Hyperbolic Geometry in Complex Networks

• Hyperbolic geometry of complex networks is a framework where hidden negative curvature underpins hierarchical organization and the emergence of scale-free, clustered structures.
• The model utilizes latent popularity–similarity paradigms and random hyperbolic graphs to accurately mirror real-world phenomena such as strong clustering and ultra-small path lengths.
• Inference methods like Mercator enable the extraction of latent hyperbolic coordinates from network topology, aiding practical visualization and efficient routing.

Hyperbolic geometry of complex networks studies network structure and function under the assumption that many networks are organized by hidden negatively curved spaces rather than by shortest-path geometry alone. In this framework, hyperbolic spaces are effective because their exponential expansion mirrors hierarchical and tree-like organization, so heterogeneous degree distributions, strong clustering, self-similarity, modular structure, and navigability can emerge as consequences of geometry rather than as independent design choices. A central result of the field is the bidirectional relation between topology and geometry: hyperbolic models naturally generate scale-free clustered networks, while networks with metric structure and heterogeneous degree distributions admit an effective hyperbolic interpretation (Krioukov et al., 2010, Boguna et al., 2020).

1. Latent negative curvature and the popularity–similarity paradigm

A foundational intuition is that hyperbolic space and trees share exponential expansion. In the hyperbolic plane, disk area grows exponentially with radius, while in a tree the number of nodes at depth $r$ also grows exponentially. This makes negative curvature a natural substrate for hierarchical and heterogeneous organization. In the geometric framework of Krioukov, Papadopoulos, Kitsak, Vahdat, and Boguñá, the hyperbolic distance between nodes governs link likelihood, and heterogeneous degree distributions and strong clustering emerge as simple reflections of negative curvature and the metric property of the underlying geometry (Krioukov et al., 2010).

The most widely used latent-space formulation is the Popularity$\times$Similarity $\mathbb{S}^1/\mathbb{H}^2$ model. Each node $i$ is assigned a hidden degree $\kappa_i$ encoding popularity and an angular coordinate $\theta_i$ encoding similarity. In the $\mathbb{S}^1$ representation, nodes $i$ and $j$ connect with probability

$$p_{ij}=\frac{1}{1+\left(\frac{d_{ij}}{\mu \kappa_i\kappa_j}\right)^\beta},$$

where $\times$0 is arc length, $\times$1 controls average degree, and $\times$2 controls clustering. The equivalent hyperbolic mapping assigns radial coordinate

$\times$3

so that popularity becomes radius and similarity remains angular separation. In this mapping, the approximate hyperbolic distance is

$\times$4

and the connection probability becomes a Fermi-Dirac-like function of $\times$5. This model can accommodate arbitrary degree distributions and reproduces pivotal properties of real networks, including self-similarity patterns (García-Pérez et al., 2019).

This latent formulation also supplies a converse interpretation. If a network has some metric structure and its degree distribution is heterogeneous, then the network has an effective hyperbolic geometry underneath. The geometric network ensemble then subsumes the standard configuration model and classical random graphs as limiting cases with degenerate geometric structures (Krioukov et al., 2010).

2. Random hyperbolic graphs, clustering, and ultra-small distances

The random hyperbolic graph model realizes these ideas generatively. In a disk of hyperbolic radius $\times$6, nodes are placed with angular coordinate uniform on $\times$7 and radial coordinate sampled from

$\times$8

or, in the more general curvature-$\times$9 formulation,

$\mathbb{S}^1/\mathbb{H}^2$0

Edges are then added either by a hard threshold, $\mathbb{S}^1/\mathbb{H}^2$1, or probabilistically through

$\mathbb{S}^1/\mathbb{H}^2$2

with inverse temperature $\mathbb{S}^1/\mathbb{H}^2$3. For finite $\mathbb{S}^1/\mathbb{H}^2$4, the model yields power-law degree distributions and strong clustering; as $\mathbb{S}^1/\mathbb{H}^2$5, it reduces to the soft configuration model and ultimately to Erdős-Rényi random graphs (Aldecoa et al., 2015, Krioukov et al., 2010).

A mathematically precise theory of clustering was developed for hyperbolic random graphs. The global clustering coefficient

$\mathbb{S}^1/\mathbb{H}^2$6

admits an explicit limiting formula in terms of model parameters. For $\mathbb{S}^1/\mathbb{H}^2$7 and $\mathbb{S}^1/\mathbb{H}^2$8, the coefficient converges in probability to a strictly positive limit $\mathbb{S}^1/\mathbb{H}^2$9; for $i$0, it converges to $i$1. This establishes that clustering can remain bounded away from zero in a sparse scale-free regime and that it is tunable independently of degree exponent and average degree (Candellero et al., 2013).

