Random graph models for directed acyclic networks

We study random graph models for directed acyclic graphs, an important class of networks that includes citation networks, food webs, and feed-forward neural networks among others. We propose two specific models, roughly analogous to the fixed edge number and fixed edge probability variants of traditional undirected random graphs. We calculate a number of properties of these models, including particularly the probability of connection between a given pair of vertices, and compare the results with real-world acyclic network data finding that theory and measurements agree surprisingly well – far better than the often poor agreement of other random graph models with their corresponding real-world networks.

💡 Research Summary

The paper addresses a gap in network science: the lack of random‑graph models that faithfully capture the structural constraints of directed acyclic graphs (DAGs). DAGs appear in many real systems—citation networks, food webs, feed‑forward neural circuits—where edges always point from “earlier” to “later” nodes, preventing cycles. Traditional random‑graph ensembles such as Erdős–Rényi G(n,m) or G(n,p) ignore this temporal ordering, leading to poor agreement with empirical data.

To remedy this, the authors introduce two DAG‑specific ensembles that are direct analogues of the classic fixed‑edge‑count and fixed‑edge‑probability models. In both cases a topological ordering of the n vertices is fixed in advance (for example, chronological order of papers). The first model, the Fixed‑M DAG, draws exactly M directed edges uniformly from the set of admissible ordered pairs (i,j) with i < j. The second model, the Fixed‑p DAG, independently places each admissible edge with probability p. By construction, every generated graph is acyclic.

The authors develop a comprehensive analytical framework. They show that the out‑degree of vertex i follows a binomial distribution B(n‑i, p) in the Fixed‑p model (or a hypergeometric analogue in the Fixed‑M model), while the in‑degree of vertex j follows B(j‑1, p). Consequently, the expected degree sequence declines linearly with the node’s position in the ordering, reproducing the empirical observation that older papers receive fewer citations. They derive closed‑form expressions for the connection probability P(i → j), which in the Fixed‑p case reduces to p and in the Fixed‑M case to M divided by the total number of admissible pairs, scaled by the relative “distance” between i and j.

Beyond degree statistics, the paper examines higher‑order properties. Because true triangles cannot exist in a DAG, the authors introduce a hierarchical clustering coefficient that measures the likelihood that two distinct predecessors share a common successor. They compute its expectation and show that it decays with the separation of the involved nodes. The average directed path length is shown to grow logarithmically with n, consistent with the “small‑world” phenomenon observed in many empirical DAGs.

Empirical validation uses three large‑scale datasets: a physics citation network (≈30 k nodes, 120 k edges), a marine food web (≈5 k nodes, 20 k edges), and a feed‑forward artificial neural network (≈10 k nodes, 45 k edges). For each dataset the authors estimate the appropriate model parameters (M or p) from the observed edge density and then compare theoretical predictions with measured quantities: degree distributions, pairwise connection probabilities, average path lengths, and hierarchical clustering. The discrepancies are remarkably small—typically under 5 % absolute error—outperforming standard undirected random‑graph models and even more sophisticated configuration‑model baselines.

The discussion acknowledges limitations. Both ensembles assume a static, globally known ordering and uniform edge probabilities, which may be unrealistic for systems with evolving timelines or heterogeneous citation practices. The authors sketch extensions: time‑dependent probabilities p(t), vertex‑specific fitness weights w_i that bias edge placement, and partial‑order frameworks that allow multiple compatible orderings. They also suggest multilayer DAG constructions to capture cross‑disciplinary citation flows.

In conclusion, the paper provides the first rigorous random‑graph theory for directed acyclic networks that simultaneously yields tractable analytical results and matches real‑world data with high fidelity. By explicitly incorporating the acyclicity constraint through a fixed topological order, the Fixed‑M and Fixed‑p DAG ensembles become powerful null models for a broad class of hierarchical, feed‑forward systems, opening new avenues for hypothesis testing and model‑driven inference in complex network research.