Contribution
Before this paper, the hyperbolic account of complex networks (Krioukov et al., LIT-865) put the geometry first: nodes are placed in a hyperbolic space and links follow from distance. This paper builds a growth model in which the geometry comes last. Simplicial complexes grown by a purely combinatorial attachment rule have the usual network signatures together (small world, high clustering, high modularity, and scale-free degrees above a dimension threshold), with the degree law derived exactly, and their natural drawing with equal links is a tiling of hyperbolic d-space by ideal simplices. It also extends the growing-network family (Barabási–Albert, random recursive trees) to higher dimension, with the flavor s and dimension d together setting the degree exponent.
Key insight
Glue each new simplex to exactly one face of the old complex and the simplices form a tree. A tree’s ball of radius D holds e^D nodes, which no flat space can hold with equal edges, and a tree of ideal simplices is exactly what tiles hyperbolic space. So “hyperbolic geometry emerges” here because “a tree of equal simplices” is a tiling of H^d; the growth rule need only make the dual a tree. What the rule adds is the statistics: which faces are chosen, and so the degrees, clustering and communities.
Assumptions
- Growth by single-face gluing. Each step adds one node and one d-simplex, glued along one existing (d − 1)-face; nothing is ever removed or rewired. This is what makes the dual a tree.
- Attachment with probability Π_α = (1 + s n_α)/Σ(1 + s n_α′), s ∈ {−1, 0, 1}, n_α the number of simplices on face α minus one; with fitness, Π_α ∝ η_α(1 + s n_α), η_α = e^{−βε_α}, ε_α the sum of node energies drawn from g(ε).
- Equal link lengths. The geometric conclusion assumes every link of the complex has the same length. The paper calls this a “proto-geometry” and says that without it the curvature is not determined.
- The embedding is chosen, not derived: ideal simplices in the Poincaré ball, initial nodes on the unit sphere with Σ r_i = 0, each new node at r_i = Σ_{j⊂α} r_j / |Σ_{j⊂α} r_j|. Its uniqueness as the space-filling embedding is asserted.
- Mean-field rates. The degree distribution is the stationary solution of the master equation for the average number of nodes of degree k, in the large-t limit.
Key results
- Degree rate (Eq. 14). A node of degree k gains a link at rate m̃(k) = [d + (d − 1 + s)(k − d)]/((d + s)t), for (d, s) ≠ (1, −1).
- Degree distribution (Eqs. 16–19). For d + s = 1, P(k) =
(d/(d+1))^{k−d}/(d+1), k ≥ d. For d + s > 1, P(k) = [(d+s)/(2d+s)]
Γ[1 + (2d+s)/(d+s−1)]/Γ[d/(d+s−1)] · Γ[k − d + d/(d+s−1)]/Γ[k − d + 1
- (2d+s)/(d+s−1)], with tail exponent γ = 2 + 1/(d + s − 1) ≤ 3. So the complex is scale-free exactly when d > 1 − s. For d + s = 0 it is a chain. Checked against single runs of N = 10⁵ (Fig. 2).
- Clustering and modularity (Table I, N = 10⁴, 20 realisations, generalised Louvain). Modularity 0.97/0.94/0.90 for d = 2 and 0.91/0.85/0.80 for d = 3 (s = −1/0/1); clustering 0.65/0.74/0.79 and 0.77/0.81/0.84. Modularity falls and clustering rises with s.
- Small world. Stated for every (d, s) except the chain, N ≃ e^D; not shown with data in the text.
- Hyperbolic embedding. The dual is a tree (immediate from the rule). The ideal-simplex drawing in the Poincaré ball is given (Eqs. 2–4, Fig. 1); for d = 3 the projection on the boundary is a random Apollonian network, and for s = −1, d = 3 the complexes are stacked polytopes whose symmetry group is a noncompact discrete subgroup of SO(3,1) (cited, refs. 51–54).
- Geometric transition with fitness (Figs. 3–4). For s = −1, d = 2, 3, as β grows the norm R of the mean node position tends to 1 and the spread σ shrinks with N: the complex grows in one preferred direction. For d = 3 the maximum degree peaks at the transition, for d = 2 it does not. Similar transitions are reported, not shown, for s = 0, 1. Simulation only (N = 2500–10⁴, 500 realisations).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | The degree distribution of a growing simplicial complex of dimension d and flavor s is geometric for d + s = 1 and power-law with γ = 2 + 1/(d + s − 1) for d + s > 1 | strong (derivation, mean field) | Methods, Eqs. 14–19; Fig. 2 |
| C2 | For d ≥ 2 these complexes have high clustering and high modularity, varying with d and s | moderate (simulation) | Table I |
| C3 | The dual of the complex is a tree | strong (immediate from the rule) | Results |
| C4 | With equal link lengths, the complex’s natural geometry is hyperbolic d-space, filled by ideal simplices | weak (a construction and an argument) | Eqs. 2–4, Fig. 1; small-world argument; no proof of uniqueness or of hyperbolicity of the graph metric |
| C5 | The hyperbolicity depends on the equal-length assumption; with unequal lengths the complex embeds in a sphere | moderate (stated with a one-line construction) | Results, first observation |
| C6 | Hyperbolic network geometry can arise from combinatorial growth that ignores any hidden metric | weak to moderate | follows from C3–C4, conditional on C5’s assumption |
| C7 | Face fitness drives a phase transition to directional growth in the hyperbolic embedding | moderate (simulation, finite-size scaling by eye) | Figs. 3–4 |
Method
A growth process and its mean-field master equation for degrees; simulation for clustering, modularity (generalised Louvain, a lower bound on maximum modularity) and the fitness transition; a geometric construction for the embedding.
