REvIEWS

Network geometry
Marián Boguñá 1,2
, Ivan Bonamassa3, Manlio De Domenico 4 ✉, Shlomo Havlin3,
Dmitri Krioukov 5,6,7,8
and M. Ángeles Serrano1,2,9
Abstract | Networks are finite metric spaces, with distances defined by the shortest paths
between nodes. However, this is not the only form of network geometry: two others are the
geometry of latent spaces underlying many networks and the effective geometry induced by
dynamical processes in networks. These three approaches to network geometry are intimately
related, and all three of them have been found to be exceptionally efficient in discovering
fractality, scale invariance, self-​similarity and other forms of fundamental symmetries in networks.
Network geometry is also of great use in a variety of practical applications, from understanding
how the brain works to routing in the Internet. We review the most important theoretical
and practical developments dealing with these approaches to network geometry and offer
perspectives on future research directions and challenges in this frontier in the study of complexity.

Many existing analytical and computational tools for the their navigability25–28, and their community29–33 and
analysis of complex networks emerged from classical multiscale34 structures. Because the group of symme-
methods in statistical physics1. Over the past two dec- tries of hyperbolic spaces is isomorphic to the Lorentz
ades, these tools have proved essential for constructing group, the latent hyperbolicity of networks was advo-
models capable of reproducing the structural properties cated to explain not only their structural self-​similarity
of many real-​world networks2–4. They have also been but also the dynamical laws of their growth35–37. The
1
Departament de Física used to quantify the importance of network structural isomorphism also establishes certain duality relations38
de la Matèria Condensada, properties for collective and critical phenomena5–8. reminiscent of the anti-​de Sitter/conformal field theory
Universitat de Barcelona,
Other fundamental insights have come from a variety correspondence in theoretical physics.
Barcelona, Spain.
of complementary approaches9–13. One such approach is In light of these advances, it is not surprising that
2
Universitat de Barcelona
Institute of Complex Systems
geometry, the focus of this Review Article. the geometric approach led to many practical appli-
(UBICS), Universitat de The first evidence that complex networks possess cations and theoretical insights. For example, in
Barcelona, Barcelona, Spain. some non-​trivial geometric properties appeared with the context of information or epidemic spreading, the
3
Department of Physics, the discovery of their self-​similarity under suitable adoption of transport-​based metrics and of the corre-
Bar-​Ilan University, scale transformations14. Initially, fractal geometry was a sponding diffusion geometries39 is helping explain the
Ramat-​Gan, Israel. major reservoir of methods and ideas. Besides boosting spatiotemporal evolution of network-​driven dynamical
4
CoMuNe Lab, Fondazione the study of transport phenomena in complex media15,16, the processes40,41, opening new research directions42,43 in
Bruno Kessler, Povo, Italy.
fractal geometric paradigm led to the definition of a many neighbouring areas of science.
5
Network Science Institute,
reversible graph-​theoretical renormalization procedure The rapid evolution of research areas related to net-
Northeastern University,
Boston, MA, USA.
that helped researchers classify networks into univer- work geometry has led to cross-​fertilization among
6
Department of Physics,
sality classes17,18 and to better understand the growth many areas of science. This rapid interdisciplinary
Northeastern University, mechanisms underlying their temporal evolution19. progress suggests that now is an apt time to ground a
Boston, MA, USA. Following the lines of this initial success, it was later milestone in network geometry research, from where
7
Department of Mathematics, found20 that network self-​similarity can be explained at to ponder future challenges. We review three major
Northeastern University, a more fundamental level in terms of latent hyperbolic research directions in network geometry: the self-​similar
Boston, MA, USA. geometry21. Latent spaces have been employed for nearly fractal geometry of network structure, the hyperbolic
8
Department of Electrical a century to model homophily in social networks22–24. In geometry of networks’ latent spaces and the geometry
and Computer Engineering,
these models, nodes are positioned in a similarity space. induced by dynamical processes, such as diffusion, in
Northeastern University,
Boston, MA, USA. Connections between them are random, but they are the networks. Distances are all different in the three geom-
ICREA, Barcelona, Spain.
9 more likely the closer the two nodes in the space, that etries, yet intimately related. In fractal geometry, dis-
✉e-​mail: is, more similar nodes are more likely to be connected. This tances are the shortest-​path distances, that is, the hop
[email protected] approach, with the latent space taken to be hyperbolic, is lengths of shortest paths in a network. In latent geome-
https://doi.org/10.1038/ the basis for a unified framework that explains the most tries, distances are the distances between network nodes
s42254-020-00264-4 common structural properties of many real networks20,21, in a latent space. In geometry induced by dynamical

114 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

proved to be a problem that prompted prolific research,

Key points
and its analysis through tools of fractal geometry54 has
• The fractal geometry of networks enables casting the self-​similar symmetries disclosed fundamental insights into the hidden sym-
underlying the organization of complex systems under the three pillars of scaling, metries, renormalization and universality classes of
universality and renormalization. complex networks.
• Latent metric spaces with hyperbolic geometry provide a natural explanation
for the architecture of real complex networks, including small-​worldness, degree Shortest-​path-​distance scaling and dimensions. The
heterogeneity, clustering, community structure, symmetries and navigability.
number ℓ of edges along any shortest path connecting
• Multiscale unfolding of complex networks is possible by means of a geometric two nodes is a well-​defined metric55. It is known as the
renormalization technique that uncovers self-​similarity at different scales.
shortest-​path distance (or, sometimes, chemical dis-
• Network dynamical processes induce kinematic distances that characterize an tance)56,57, and it can be adopted to observe networks
effective geometry of a system’s function, which cannot be obtained by purely at different length scales. The process of zooming out
topological latent geometry.
on networks can be performed analogously to that on
• Network geometry enhances our understanding of complex systems across their ordinary fractals by repeatedly coarse-​graining the
multiple scales of organization and of collective phenomena emerging from
structure through optimal coverings of nodes made
their information exchange.
by non-​overlapping boxes of diameter ℓB (Box 1) 14.
Under this renormalization group (RG) transfor­
processes, distances are the kinematic distances, such mation58,59, a surprising variety of real-​world and syn-
as those emerging from spreading dynamics in a net- thetic networks remain statistically self-​similar, in the
work such as reaction–diffusion or random-​search pro- sense that their degree distribution14 and other mix-
cesses. We discuss open questions and future challenges, ing patterns60 are invariant (Fig. 1a) over the available
relating to both theory and applications of these three length scales.
flavours of network geometry. A primary consequence of the discovery of networks’
Understood broadly, network geometry encompasses self-​similarity was their characterization through a set of
diverse research directions, many of which are as impor- fractal dimensions61 intimately related to their multiscale
tant as the ones covered here, thus deserving separate organization and functioning. In what follows, we review
reviews. Therefore, in a short Review Article like this an essential selection of these network dimensions and
one, most of these directions must unfortunately be their significance.
omitted. The most notable omissions include spatial net- We start from two crucial quantities: the number of
works, quantum gravity, graph curvature, geometrogen- boxes (supernodes) N′ optimally covering a network
esis, graph embedding, topological data analysis, and the degree sequence {k i′}i =1,…, N ′ of the boxes, where
topological graph theory and Gromov’s δ-​hyperbolicity. each k′ corresponds to the degree of the hub contained
However, in the outlook we briefly comment on some in the ‘parent’ box. Performing a shortest-​path-​distance
of these topics that we find particularly important, RG transformation step, R B , these quantities have
promising or intriguing. power-​law scaling (Fig. 1a)
RB RB
Fractal geometry of network structure N → N ′ ∼ ℓ −BdBN , k → k ′ ∼ ℓ −Bdk k . (1)
Many natural objects are best described geometrically as
fractals44, with self-​similar patters on all their available The exponents (dB, dk) are the box and degree dimen-
length scales and well-​defined fractal dimensions. From sions, respectively (Fig. 1b, top and middle panels);
surfaces and snowflakes to turbulent flows45 or critical they characterize the (RG invariant) scale-​free degree
phenomena near phase transitions46,47, this scale-​free distribution through the scaling identity γ = 1 + dB/dk
property is among the basic hallmarks of complexity, (ref.14). Besides unveiling a fundamental connection
reflecting certain laws and organizing principles48,49 between scale-​free and RG-​invariant properties, this
underlying nature’s evolution. fractal analy­sis yields a birds-​eye view over the spec-
The discovery of the scale-​free nature of many real trum of different organization mechanisms that under-
networks has added a new paradigm to this understand- lie self-​similar networks. It does so by identifying two
ing, providing scientists with a versatile proxy to tame limiting cases62,63: fractal structures when dB and dk
complexity on combinatorial grounds50. In this context, are both finite and non-​zero (in biological systems,
scale-​freeness refers to the absence of a characteristic the world wide web (WWW) or social networks, for
number of links k per node (rather than a length scale), example), and small worlds when dB, dk → ∞ (in the
as described by a fat-​tailed degree distribution of the Internet at the router level or synthetic networks, for
form P(k) ∼ k−γ as k ≫ 1, with the scale-​free exponent example), implying an exponential decay in the N and
γ ∈ (2, 3). This scale-​free property hinted at the exist- k distributions (Fig. 1b, top panel) instead of the scaling
ence of some degree of structural invariance under a of equation 1.
suitable rescaling. However, the equally ubiquitous A more comprehensive picture is reached by analy­
small-​world2 property — that is, an average distance sing structural fractality in terms of the profiles of
between nodes growing logarithmically or slower51–53 degree–degree correlations14,64. Fractal networks feature
with the system size N — contextually hindered this a strong ‘hub–hub repulsion’, which leads to disassor-
possibility, implying a diverging Haussdorf dimension. tative structures, with hubs spread uniformly instead
This led to the common belief that networks cannot be of being crumpled in a core as they are in pure small
self-​similar. Fortunately, this apparent contradiction worlds. In light of their RG invariance60, these correlation

NaTure RevIewS | PhySicS volume 3 | February 2021 | 115

Reviews

patterns can be captured by a hub–hub dimension de, symmetries that leave invariant the network mesoscale
defined by the scaling factor E(ℓB) ∼ ℓ −Bde , which is organization. In fact, the RG transform identifies a hier-
proportional to the probability that two RG boxes are archy of modular configurations into which networks
connected through their hubs (Fig. 1b, bottom panel). can be optimally partitioned at increasing length scales.
Scaling arguments analogous to those adopted for the The scale invariance of such a tiling can be quantified
triple (dB, dk, γ) make it possible to relate the correlation via the power-​law scaling Q(ℓB) ∼ ℓ dBM (refs15,65), where
exponent ε > 1 — defined by the RG invariant60 degree– dM is the modular dimension and Q(ℓB) is a modu-
degree distribution P(k, q) ∼ k−γ+1q−ε — to the hub–hub larity factor66,67 maximized by the box covering. The
dimension through the identity ε = 2 + de/dk (ref.60). These value dM = 1 (characterizing regular lattices) is the bor-
results allow the classification of many networks accord- der line between modular structures (dM > 1, typical of
ing to RG-​invariant properties related to their large-​scale biological networks, for instance) and non-​modular
organization (Fig. 1c). ones (dM < 1); networks with lower values of dM are
A final essential feature of the fractal geometry of net- more small-​world-​like. We will see in what follows
works is the existence of a larger group of self-​similar that dM, unlike the other ‘structural’ dimensions, is

