Poisson distribution
A Poisson distribution is a discrete probability distribution that expresses the probability of a given number of events occurring in a fixed interval of time or space, assuming the events occur independently at a constant average rate. If events happen with rate \( \lambda \), the probability of observing exactly \( k \) events is:
\[ P(k) = \frac{\lambda^k e^{-\lambda}}{k!} \]
The Poisson distribution is the default model for randomness without structure. It arises whenever events are independent, uniformly distributed in time or space, and rare relative to the observation window. It is mathematically elegant, analytically tractable, and — in the study of networks — almost universally wrong as a description of real systems.
The Poisson as a Null Model
In graph theory, the Erdős–Rényi model generates random graphs in which each possible edge exists with a fixed probability \( p \). The resulting degree distribution is approximately Poisson: most nodes have degrees near the mean \( \langle k \rangle = p(N-1) \), and deviations from the mean decay factorially fast. For decades, this was the default model of what a "random" network looks like.
The Poisson assumption underlies much of classical network theory, including percolation thresholds, epidemic thresholds, and phase transitions in random graphs. The mathematics is beautiful: exact solutions exist for connectivity, component sizes, and path lengths that have no closed-form counterparts for non-Poisson networks. The price of this beauty is empirical irrelevance.
Virtually no real network has a Poisson degree distribution. Social networks, the Internet, protein interaction networks, citation networks, power grids, and ecological food webs all exhibit heavy-tailed degree distributions — power laws, log-normals, or stretched exponentials — in which extreme-degree nodes are orders of magnitude more common than the Poisson model permits. The Poisson is not merely a simplification; it is a structural mischaracterization of the systems it purports to describe.
Why Real Networks Are Not Poisson
The independence assumption fails in almost every domain where networks matter. In social networks, people do not form friendships randomly; they meet through shared contacts, shared spaces, and shared interests. On the web, pages do not link randomly; they link to related content, popular destinations, and previously linked pages. In biological networks, genes do not interact randomly; they are co-regulated, co-expressed, and constrained by cellular architecture.
The Poisson distribution assumes that the rate \( \lambda \) is constant across the population. In real networks, the "rate" at which nodes acquire connections is heterogeneous: some nodes are intrinsically more attractive, appeared earlier, occupy more central positions, or are promoted by algorithms. This heterogeneity — whether from preferential attachment, fitness differences, duplication processes, or optimization — produces distributions with fatter tails than the Poisson allows.
Even when a network's degree distribution looks roughly unimodal and symmetric, careful statistical testing often reveals that the tails are heavier than Poisson. The degree sequence of a real network may have a mode near the mean, but the variance-to-mean ratio (the Fano factor) is typically much larger than 1, whereas the Poisson distribution has variance equal to mean.
When the Poisson Is the Right Model
The Poisson distribution is not useless. It remains the correct model in specific contexts:
- Random geometric graphs, where nodes are placed uniformly in space and connect to neighbors within a fixed radius. In the limit of large networks with constant density, the degree distribution is approximately Poisson.
- Neural spike trains under certain experimental conditions, where action potentials are triggered by independent Poisson-like inputs.
- Queueing theory, where arrivals to a service system are often modeled as Poisson processes.
- Null models in statistical network analysis, where the configuration model with a fixed degree sequence is compared against a Poisson baseline to measure deviation from randomness.
In each case, the Poisson is appropriate because the generative process genuinely involves independent, uniform events. When these conditions fail — as they do in virtually all complex networks — the Poisson is not merely imprecise; it is actively misleading.
The Poisson as Pedagogical Danger
The persistence of Poisson models in network science education is a problem. Students learning graph theory are often taught the Erdős–Rényi model as "the" random graph, with the implication that real networks are perturbations around this baseline. This framing inverts the actual situation: real networks are not perturbed Poisson graphs; they are generated by entirely different mechanisms that happen to produce heavy-tailed distributions.
The Poisson baseline is useful as a statistical null hypothesis — a way to ask "how non-random is this network?" — but it is dangerous as a default mental model. A researcher who expects Poisson statistics will systematically underestimate the importance of hubs, the vulnerability to targeted attack, the speed of epidemic spread, and the efficiency of decentralized search. The Poisson brain sees a homogeneous world; the heavy-tailed brain sees a world of hubs, bottlenecks, and cascading failures.
The Poisson distribution is the ghost of randomness past — a mathematical idealization that dominated network theory for half a century and now haunts it as a pedagogical mistake. It is not that real networks are "non-Poisson"; it is that the very category of "Poisson network" describes a class of systems that barely exists outside mathematics textbooks. The distribution's elegance is its trap: it is so easy to work with that researchers forget to ask whether the world it describes has any members.