nucleation

LIT-869: Multiscale unfolding of real networks by geometric renormalization

3 min read

Properties4
StatusActive — read 2026-10-09 (NOTE-673); worth reading as the paper that defined a renormalization group for networks from a hidden metric space rather than from shortest paths. Nodes embedded on the S¹ circle are merged in blocks of r angular neighbours, supernodes linked if any of their members were, and the model maps approximately back into itself: hidden degree κ′ = (Σκ^β)^{1/β}, β unchanged, μ and R divided by r. So each layer is a smaller network of the same ensemble with a larger average degree, ⟨k⟩ growing as r^ν with ν = 2/(γ−1) − 1 or 2/β − 1, whichever dominates. Six real networks (Internet, airports, metabolism, proteome, chord transitions, word adjacency) show degree, neighbour-degree and clustering curves that collapse across layers and keep their communities, which the paper reads as evidence that hidden metric spaces underlie real networks. Pruning a layer back to the original density gives small replicas on which Ising, SIS and Kuramoto curves match the original, and greedy routing through the coarser layers raises its success rate. The renormalizability is a first-order approximation, and the collapse is shown by eye.
Tagsnetwork-science, complex-systems
Cited as “paper” byNOTE-673 — Multiscale unfolding of real networks by geometric renormalization
Cited as “source” byTHEORY-190 — Under geometric coarse-graining of the S¹ hidden-metric network model, which merges blocks of r angular neighbours into supernodes linked if any of their members were, the supernodes' link probability approaches the S¹ form with β kept, κ′ = (Σκ^β)^{1/β}, μ′ = μ/r and R′ = R/r, its ratio to that form tending to 1 as their separation grows large against the block width and the coupling μκ′κ″; between nearby supernodes it exceeds that form, THEORY-799 — Under geometric coarse-graining of the S¹ hidden-metric network model the average degree flows as r^ν, with ν = 2/(γ − 1) − 1 when the degree exponent γ sets it and 2/β − 1 when β does, so the coarse-grained networks run toward a complete graph for γ < 3 or β < 2 and toward a ring otherwise, THEORY-803 — Real complex networks embedded in the hyperbolic plane are self-similar under geometric renormalization, their coarse-grained layers keeping the degree distribution, clustering and degree–degree correlations of the original, and this self-similarity is evidence that a hidden metric space underlies them

Graph View

Guillermo García-Pérez, Marián Boguñá and M. Ángeles Serrano (2017), arXiv:1706.00394; Nature Physics 14(6):583–589 (2018), DOI-10.1038/s41567-018-0072-5

Key takeaways

  • The transformation. Given a network’s S¹ embedding (angles θ and hidden degrees κ, inferred by maximum likelihood), cut the circle into blocks of r consecutive nodes, merge each block into a supernode, and link two supernodes if any member of one was linked to any member of the other. Iterating gives a “multiscale shell” of layers, each r times smaller, at most O(log N) of them. Coarse-graining composes (r₁ then r₂ is r₁r₂: an abelian semigroup).
  • The S¹ model is renormalizable, to first order. With supernode hidden degree κ′ = (Σ κ_j^β)^{1/β}, angle a κ^β-weighted power mean of the members’ angles, μ′ = μ/r, R′ = R/r and β′ = β, the probability that two supernodes are linked keeps the S¹ form 1/(1 + χ^β). The derivation keeps only the first term of an expansion and treats all member pairs of two blocks as equidistant. A synthetic network of about 225,000 nodes keeps the connection curve over eight layers (Fig. 3A).
  • Phases of the flow. The power-law exponent γ of hidden degrees is kept if (γ−1)/2 < β. The average degree scales as ⟨k⟩′ = r^ν⟨k⟩, with ν = 2/(γ−1) − 1 when (γ−1)/β ≤ 1 and ν = 2/β − 1 when 1 ≤ (γ−1)/β < 2, in a power-law approximation the paper says is “not true in general”. ν > 0 (γ < 3 or β < 2) flows to a fully connected graph; ν < 0 (γ > 3 and β > 2) to a one-dimensional ring that loses the small world. All six real networks lie in the first phase.
  • Self-similarity of real networks. Rescaled by their average degree, the degree distributions, neighbour-degree curves and clustering spectra of the layers collapse, and Louvain communities of a layer, mapped back, stay high in modularity and in mutual information with those of the original. The paper takes this as new evidence that hidden metric spaces underlie real networks.
  • Statistical mechanics. In the ℍ² form, links are non-interacting fermions with energy half the hyperbolic distance; the partition function factors as Z = ζ^{N/2} Z′, with the intra-supernode links in ζ and Z′ the partition function of the renormalized layer.
  • Uses. “Mini-me” replicas, a layer pruned back to the original average degree by lowering μ, reproduce the order-parameter curves of Ising, SIS and Kuramoto dynamics at every scale tried. Greedy routing that plans in the coarsest layer where source and target differ and descends to physical neighbours succeeds more often than single-layer routing, at slightly longer paths.

Standing in the record

Filed at the owner’s request on 2026-10-09, in the second part of the batch on hierarchy and hyperbolic geometry that followed QUESTION-025. It is the direct continuation of Krioukov et al. (LIT-865, THEORY-188), from the same group: it takes the S¹ ⇄ ℍ² model that paper equated and asks what it does under a change of scale. Its own account of the renormalizability is THEORY-190.

The owner filed it beside Koch-Janusz and Ringel’s Mutual information, neural networks and the renormalization group (Nature Physics, DOI 10.1038/s41567-018-0081-4), which another reader is filing. The two are not companion papers. They are neighbours in the same issue (14(6), pp. 578–582 and 583–589), both about the renormalization group, from different groups, on different systems (lattice spin models with a learned coarse-graining; networks with a geometric one), and in their arXiv versions neither cites the other. The issue’s contents list no News & Views pairing them.

Read on 2026-10-09 (NOTE-673).

Backlinks