Box 1 | Zooming out on complex networks

Shortest-​path-​distance renormalization a Shortest-path-distance b
In the box-​covering technique (see the figure, part a), renormalization
the original network (first column) is tiled with the
minimum number of boxes of diameter ℓB — using ℓB = 1
the maximum excluded mass burning algorithm272,
for instance — which are contracted into supernodes
(second column) and then connected though at least ℓB = 2
one link existed between the ‘parent’ tiles. This coarse
graining is repeatedly applied — albeit for a limited
number of steps due to small-​worldness — until the ℓB = 3
network is reduced to a single ‘ancestral’ node. As an
example, the shortest-​path-​distance renormalization
group can be applied to the entire world wide web, here
with with ℓB = 2 (see the figure, part b). The network’s c Degree-thresholding d
structure remains statistically invariant under the renormalization
renormalization (Fig. 1). 100 100

Degree-​thresholding renormalization

c– (ki/kT)
c– (ki/kT)

Subgraphs are obtained by removing all nodes with 10–1 10–1 Internet
Internet BGP map
degrees below a given threshold kT (see the figure, part c). BGP map randomized
kT = 0 kT = 1
Doing so defines a hierarchy of nested subgraphs that 10–2 10–2
in real complex networks are found to be self-​similar. 10–1 100 101 102 10–1 100 101 102
Applying this procedure to the Internet border gateway ki/<ki(kT)> ki/<ki(kT)>
protocol (BGP) graphs (see the figure, part d, left
column) and its randomized version that preserves
kT = 5 kT = 10 kT = 20 kT = 50 kT = 100
the degree sequence (see the figure, part d, right kT = 2 kT = 3
column) shows that the clustering spectrum (average
clustering coefficient c versus rescaled internal degree
ki/〈ki(kT)〉) exhibits data collapse for the original network e Geometric renormalization f
but not for the randomized network. The nice collapse I=0 I=1 I=2 Human metabolic
of the clustering spectrum for real complex networks network
finds a natural explanation in their underlying
geometry20. Renormalized
layer I = 2
Geometric renormalization
Like shortest-​path renormalization, the geometric
renormalization transformation zooms out by changing
the minimum length scale from that of the original
network to a larger value, this time in the similarity
space34 (see the figure, part e). For each layer l of the
renormalization transformation, first, non-​overlapping
blocks of consecutive nodes are defined along the
l

similarity circle. Second, the blocks are coarse-​grained

into supernodes. Each supernode is then placed within
the angular region defined by the corresponding block structure detected by the Louvain algorithm273. The renormalized network
so that the order of nodes is preserved. Finally, two supernodes are preserves the original community structure despite being four times
connected if any of their constituents were in the precursor layer. smaller.
This procedure has been performed on the hyperbolic embedding of Parts a,b reprinted from ref.14, Springer Nature Limited. Parts c,d adapted with
the human metabolic network and its renormalized layer l = 2 (ref.34; permission from ref.20, American Physical Society. Parts e,f adapted from ref.34,
see the figure, part f). The colours of the nodes correspond to the community Springer Nature Limited.

116 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

a b
100 106
ℓB = 1
10–2 105
γ ℓB = 3 dB
104
ℓB = 5

NB(ℓB)/N
10–4

P(k)
100
k 103
10–6
102
50 k′
10–8 101
0
0 100 200
10 –10
100
100 101 102 103 104 1 10
k ℓB
20

c 2–2 dk

s(ℓB)
4 2–4
Metabolic 2–6
2–8
2 4 8 16 32
3 III ℓB
PPIN (human) 2–1
ε=γ WWW
ε

Homology Taxonomy Random 2–2

PPIN (yeast) γ=3
2 2–3
Internet Guestbook

ε (ℓB)
Citations 2–4 de
(AS) Random
II DIMES γ = 2.5 Cond-mat
I
1 2–5
ε –
= γ Hospitals Blogs
1 IMDb Internet (router) 2–6
2.00 2.25 2.50 2.75 3.00 2–7
γ
2 4 8 16 32
ℓB

d Large-world Middle-world networks Small-world

fractal network (fractal with shortcuts) random network

prew = 0 prew = 0.01 prew = 0.1 prew = 1

e f g

Fractal + λ Unstable
G0 (fractal fixed region
point, α > 2dB) shortcuts
fixed s=2 1 WWW λ=2–s
G' (fractal point s>2
logNb

with shortcuts, Metabolic

α ≤ 2dB) Fractal Unstable
fixed direction
1 2
PPIN s
point s > 2 Navigability IMDb
logb s<2 Complete PHN
graph transition
Stable
G' (pure small-world, small α) Stable fixed Fractal/small-world region
direction point transition
Small-world Fractal
phase phase

directly connected to other dynamical exponents that stability against small perturbations enshrines universal
characterize transport in complex media. features.
A generalization of finite-​size scaling methods17 has
RG flow. Repeatedly applying the shortest-​path-​distance been used to analyse the RG flow beyond the limiting
RG transformation, R B, identifies a flow in the space constraint of a few RG steps, a constraint that arises
of graph configurations. Much like the case of critical because of the networks’ small-​worldness. For a large
phenomena58,59,68, such a flow enables a classification set of real-​world and artificial networks69, a coher-
of network topologies into universality classes. Indeed, ent picture about their stability under the RG flow
scale invariance and self-​similarity are natural symme- emerged after looking at the critical exponents govern-
tries featured by the fixed points of the RG flow, whose ing their fluctuations (γ′) and correlation length (ν)70.

NaTure RevIewS | PhySicS volume 3 | February 2021 | 117

 Reviews

◀ Fig. 1 | Structural self-similarity, renormalization-group flow and universality. the exponent α > 0. These outcomes can be quantified by
a | Renormalization group (RG) invariance of the world wide web (WWW) degree focusing on the renormalized distribution of shortcuts
distribution P(k) ∼ k−γ (γ is the scale-​free exponent) and of its degree sequence (k′ ∼ ℓB−d kk, after one RG step, which, in the formal limit of ℓB → ∞,
see equation 1) for different box diameters ℓB (inset). b | RG scaling factors of box sizes leads to the fixed point equation
NB(ℓB)/N (NB is the number of boxes and N the number of nodes) (top), their degree
sequence s(ℓB) ≡ k′/k (middle) and the hub–hub correlations E(ℓB) (bottom). Fractal p∗ (ℓ) ≡ 1 − lim exp[− C (ℓ)x 2dB/α −1], (2)
networks (such as the WWW, open circles) feature well-​defined dimensions dB, dk and de; x→∞
non-​fractal ones (such as the Internet, filled circles) show an exponential (or faster) decay,
that is, dB, dk → ∞ and de → 0. c | Classification of self-​similar networks in the (γ, ε) plane where C(ℓ) ≡ A2dB/αℓ −2dB and x ≡ A−1(ℓBℓ)α. Equation 2
(ε is the correlation exponent). At the line ε = γ − 1, the network structure is random. Below has three distinct solutions (Fig. 1f), depending on the
the line (region I), hub correlations are stronger than in random networks; above the line, value of the parameter s ≡ α/dB:
hub correlations are weaker than random. The scaling identity ε = 2 + de/dk = 2 + (γ − 1)de/dB, • if s > 2, then p∗(ℓ) = 0 and the RG flow converges again
for which dB → ∞ yields ε = 2, distinguishes non-​fractal (region II) from fractal networks towards the fractal network G0;
(region III). Within region III, the subregion defined by ε ≥ γ identifies the realm of • if s < 2, then p∗(ℓ) = 1 and the RG flow converges
hierarchical scale-​free graphs. A description of the databases and acronyms used can be towards the complete graph fixed point;
found at http://jamlab.org. d | Fractal to small-​world transition, obtained by randomly
• if s = 2, then the RG flow has one non-​trivial stable
rewiring a fraction prew of links. Whereas modularity is rapidly lost, small-​worldness is
fixed point G′, consisting of G0 dressed with shortcuts
rapidly gained, emphasizing the trade-​off between these network phases. e | Crossover
from the power-​law scaling to an exponential decay in fractal networks upon the random following p∗ (ℓ) = 1 − exp(−Aℓ −2dB).
addition of shortcuts according to the probability P(ℓ) ∼ℓ−α. All Nb nodes within a box are
within a shortest-​path distance ≤b. f ∣ RG flow diagram in the space of configurations. Analysing the flow of the difference between the
In the stable phase (s ≡ α/dB > 2), the RG flows towards the fractal fixed point, whereas in average degree z0 in G0 and the average degree zB in
the unstable phase (s < 2) it flows towards a complete graph. For s ≈ 2, the scaling has an the renormalized network G B = R B(G ′) yields further
exponential cut-​off at a characteristic scale (part e): globally the network is small world insights into these three network phases. Such analysis
and at small scales it is fractal. g | Stability analysis of the RG flow (in terms of stability shows that z B − z 0 = (z ′ − z 0 )D(x B ) where, in the thermo-
exponent λ) leads to equations 2 and 3. A navigability transition exists at s = 1 and dynamic limit, the function D(x B ) scales with the rela-
a fractal to small-​world transition at s = 2. Real-​world networks can be classified
tive network size as D(x B ) ∼ x Bλ and the RG exponent
accordingly. See ref.18 for additional information regarding the databases used and
λ depends on the shortcut exponent α as
their corresponding acronyms. Part a reprinted from ref.14, Springer Nature Limited.
Part b (top) adapted with permission from ref.85, Elsevier; (middle) adapted from ref.75, 
1, if s ≤ 1,
Springer Nature Limited; (bottom) adapted from19, Springer Nature Limited. Part c λ= 
 (3)
adapted with permission from ref.60, American Physical Society. Part d adapted with  2 − s,

