Maximum lilkelihood estimation in the $beta$-model

We study maximum likelihood estimation for the statistical model for undirected random graphs, known as the $\beta$-model, in which the degree sequences are minimal sufficient statistics. We derive necessary and sufficient conditions, based on the polytope of degree sequences, for the existence of the maximum likelihood estimator (MLE) of the model parameters. We characterize in a combinatorial fashion sample points leading to a nonexistent MLE, and nonestimability of the probability parameters under a nonexistent MLE. We formulate conditions that guarantee that the MLE exists with probability tending to one as the number of nodes increases.

💡 Research Summary

The paper investigates the existence of the maximum likelihood estimator (MLE) for the β‑model of undirected random graphs, a model in which the degree sequence of a graph serves as the minimal sufficient statistic. The β‑model assigns a real parameter β_i to each node i and defines the edge probability between nodes i and j as p_{ij}=exp(β_i+β_j)/(1+exp(β_i+β_j)). When each possible edge (i,j) is observed N_{ij} times, the observed counts x_{ij} follow independent binomial distributions Bin(N_{ij}, p_{ij}). The authors introduce a normalized sufficient statistic ˜d(x)=A·(x./N), where A is the node‑edge incidence matrix of the complete graph and “./” denotes element‑wise division.

The central result (Theorem 3.1) states that the MLE exists if and only if the vector ˜d(x) lies in the interior of the degree‑sequence polytope P_n, defined as the convex hull of all degree sequences of simple graphs on n vertices. In geometric terms, P_n is a well‑studied polytope in graph theory; its boundary corresponds to degree sequences that satisfy the Erdős‑Gallai inequalities with equality. If ˜d(x) falls on the boundary, some β_i must diverge to ±∞, forcing the associated edge probabilities to be exactly 0 or 1. Consequently the likelihood cannot be maximized by any finite β vector, the MLE does not exist, and only a subset of the probability parameters (those corresponding to the “fixed” edges) remain estimable. The authors give a combinatorial description of the forbidden configurations that lead to boundary points, such as subgraphs that are simultaneously complete and independent in complementary vertex sets.

Beyond the deterministic condition, the paper derives probabilistic sufficient conditions guaranteeing that, as the number of nodes n grows, the MLE exists with probability tending to one. Assuming the expected degree sequence lies strictly inside P_n and is separated from the boundary by a margin that grows faster than the typical √n fluctuations, the authors prove that the probability of MLE existence converges to 1. This result improves on earlier work by Chatterjee, Diaconis, and Sly (2011) and Barvinok & Hartigan (2010), which required stronger “tameness” assumptions.

The methodology is rooted in the geometry of discrete exponential families under product‑multinomial sampling. The design matrix A defines a polyhedral cone; the existence of the MLE is equivalent to the observed sufficient statistic belonging to the relative interior of the associated marginal polytope. This perspective is not specific to the β‑model. The authors illustrate how the same polyhedral analysis applies to related models such as the p₁‑model, the Rasch model, the Bradley‑Terry model, and other exchangeable binary array models. In each case, the relevant polytope is defined by linear constraints analogous to the Erdős‑Gallai inequalities, and the interior‑point condition characterizes MLE existence.

From a computational standpoint, the paper proposes a practical algorithm for checking the interior condition. Since P_n can be described by a finite set of linear inequalities (the Erdős‑Gallai constraints), one can test whether ˜d(x) satisfies all inequalities strictly. This reduces the problem to a linear feasibility test, which is far more efficient than the previously used importance‑sampling or Markov‑basis methods for sampling graphs with a given degree sequence.

The authors also discuss extensions. They suggest studying more general sampling schemes where edges are observed a random number of times, refining the finite‑sample bounds on the non‑existence probability, and developing fast Newton‑type algorithms that exploit the polyhedral structure for parameter estimation when the MLE exists.

In summary, the paper provides a complete geometric characterisation of when the MLE for the β‑model exists, links non‑existence to explicit combinatorial patterns, offers milder asymptotic conditions for high‑probability existence, and presents an efficient diagnostic tool based on the degree‑sequence polytope. This work deepens the theoretical understanding of exponential‑family network models and supplies practitioners with concrete criteria to assess the feasibility of likelihood‑based inference on large network data.