🔢 mathematics

A Sklar-Type Decomposition for Positive Binary Distributions and a Gibbs Interpretation

This paper introduces a Sklar-type decomposition for strictly positive multivariate binary distributions that uniquely separates one-body marginals from a dependence nucleus characterized by higher-order Walsh log-linear coefficients, providing a Gibbs interpretation where these coefficients represent many-body interaction couplings.

Original authors: Yi-Yu Lin

Published 2026-10-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yi-Yu Lin

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 probability, where scientists try to understand how chance shapes the world, there is a fundamental puzzle: how do we tell the difference between the behavior of a single thing and the way that thing behaves when it is part of a group? Imagine a room full of people. Each person has their own habits, their own likelihood of speaking up or staying quiet. But when they are together, they might start a conversation, form a riot, or sit in silence. The total picture of the room is a mix of these individual tendencies and the complex web of connections between them. For continuous data, like height or temperature, mathematicians have long had a tool to separate these two ingredients, allowing them to study the connections without the noise of individual quirks. However, when the data is discrete—when things can only be in one of two states, like a coin flip or a light switch being on or off—this separation has been much harder to define with the same clarity.

This is the territory explored by Yi-Yu Lin in a new study focused on strictly positive binary distributions. In this context, "binary" simply means variables that can only take two values, and "strictly positive" means that every possible combination of these values has a non-zero chance of happening. The researcher set out to answer a precise question: can we take a complex table of probabilities for many such variables, strip away the individual behaviors of each variable, and be left with a pure, unchanging record of how they depend on one another? Furthermore, if we have that record of dependence and we pick any new set of individual behaviors, can we rebuild the exact original table of probabilities? The paper proves that the answer is yes, and it does so by identifying a specific set of numbers that act as the fingerprint of the relationship itself, independent of the individual parts.

The journey begins with a simple observation about how probability tables work. If you have a table showing the chances of different outcomes for two variables, you can change the table by simply adjusting the weights of the rows and columns. This process, which the author calls local positive reweighting, is like turning the volume up or down for specific rows and columns independently. It changes the overall balance of the table, altering the individual probabilities for each variable, but it does not change the fundamental way the variables relate to each other. Think of it as changing the size of the pieces in a puzzle without changing the shape of the picture they form together. The researcher shows that if you take any two tables that can be transformed into one another through this kind of weight adjustment, they share a deep, invariant core. This core is the "dependence nucleus," a concept that captures the essence of the relationship while ignoring the specific statistical laws of the individual parties.

To make this abstract idea concrete, the study turns to a mathematical language known as log-linear expansion. Instead of looking at the raw probabilities, the researcher looks at the logarithm of those numbers. When this is done, the complex web of interactions breaks down into a sum of simple terms. Some of these terms describe the behavior of single variables, while others describe how pairs of variables interact, how triplets interact, and so on. The crucial discovery is that the terms describing the individual variables are the only ones that change when you apply local reweighting. The terms that describe the interactions between two or more variables remain exactly the same. These higher-order interaction terms act as a complete coordinate system for the dependence nucleus. In other words, if two probability tables have the exact same set of these higher-order interaction numbers, they belong to the same family of relationships, regardless of how their individual probabilities differ.

The paper then tackles the second half of the puzzle: reconstruction. Having separated the relationship from the individual behaviors, can we put them back together? The author proves that if you fix the relationship (the set of higher-order interaction numbers) and then specify any new set of individual probabilities, there is one and only one way to combine them to form a valid joint probability table. This is a powerful result because it means the relationship and the individual behaviors are truly independent ingredients. You can swap out the individual habits of the variables without breaking the relationship, and you can change the relationship without forcing the individual habits to change. The two pieces of information can be specified separately and then uniquely recombined to recover the whole.

This mathematical structure finds a surprisingly direct home in the physical world of statistical mechanics, the branch of physics that studies how large collections of particles behave. In a system of tiny magnets, or spins, that can point up or down, the probability of finding them in a certain configuration is governed by an energy function. The higher-order interaction numbers identified in the paper correspond exactly to the forces, or couplings, that link these spins together. The individual probabilities correspond to the local magnetic fields acting on each spin. The study shows that for any fixed pattern of connections between the spins, you can adjust the local magnetic fields to achieve any desired pattern of individual magnetizations, and there will be exactly one way to do it. This provides a physical realization of the abstract mathematical proof: the "dependence nucleus" is the interaction structure of the system, and the "marginals" are the local fields.

The significance of this work lies in its clarity and its explicitness. While similar ideas have existed in the theory of contingency tables and in the study of continuous distributions, this paper brings them together in a transparent, binary setting. It moves beyond abstract definitions to show exactly how the separation works and how the reconstruction happens. By proving that the higher-order interaction coefficients completely characterize the relationship and that they can be uniquely paired with any set of individual probabilities, the study offers a rigorous framework for understanding complex dependencies. It confirms that the statistical behavior of a whole can be elegantly split into the behavior of its parts and the rules of their connection, and that these two aspects can be manipulated independently. This is not just a theoretical curiosity; it provides a clear, explicit method for disentangling the signal of interaction from the noise of individual variation in any system of binary variables.