 if s > 1.
permission from ref.83, PNAS. Parts e–g reprinted with permission from
ref.18, American Physical Society. Equation 3 identifies two transitions in the space of
network configurations (Fig. 1g). One is a small-​world-​to-​
fractal transition at s = 2, equivalently at α = 2dB, separat-
All small-​world networks, such as Erdős–Rényi graphs, ing the stable (λ < 0, s > 2) phase of compact topologies
Watts–Strogatz networks or the preferential attachment from the unstable phase (λ > 0, s < 2) of modular struc-
model, are characterized by v = γ′ = 2. Conversely, frac- tures. The other transition is a navigability transition
tal structures, such as the Song–Havlin–Makse (SHM) at s = 1, equivalently at α = dB, which is the network
model, Apollonian networks or percolating clusters, analogue of Kleinberg’s optimal point71.
feature different values of ν and γ′. The SHM model15,19 Besides raising theoretical questions regarding the
is generated by a multiplicative process of network characterization of these configurational transitions,
growth, the dynamics of which follow the inverse of the the RG theory sketched above provides an indirect
shortest-​path-​distance RG transform. For this model, method for extracting information about the distri-
in particular, v = γ′ = 1 for every scale-​free exponent, bution of shortcuts in fractal real-​world networks —
whereas v = γ′ = 2 after any arbitrarily small random a crucial ingredient for understanding information
rewiring69, suggesting that fractal topologies are indeed flow and optimal search72 — and for determining their
unstable fixed points of the shortest-​p ath-​distance approximate location in the space of network config­
RG flow. For increasing fractions of randomly drawn urations. Such analysis has been applied to the WWW,
shortcuts, in fact, fractal networks rapidly cross over to the metabolic network of Escherichia coli, a yeast protein–
more and more compact architectures (Fig. 1d) that have protein interaction network, the actors network of the
weaker and weaker modularity. This crossover leads to Internet Movie Database and the protein homo­logy
networks with fractal scaling ℓ̄ ∼ N01/dB that is observed network (Fig. 1g). The results show, in particular, that the
only up to a certain cut-​off length scale before a global WWW (Fig. 1b) is fractal up to a given length scale, but it
small-​world behaviour ℓ̄ ∼ N01/dB is found (Fig. 1e). is also sufficiently randomized to host an optimal flow,
These promising results prepared the ground for the as manifested by its proximity in the (λ, s) plane to the
firm RG theory presented in ref.18, which elucidates navigability threshold.
the universal features underlying the fractal to small-
​world transition in complex networks. Adding short- Network functionality and evolution. Although the
cuts to a fractal network G0 with a probability p(ℓ) = A shortest-​path-​distance RG placed the mixing patterns
ℓ−α, where A is a normalization constant, causes the RG of networks (such as the degree distribution) on more
trajectories to either converge towards G0 or to transform fundamental grounds, it also raised several puzzles of
the network into a complete graph — a trivially stable interpretation73. Questions arose about the functional
fixed point of the RG flow — depending on the value of significance encoded in the modules detected by the RG.

118 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

Furthermore, it soon proved impossible to understand nodes, subject to the noisy appearance (as exempli-
their emergence in terms of popular mechanisms of fied in the Watts–Strogatz model) of randomly placed
network growth74, such as the ‘rich get richer’ principle shortcuts. The net result is that hubs are buried deep in
of preferential attachment, or the ‘democratic’ wiring of modules whose low-​degree nodes are the inter-​module
Erdős–Rényi networks, which are generally dominated connectors, resulting in scale-​free networks that are frac-
by small-​worldness. tal and modular up to a cut-​off scale above which they
The first issue was soon clarified by analysing biologi­ become global small worlds (Fig. 1e). The SHM model
cal networks15,19,65, for which the RG transform has led to represents a theoretical benchmark for understanding
identification of hierarchies of modules closely related the self-​similar patterns observed in real-​world systems
to their known biochemical annotations, with a resolu- in terms of the microscopic growth rates controlling
tion as accurate as that of other clustering algorithms4,66,67. their dynamics75, and it has raised important evolution-
An important example of what RG analysis has enabled ary implications. Besides highlighting the evolutionary
is an integrated multiscale view of the network of human drive of many biological networks19 towards fractal
cell differentiation65, raising the possibility of identify- modular structures — which maximize their robust-
ing hitherto unknown functional relations between ness against random failures76 — the SHM model has
previously unrelated cellular domains. been used to reveal the inherently fractal geometric
The problem of growth, instead, has been elegantly nature of the duplication-​divergence mechanisms in
solved by building on the observation that modular genetics, suggesting that fractality and multiplicative
networks are also fractal (though the converse is not growth are essential features of the evolution of bio-
necessarily true). The SHM model (Fig. 2a,b) is built on logical networks. A notable result in this respect is the
this property15,19. In this model, hubs acquire new con- successful reconstruction of the present-​day structures
nections by linking preferentially with less connected of several protein–protein interaction networks (Fig. 2c),

a b c
1
x= LUCA
Present-day

Ancestral
network

node x=
2
Ancestral
eukaryote
Ancestral
fungi

Archaea and Other Other Present-day

bacteria eukaryotes fungi yeast

d e f
10 2
15 PPIN network Metabolic PPIN
3.0
101
Model networks: High modularity
R/k2ζ/dk

ζ=1 Sub-diffusion
100 ζ = ln(3/2)/ln(3) 2.5
10
ζ=0
〈T〉/N

dw

104 2.0
T/k2dw/dk

10 2 5
2 4 8 16 32 1.5 ζ=1
2
4 ζ = ln(3/2)/ln(3)
100 8 Low modularity
16 Super-diffusion ζ=0
0 1.5
10–1 100 101 0 5 10 15 20 25 0 1 2
ℓk21/dk Source–target distance dM

Fig. 2 | Network evolution and transport theory. a | Dendrograms process. Colours represent functional categories. d | Data collapse of the
demonstrating the network evolution (top tree) as inverse of the resistance R (top) and the diffusion time T (bottom) according to equation 4
shortest-​path-​distance renormalization group (RG) procedure (bottom tree) for the yeast PPIN (open symbols) and the SHM model with e = 1 (filled
(Box 1). The arrow of time increases from right to left. b | Seeds of the Song– symbols). Markers indicate ratios of scaling parameters k1/k2, and colours
Havlin–Makse (SHM) model for network growth. At each time step, a node denote values of k1. dw, walk dimension; dk, degree dimension; ζ, conductivity
produces m offspring for every link. The original link is then removed with dimension. e | Comparison between the finite-​size scaling law equation 5
probability e, and x new links between randomly selected nodes of the new and simulated diffusion times normalized by the network size N (markers)
generation are added. Illustration: m = 3 and e = 1. e tunes between pure in the PPIN and SHM model with m = 3 and e = 1. f | Modularity versus
fractals (e = 1) and pure small worlds (e = 0) and x controls the degree of transport in real-​world and SHM model networks in the (dw, dM) plane (dM is
modularity, so that x = 1 yields tree-​like structures and shortcuts among the modular dimension). Part a, adapted with permission from ref. 85,
modules appear for x > 1. c | Phylogenetic tree showing the evolutionary Elsevier. Part c is adapted from ref.77, CC BY 4.0. (https://creativecommons.
path of the Saccharomyces cerevisiae protein–protein interaction network org/licences/by/4.0/). Part d,f adapted with permission from ref.15, PNAS.
(PPIN), reconstructed from the last universal common ancestor (LUCA) to Copyright (2007) National Academy of Sciences, U.S.A. Part e adapted from
its present-​day structure by combining the SHM growth model with the RG ref.16, Springer Nature Limited.

NaTure RevIewS | PhySicS volume 3 | February 2021 | 119

Reviews

starting from their primitive ancestors identified via the Open questions. One pragmatic question raised by the
shortest-​path-​distance RG technique77. fractal geometry of network structures involves
the very same notion of self-​similarity and fractality
Transport in self-​similar media. Discovering the frac- in complex networks. Although traditional fractal the-
tal geometry of networks has also had implications for ory does not distinguish between fractality and self-​
the study of transport in complex media, a notoriously similarity, these two properties are distinct under the
hard theoretical problem78. In fact, the existence of an lens of the shortest-​path-​distance RG85. Whereas both
underlying self-​similar symmetry makes it possible to fractal and pure small-​world structures are statistically
circumvent the search for exact solutions and to attack self-​similar under the shortest-​path-​distance RG (see,
the problem by means of scaling arguments. A first for example, Fig. S6 in ref.19 and discussions therein),
result in this direction15 involved the scale-​invariant only the fractal structures feature a well-​defined set of
forms featured by the mean diffusion time, T ∼ ℓ d w, and (finite) fractal exponents. The divergence of dB in small-​
the average resistance, R ∼ ℓζ, experienced by a blind world networks (Fig. 1b) can be interpreted (using field
ant travelling a shortest-​path distance ℓ, where dw is the theory jargon) as an ‘ultraviolet’ limit above which the
walk dimension and ζ is the conductivity dimension79 fractal geometric approach based on the shortest-​path
of the underlying medium. Combining these relations RG fails to quantify the self-​similar symmetry. In this
with equation 1, the observables T and R re-​scale as perspective, identifying suitable embeddings or duality
T ′/T ∼ (N ′/N )d w/dB and R/R′ ∼ (N ′/N )ζ /dB under RG transformations86 that can regularize this ‘small-​world
transformations. A direct evaluation of dw and ζ in divergence’ is an open theoretical challenge that could
biological and synthetic (SHM) networks with finite dB raise the opportunity of designing a unified geometric
reveals15 that dB, dw and ζ obey the Einstein identity80 framework for the study of fundamental problems such
ζ = dw − dB, elegantly relating static and dynamic pro­ as transport78, evolution37, navigability25 or RG-​based
perties of transport in complex media. Thanks to this classification of the universality classes of networks18.
insight, it proved possible to predict the dependence In this respect, latent-​geometric approaches or the adop-
of T and R on microscopic network features — such as tion of generalized topological metrics and embeddings87
nodes’ degrees and the shortest-​path distance between could offer hints for solutions. Hyperbolic geometry,
them — leading to the scaling relations in particular, shares profound connections with self-​
similar metric spaces88–90, suggesting the possibility that
T (ℓ; k1, k 2 ) = k 2d w/dk fT (ℓ/k 21/dk ), the fractal exponents featured by self-​similar networks
(4) may have suitable counterparts in their corresponding
R(ℓ; k1, k 2 ) = k 2ζ /dk f R (ℓ/k 21/dk ), latent spaces and could further be extended to pure
small-​world structures.
where fT,R are scaling functions. These scaling relations On a more fundamental level, understanding how
agree well with the collapse of real-​world networks data the dynamical properties of network growth processes
(Fig. 2d). Furthermore — and more completely than the influence their asymptotic self-​similar patterns is an
above scenario — the assumption of self-​similarity has intriguing and as yet unexplored territory in the study
been used to develop a general theory of transport in of network geometry. The process of zooming out
complex media16. This theory culminates in the exact induced by the network RG transform finds its (statis-
finite-​size expression tically equivalent) inverse in dynamics19,37, so that the
varying structures observed at increasing length scales
 N (a + sgn(ζ )bℓ ζ ), ζ ≷ 0, correspond to the evolution of certain dynamical varia-
T (ℓ) ∼  (5)
 N (a + bln ℓ), ζ = 0, bles (Fig. 2a). A profound link between self-​similarity and

ergodicity has been explored in mathematics91, showing
where a and b are positive constants. Equation 5 that the notions of fractal dimensions and self-​similarity
describes a universal scaling law for transport in can be interpreted in terms of ergodic averages of some
complex media81 with finite dB (Fig. 2e). appropriate measure-​preserving dynamical systems92.
Besides their importance for theory, these results In the simple case of growing trees, it has further
have offered new methods of fractal geometry to relate been proved91 that this connection is a consequence
the kinetics of transport-​limited processes in real-​world of the explicit dependence of the fractal dimension on
systems to their interaction topologies. In particular, the growth rates ruling the system’s evolution. An anal-
equation 4 enables bridging of the degree of modularity ogous result holds for the self-​similar structure of the
of networks to their efficiency of transport15. This rela- SHM model in the thermodynamic limit19: its char-
tionship is epitomized by the identity dw = 1 + dM (Fig. 2f): acteristic fractal exponents dB, dk, dM, … depend only
high levels of modularity (dM > 1) generally result in on the process’ growth rates (Fig. 2b). A relevant and
subdiffusive dynamics (dw > 2). This simple result has challenging question, in this respect, is to understand
had significant implications in applications such as the weather the SHM model and/or other more popular
characterization of the flux responses of metabolites15 or growth processes19,37,93 can be themselves interpreted as
the community detection in global small-​world social ergodic dynamical systems with respect to some suitable
networks82, and it has provided intriguing ideas for invariant measure. Besides its theoretical relevance, find-
addressing the long-​standing conundrum of the highly ing viable directions to tackle this overarching problem
modular yet globally optimal organization of functional could help establish fundamental connections among
brain networks83,84. the static and dynamic facets of network geometry.

