information functions F that are applied to probability mass functions. We show that if the probability distributions of a set of random variables are Markov random fields for the same graph, then we formally recover the notion of an F-Markov random field for that graph. We then study the Kullback-Leibler diagrams on specific Markov chains, leading to a visual representation of the second law of thermodynamics and a simple explicit derivation of the decomposition of the evidence lower bound for diffusion models.

💡 Research Summary

The paper presents a comprehensive generalization of Markov random fields (MRFs) by extending the classical information‑diagram (I‑diagram) framework to a broader class of information‑like functions, denoted F. Functions F are required only to satisfy a chain rule—examples include Shannon entropy, Kullback‑Leibler (KL) divergence, cross‑entropy, Tsallis entropy, and even Kolmogorov complexity. The authors introduce F‑diagrams, which retain the Venn‑type region structure of I‑diagrams but assign to each region a value computed from the chosen F.

Key definitions are introduced:

• F‑independence – two variables X and Y are F‑independent when the second‑order F‑information (the analogue of mutual information) vanishes.
• F‑mutual independence – a collection of variables is F‑mutually independent if all higher‑order F‑interactions (I_q^F for any q) are zero.
• F‑dual total correlation – a generalized total correlation defined as the sum of first‑order F‑terms minus the joint F‑term; its vanishing is proved to be equivalent to F‑mutual independence.

Using the algebraic notion of a separoid (a structure capturing conditional independence axioms) together with a commutative idempotent monoid of random variables, the authors show that F‑conditional independences satisfy the separoid axioms. This provides a solid abstract foundation for reasoning about independence in the F‑framework.

The central contribution is the definition of F‑Markov random fields. A set of random variables indexed by a graph G forms an F‑MRF if the global Markov property holds with respect to the F‑conditional independence: whenever two vertex sets are separated by a third in G, the corresponding F‑conditional independence holds. The authors prove a diagrammatic characterization: the regions of the F‑diagram that correspond to disconnected vertex sets in G are exactly those that must vanish. This mirrors the classic result for I‑diagrams but now applies to any chain‑rule‑satisfying F.

The paper then specializes the theory to probability mass functions. When the function F is applied to probability distributions (e.g., entropy, KL‑divergence, cross‑entropy), the same graphical criteria hold: if several distributions share the same underlying graph and each individually satisfies the Markov property, then the collection of distributions together constitutes an F‑MRF on that graph. In other words, the “distribution‑level” Markov structure mirrors the “variable‑level” Markov structure under the chosen F.

Two concrete applications illustrate the power of the framework:

1. KL‑diagrams and the second law of thermodynamics – By treating a discrete‑time Markov chain as a sequence of random variables and choosing F to be KL‑divergence, the authors construct a KL‑diagram that visualizes the monotonic decrease of KL‑divergence over time. This provides an information‑theoretic diagrammatic representation of the weak form of the second law: entropy (or KL‑divergence from equilibrium) cannot increase in a closed Markov process.

2. Evidence Lower Bound (ELBO) decomposition for diffusion models – Diffusion models are built on a forward noising Markov chain and a reverse denoising chain. The training objective is the ELBO, which involves a sum of KL‑divergences across time steps. Using the KL‑diagram of the underlying Markov chain, the authors derive an explicit, step‑by‑step decomposition of the ELBO. This not only clarifies the contribution of each timestep but also yields a compact visual proof of the standard ELBO formula used in modern generative modeling.

The discussion highlights several promising directions: extending F‑diagrams to continuous spaces (requiring measure‑theoretic care), exploring richer algebraic structures beyond monoids (e.g., groups or rings), and applying the framework to complex systems in physics, biology, and neuroscience where non‑Shannon information measures are natural.

In summary, the paper establishes a unifying, diagram‑based theory that lifts the classic Markov random field characterization from Shannon entropy to any information‑like functional satisfying a chain rule. By doing so, it bridges information geometry, graphical models, and modern machine‑learning objectives, offering both deep theoretical insights and practical tools for visualizing and decomposing complex probabilistic systems.