Technical Summary: A Sklar-Type Decomposition for Positive Binary Distributions and a Gibbs Interpretation

Problem Statement
The paper addresses the structural decomposition of strictly positive multivariate binary probability distributions. The central problem is to rigorously separate the statistical laws governing individual variables (one-body marginals) from the statistical relations among them (dependence). While Sklar's theorem provides a well-established framework for continuous distributions—decomposing a joint distribution into marginals and a copula—the discrete case presents challenges. Specifically, discrete cumulative distribution functions are step functions, preventing a unique continuous copula extension. The author aims to establish a strict, explicit decomposition for finite binary systems that mirrors the spirit of Sklar's theorem without relying on arbitrary continuous extensions.

Methodology
The author adopts a framework based on local positive reweighting and log-linear (Walsh) expansions.

Local Reweighting: The paper defines an equivalence relation where two probability tables are considered equivalent if one can be transformed into the other by multiplying entries by positive factors depending only on individual variables (row/column scaling in the bipartite case, generalized to $N$ -partite systems). This transformation changes the one-body marginals but is intended to leave the "relational structure" invariant.
Log-Linear Expansion: For a strictly positive joint probability table $p(x)$ over binary variables $x_i \in \{-1, +1\}$ , the logarithm of the probability is expanded uniquely using Walsh characters (products of variables):
$\log p(x) = \sum_{S \subseteq [N]} J_S \chi_S(x)$
where $\chi_S(x) = \prod_{i \in S} x_i$ .
Coordinate Separation: The expansion separates coefficients into:
- One-body terms ( $J_i$ ) and the constant ( $J_\emptyset$ ): These are shown to be the only terms affected by local reweighting.
- Higher-order terms ( $J_S$ for $|S| \ge 2$ ): These coefficients are proven to be invariant under local reweighting.

Key Contributions and Results

Definition of the Dependence Nucleus:
The paper defines a dependence nucleus $[p]$ as the equivalence class of all strictly positive binary tables related by local positive reweighting. This abstract object represents the pure relational structure, independent of the specific marginal distributions of the individual parties.
Complete Coordinates for Dependence:
The author proves that the set of higher-order Walsh log-linear coefficients $\{J_S : |S| \ge 2\}$ serves as a complete set of coordinates for the dependence nucleus.
- Invariance: If two tables are related by local reweighting, they share identical higher-order coefficients.
- Completeness: Conversely, if two strictly positive tables share identical higher-order coefficients, they must belong to the same equivalence class (i.e., they differ only by local reweighting).
- This establishes a bijection: $[p] \leftrightarrow \{J_S : |S| \ge 2\}$ .
The Discrete Sklar Decomposition (Theorem 4.1):
The core result is the proof that the joint distribution can be uniquely recovered from its marginals and its dependence nucleus.
- Statement: For any fixed dependence nucleus (defined by $\{J_S : |S| \ge 2\}$ ) and any collection of strictly positive one-body marginals, there exists exactly one joint probability table that realizes both.
- Proof Strategy: The proof utilizes the properties of the log-partition function $A(h)$ $A (h)$ in the exponential family representation $p_h(x) \propto \exp(R(x) + h \cdot x)$ $p_{h} (x) \propto exp (R (x) + h \cdot x)$ , where $R(x)$ $R (x)$ contains the fixed higher-order terms and $h$ $h$ represents the one-body natural parameters.
  - Injectivity: The map from natural parameters $h$ to marginal expectations (magnetizations) $m$ is injective because the Hessian of the log-partition function (the covariance matrix) is positive definite.
  - Surjectivity: The map is surjective because the auxiliary function $\Phi_m(h) = A(h) - m \cdot h$ is strictly convex and coercive, guaranteeing a unique global minimum for any admissible target $m$ .
- Consequently, the decomposition $p \leftrightarrow (\{p_i\}, [p])$ is a bijection.
Canonical Representative:
Every equivalence class contains a unique representative with uniform one-body marginals (zero magnetization). This table serves as the discrete analog to the continuous copula.

Statistical-Mechanical Interpretation
The paper provides a direct physical interpretation of these mathematical objects within the context of finite-spin Gibbs systems:

The higher-order coefficients $J_S$ ( $|S| \ge 2$ ) correspond precisely to many-body interaction couplings in a Gibbs Hamiltonian.
The one-body coefficients correspond to local external fields.
The one-body marginals correspond to local magnetizations.
The decomposition implies that for any fixed interaction structure (couplings) and any desired configuration of local magnetizations, there exists a unique set of external fields that realizes that state. This clarifies that the "dependence structure" (interactions) is distinct from the observed correlations, which vary with the external fields.

Significance and Claims
The author explicitly states that their goal is not to develop a general theory for all discrete distributions or to define a specific function on $[0,1]^N$ as "the" discrete copula. Instead, the paper's significance lies in:

Explicitness: Providing a completely explicit, non-abstract formulation of the Sklar decomposition for the specific but structurally transparent case of strictly positive binary distributions.
Unification: Bridging concepts from contingency table theory (odds ratios, log-linear models), discrete probability (dependence nuclei), and statistical mechanics (Gibbs interactions).
Conceptual Clarity: Demonstrating that "relations" and "individual statistics" can be mathematically separated, varied independently, and uniquely recombined. The dependence nucleus is defined as an equivalence class under a specific transformation group, capturing a well-defined notion of relational invariance.

The work is motivated by problems in quantum-state reconstruction but is presented as a purely classical result in discrete probability, independent of quantum mechanics.

More like this