120 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

Hyperbolic geometry of latent spaces boundary at infinity that is always a self-​similar metric
Deep connections between self-​s imilarity and space94. Another connection is that self-​similar groups
hyperbolic geometry have been well explored by can always be represented as the groups of automor-
mathematicians88,90,94. One connection goes via the phisms of trees, which are the simplest example of
observation that any Gromov-​hyperbolic space has a discrete hyperbolic spaces90. The connections between

a Model 1 κB1 da,B b Model 2 c

κB2
B2
B1 B
da,A κA2 2 BRA
κA1 A2 A2 THI
SAF ROK
B1 JPN SIN
A1 κC2 AUL
A1 GAB
C2 UAE CHN
C2 IND TRI
USA
κC2
C1 C1
BEL SAU
UKG
NTH GMY
RUS FRN
ITA NIG
SPN
da,C

Murein R
d
Murein B

Lipopolysaccharide
Histidine

Glycolysis/gluconeogenesis
Nucleotide salvage D-alanine
Pentose phosphate D-glucose
UMPUDP

Ammonium
Glutamate
Pyruvate
Alanine and aspartate L-glutamate
CTP CMP
L-serine
Purine and pyrimidine
Glycine and serine CO2 Glycerol 3-phosphate
Glyoxylate
Tyrosine, tryptophan Glycerophosphoglycerol
and phenylalanine CoA Acyl carrier Glycerophospholipid
Oxidative phosphorylation Acetyl-CoA
Valine, leucine and isoleucine O Malonyl-acyl carrier
2
Pyruvate/citric acid cycle
Anaplerotic reactions FAD
Methionine
Cell envelope B
Arginine and proline
Threonine and lysine
Nitrogen
Cysteine Folate
Methylglyoxal
Inorganic ion transport
Membrane lipid

e f g

Source Path

Geodesic

Destination

NaTure RevIewS | PhySicS volume 3 | February 2021 | 121

 Reviews

hyperbolicity and self-​similar sets, fractals and sim- In this ensemble, edges are fermions with energies
ilar objects also go through the idea that rescaling is
(approximately) an isometry transformation in (coarse) εij = f (xij ), (7)
hyperbolic geometry95.
Yet hyperbolic geometry is not the geometry of the where f(xij) can be any function of distances xij between
observable structure of real-​world networks discussed nodes on the circle, β fixes the average energy and μ is
above, but the geometry of their latent spaces that the chemical potential controlling the expected number
we discuss here. The two geometries are intimately of edges (fermions) and thus the average degree.
related because network paths that follow hyperbolic The choice of f(x) defines network properties in the
geodesics in the latent space are the shortest paths in ensemble. The necessary and sufficient conditions for
the network with high probability21,25,96,97 — the network networks in the model to be sparse small worlds with
is said to be congruent with its underlying latent non-​vanishing clustering are f (x ) ∝ ln x and β ∈ (1, 2)
geometry. (ref.99). More precisely, if f (x ) ∝ ln x , then clustering
in the n → ∞ limit is zero for β ≤ 1 but an increasing
Models. Models in which nodes are connected based positive function of β > 1; networks are small worlds
on their proximity in a latent space are known as (soft) whenever β < 2.
random geometric graphs in mathematics, where they The distribution of node degrees in such models is
have been extensively explored98. In the simplest random homogeneous, but the model can be modified to yield
geometric graph model, n nodes are placed uniformly at any degree distribution. Such a modification, known
random on the interval [0, n] with periodic boundary as the S1 model (Fig. 3a), sets the edge energy given in
conditions, that is, on the circle S1. The node pairs are equation 7 to
then connected if the distance between them on the cir- xij
cle is less than a parameter μ > 0 controlling the average εij = ln , (8)
degree (〈k〉 = 2μ). The model yields networks with non-​ κ iκ j
vanishing clustering (〈c〉 = 3/4) and an average shortest-​
path length that scales linearly with network size. Such and hence the connection probability in equation 6 to
networks are thus large worlds.
From the statistical physics perspective, this sim- 1 1
pij = = ,
1 + χ ijβ
( )
β
plest possible latent-​space network model is the zero xij (9)
1+
temperature (that is, the inverse temperature β → ∞) μκ̂ iκ j

limit of a more general entropy-​maximizing probabil-

istic mixture of grand canonical ensembles with the where κi is the expected degree of node i = 1, …, n in
Fermi–Dirac probability of connection between nodes i the ensemble20 and μ̂ is a constant that fixes the aver-
and j: age degree. The values of parameters κi can be either
fixed or random, sampled from any desired distribu-
1 tion. If they are sampled from the Pareto distribution
pij = β (εij− μ )
. (6)
e +1 ρ(κ ) = (γ − 1)κ 0γ −1κ −γ , the resulting degree distribution is
Pareto-​mixed Poisson100, which for k ≫ 1 is a power law
◀ Fig. 3 | Networks in latent geometry. a | Model S1 (1D sphere). The S1 distances da P(k) ∼ k−γ observed in many real-​world networks101.
between pairs of nodes A1–A2, B1–B2 and C1–C2 (coloured in blue) are highlighted. The As in the simplest random geometric graph model, clus-
size of a node is proportional to its expected degree κ. b | Model H2 (2D hyperbolic disk). tering is zero for β ≤ 1 and an increasing positive func-
The same highlighted pairs of nodes (in blue) are at the same hyperbolic H2 distance. tion of β > 1. The networks in the model are ultrasmall
1
Higher-​degree nodes are positioned closer to the centre. The angular S distances worlds whenever γ < 3. The definition of edge energy
|
between the corresponding node pairs are the same in parts a and b. c World trade map in equation 8 combines the popularity (degrees κi)
in 2013. Node colours correspond to communities detected by the critical gap method and similarity (distances xij) dimensions into a single
(part e). See Supplementary Table S1 in ref.30 for the country associated with each measure, whereas the connection probability takes the
acronym. d | Metabolic network of bacterium Escherichia coli. Yellow circles represent
gravity-​law form in equation 9, decreasing with the sim-
reactions and blue squares are metabolites. The name of a pathway is located at the
average angular position of all the reactions belonging to it. Acronyms for metabolites ilarity distance xij and increasing with the popularity
are defined in the BiGG database (http://bigg.ucsd.edu/). e | Critical gap method. product κiκj.
Nodes are partitioned into groups separated by void angular gaps. The modularity of the
partition is computed by comparing the number of links within the communities (purple An intermediate step to go from the S1 to hyperbolic
links) to the number of links between nodes in different communities (green links). geometry is to map κ to y by
The partition with the highest modularity is selected. f | Proximity of shortest paths in
hyperbolic networks to hyperbolic geodesics. Blue arrows show the paths that geometric κ ↦ y = κ 2. (10)
routing finds for two source–destination pairs in a hyperbolic network. The found paths
are also the shortest paths in the network in terms of the number of hops. The hyperbolic This change of variables places all nodes i at coordi-
geodesics between the corresponding sources and destinations in the hyperbolic plane
nates (xi, yi), xi ∈ R, yi > κ 02, in the upper half-​plane model
are shown as dashed red curves. g | Structural network of the human brain with links
coloured by whether they belong (magenta) or do not belong (cyan) to the minimal of the hyperbolic plane H2 (ref.102), which has a long
network that enables maximal navigability in the brain (the navigation skeleton). Parts history in relation to networks88,103–108. If ρ(κ) is Pareto
a,b,c are adapted from ref.30, CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). with γ = 3, then nodes are distributed uniformly on the
Part d adapted with permission from ref.139, Royal Society of Chemistry. Part g adapted hyperbolic plane, where the metric is ds2 = (dx2 + dy2)/y2.
from ref.26, Springer Nature Limited. The group of distance-​preserving isometries of the