Hyperbolic geometry also yields a distinctive path-length regime. In the geometric model studied by Abdullah, Bode, and Fountoulakis, when $i$2 the graph distance between two uniformly chosen connected vertices is doubly logarithmic in $i$3:

$i$4

The same asymptotic constant appears in the analogous Chung-Lu regime, but the hyperbolic model retains geometric clustering while Chung-Lu is locally treelike. In this sense, the hyperbolic model is an ultra-small world with clustering rather than merely a heavy-tailed sparse graph (Abdullah et al., 2015).

Beyond local curvature in the Riemannian sense, nonassociative geometry introduces elementary holonomy as a nonlocal curvature measure. In that framework, nonlocal curvature controls the small-world property and community formation, and a model with nonzero holonomy explains Internet connectance anomalies, including local minima and maxima in connection probability, that the featureless sigmoidal $i$5 hyperbolic fit cannot capture (Nesterov et al., 2018).

3. Inference of hidden coordinates and faithful embeddings

A major research direction concerns inverse problems: inferring latent hyperbolic coordinates from observed topology. Mercator is a reliable embedding method that assumes network structure is described by the Popularity$i$6Similarity $i$7 model and combines machine learning with maximum likelihood. In fast mode, it performs a model-adjusted dimensional reduction based on modified Laplacian Eigenmaps, computes the first two nontrivial eigenvectors, and assigns angular coordinates through $i$8. In refined mode, it uses the fast embedding as initialization for maximum-likelihood optimization with node updates guided by onion decomposition and with objective

$i$9

Mercator systematically infers angular positions, hidden degrees, and global model parameters, and it can embed networks with arbitrary degree distributions (García-Pérez et al., 2019).

The empirical validation of embedding methods has two complementary forms. One is internal statistical congruence: synthetic networks generated from inferred parameters reproduce empirical degree distributions, clustering spectra, neighbor-degree structure, assortativity, and related observables. In Mercator, this was illustrated on synthetic benchmarks and on real systems including the world airports network, Internet AS-level topology, metabolic networks, and trade networks; in the airports network, angular coordinates correlate with geography even though geography is not used as input (García-Pérez et al., 2019).

The second form is correspondence with external metadata. Coalescent embedding provides an unsupervised, topology-only family of dimensionality-reduction methods that map networks into hyperbolic space and recover latent angular structure. On human structural connectomes, this methodology reconstructed hemispheric segregation, anterior–central–posterior patterning, and lobar organization from connectivity alone, and it detected geometrical pathological changes in Parkinson's Disease through hyperbolic markers such as mean hyperbolic distance and hyperbolic shortest-path measures (Cacciola et al., 2017).

These inference results support a strong but qualified interpretation of latent geometry. They show that hidden hyperbolic structure can be statistically reconstructed from topology alone, yet the quality of the reconstruction depends on model congruence and on whether the observed network is well described by a popularity–similarity mechanism rather than by a different source of clustering or modularity (García-Pérez et al., 2019, Cacciola et al., 2017).

4. Hyperbolicity as a metric property and why triangles are not enough

Latent hyperbolic geometry is distinct from Gromov-style hyperbolicity of the observed shortest-path metric, but the two are often studied together. For graphs, $\kappa_i$0-hyperbolicity can be defined through slim triangles or the four-point condition. Given four points $\kappa_i$1, let $\kappa_i$2, $\kappa_i$3, and $\kappa_i$4 be the largest, middle, and smallest of the three opposite-pair distance sums; the graph is $\kappa_i$5-hyperbolic if

$\kappa_i$6

Through statistical curvature plots on many real data sets, Kennedy, Narayan, and Saniee found that communication and social networks are strongly hyperbolic, with sampled $\kappa_i$7 values small compared with graph diameter, whereas road networks are not hyperbolic and behave more like square lattices. They also showed that renormalization preserves and often amplifies hyperbolicity, allowing detection on much smaller coarse-grained graphs (Kennedy et al., 2013).

Negative curvature has direct topological implications. For biological and social networks, Albert, DasGupta, and Mobasheri adapted the combinatorial four-point definition to finite parameterized networks and showed that many such networks are hyperbolic. They derived bounds on the distance between shortest or approximately shortest paths; for example, if $\kappa_i$8 and $\kappa_i$9 are two shortest paths between the same endpoints, then for every $\theta_i$0 there exists $\theta_i$1 with

$\theta_i$2

This leads to two network-level consequences emphasized in the paper: crosstalk between long pathways in biological regulation and the existence of central, influential neighborhoods that control many efficient routes in biological and social systems (Albert et al., 2014).

A frequent misconception is that many triangles or a high clustering coefficient are sufficient evidence of latent hyperbolic geometry. Recent work shows that this is not generally true. In heavy-tailed regimes, triangle counts in geometric and non-geometric inhomogeneous random graphs can be asymptotically indistinguishable, and average clustering can remain misleadingly high even without geometry. To address this, a weighted triangle statistic was introduced:

$\theta_i$3

In the non-geometric IRG model, $\theta_i$4 with high probability, whereas in the geometric GIRG model, $\theta_i$5 with high probability. The point is not that triangles are irrelevant, but that they must be weighted so that degree heterogeneity does not masquerade as geometry (Michielan et al., 2022).