Concepts
- flavor s: the parameter in Π_α ∝ 1 + s n_α; −1 forces a discrete manifold (each face on at most two simplices), 0 is uniform, 1 is preferential attachment to faces.
- incidence number n_α: the number of d-simplices on the (d − 1)-face α, minus one.
- dual network: simplices as nodes, linked when they share a (d − 1)-face; a tree for these complexes.
- ideal simplex: a simplex of the Poincaré ball with every vertex on the boundary sphere, so all its edges are equal (and infinite).
- emergent geometry: here, geometry as the outcome of network dynamics rather than as the space in which the dynamics happens.
Connections
The paper sets itself against the hidden-metric models of Krioukov et al. (LIT-865) and their “popularity versus similarity” growth model: those place nodes in a hyperbolic space and link by distance, so the geometry causes the network. Here the geometry is read off the network. The degree-rate equation is the standard master-equation treatment of growing networks (Dorogovtsev and Mendes; Krapivsky, Redner and Leyvraz), and the fitness is Bianconi and Barabási’s. The model is the authors’ “network geometry with flavor” of refs. 9–10, of which this paper is the hyperbolic reading. The Battiston et al. review (LIT-091, NOTE-057) cites this school for the spectral dimension of the complexes. The quantum-gravity motivation (causal dynamical triangulations, causal sets, quantum graphity) is stated, not developed.
Bearing on the record
- THEORY-188 (Krioukov et al.). The two are converses in framing only. Krioukov et al. derive network statistics from a given hyperbolic geometry; this paper derives network statistics from growth and then supplies a hyperbolic drawing. It does not show that its complexes have the statistics of the hyperbolic random graph, nor fit them to it, so it neither supports nor challenges THEORY-188’s mechanism. Its power-law exponents, γ = 2 + 1/(d + s − 1), lie in (2, 3], the range the hyperbolic model also reaches, but by a different route.
- THEORY-196 (Sala et al.) holds that trees embed in the hyperbolic plane with distortion near 1. This paper is a higher-order case of the same fact: a complex whose dual is a tree has a hyperbolic drawing. Neither paper says the equal-link drawing here has bounded distortion; in the ideal drawing the distances are infinite, so it is a tiling, not a metric embedding of the graph.
- QUESTION-025. No co-occurrence, no attributes, no embedding learned from data, so it does not answer the question. It offers one thing: a construction in which a node’s position is the normalised sum of its parent face’s positions, Eq. 4, so a descendant’s vector is an additive combination of its ancestors’ directions and the hierarchy is carried in directions on the sphere as well as in the tree of simplices. That is a placement rule chosen by the authors, not a derivation from statistics, and it says nothing about whether a learned embedding does the same. The paper also bears on the record’s reading of “hyperbolic geometry as evidence of hierarchy”: here the curvature is set by the equal-length convention, and with unequal lengths the same combinatorial hierarchy fits a sphere, a caution to set beside Yang et al. (LIT-871).
- THEORY-195 (random hierarchy model) is a different generative hierarchy, of symbols rather than simplices; the two share only the tree.
- No instruction for machine-learning practice; nothing for the anthology.
The reading is the source of THEORY-193.
Limitations
- Hyperbolicity is not measured on the graph: no Gromov δ, no curvature (Ollivier or Forman, both cited), no distortion of a finite-length embedding. It is a drawing plus the small-world argument, which the paper admits is not sufficient.
- The ideal-simplex embedding puts every node at infinite distance from every other, and for s = 0, 1 maps sibling nodes to one point, so it is a picture of the dual tree more than a metric for the nodes.
- “Only one embedding can fill the entire space” and the link to Farey sequences are asserted without proof.
- The fitness transition is from simulations at N ≤ 10⁴; its order, its critical β, and the analytic link to the authors’ earlier Bose–Einstein-type transitions are not given.
- The real-network signatures are compared qualitatively (“as most complex networks”); no real network is fitted.
Open questions
- Is the 1-skeleton’s graph metric Gromov-hyperbolic with δ bounded uniformly in N, for each flavor? That would make the hyperbolicity a property of the network rather than of the chosen drawing.
- Does the equal-length complex admit a finite-length embedding in H^d of bounded distortion, and how does the distortion scale with d and s?
- Is the fitness transition the face-level analogue of Bose–Einstein condensation in the fitness model, and what is its order?