122 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

half-​plane is isomorphic to the Lorentz group SO(1, 2). networks112, multilayer networks113,114 and to networks
The Lorentz boosts, which are hyperbolic rotations in the with community structure29,32,115. The model has yet
3D Minkowski space, act on the upper half-plane as to be extended to directed networks, because it is not
space-rescaling transformations x ↦ x ′ = ξx, y ↦ y ′ = ξy, clear how to reconcile the intrinsic symmetry of metric
where ξ > 0 (ref.109). The energies distances with asymmetric interactions among nodes.
xij xij The equivalence between the two models in equa-
εij = ln = ln (11) tions 8 and 13 is a reflection of the isomorphism between
κ iκ j yi yj the Lorentz group SO(1, D + 1) and the Möbius group
acting on sphere S D as the group of its conformal trans-
are thus manifestly invariant with respect to rescaling formations. This isomorphism is a starting point of the
Lorentz boosts, as is the model. Lorentz boosts form a anti-​de Sitter/conformal field theory correspondence in
non-​compact subgroup of all isometries of the hyperbolic string theory116.
plane. However, the model is not invariant with respect It is important to re-​emphasize that the latent space
to all isometries of the full Lorentz group because energy in the described models can indeed be any compact
is not exactly a function of the hyperbolic distance, homogeneous space of any dimension D (ref.99). This
and hence neither is the connection probability. space can also be flat, or positively or negatively curved.
This problem is fixed in a slightly different but asymp- However, the effective space of the higher dimension
totically equivalent model, known as the H2 model21, D + 1 is always hyperbolic, that is, negatively curved.
given by the map What makes this space hyperbolic is the simple change
of variables in equations 10 and 12, which maps the
κ ↦ r = R − 2ln κ , (12)
expected degrees of all nodes to their D + 1th (radial)
where R = 2ln (n/c ) and c is the parameter controlling coordinates. On this change of variables, the probabil-
the average degree. This map places nodes i at polar ity of connections is always a function of hyperbolic
coordinates (ri, θi), θi = 2πxi/n, on the hyperbolic disk D + 1-​dimensional distances only (equations 13 and 14).
of radius R in the hyperboloid model of the hyperbolic However, the nodes are distributed uniformly accord-
plane with metric ds 2 = dr 2 + sinh2r dθ 2 (ref.102; Fig. 3b). ing to the metric in this D + 1-​dimensional hyperbolic
The edge energy becomes space only if the degree distribution is a power law with
exponent γ = 3, that is, only if ρ(κ) is Pareto with γ = 3.
1  θ 
εij = r + r + 2ln ij  ≈ 1 d , (13)
Another important observation is that, as was
 i j  ij
2 2  2 shown recently in ref. 99, the described models are
unique latent-​space network models that satisfy cer-
where θij and dij are the angular and hyperbolic dis- tain maximum-​entropy requirements and that produce
tances between the two nodes. The approximation holds sparse heterogeneous uncorrelated small worlds with
for a fraction of node pairs that converges to 1 in the non-​zero clustering. At present, these models are also
n → ∞ limit21. With this approximation the connection the only known class of network models that capture
probability reads all the follow­ing properties of many real-world networks:
sparsity, self-similarity, small-worldness, heterogeneity,
1 non-vanishing clustering and community structure.
pij = β
. (14)
1+e 2 (dij− R ) As a result, this ensemble of random graphs has attracted
substantial research attention in mathematics and theo-
The simplest formulation of the model is when ρ(κ) retical computer science, fields in which many basic and
is Pareto with exponent γ = 3 and β → ∞: sprinkle n advanced properties of random hyperbolic graphs have
points uniformly at random over a hyperbolic disk of been (re)derived rigorously117–128.
radius R, and then connect all pairs of points located at
hyperbolic distance dij < R from each other. The equiva- Hyperbolic maps of real-​world networks. A collection of
lence between the S1 model and this H2 representation is methods have been developed to infer the coordinates
shown in Fig. 3b. Because energy is a function of the dis- of the nodes of a real-​world network in its latent space.
tance in equation 13, the model is fully Lorentz invariant Many generative-​model-​based methods perform
in the n → ∞ limit for any β. statistical inference using Monte Carlo sampling
The latent space in the model does certainly not have and maximum likelihood estimation97,129–131. Data-​
to be the circle S1. It can be any compact homogene- driven methods vary in flavour. One such method
ous space of any curvature and dimension D (ref.99). is unsupervised machine learning, which is used, as
The higher the dimension, the lower the clustering in coalescent embedding132, to implement nonlinear
for the same value of β (refs34,110). On the hyperboliz- dimension reduction133. Other data-​driven methods rely
ing change of variables in equations 10 and 12, these on the network community structure31,134. Mechanistic-​
D-​dimensional spaces turn into hyperbolic spaces model-​based methods map the network while unfold-
of dimension xij
D + 1 and the edge energy becomes ing the similarity space135. Generative models and
εij = ln 1/ D ≈ d ij/2 (ref. ).
99
data-​driven models both have advantages and disadvan-
(κ iκ j)
The models have been also adapted to growing tages in terms of accuracy and speed, so hybrid models
networks35 — in which case the latent space is not hyper- have been developed136,137. Hybrid methods obtain initial
bolic but de Sitter space d S1, D with the same Lorentz coordinate estimates using machine learning techniques,
group SO(1, D + 1) of symmetries111 — and to weighted and then refine the results via maximum likelihood

NaTure RevIewS | PhySicS volume 3 | February 2021 | 123

Reviews

estimation. In general, the main challenge that hyperbolic in the network ensemble. In the ‘vanilla’ version of the
network mapping methods face is the abundance of local model described above with the homogeneous distribu-
maxima in highly non-​convex likelihood landscapes. tion of angular coordinates, these properties are not tun-
An efficient way to escape from these maxima — thus able but are fixed by structural constraints147,148. However,
boosting the inference accuracy — is to shake the sys- if these coordinates are not homogeneously random
tem by adding decreasing levels of noise to it138, a method but instead set by their heterogeneous values inferred
conceptually similar to simulated annealing. in a real network with communities, the ensemble of
The application of these methods to real-​world net- random networks generated by the model using these
works makes it possible investigate them at different heterogeneous inferred coordinates accurately repro-
resolutions using the latent-​space RG, which unfolds duces the degree correlations and clustering in the real
the network in a sequence of self-​similar layers, or network137,138. This observation suggests that, to a great
renormalization shells34. The maps of real networks30,139 extent, degree correlations are a consequence of latent
(Fig. 3c,d) also revealed the existence of geometric com- geometry coupled with inhomogeneous distributions
munities and helped decode mechanisms that govern of nodes in it.
network evolution, such as globalization, localization
and hierarchization, which drive the evolution of inter- Navigability. Latent space guides navigation in the
national trade30. Such maps have also shed light on many network based on distances between nodes in
dynamical processes in real-​world networks, showing, the space25. That is, instead of finding shortest paths
for instance, that cooperation in social networks is con- in the network — a computationally intensive combi-
trolled more strongly by the latent-​space organization natorial problem in a network that changes dynami-
than by highly connected hubs in the system140. Another cally, such as the Internet149 — a transport process can
class of applications of hyperbolic maps of real networks be geometric, relying only on geodesic distances in the
is link prediction. Because the connection probability space. The efficiency and robustness of such processes28
in the described models is a decreasing function of the determines how navigable a network is. Navigability is
latent hyperbolic distance, the models predict that links improved for smaller γ and larger β, thereby defining
are more likely to exist between hyperbolically closer a navigable parameter range to which many real-​world
pairs of nodes. Link prediction using hyperbolic geom- networks belong25. Networks in the hyperbolic model
etry has been analysed from different angles135,138,141,142. described above are nearly maximally efficient for such
It appears to be particularly powerful when it comes geometric navigation21, which has recently been proven
to predicting links that are difficult to predict, such as rigorously121.
links between nodes without common neighbours138. The main reason behind this phenomenon is the
Finally, one of the most practical applications of map- existence of shortest paths close to the correspond-
ping real-​world networks to their latent geometries is ing geodesics in the underlying hyperbolic geometry
the design of efficient routing protocols for the Internet97 (Fig. 3f) for any pair of nodes in hyperbolic networks.
and for emerging Internet-​of-​Things telecommunication Another critical factor is the existence of superhubs that
networks143. We discuss some of these applications below. interconnect all parts of the network, present as soon
as γ < 3 (ref.21), in which case the networks are known as
Geometric communities. In the above models, the angu- ultrasmall worlds52. Navigation in hyperbolic networks
lar distribution of nodes is uniform. However, nodes in with γ < 3 can always find these ultrashort paths96; thus,
maps of real networks are clustered in regions that define navigation in these networks is asymptotically optimal.
geometric communities. Such communities have been Conversely, networks that are maximally navigable by
observed in many real networks, including the Internet97, design are similar to hyperbolic networks26, and many
metabolic networks in cells139, trade networks30 and brain real-​world networks, such as the human brain (Fig. 3g),
connectomes144. Non-​overlapping communities can be contain large fractions of their maximum-​navigability
detected in the geometric domain using purely geometric skeletons26. Assuming that real-​world networks evolve
methods. One definition considers soft communities as to have a structure that is efficient for their functions,
groups of nodes in similarity space separated from the these findings provide an evolutionary perspective
rest by angular gaps that exceed a certain critical value139. on the emergence of the latent geometry that leads to
An alternative, known as the critical gap method30, finds structural commonalities observed in many different
the communities by changing the gap and selecting the real-​world networks.
soft community partition that maximizes the standard
modularity measure (Fig. 3e). These distance-​based com- Renormalization and self-​similarity. Networks in the
munities show strong correlations with groups defined discussed models are purely scale invariant in the ther-
by metadata, such as geographical location of the auto­ modynamic limit, and thus contain an infinite hierarchy
nomous systems in the case of the Internet97, biochemical of self-​similar nested subgraphs induced by nodes that
pathways of reactions in metabolic networks139 or ana- have degree exceeding a given threshold (Box 1). This
tomical brain regions in structural brain networks27,145. observation applies to many real networks, for which
Geometry-​based communities also overlap substantially the average degree of the subgraphs increases as a func-
with topology-based communities, making the geometric tion of the degree threshold150. This property makes it
nature of complex networks even more evident30,146. possible to prove the absence of percolation or epidemic
Geometric communities affect degree–degree correla- thresholds in such networks150, independently of the
tions and the functional form of the clustering spectrum commonly used tree-​like or scale-​free assumptions151.

124 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

The proof is general and is based solely on a symmetry geometry168–170. Euclidean distances certainly do play a
principle. Thus, it also applies to any phase transition role in the brain. However, they are not the only factor
for which the critical point is a monotonic function of determining similarity, and thus connectivity, between
the average degree150. Examples of models for which the brain regions27,144. More recently, data-​driven dimen-
proof holds include susceptible–infected–susceptible sional reduction techniques145,171 and local curvature
(SIS)-​type epidemic spreading, which in scale-​free net- measures172 have been proposed as alternative geomet-
works lack a healthy phase152 or the Ising model, which ric descriptions that avoid the definition of connectivity
in scale-​free networks lacks a disordered phase153–155. laws. It remains unknown how the hyperbolic geometry
Self-​similarity is also observed in the multiscale of the brain relates to the widely accepted optimization
organization of networks, and can be explored at dif- principles that Santiago Ramón y Cajal hypothesized
ferent resolutions by applying a geometric renor- more than a century ago as the two forces ruling the
malization transformation 34 inspired by concepts evolution of mammalian brain connectivity: minimizing
from the real-​space RG in statistical physics 58,59,156. wiring costs and maximizing conductivity speed173.
The method takes a different approach to that of the In the context of machine learning, primarily after
shortest-​path-​distance RG discussed above, as it relies the publication of ref.174, hyperbolic spaces have ignited
on distances in the similarity subspace to coarse-​grain vigorous research activity. This activity has been applied
neighbouring nodes into supernodes that define a new to many settings and tasks including embedding graphs
rescaled map (Box 1). The iteration of the transformation and other data, such as images and texts, as well as in
unfolds a network into a multiscale shell that progres- the design of neural networks, attention networks and
sively selects longer-​range connections, revealing the knowledge graphs, with applications including data
co-​existing scales and their interactions. Self-​similarity classification, image recognition, natural language
under geometric renormalization is a ubiquitous sym- processing, link prediction and scalable recommender
metry in real-​world networks (Fig. 4a–c), in good agree- systems175–185. Overall, the main flavour of these results
ment with the prediction given by the renormalizability confirms one of the main points in ref.21: compared
of the underlying S1 model34 (Fig. 4d–f). This result sug- with Euclidean geometry, hyperbolic geometry appears
gests that the same connectivity law rules short- and to be a better (embedding space) model for highly
long-​range connections, and operates at different length heterogeneous networks and other data.
scales. Interestingly, the structure of the human brain In terms of open questions, perhaps one of the most
remains self-​similar when the resolution length is pro- common ones is how to tell whether a given (real world)
gressively decreased by hierarchical coarse-​graining of network has a latent (hyperbolic) space. This question
the anatomical regions, a symmetry that is predicted by is a variant of the more general question of how to tell
geometric renormalization144. From a practical point whether a given model is a good model for a given net-
of view, applications of self-​similarity under geometric work. Such questions can never find positive answers
renormalization include a multiscale navigation pro- because one can never learn a model or a distribution
tocol that takes advantage of the increased navigation based on one sample from it. Neither can one know for
efficiency at higher scales, and scaled-​down network sure whether any given network has typical values of all
replicas. In these network replicas, the behaviour of possible network properties in any given model, simply
dynamical processes — such as SIS epidemic spread- because the number of such properties is infinite186. One
ing, the Ising model or the Kuramoto model157 — is can always check if any finite collection of properties of
preserved across layers (Fig. 4g–i). the network have the typical values in the model, and
as soon as an atypical property is found, one may raise
Open questions. In neuroscience, geometric naviga- doubts how good the model is for the network at hand.
tion as discussed above offers a possible explanation Such a finding does not always render the model useless.
and mechanism for the routing of information in the For instance, clustering is zero in stochastic block mod-
brain. This hypothesis has been investigated at different els and non-​zero in real-​world networks, but this fun-
depths and from different perspectives158–167. Geometric damental mismatch does not appear to diminish much
navigation is effective only when the network topo­logy the interest in stochastic block models, owing to their
is congruent with the underlying latent geometry so that simplicity and tractability.
following geodesic paths in the latent space is equiva- The latent-​space models described above repro-
lent to navigating through topological shortest paths. duce most of the important properties of real-​world
This equivalence seems to hold for structural brain net- networks. What this means is that any network that
works whose many structural and navigability proper- has these properties can be mapped to a latent space
ties are well described by the S1/H2 geometric network of any dimension, yielding some non-​trivial results.
model27,145, for which the same connectivity laws apply What space dimension one should choose for the
both to short-​range and long-​range connections and at embedding132,187 is an interesting open problem. Solving
different scales144. Thus, simplicity might be one of the it requires identifying a network property that would
main organizing principles of human structural brain depend on the space dimension in a known way. At
networks — at least at the macroscale level, which dis- present, it is known that clustering is a decreasing
plays a self-​similar architecture across different anatom- function of dimension34,110, but it is also a decreasing
ical length scales, in good agreement with the discussed function of temperature20,21, so that the value of cluster-
geometric models144, unlike traditional approaches ing by itself cannot tell us the likely value of the space
that describe brain connectivity using Euclidean dimension. In the absence of such understanding, and

