Non-Poissonian Temporal Hypergraphs

• Non-Poissonian temporal hypergraphs are mathematical frameworks that capture time-varying, memory-driven group interactions beyond simple Poisson processes.
• They employ models like DARH, cDARH, and AR(1) to quantify long-range correlations, cross-order dependencies, and heavy-tailed inter-event times.
• Empirical validations on social and communication data demonstrate their effectiveness in reproducing burstiness, hierarchical decay, and community structure.

Non-Poissonian temporal hypergraphs are mathematical objects designed to capture time-varying group interactions that do not conform to simple memoryless (Poissonian) activation statistics. In contrast to static or pairwise temporal network representations, this framework enables rigorous characterization and modeling of complex memory, higher-order dependencies, and bursty event patterns observed in empirical data ranging from face-to-face interactions to collaborative projects. Recent research formalizes these concepts, introduces models with and without explicit memory, and provides extensive analysis of emergent statistical features, inference methods, and real-world applications (Gallo et al., 2023, Mancastroppa et al., 1 Jul 2025, Jo et al., 9 Apr 2026, Zhu et al., 20 Jun 2025).

1. Formal Definition and Distinction from Poissonian Models

A temporal hypergraph is defined as a sequence of hypergraphs

$$(\mathcal V,\mathcal H(1)),(\mathcal V,\mathcal H(2)),\dots,(\mathcal V,\mathcal H(T))$$

where $\mathcal V$ is a fixed set of $N$ nodes and $\mathcal H(t)$ is the set of hyperedges (groups of arbitrary size) active at discrete time $t$. A hyperedge of order $d$ is a subset of $d$ nodes.

Poissonian group interactions are memoryless: each hyperedge's activations are independent in time, resulting in exponential inter-event time distributions

$P(\Delta t=\tau)\propto e^{-\lambda \tau}$

and vanishing autocorrelations at any lag $\tau>0$. In contrast, non-Poissonian temporal hypergraphs encode statistical memory. Here, the probability that a group interaction occurs at time $t$ depends on its activation history, producing broadly distributed (often heavy-tailed) inter-event times and significant long-range temporal correlations (Gallo et al., 2023, Jo et al., 9 Apr 2026, Mancastroppa et al., 1 Jul 2025, Zhu et al., 20 Jun 2025).

2. Empirical Phenomena Characterizing Non-Poissonian Temporal Hypergraphs

Analysis of high-resolution social interaction data embedded into temporal hypergraphs reveals several salient non-Poissonian features (Gallo et al., 2023, Jo et al., 9 Apr 2026, Mancastroppa et al., 1 Jul 2025):

• Long-range autocorrelations: Intra-order correlation functions $\mathcal V$0 for fixed group size $\mathcal V$1 decay slowly, often as power-laws ($\mathcal V$2), in stark contrast to the rapid decay of memoryless processes.
• Hierarchical decay: Larger groupings ($\mathcal V$3 higher) exhibit shorter memory horizons, with their autocorrelation functions decaying more rapidly.
• Cross-order temporal dependencies: The activation of a group of size $\mathcal V$4 is temporally correlated with the future activation of groups of size $\mathcal V$5. This is quantified via cross-order correlation functions and is inconsistent with independent Poissonian group formation.
• Asymmetry and cross-order gaps: Many systems display positive cross-order gaps $\mathcal V$6, where formation of a larger group from a smaller one is more likely than the reverse, a feature absent from memoryless models.
• Heavy-tailed event statistics: Duration and inter-event time distributions of group activations and node participation are heavy-tailed: $\mathcal V$7 and $\mathcal V$8, with exponents outside those expected for Poissonian processes.

These observations collectively necessitate models with dynamic memory and higher-order dependency structure.

3. Mathematical Formalisms and Correlation Measures

To characterize and quantify temporal memory and higher-order dependencies, several correlation functions are introduced (Gallo et al., 2023):

• Intra-order correlation matrix $\mathcal V$9: Measures temporal dependencies of hyperedge activations within fixed order $N$0. Its trace yields the intra-order correlation function $N$1.
• Cross-order correlation matrix $N$2 and function $N$3: These quantify how group interactions of different sizes are temporally correlated.
• Normalized interaction matrices and cross-order gap $N$4: These normalize cross-order relations and capture asymmetry in group dynamics.

Such metrics enable rigorous empirical quantification of memory depth, cross-order coupling, and statistical hierarchies in real datasets.

