📊 statistics

Graphical Models for Multivariate Count Data

This paper introduces a unified parametric framework for modeling multivariate count data by extending classical sampling schemes to decomposable graphs through the addition of graphical hypergeometric and negative hypergeometric distributions, thereby enabling tractable Bayesian inference for data subject to exclusion or incompatibility constraints.

Original authors: Iza Danielewska, Bartosz Kołodziejek

Published 2026-08-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Iza Danielewska, Bartosz Kołodziejek

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 trying to organize a chaotic party where certain guests simply cannot be in the same room together. Maybe two people are rivals, or two devices interfere with each other's signals. In the world of statistics and data science, this is a classic puzzle: how do you count things when the things you are counting have strict rules about who can hang out with whom? This field is called graphical modeling. Think of a "graph" not as a chart on a spreadsheet, but as a map of connections. The dots (called vertices) are your items, and the lines (called edges) show which items are friends and which are enemies. If two items are enemies, they cannot appear together in a valid group.

For a long time, statisticians had great tools for counting when there were no rules at all, or when the rules were very simple. They had formulas for "sampling with replacement" (like drawing a card from a deck, looking at it, putting it back, and drawing again) and "sampling without replacement" (drawing a card and keeping it out). They also had ways to stop counting after a fixed number of tries, or after a specific number of "failures" (like drawing until you get a red card). But when the rules got complicated—like a complex web of enemies in a large party—scientists lacked a unified way to describe the counts. They needed a new set of mathematical tools that could handle these complex "incompatibility" rules while still being easy to calculate and understand.

This paper, written by Iza Danielewska and Bartosz Kołodziek, introduces a fresh and complete set of four mathematical families to solve exactly this problem. The authors take the four classic ways of counting (with/without replacement, fixed draws/fixed failures) and build a "graphical" version of each one. They show how to count groups of items that obey a specific map of "no-go" zones.

The core idea is surprisingly visual. Imagine your party guests are dots on a map. The "forbidden" pairs are connected by red lines. A valid group of guests is one where no two people in the group are connected by a red line. In math-speak, this is called an "independent set." The authors prove that you can treat these valid groups as the basic building blocks for counting. They create four distinct models:

Graphical Multinomial: You pick valid groups over and over again, putting them back each time (sampling with replacement), and count how many times each guest appears.
Graphical Negative Multinomial: You keep picking valid groups until you hit a specific "failure" condition, then you count the results.
Graphical Hypergeometric: You have a fixed, finite pool of valid groups. You pick a certain number of them without putting them back, and count the results.
Graphical Negative Hypergeometric: You pick from a finite pool without replacement, but you stop as soon as you hit a specific failure condition.

The beauty of this work is that these four models fit together perfectly like a puzzle. They all rely on the same underlying map of rules. If the map has no rules (everyone is friends), the models turn into the standard, simple counting formulas we already know. If the map is completely full of rules (everyone is enemies with everyone), the models turn into the complex classical formulas for those specific cases. In between, they offer a smooth, flexible way to handle any level of complexity.

The authors didn't just invent these formulas; they gave them a story. They showed that these distributions arise naturally from specific "sampling stories." For example, the "Hypergeometric" version isn't just a random equation; it describes exactly what happens if you take two independent groups of party-goers, mix them together, and then look at just one of the groups. This connection makes the math feel less like magic and more like a logical consequence of how the sampling works.

To prove their ideas work in the real world, the team tested their models on data from a physics experiment involving Rydberg atoms. In this experiment, scientists excite atoms to a high energy state, but there's a catch: if two atoms are too close, they can't both be excited at the same time (this is the "blockade" effect). The researchers mapped the atoms and their "too close" relationships onto a graph. They found that the "Graphical Multinomial" model perfectly described the patterns of excited atoms that followed the rules. Even though the real experiment had some messy errors (atoms that broke the rules due to measurement noise), the model was incredibly accurate at describing the valid patterns.

The paper also builds a "Bayesian hierarchy," which is a fancy way of saying they created a system for learning from data. If you start with a guess about how likely different valid groups are, and then you see some data, this system tells you exactly how to update your guess. It provides a clear path from "what we think might happen" to "what actually happened," all while respecting the complex rules of the graph.

In short, this paper completes a missing piece of the statistical puzzle. It provides a unified, flexible, and mathematically sound toolkit for counting things that have to play by strict social rules. Whether you are scheduling wireless signals, studying which genes mutate together in cancer, or packing particles into a box, these new models offer a way to understand the counts that respect the underlying structure of the problem. The authors have shown that by organizing these four families of distributions around a single graph, we can handle complex dependencies with the same ease we once handled simple ones.

Technical Summary: Graphical Models for Multivariate Count Data

Problem Statement
Multivariate count data frequently arise in applications where dependence between counts is structured by exclusion or incompatibility constraints (e.g., hard-core systems, mutually exclusive biological events). While classical parametric families exist—specifically the multinomial, negative multinomial, multivariate hypergeometric, and multivariate negative hypergeometric distributions—extending these to incorporate structured conditional independence via graphs has been challenging. Previous work by Danielewska et al. (2025) introduced graphical counterparts for the multinomial and negative multinomial families (sampling with replacement) but left the without replacement analogues undefined. The central challenge is to construct a unified framework that enriches classical families with graph-based dependence while preserving explicit probabilistic interpretations, tractable marginal/conditional distributions, and feasible statistical inference.