NaTure RevIewS | PhySicS volume 3 | February 2021 | 125

Reviews

a b c
100 100 100

10–2 (knn,n
(l)
(kres
(l)
)
10–1

)
101

)
Pc(l) (kres
(l)

c– (l) (kres
(l)
P (χij )

10–2
(l)

10–4 100

l=0 l=3 10 –2
10–1
10–6
10–4 l=1 l=4 10–2
l=2 l=5 10–1 101 10–1
10–8 10 –3

10 –6
10 –2
10
2
10 6
10
–2
10 –1
10
0
10
1
10 2
103
10 –2
10 –1
10 0
101 102 103
(l)
χij (kres
(l)
) (kres
(l)
)
d e f
2.5
10 0
I II
102
10–3 2.0
1.00
Airports
P (χij )

〈k〉 (l)
(l)

β
C(l)

0.50
10–6 101
0
1.5 Internet
Synthetic
0 2 4 6 8 Metabolic
l Drosophila
10–9 Music Enron
10–6 10–3 100 103 106 Words Proteome 0 1 2 3 4 5
1.0
(l) l
χij 2.0 2.5 3.0 3.5
γ
g h i
1.0 1.0 1.0

0.8 0.8 0.8

0.6 0.6 0.6

〈|m|〉 (l)

〈|ρ|〉 (l)

〈r〉 (l)
0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.5 1.0 1.5 2.0 0 0.5 1.0 1.5 2.0 0 0.05 0.10 0.15 0.20
1/T λ σ

Fig. 4 | Multiscale unfolding of network structure and function by also shown. Darker blue (green) in the shaded areas represent higher values
geometric renormalization. a–c | Self-​similarity of the structural and of the exponent c controlling the flow for the average degree 〈k〉(l+1) = rc〈k〉(l).
dynamical properties of the multiscale five-​l ayered geometric The dashed line separates the γ-​dominated region from the β-​dominated
renormalization shell of the Internet, in terms of the empirical connection region. In phase I, c > 0 and the network flows towards a fully connected
probability in layer l of renormalization, measured as the fraction of graph. In phase II, c < 0 and the network flows towards a 1D ring. The red line
connected pairs of nodes (denoted i and j) as a function of the rescaled c = 0 indicates the transition between the small-​world and non-​small-​world
distance χij(l ) as defined in equation 9 (part a); complementary cumulative phases. In region III, the degree distribution loses its scale-​freeness along
distribution Pc(l ) of rescaled degree k res
(l )
(part b); and degree-​dependent the flow. f | Exponential increase of the average degree of the renormalized
(l )
clustering coefficient c over rescaled-​degree classes (part c; inset shows real networks 〈k〉(l) with respect to l. g–i ∣ Simulation of the Ising (part g),
(l ) susceptible–infected–susceptible (SIS) (part h) and Kuramoto (part i)
normalized average nearest-​neighbour degree knn,n ). d | The same as in part a,
but for a synthetic S1 network with N ≈ 225,000 nodes, scale-​free exponent models in scaled-​down replicas of the real Internet. The dynamics is
γ = 2.5 and inverse temperature β = 1.5. The black dashed line shows the preserved even in small replicas 25 times smaller than the original network.
theoretical curve given in equation 9. Inset: invariance of the mean local m, magnetization; T, temperature; ρ, prevalence; λ, infection rate; r,
clustering C(l) along the flow. e | Real networks in the connectivity phase coherence; σ, coupling strength. Results in parts g–i are averaged over 100
diagram of the S1 renormalization flow. The synthetic network of part d is simulations. Parts reprinted from ref.34, Springer Nature Limited.

given that space in hyperbolic geometry expands expo- can be formulated as follows: is a model that reproduces
nentially fast in any dimension, which minimizes any clustering but produces otherwise maximally random
distortions of low-​dimensional mappings, there are networks equivalent to a latent-​space model? Such ques-
no reasons not to choose the simplest case D = 1 for tions cannot be answered without additional stringent
the embedding, unless overparametrization may be assumptions and requirements, such as the requirement
beneficial, as in deep learning188. of maximum entropy. The first steps in this directions
Another important open problem is to identify a have been made in refs99,189.
collection of network properties that are not only nec- Other open problems, which are interesting from
essary but also sufficient conditions for latent geomet- the practical perspective, are to study the coupling
ricity. This problem is not about any given network, in of dynamical processes with the latent space and to
which case it can never be solved for the reasons given generalize latent-​space network models to temporal
above, but about network models. For instance, the networks190. It has been observed that clusters in the
question about whether clustering is such a property underlying metric space emerge in evolutionary games

126 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

on scale-​free networks140. Other processes could present network-​driven processes. From the spread of rumours
geometric patterns in their dynamical states as well. and opinions in sociotechnological systems to the
A key point is that essentially all real-​world networks are global spread of innovations and epidemics, combin-
highly dynamic at different timescales, suggesting that ing dynamics with the self-​similar structure of complex
the positions of nodes in the latent space cannot be fixed networks and their latent geometry results in heteroge-
but must constantly change. However, the development neous processes that cannot be easily understood when
of dynamic latent-​space models is challenging, because investigated in the Euclidean space where they are often
there is no consensus or even general understanding embedded. However, when the same dynamical pro-
concerning the exact set of requirements for such mod- cesses are analysed through the lens of the geometries
els. The main high-​level problem is that there are too they induce, one often discovers simple and elegant
many ‘degrees of freedom’ in designing these models, arguments to better understand the complex spatiotem-
that is, arbitrary choices can be made, and there are no poral patterns observed in a broad spectrum of complex
generally agreed guiding principles concerning what systems.
choice is better or worse. Because it is possible to define multiple dynamical
At a more fundamental level, the fact that hyperbolic processes on the network — such as epidemic spread-
networks are Lorentz invariant in the thermodynamic ing (Fig. 5) or random searches (Fig. 6) — one does not
limit suggests the need to re-​examine the role of pro­ expect to find a unique latent geometry for the system’s
babilistic symmetries191 in the theory of graph limits192. functioning: in principle, there could be as many hidden
Traditionally, the main symmetry of interest in that geometries as the number of plausible network-​driven
context has been exchangeability193 — the requirement processes.
that the probability of a graph in an ensemble does not In this universe of dynamical processes, the dynam-
depend on how nodes in the graph are labelled, remi- ics of information exchange has mostly been modelled
niscent of gauge invariance in physics. This requirement through diffusive processes, for which the amount of
is stringent for deep statistical reasons193, but it is also matter being diffused is conserved, and spreading pro-
easy to appreciate intuitively: if node labels 1, …, n are cesses, in which the amount of matter being diffused is
just random ‘coordinates’ used to represent an otherwise duplicated at each step (as in the spreading of viruses or
unlabelled graph as an adjacency matrix, then the prob- ideas) and consequently is not conserved. Such dynam-
ability of the graph in the ensemble cannot depend on ics have been successfully exploited to define metric
these meaningless coordinates. However, the Aldous– or quasi-​metric measures to probe the corresponding
Hoover theorem194,195 states that graphs in the thermo- latent geometry. The difference between metric and
dynamic limit of any exchangeable graph ensemble quasi-​metric measures is that, in the latter, the symmetry
are either dense or empty. This fact is in an unsettling axiom is relaxed.
disconnect with real-​world networks, a vast major- Remarkably, these classes of network geometry,
ity of which are sparse. Therefore, different notions of induced by kinematic distances, provide results about
exchangeability-​like probabilistic symmetries have been a system’s function that cannot be obtained by geomet-
investigated for sparse graph limits196–199. Because of the ric approaches discussed in the previous sections. An
Aldous–Hoover theorem, they all depart from the 1, …, n emblematic example concerns the mesoscale organiza-
‘coordinate system’ for node labels, and rely on different tion of interconnected components that exchange infor-
systems of node or edge labels. The Lorentz invariance mation during collective phenomena — such as coupled
of hyperbolic networks suggests that the labels of nodes oscillators trying to synchronize or people with social
can be indeed their coordinates in a latent space, with relationships attempting to reach consensus — which
exchangeability replaced by invariance with respect can be characterized by mapping the interplay between
to the space isometries. The hope is that such ensembles structure and function to a geometric space induced by
may have some interesting and tractable thermodynamic diffusion dynamics42. The resulting functional modules
limits. differ from the ones obtained by other geometric tech-
niques such as functional modularity maximization65,
Dynamic geometry of network processes because the latter finds an optimal partition of the sys-
The existence of a hidden geometry of network structure tem while minimizing the possible number of modules
naturally raised the question of whether a hidden geom- of given topological size ℓ, which in turns defines the
etry of network dynamics was plausible. The motivation characteristic distance between nodes within modules.
for this question is to identify the latent space due to Conversely, modules identified on the diffusion man-
system function that arises from the interplay between ifold at time τ, determining the scale of dynamics, are
structure and dynamics. Note that network dynamics characterized by groups of nodes that easily exchange
is broadly defined, and includes the dynamics of ver- information — for instance, in terms of random searches
tex and edge creation or destruction — exemplified by — within Markov time τ, as we discuss here.
network growth processes or time-​varying topologies190,
for example — and the dynamics of processes on the Geometry induced by resistance. In first approximation,
network. The hidden geometry induced by dynamics one can model information exchange in a network with
has been mostly explored for the latter, except for a few a model similar to that of current flow in a circuit. This
notable cases200. approach is probably among the first historical attempts
Geometric tools have proved fruitful for unravelling to quantify a distance between nodes in terms of a
hidden patterns lurking in the complex behaviours of (simple) dynamics. This resistance distance Rpq (ref.201)