4. Model Classes Generating Non-Poissonian Temporal Hypergraphs

Several models have been developed to generatively capture and analytically study non-Poissonian features.

a. Discrete Auto-Regressive Hypergraph (DARH) Model (Gallo et al., 2023):

Each hyperedge follows a discrete auto-regressive binary process: $N$5 with $N$6 a “copy-from-memory” Bernoulli switch, $N$7 a lag in $N$8, and $N$9 an independent Bernoulli variable. DARH captures long-range intra-order correlations but no cross-order dependencies.

b. Cross-memory DARH (cDARH) (Gallo et al., 2023):

Hyperedges sometimes copy their state from overlapping hyperedges of other orders, controlled by a cross-memory parameter $\mathcal H(t)$0. This extension reproduces empirical cross-order dependencies, asymmetries, and hierarchical memory decay.

c. Node-driven Markovian Models (Jo et al., 9 Apr 2026, Mancastroppa et al., 1 Jul 2025):

Each node undergoes a two-state Markov process (low/high activity), and hyperedge events are generated based on the configuration of constituent nodes (e.g., all nodes “active” for the AND rule; mixtures possible for linear rules). While node states are Markovian, the joint process over groups is inherently non-Poissonian, yielding heavy-tailed group inter-event times due to rate mixing.

d. Emerging Activity Temporal Hypergraph (EATH) (Mancastroppa et al., 1 Jul 2025):

Nodes switch between high and low activity phases with system- and node-specific modulation. Hyperedge generation is governed by instantaneous node activities, long-term tie memory, and short-term group reinforcement. Group continuation probabilities are modulated by group duration and tie strengths. EATH replicates real group-level burstiness, heavy tails, and higher-order correlations without imposing explicit non-exponential kernels.

e. AR(1) Hypergraph Model (Zhu et al., 20 Jun 2025):

Each hyperedge follows a first-order Markov (AR(1)) process: $\mathcal H(t)$1 with transition probabilities $\mathcal H(t)$2 controlling appearance/disappearance. This framework yields temporal dependence (non-independent increments), exponential decay of autocorrelation, and stationary non-Poissonian group activation statistics.

5. Theoretical Properties and Inference

Models for non-Poissonian temporal hypergraphs admit both analytical tractability and practical inference:

• Closed-form expressions: Many models yield explicit formulas for autocorrelation, inter-event distributions, and Hamming distances in terms of model parameters (Zhu et al., 20 Jun 2025, Jo et al., 9 Apr 2026).
• Likelihood-based estimation: The AR(1) hypergraph provides maximum likelihood estimators for transition parameters with uniform error bounds and asymptotic normality (Zhu et al., 20 Jun 2025).
• Diagnostic tools: Permutation-based residual analysis for independence and $\mathcal H(t)$3–statistic for model checking are introduced (Zhu et al., 20 Jun 2025).
• Block models and community detection: The AR(1) hypergraph stochastic block model (HSBM) enables explicit community recovery via Laplacian spectral clustering, with theoretical guarantees for recovery under signal-strength and separation conditions (Zhu et al., 20 Jun 2025).
• Change-point detection: Likelihood-based estimators detect statistically significant structural breaks in temporal hypergraph processes (Zhu et al., 20 Jun 2025).

6. Model Validation, Empirical Fit, and Applications

These models have been validated extensively on empirical datasets:

• Face-to-face and proximity interaction data: DARH/cDARH, node-driven Markovian, and EATH models accurately reproduce intra- and cross-order correlation functions, inter-event time distributions, burstiness parameters, and higher-order topological statistics observed in human interactions (Gallo et al., 2023, Mancastroppa et al., 1 Jul 2025, Jo et al., 9 Apr 2026, Zhu et al., 20 Jun 2025).
• Contagion processes: Simulations of higher-order SIR dynamics on EATH-based surrogates yield epidemic curves indistinguishable from real data, outperforming memory-free models (Mancastroppa et al., 1 Jul 2025).
• Community structure and change detection: AR(1) HSBM accurately detects communities and change-points in both educational and corporate communication hypergraphs, with improved clustering performance over static baselines (Zhu et al., 20 Jun 2025).
• Synthetic/Hybrid data generation: Models like EATH can interpolate or compose burstiness and memory structure from multiple datasets, allowing the synthesis of “hybrid” hypergraphs for controlled experiments (Mancastroppa et al., 1 Jul 2025).

7. Mechanisms and Interpretation of Non-Poissonian Memory

Non-Poissonian statistics in temporal hypergraphs emerge from either explicit memory mechanisms (e.g., auto-regression, cross-order copying, group reinforcement) or from the superposition of Markovian node dynamics yielding rate-mixed group event streams (Gallo et al., 2023, Mancastroppa et al., 1 Jul 2025, Jo et al., 9 Apr 2026, Zhu et al., 20 Jun 2025). Memory may be hierarchical (group-size dependent), cross-order (linking different group cardinalities), and modular (individually parameterized per node or group). All minimal generative rules reproducible in current models suffice to produce the heavy-tailed, bursty, and hierarchically correlated phenomena documented in empirical group interaction data.

In conclusion, non-Poissonian temporal hypergraphs provide a mathematically principled framework for modeling, analyzing, and simulating the rich, memory-driven structure of real-world group dynamics. Rigorous higher-order correlation metrics, tractable dynamic generative processes, and inference methodologies collectively enable deeper theoretical and applied understanding of bursty phenomena in complex systems (Gallo et al., 2023, Mancastroppa et al., 1 Jul 2025, Jo et al., 9 Apr 2026, Zhu et al., 20 Jun 2025).