The Fermionic Network

Network models in hyperbolic space reproduce the fundamental structural properties of real-world networks: sparsity, small-world property, heterogeneous degree distributions, high clustering, and scale invariance under renormalization. These models have been developed across different papers with different derivations. This review consolidates them within a single framework: maximum-entropy statistical mechanics with fermionic constraints.

The fermionic analogy is exact, not metaphorical. Network links are binary — present or absent — like fermion occupation numbers. The maximum-entropy principle, subject to constraints imposed by observed network properties, yields a probability distribution over graphs. The resulting ensemble is the least-biased prediction consistent with the observations, in the same sense that the canonical ensemble is the least-biased prediction consistent with a given average energy.

The hyperbolic geometry enters through the constraint structure. Nodes are embedded in hyperbolic space, and the connection probability between any two nodes depends on their hyperbolic distance. The curvature of the space naturally produces the heterogeneous degree distributions and hierarchical organization observed in real networks — properties that require fine-tuning in Euclidean models but emerge generically from hyperbolic geometry.

The framework reveals an anomalous phase transition, dependent on temperature, between a geometric phase (where the hyperbolic embedding meaningfully structures the network) and a non-geometric phase (where the spatial structure is irrelevant). This transition is not a failure of the model but a prediction: some networks have latent geometry, and some don’t, and the temperature parameter determines which regime the system occupies.

The least-biased model is the one that assumes nothing beyond what the data requires. For networks, that model turns out to be hyperbolic.