The power of a null model in network science lies in what it preserves and what it randomizes. To determine whether a feature of an observed network—say, an unusual number of triangular interaction patterns among contacts in a disease-spread network—is genuinely surprising, researchers compare it against ensembles of random networks that share the observed network’s fundamental constraints, such as the degrees of every node. The subtlety is that a poorly chosen null model can produce bogus statistical conclusions: if the null ensemble contains networks that are structurally impossible for the phenomenon under study, or structurally unrealistic, a hypothesis may be wrongly accepted or rejected. The DiNgHy work, by Maryam Abuissa and Eli Upfal of Brown University and Matteo Riondato of Amherst College, targets exactly this problem for directed hypergraphs.

Their central concern is what the authors call degeneracy. In a non-degenerate directed hypergraph, no node can appear simultaneously in the tail and the head of the same hyperedge. This constraint is natural in many real systems: in a bill cosponsorship network, a sponsor is not also a cosponsor of the same bill; in a chemical reaction network, a gene is either a reagent or a product of a given reaction, not both; in a citation hypergraph connecting the authors of a citing paper to the authors of the cited one, self-citations are typically excluded. Yet every existing null model for dihypergraphs permits degeneracy, and—more damningly—the vast majority of samples drawn from those models are degenerate, even when the starting, observed network is not. Using such an ensemble as the benchmark for hypothesis testing is like asking whether a coin is fair by comparing it to flips of a biased one.

DiNgHy, which stands for a suite of Markov chain Monte Carlo (MCMC) sampling algorithms, is designed to sample uniformly from null models of non-degenerate directed hypergraphs. The underlying idea is conceptually familiar but technically demanding. The algorithms perform random “double-edge swaps” that reshuffle which nodes belong to the tails and heads of hyperedges, while strictly preserving the degree sequence of the network—every node keeps exactly the same number of appearances in tails and heads, and each hyperedge keeps its dimensions. Crucially, the swaps are engineered so that no intermediate or final state ever violates the non-degeneracy constraint. The team introduces two families of moves, dubbed HNDS and TNDS, which operate respectively on the head-side and tail-side membership of pairs of hyperedges, with a probabilistic choice between them at each step of the chain.

Verifying that such a Markov chain actually samples the target distribution is far from trivial, and the authors supply a complete theoretical treatment. The chain’s transition probabilities are symmetric, which makes the transition matrix doubly stochastic and guarantees that the unique stationary distribution is uniform over the entire space of non-degenerate dihypergraphs with the fixed degree sequence. This means that after a sufficient number of steps—known as the mixing time—the chain delivers samples from exactly the ensemble statisticians need, with no hidden bias toward particular networks.

An even more delicate issue is irreducibility: the property that the chain can reach any valid network from any other. Without it, the sampler might be trapped in a subset of the state space, silently narrowing the ensemble. Remarkably, DiNgHy’s underlying Markov chain is not irreducible in general—a departure from many classical swap-based samplers—so the authors prove mild sufficient conditions on the degree sequences under which irreducibility is guaranteed. These conditions, expressed through inequalities involving the sums of ordered degrees and hyperedge dimensions, are shown to hold for the real-world networks examined in the study. The proofs, presented in the appendices, build on combinatorial arguments bounding the availability of valid swap partners for every edge in the chain’s bipartite representation.

To understand why existing tools fall short, it helps to see how a directed hypergraph is represented computationally: as a pair of bipartite graphs, one linking nodes to the hyperedges whose tails they join, the other linking nodes to the hyperedges whose heads they belong to. Sampling degenerate-free structures therefore requires simultaneously randomizing both bipartite graphs under coupled constraints. Previous samplers, including configuration-model approaches based on curveball-style algorithms and methods for bipartite graphs, either permit degeneracy, cannot handle the directed, two-sided structure, or—as the authors show rigorously in earlier related work—provably cannot sample certain microcanonical bipartite ensembles via MCMC at all. DiNgHy is the first correct MCMC sampler for the configuration model of non-degenerate directed hypergraphs previously proposed in the literature, closing a genuine methodological gap.

The empirical evaluation is extensive, spanning both real and synthetic networks. The real datasets are strikingly diverse: email networks with senders in the tail and recipients in the head; MathOverflow question-and-answer threads where the original poster anchors the tail and responders populate the head; citation hypergraphs from DBLP in which a hyperedge links the authors of a citing paper to those of the cited work; metabolic and chemical reaction networks from Escherichia coli pathways and the Open Reaction Database, with genes or molecules as nodes; and a congressional hypergraph built from Senate bill sponsor-cosponsor relationships in the 107th U.S. Congress, a classic testbed for studying bipartisanship. Several of these datasets originally contained degenerate hyperedges—self-citations, self-responses, or genes appearing on both sides of a reaction—which were removed in a principled preprocessing step, producing clean networks that embody the non-degenerate settings DiNgHy is designed for.

The experiments quantify two things. First, the practical consequence of choosing a degeneracy-permitting null model: when testing hypotheses about motifs such as directed triangles and reciprocal two-cycles, the p-values obtained with the degenerate ensemble of prior work can differ sharply from those obtained with DiNgHy’s non-degenerate ensembles. In several cases a feature that appeared statistically ordinary under the old model turns out to be significant—or vice versa—under the correct null, underscoring how much a mismatched baseline can distort scientific inference. Second, the efficiency of the sampler itself: the authors measure the empirical mixing time of their chains, showing that the number of swap steps needed to reach the stationary distribution is manageable for networks of realistic size. The synthetic benchmarks, generated with varying hyperedge sizes and densities, confirm that the sampler’s performance degrades gracefully as networks grow denser.

The work also contributes conceptual clarity to a literature in which terminology has drifted. Different research groups have used conflicting conventions for “head” and “tail,” and some recent formulations treat node memberships as multisets, allowing repeats that blur the boundary between degeneracy and self-loops. DiNgHy adopts the convention from standard directed graphs—arrows point from tail to head—and defines degeneracy as any node appearing in both head and tail of the same hyperedge, a definition that connects cleanly to prior work on annotated hypergraphs and to the undirected case. Beyond the specific algorithms, the authors derive a striking structural result: random directed hypergraphs drawn from simple generative processes are, with overwhelming probability, degenerate. The probability that a sampled hypergraph is entirely non-degenerate decays exponentially with the network’s size-to-density ratio, which is precisely why naive rejection sampling fails and a constrained chain like DiNgHy’s is necessary.

The implications reach across the sciences. Higher-order network models have been used to identify genes critical to pathogenic viral response, to model epidemic and information spreading, to study percolation on multiplex structures, and to analyze the social dynamics of group interactions. In all of these settings, statistically sound null models are the backbone of any claim that an observed pattern is meaningful rather than a byproduct of degree structure. By making the non-degenerate, directed, degree-preserving ensemble accessible for the first time through a correct, efficient sampler—with code publicly released on GitHub—the DiNgHy team gives researchers a sharper instrument for those claims. As hypergraph mining matures into a core tool of data science, the lesson of this work is clear: the ensembles we compare against must respect the structural rules of the systems we study, and with DiNgHy, for directed higher-order networks, they finally can.