Methodology and Framework
The authors complete the graphical analogue of the classical sampling scheme by introducing two new distributions: the graphical hypergeometric ($hgG$) and the graphical negative hypergeometric ($nhgG$). These are defined for decomposable graphs (chordal graphs), a class that ensures the existence of clique-separator factorizations.

The methodology relies on the following core components:

Graph-Multinomial Coefficients: The distributions are constructed using graph-multinomial coefficients of the first and second types, derived from the multivariate independence polynomial $\delta_G$ and the polynomial $\Delta_G$ of the complement graph $G^*$ .
Sampling Schemes: The four families are organized by two sampling features:
- With Replacement: Graphical Multinomial ($multG$) and Graphical Negative Multinomial ($nmG$).
- Without Replacement: Graphical Hypergeometric ($hgG$) and Graphical Negative Hypergeometric ($nhgG$).
- Stopping Rules: Fixed number of draws vs. fixed number of failures.
Probabilistic Interpretation (Cartier–Foata Normal Form): The paper provides a unified probabilistic interpretation based on the Cartier–Foata normal form. Admissible sequences of cliques of the complement graph $G^*$ $G^{*}$ (equivalently, independent sets of $G$ $G$ ) serve as the fundamental combinatorial objects.
- $multG(r, y)$ represents the aggregate counts from $r$ independent draws of cliques from $G^*$ with replacement.
- $nmG(r, x)$ represents counts accumulated from a "sample cliques until failure" Markov chain on $G^*$ , where transitions follow specific Cartier–Foata admissibility relations.
- $hgG$ and $nhgG$ are derived as conditional distributions of $multG$ and $nmG$ vectors, respectively, given their sums, effectively modeling sampling without replacement from a finite population of admissible configurations.
Bayesian Hierarchies: The authors develop a unified Bayesian framework using graphical Dirichlet and graphical inverted Dirichlet distributions as conjugate priors. Integrating out the latent graphical parameters yields compound distributions (G-Dirichlet-multinomial and G-Dirichlet-negative multinomial) and provides explicit posterior and predictive laws for the unobserved remainder of the population.

Key Contributions

Completion of the Four-Family Scheme: The paper introduces $hgG$ and $nhgG$, completing the set of four graphical count distributions that correspond to the classical sampling matrix (Table 1).
Unification of Interpretations: It establishes a common probabilistic framework where all four models are viewed through the lens of clique sampling on the complement graph $G^*$ , clarifying the relationship between the "with replacement" and "without replacement" mechanisms.
Explicit Factorizations: For decomposable graphs, the probability mass functions admit explicit clique–separator and DAG factorizations. This ensures the models satisfy the global Markov property and possess tractable marginal and conditional distributions (which belong to the corresponding classical families).
Conditioning Relationships: The paper rigorously demonstrates that $hgG$ and $nhgG$ arise naturally by conditioning independent graphical multinomial/negative multinomial vectors on their sum, preserving the classical relationships between sampling with and without replacement.
Extension to Non-Decomposable Graphs: While the general four-family construction is restricted to decomposable graphs, the paper notes that the graphical Bernoulli case ($multG(1, y)$) is well-defined and globally Markov for arbitrary graphs, coinciding with the hard-core model from statistical physics.

Results and Applications

Theoretical Properties: The authors prove that the new distributions possess the global Markov property with respect to $G$ and that their clique marginals are classical multivariate hypergeometric and negative hypergeometric distributions.
Empirical Illustration: The framework is applied to Rydberg-atom excitation data (Kim et al., 2019). In this physical system, the "blockade constraint" implies that simultaneous excitation of adjacent atoms is forbidden, making admissible configurations independent sets of an interaction graph.
- The study fits the graphical multinomial model ($multG(1, y)$) to the subset of experimental shots that satisfy the blockade constraint (support-restricted analysis).
- Goodness-of-fit tests (likelihood-ratio statistics with parametric bootstrap) indicate that the graphical multinomial model adequately describes the admissible excitation patterns for various graph topologies (path, star, cycle, and paw graphs), even when the underlying graph is non-decomposable (e.g., $C_4$ ).
Potential Applications: The paper discusses several domains where these models are naturally applicable:
- Wireless Networks: Modeling collision-free schedules (independent sets) in CSMA networks.
- Cancer Genomics: Analyzing mutually exclusive genomic alterations across tumor samples.
- Spatial Statistics: Modeling hard-core configurations, adsorption, and packing processes.
- Loss Networks and Polymer Models: Representing resource-sharing constraints and compatible polymer configurations.

Significance and Claims
The paper claims to provide a unified parametric framework for graphical modeling of multivariate count data where dependence and admissible configurations are encoded by a graph. Its significance lies in:

Bridging Classical and Graphical Models: It interpolates between products of univariate distributions (for the empty graph) and classical multivariate distributions (for the complete graph), retaining explicit Markov factorizations.
Handling Exclusion Constraints: The framework is particularly natural for data arising under exclusion or incompatibility constraints, offering a principled way to model "hard" constraints (like the Rydberg blockade) alongside statistical dependence.
Statistical Tractability: By restricting to decomposable graphs (with the noted exception of the Bernoulli case), the models maintain computational feasibility for inference, including explicit posterior updates and predictive distributions within a Bayesian hierarchy.

The authors emphasize that these four distributions should not be viewed as models for unrelated data types, but rather as describing different observation schemes or inferential questions (e.g., fixed sample size vs. fixed population size) for the same underlying graph-constrained mechanism.

Technical Summary: Graphical Models for Multivariate Count Data

More like this