NaTure RevIewS | PhySicS volume 3 | February 2021 | 127

 Reviews

a Fig. 5 | Geometry induced by spreading dynamics and

universal temporal distance. a | Evolution of a simulated
disease spreading originating in Hong Kong: red symbols
encode the prevalence. Top panels show the evolution in
the latent space; bottom panels show the same dynamics
in the natural space. When the spreading dynamics is
depicted by exploiting the induced geometry, complex
T = 41 d T = 51 d T = 62 d T = 72 d
spatial patterns are mapped to homogeneous wave fronts
propagating at constant effective speed. b,c | Relationship
between epidemic arrival times (Ta) and the two distances
for the spreading of influenza A virus subtype H1N1.
b c The relation is nonlinear when geographic distance Dg is
150 H1N1 (2009) 160 H1N1 (2009) used (part b) whereas it is nicely reproduced by a straight
Afghanistan 140 line when effective distance Deff is considered (part c).
Latvia Latvia d,e | A signal travelling from vertex j to the rest of the
120 Algeria system exhibits different spreading patterns (part d),
100 Slovenia
Barbados 100 Barbados Afghanistan captured by the universal temporal distance L(j → i),
Japan Argentina impacting different nodes (such as nodes 1 and 2) in
Ta (d)

Ta (d)

Norway 80 Slovenia
Italy different ways (part e) depending on the type of dynamics,
France India Japan
50 Argentina 60 India described by the parameter θ: distance limited (θ = 0; top);
France
Norway degree limited (θ > 0; middle); or composite (θ < 0; bottom)
China 40
UK Italy UK China Aij indicates the corresponding entries of the adjacency
USA 20 Spain matrix; xi(t) indicates the ith entry of the state vector at
Spain Germany Netherlands
0 Mexico USA Germany time t. f–h | The homogeneous propagation of concentric
0
0 5 10 15 20 Mexico 5 wave fronts emerge from the analysis of a broad spectrum
10 15 20
Dg (103 km) Deff of synthetic and empirical systems. R1 and R2 indicate
models capturing regulatory dynamics, M captures
d e
1, 2 ecological interactions, E captures epidemic
spreading, N captures neuronal activation and P
Δxi(t)

θ=0
L(j → 1) = L(j → 2) captures population dynamics. Parts a–c adapted with
permission from ref.40, AAAS. Parts d–h adapted from
Aij t ref.43, Springer Nature Limited.
Δxi(t)

1 θ>0
4 L(j → 1) > L(j → 2) 1 where Ω = L† + N −1U, and N is the size of the network, U is
2
j
3 a matrix with all entries equal to 1 and L† is the Moore–
2 t Penrose inverse of the Laplacian matrix of the network.
On the one hand, this metric has been successfully used,
for instance, to analyse specific isomers202 and genetic
θ<0
Δxi(t)

L(j → 1) < L(j → 2)

1
2
differentiation203. On the other hand, it has been also
shown that, for some classes of graphs, the resistance
t distance converges to the trivial thermodynamic limit
k p−1 + k q−1, where k indicates the node degree204, reduc-
f g h
R1 N R2 P M E ing the usefulness of this metric for the analysis of most
empirical systems.

Geometry induced by communicability. One of the

first metrics based on dynamics was the communica-
bility distance87,205. This metric can be introduced by
Time

starting from the concept of communicability between

two nodes i and j of a network, which is defined as
Gij = exp(A)ij , where A is the underlying adjacency
matrix. From a mathematical point of view, communi-
cability quantifies how well a pair of nodes exchanges
information by all possible walks between them, giv-
ing more weight to the shortest ones, as can be under-
stood by considering the Taylor expansion of the
between two nodes p and q of a network is calculated communicability matrix G (refs87,205).
by assuming fixed resistors on each network edge: the Communicability can also be understood from a
corresponding circuit — under the further assumption physical perspective. Consider a network where nodes
that p and q are directly connected by a battery to allow are harmonic oscillators and links are springs, and the
electric currents to flow — allows for the calculation of system is submerged into a thermal bath with inverse
effective resistances as temperature β = 1/kBT, where kB is the Boltzmann con-
stant. Then communicability provides a representation
R pq = Ω pp + Ωqq − Ω pq − Ωqp , (15) for the thermal Green’s function of the system, indicating

128 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

a Original network b Induced geometry c e Rhesus macaque

1.00 Anatomical Spin glass
Slow
0.75 shrinking

Diffusion
distance
0.50
τ=1
0.25
Fast
shrinking
0
0 10 20 30 40
τ
Diffusion geometry Configurational

d
τ = 10

τ = 20
τ=1

Mean diffusion distance

1.00

0.75

0.50
Data
Temporal
0.25
average
1.0 0 Configuration model
Diffusion
distance

0.5 0 50 100
Modules
0

Fig. 6 | Diffusion geometry of complex networks and multilayer networks. a | Euclidean embedding of a network with
four clusters. The embedding provides a representation of the latent diffusion geometry with Markov time τ = 1. b,c | In the
diffusion space, two nodes from the same functional cluster are (and remain) closer across time, τ, than nodes belonging
to different clusters. Diffusion-​distance matrices corresponding to different times are shown at the bottom and allow
identification of the underlying functional organization at different scales. d | The functional modules that maximize the
average diffusion distance define the mesoscale structure that favours the overall information exchange. The significance
of this structure can be quantified by comparing against the result obtained from a configuration model preserving the
degree distribution of the original data while destroying other correlations. e | Diffusion geometry analysis of the anatomical
connectivity (335 visual, 85 sensorimotor and 43 heteromodal) from 30 visual cortical areas and 15 sensorimotor areas
in the macaque monkey. Clusters identified by structural analysis of the connectome using the spin-​glass approach are
different, as the anatomical organization and the mesoscale organization obtained from the configuration model.
Figure adapted with permission from ref.42, American Physical Society.

how a thermal oscillation propagates between nodes. as effective distance, and can be used to gain insights
The difference between the absorbed and transmit- about reaction–diffusion processes such as the spread-
ted excitation between two nodes due to such thermal ing of infectious diseases through mobility networks40,214.
disturbances is quantified by the communicability Given a network of geographic areas (such as airports)
distance205 and edges encoding direct air traffic (in units of passen-
gers per day) from node i to node j, let Fji indicate the
ξ pq(β) = Gpp(β) + Gqq(β) − 2Gpq(β), (16) corresponding mobility flow. The connectivity matrix
P, with components Pij = Fij/∑iFij, quantifies the fraction
which enables building a hyperspherical embedding of of this flow originating from node j directed towards
a complex network206 at different temperatures. This node i. Despite the structural complexity of the network,
graphical embedding yields a representation of a geom- involving multiple and often redundant pathways for
etry that is able to capture, for instance, spatial efficiency the transmission of contagion phenomena, the effective
of networks207, traffic flows in cities208 and constrained distance defined by
diffusion in coupled networks209 such as multi­layer
systems210. Because the topic is vast and beyond the δij = 1 − log Pij , (17)
scope of the present Review Article, we refer the inter-
ested reader to ref.211 for a thorough review of the mathe­ reveals hidden geometric patterns in which the dynam-
matical and physical properties of communicability ics of epidemic spreading is elegantly mapped into the
distances in complex networks. We also stress that, as propagation of wave fronts with an effective speed. This
well as the communicability distance, other important latent dynamic geometry can be used to better pre-
generalizations of the traditional shortest-​path distance dict the arrival times of empirical contagion processes
have been investigated in mathematics212,213. in distinct geographic areas and to reconstruct, with
reasonable accuracy, the origin of outbreaks (Fig. 5a–c).
Geometry induced by reaction–diffusion. Another Similarly, the collective dynamics of spatiotemporal
important class of geometry that is induced by network propagation of signals on networks can be interpreted
dynamics is obtained by considering a quasi-​metric by looking at the effect of the geometry induced by their

NaTure RevIewS | PhySicS volume 3 | February 2021 | 129

Reviews

dynamics on the definition of the universal temporal As a direct consequence, the mesoscale functional organ-
distance43 ization of the network is mapped into spatial clusters in
the corresponding diffusion manifold, with Markov time
 
 playing the role of a multiresolution parameter, that is, a
L(j → i ) = min  ∑ S pθ 
. (18) temporal length scale playing the dynamic analogue of
Π (j →i ) p ∈ Π (j → i ), p ≠ j 
 
 the shortest-​path distance or the similarity metrics used
in latent spaces.
In equation 18, Π(j → i) = j → q → . . . → i indicates This latent diffusion space has many applications.
the shortest path from node j to node i. The delay τp in Studying it at increasing values of τ allows identification
signal propagation occurring on each node p ∈ Π(j → i) of functional hierarchies at multiple temporal resolu-
is assumed to scale as τp ∼ S pθ , where Si = ∑kAik is the tions, the persistence of which across time identifies the
weighted degree of node i and θ = −2 − Γ(0), and where mesoscale structure that favours the overall information
Γ is a parameter determined by the system’s dynam- exchange, providing the best coarse-​graining of the sys-
ics. This metric yields excellent predictions about the tem in functional modules (Fig. 6a–e). Microscale, meso­
actual propagation times T(j → i) for a range of non- scale and large-​scale structures can be probed for small,
linear dynamic models (Fig. 5d,e). Furthermore, despite increasing and large τ, respectively. Geometry induced
their diversity, disparate propagation patterns actually by diffusive processes reveals physical insights about col-
condense into three distinct dynamic regimes (Fig. 5f–h) lective phenomena in structured populations, by estab-
characterized by the interplay between network paths, lishing a formal relationship with complex dynamics
degree distribution and the interaction dynamics43. responsible for synchronization in the metastable state
In the same spirit, spreading processes on noisy and emergence of consensus. The recent application to
geometric networks215,216 have been recently investi- anatomical connectivity within and between visual cor-
gated to understand how contagion dynamics is driven tical and sensorimotor areas in macaque brain reveals
by the underlying topology41. Noisy geometric networks a hierarchical functional organization of cortical units,
provide a suitable framework to model systems charac- not identified by existing methods and not compatible
terized by both short-​range (that is, large world) and with null models42 (Fig. 6e). The network embedding in a
long-​range (that is, small world) interactions among geometric space induced by diffusion distance, together
nodes. Contagion maps, which have a manifold struc- with statistical data depth, allows for the statistical and
ture that reflects the interplay between local structure, most natural generalization of the concept of median
non-​local structure and the epidemic spreading process, to the realm of complex networks. This generalization
provide a suitable tool to recover the geometric features has advantages for defining the centre of the system and
of a network’s underlying manifold and describe wave percentiles around that centre to identify vertices that
front propagation in the corresponding geometric space. are socially or biologically relevant217.
This geometric framework yields physical insights on More recently, it has been proposed to extend the
contagions by exploiting computational topology and framework to the realm of multilayer networks 218,
allows identification of low-​dimensional structure in capitalizing on the generalization of different random
complex networks41. walk dynamics to multilayer systems210,219,220. In fact,
layer–layer topological correlations might alter infor-
Geometry induced by random search. More recently, mation exchange among state nodes, and the presence
random walk dynamics has been proposed to define a of different interlayer connectivity patterns might lead
diffusion distance between pair of nodes. For a random to distinct geometric regimes. It appears that when
walk that starts in node i with probability 1, the evo- the fraction of interlinks is small, flow is segregated
lution over time τ of the probability to find the walker within layers and the diffusion manifold consists of
in any node is described by a master equation81, the two well-​separated submanifolds, corresponding to the
solution of which is given by p(τ |i ) = e i exp(−τ L), where functional organization of each layer separately, con-
ei ≡ (0, 0, . . . , 1, . . . , 0) is the ith canonical vector in the nected by weak geometric pathways; when the fraction
Euclidean space with dimension N, the size of the sys- is sufficiently high, the flow is integrated, new geometric
tem, and L = I − T is the normalized Laplacian matrix, pathways are made available for information to be
with Iij the Kronecker delta and Tij the probability for exchanged across layers and those submanifolds mix up.
the random walker to move from node i to node j. The Different multilayer diffusion manifolds have been
hidden geometric space induced by the random walk- used to better understand the functional organization
er’s Markov dynamics42 is characterized by the diffusion of multimodal transportation and multiplex social
distance between nodes i and j, defined by systems218.
d i , j (τ ) = ∣∣p(τ i ) − p(τ j )∣∣ , (19) Open questions. The results discussed above make the
geometry induced by network-​driven processes perhaps
which provides the starting point to build diffusion the most suitable framework for several practical appli-
maps which are widely adopted for low-​dimensional cations, such as predicting the time course of dynamics
embedding of high-​dimensional data39, among other for forecasting and control of spreading processes, and
applications. Two nodes are close in their latent diffu- locating their origin. Indeed, diffusion geometry defines
sion space if they are connected by multiple pathways a class of models that have the desirable advantages of
that facilitate information exchange in less than τ steps. being mathematically tractable and readily interpretable.