5. Communities, navigation, and geometric visualization

Hyperbolic embeddings and community structure are closely related but not identical. A systematic comparison of real and synthetic networks showed that community structure can be viewed as a coarse version of hidden-space hyperbolic embedding: radial coordinates correspond to popularity, angular coordinates to similarity, while in a degree-corrected stochastic block model the analogous pair is degree and community label. Nodes in the same community tend to have similar angular coordinates, and a partition angular coherence can quantify this alignment. On that basis, results first expressed in terms of continuous hyperbolic coordinates can be reinterpreted using only mesoscopic community structure (Faqeeh et al., 2018).

This analogy has algorithmic consequences. In multiplex networks, robustness under targeted attacks depends on correlations between layers: when community assignments are strongly correlated across layers, percolation transitions are smooth or continuous, while destruction of these correlations yields abrupt breakdowns. The same paper proposed community-based greedy routing in which a packet at node $\theta_i$6 is forwarded to the neighbor $\theta_i$7 minimizing

$\theta_i$8

where $\theta_i$9 is shortest-path distance in the supernetwork of communities. Its success rates are close to those of hyperbolic greedy routing as long as communities are not excessively large (Faqeeh et al., 2018).

The original hyperbolic routing picture remains central. In the geometric framework of Krioukov and collaborators, greedy forwarding sends a packet to the neighbor closest in hyperbolic space to the destination. In networks with strongest heterogeneity and clustering, targeted transport processes without global topology knowledge are maximally efficient according to all efficiency measures, and this efficiency is remarkably robust with respect to even catastrophic disturbances and damages to the network structure (Krioukov et al., 2010).

Visualization methods exploit the same geometry. Browser-based systems now implement three hyperbolic visualization strategies: inverse projections, generalized force-directed algorithms, and hyperbolic multidimensional scaling (H-MDS). Hyperbolic visualization provides a natural focus+context effect in the Poincaré disk because peripheral regions are exponentially compressed. Among these methods, H-MDS yields lower distortion than Euclidean MDS for tree-like and hierarchical graphs, whereas Euclidean spaces perform better for lattice- and cycle-like graphs; inverse projection is the most scalable, but it improves visualization rather than embedding fidelity (Miller et al., 2022).

6. Scalable generation, emergent geometry, and current methodological debates

The practical study of hyperbolic network geometry depends on scalable generators. A first major advance used a polar quadtree adapted to the Poincaré disk to generate random hyperbolic graphs in

$\mathbb{S}^1$0

time with high probability, improving earlier quadratic approaches by at least two orders of magnitude and enabling graphs with billions of edges to be generated in a few minutes (Looz et al., 2015). A later generator replaced quadtrees by concentric ring-shaped slabs, achieved empirical running time $\mathbb{S}^1$1, reported speedup factors of $\mathbb{S}^1$2–$\mathbb{S}^1$3 over the best previous implementation, and added a dynamic extension in which node motion preserves the point position probabilities at each step; one billion edges can then be generated in under one minute on a shared-memory workstation (Looz et al., 2016).

Hyperbolic geometry need not be imposed a priori. In growing simplicial complexes, purely combinatorial nonequilibrium growth rules can produce emergent hyperbolic network geometry spontaneously. These models generate small-world behavior, scale-free degree distributions, high clustering, and community structure, with dimensionality and flavor controlling the degree exponent

$\mathbb{S}^1$4

When faces have heterogeneous fitness, the growing simplicial complex undergoes phase transitions reflected in changes of network geometry, including symmetry breaking and directional condensation in the Poincaré ball representation (Bianconi et al., 2016).

The same geometric bias now appears in machine learning. Hyperbolic Attention Networks impose hyperbolic geometry on neural activations and re-express attention through operations in the hyperboloid and Klein models, improving generalization on neural machine translation, graph learning, and visual question answering while keeping neural representations compact (Gulcehre et al., 2018). Yet recent work on Hyperbolic Graph Neural Networks argues that graph hyperbolicity alone is not enough: the decisive criterion is geometry–task alignment. On synthetic regression problems that require preserving graph metric structure, HGNNs recover lower-distortion representations than Euclidean models; on real tasks, link prediction is geometry-aligned, whereas node classification is not, so the hyperbolic advantage vanishes outside aligned settings (Naddeo et al., 2 Feb 2026).

This contemporary debate extends to classical network inference. In link prediction, the Cannistraci-Hebb local automaton was found to be the best predictor on synthetic networks generated by the Popularity-Similarity-Optimization model and to degrade on non-hyperbolic synthetic networks, suggesting that some successful local rules exploit the rise of hyperbolic geometry even when no explicit embedding is used (Muscoloni et al., 2017). A plausible implication is that hyperbolic geometry functions both as a generative principle and as a benchmark for evaluating when topological heuristics, latent-space inference, and learning architectures are actually solving the same structural problem.