Dmitri Krioukov, Fragkiskos Papadopoulos, Maksim Kitsak, Amin Vahdat and Marián Boguñá (2010), Physical Review E 82, 036106 — ARXIV-1006.5169, DOI-10.1103/PhysRevE.82.036106
Key takeaways
- Hyperbolic space as a continuous tree. In the hyperbolic plane of curvature −ζ², circle length and disk area grow as e^{ζr}, as the node count of a b-ary tree grows as b^r; with ζ = ln b the two are metrically alike. The paper’s rationale is that the similarity of nodes is organised by an approximately tree-like hierarchy of groups, so the space of similarities is negatively curved. That step is an argument, backed by a cited fact (overlapping disks in ℝ² correspond to points of the upper half-space H³, with nested disks forming the tree and partial overlaps adding cycles that keep it hyperbolic), not a derivation.
- Geometry gives heterogeneity. N nodes in a disk of radius R ~ ln N, joined when their distance is at most R: a node at radius r has expected degree ∝ e^{−ζr/2}, the node density grows as e^{αr}, and the two exponentials give P(k) ~ k^{−γ} with γ = 2α/ζ + 1 when α/ζ ≥ ½ (γ = 3 for uniform density, α = ζ), and γ = 2 below that. Only α/ζ matters: branching factor against curvature.
- Heterogeneity gives an effective geometry (§V). The S¹ model (nodes on a circle, power-law hidden degrees κ, connection probability a function of distance over κκ′) becomes the H² model under κ = κ₀e^{ζ(R−r)/2}: degree is radial depth, and the circle is the boundary at infinity. This is an equivalence of two model ensembles, in the large-distance approximation, given that the network already has a similarity metric; it is not a test that any real network does.
- Statistical mechanics. With p(x) = 1/(e^{β(ζ/2)(x−R)} + 1), edges are non-interacting fermions with energy the hyperbolic distance, chemical potential R and temperature T = 1/β; the link fields of the exponential random-graph formalism are linear in distance. At T = 0 the step function returns and clustering is largest; clustering falls nearly linearly to zero at T = 1, where the chemical potential diverges; for T > 1 clustering vanishes in large networks.
- Degenerate limits. Letting ζ, T → ∞ with ζ/T fixed kills the angular term of the distance, x_ij = r_i + r_j, so link probabilities depend only on the product of expected degrees: the configuration model. Heating with α and ζ fixed gives Erdős–Rényi graphs. The metric structure is the angular term.
- Navigation. Greedy forwarding toward the destination’s hyperbolic coordinates, with no global knowledge, reaches 99.92% (original) and 99.99% (modified) of destinations at γ = 2.1, T = 0 (N = 10⁴, k̄ = 6.5), every greedy path being a shortest path; it stays above 99% with 10% of links removed, worsens with γ and T, and fails in the configuration model (≤ 40%) and Erdős–Rényi graphs (≈ 0.2%). Simulation only.
Standing in the record
Filed on 2026-10-09 at the owner’s request, in a batch of six works on hierarchy and hyperbolic geometry asked for right after the record opened QUESTION-025 (with Cagnetta et al.’s random hierarchy model, Sala et al. 2018, Lin et al. 2023, Zhang et al. 2023 and Yang et al. 2023).
Read the same day (NOTE-680). It is the first work in the record on hyperbolic geometry as the latent space of a network; the nearest holdings are the discrete curvatures of LIT-004 and LIT-020. The reading is the source of THEORY-188. It does not answer QUESTION-025, which asks about linear attribute directions in co-occurrence embeddings, but it supplies a generative model in which hierarchy enters link probabilities through a curved distance rather than through independent factors; the NOTE says how far that reaches.