130 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

The research programme for the future is broad, with of quanta of space — is discrete and the extrinsic cur-
open challenges of both theoretical and practical rele- vature is fuzzy because of Heisenberg’s uncertainty
vance. On the one hand, the topological organization of principle237. The main challenge is to assign a classi-
empirical complex systems can be characterized in terms cal geometrical interpretation to such states. Advances
of hierarchies221,222 and mesoscale structures such as have been made in this direction based on operators
bow-​tie223,224, k-​cores225,226 and core–periphery227,228; thus that quantize both scalar and mean curvature when
it will be interesting to identify and characterize their spin network edges run within the surfaces of the quan-
functional counterparts in terms of diffusion geometry. tized geometry238. In quantum graphity235, the space is
The main challenge is to define mesoscale objects such a dynamical graph that evolves under the action of a
as functional giant components and functional cores. Hamiltonian. In causal dynamical triangulation239,
On the other hand, research is needed to better under­ a non-​perturbative path integral approach is used to
stand the deep relationships among the aspects of build a connection with Hovrava–Lifshitz gravity in
network geometry discussed in this Review Article. 2 + 1 dimensions240. The spectral dimension, defined as
A promising step towards building a theoretical bridge the scaling exponent of the average return probability of
is provided by the concept of communicability, which diffusion processes (such as random walks), is used in
can be understood in purely combinatorial terms as causal dynamical triangulation to measure the effective
an effective pathway — weighting in a specific way the dimension of the underlying geometry241. It can provide
contributions of walks of different lengths — between an interesting bridge to geometry induced by network-​
nodes in the network space. In this regard, it is intimately driven processes, for which one expects that this geome-
related to the concept of geometrical navigability, for try characterizes the underlying diffusion manifold. More
instance, which is determined by selecting topological recently, a model242 in which random graphs dynamically
paths that follow geodesics in the latent space. In gen- self-​assemble into discrete manifold structures has been
eral, those paths happen to be topological shortest paths proposed as an alternative to approaches based on sim-
when networks are sufficiently congruent with the latent plicial complexes and Regge calculus. The Ollivier cur-
space. Note that congruency can be defined in different vature, defined for generic graphs — and similar in spirit
ways, not only in terms of greedy shortest paths21,27,97,229 to Ricci curvature — is used to discretize the Euclidean
but also by taking into account all topological shortest Einstein–Hilbert action and to provide a new ground for
paths230. emergent time mechanisms242.
As for further open challenges, developing RG tech- In network science, a step forwards231,232,243–245 in
niques in the space induced by diffusion distances is a addressing the geometrogenesis problem has been to
fundamental open problem, the solution of which could define models of growing random graphs and simplicial
shed light on the self-​similar symmetries of dynamical complexes. For a wide range of parameters, these models
processes evolving on top of complex networks. Such lead to an effective preferential attachment and, thus,
techniques could further pave the way to the analysis of heterogeneous degree distributions243. In addition, some
co-​existing temporal scales due to the interplay between growth processes can be mapped to trees, leading to
structure and dynamics. Finally, a natural development emergent hyperbolic geometry in the resulting graphs245.
of geometries induced by network-​driven processes is For a more exhaustive introduction to the topic, we refer
the identification and characterization of a more general the interested reader to the review in ref.232.
framework in which diffusive dynamics are replaced by
more complex ones. This advance would improve our Graph curvature. Curvature is one of the most basic
understanding of complex dynamical processes and geometric notions, and a key player in the Einstein–
has the potential to enhance the control and forecasting Hilbert action, whose least-​action variation leads to
of the evolution of empirical systems. Einstein’s equations in general relativity. It is thus not
surprising that graph curvature appears in many flavours
Other flavours of network geometry of geometrogenesis and combinatorial approaches to
Network geometry comes in flavours other than those quantum gravity242. More surprising is that there is not
discussed in detail in this Review Article. Here, we dis- one but many successful attempts to port the notion of
cuss advances and open challenges in some of those curvature to the realm of networks, resulting in many
areas. non-​equivalent definitions of graph curvature105,246–252.
Unfortunately, none of these definitions of graph curva-
Geometrogenesis. Perhaps one of the most funda- ture is rigorously known to converge to any traditional
mental open problems in network geometry is that of curvature of smooth space in the continuum limit of any
geometrogenesis, that is, the emergence of continuous random graph ensemble, with the exception of Ollivier
geometric spaces from discrete combinatorial rules. The curvature of random geometric graphs recently shown to
study of emergent geometries has gained momentum231,232 converge to Ricci curvature in any Riemannian manifold
from its intimate connections with long-​standing combi- under some strong assumptions253.
natoric problems in several approaches to quantum grav- Nevertheless, many of these graph-​curvature defi-
ity, such as causal sets233,234, quantum graphity235 and causal nitions have recently found applications in diverse
dynamical triangulations236. In loop quantum gravity, for applications such as differentiating cancer networks254,
example, the basis of states is formed by spin networks that characterizing human brain structural connectivity172
have support on a graph, determining a sort of quantum and mesoscale characterization of complex networks
geometry in which the intrinsic geometry — consisting based on Ollivier–Ricci curvature255 and Ricci flow256.

NaTure RevIewS | PhySicS volume 3 | February 2021 | 131

Reviews

The topic is receiving increasing attention because com- Topological data analysis. Another interesting bridge
binatorial notions of network curvature have disclosed with topology has been established by generalizing net-
profound connections between network measures, works to higher dimensions via simplicial complexes.
such as the graph Laplacian, and exquisitely geometri- Doing so allows for the application of persistent homol-
cal quantities such as the Laplace–Beltrami operator in ogy methods267 from topological data analysis (TDA)268.
Riemannian manifolds257,258 or the Fisher–Rao metric Persistent homology relies on the filtration of a simpli-
in information geometry259. cial complex to uncover topological features that recur
An important combinatorial definition of curva- over multiple scales and are thus likely to represent some
ture that is directly applicable to networks, is Gromov’s true features of the underlying space. The TDA in gen-
δ-​hyperbolicity88, which has been measured for a vari- eral and persistent homology in particular are vigorous
ety of real-​world networks260–263. It is a rough measure research areas in data science that find applications in
of how far a metric space is from a tree264. It applies problems including spreading processes in networks41
to any metric space, including the metric spaces of and the detection of geometric structure in neural
shortest-path distances and the latent spaces. Any activity269. For thorough reviews of advances in TDA,
hyperbolic space is also δ-​hyperbolic94, but a network we refer the interested reader to refs270,271.
is called δ-​hyperbolic if its shortest-​path metric space is Taken together, the advances in network geometry
such. This terminology often causes confusion because offer a new theoretical framework to gain deep insights
networks in the hyperbolic latent-​space models dis- into the fundamental principles of complex systems and,
cussed in this Review Article, which are often called more generally, into physical reality. It is not excluded
random hyperbolic graphs, are actually unlikely to that the existing results and future advances in this area
be δ-​hyperbolic because their two different limits are not will lead to fruitful cross-​fertilization with other areas
δ-​hyperbolic265,266. However, at present, it is not exactly of physics.
known how δ-​hyperbolic the random hyperbolic graphs
are — another open problem. Published online 29 January 2021

1. Cimini, G. et al. The statistical physics of real-​world 19. Song, C., Havlin, S. & Makse, H. A. Origins of fractality 37. Zheng, M., García-​Pérez, G., Boguñá, M. &
networks. Nat. Rev. Phys. 1, 58–71 (2019). in the growth of complex networks. Nat. Phys. 2, Serrano, M. A. Scaling up real networks by geometric
2. Watts, D. J. & Strogatz, S. H. Collective dynamics 275–281 (2006). branching growth. Preprint at https://arxiv.org/abs/
of ‘small-​world’ networks. Nature 393, 440–442 20. Serrano, M. Á., Krioukov, D. & Boguñá, M. Self-​ 1912.00704 (2019).
(1998). similarity of complex networks and hidden metric 38. Krioukov, D. & Ostilli, M. Duality between equilibrium
3. Barabási, A.-L. & Albert, R. Emergence of scaling in spaces. Phys. Rev. Lett. 100, 078701 (2008). and growing networks. Phys. Rev. E 88, 022808
random networks. Science 286, 509–512 (1999). 21. Krioukov, D., Papadopoulos, F., Kitsak, M., Vahdat, A. (2013).
4. Ravasz, E., Somera, A. L., Mongru, D. A., Oltvai, Z. N. & Boguñá, M. Hyperbolic geometry of complex 39. Coifman, R. R. et al. Geometric diffusions as a tool
& Barabási, A.-L. Hierarchical organization of networks. Phys. Rev. E 82, 036106 (2010). for harmonic analysis and structure definition of data:
modularity in metabolic networks. Science 297, 22. Sorokin, A. P. Social Mobility (Harper, 1927). diffusion maps. Proc. Natl Acad. Sci. USA 102,
1551–1555 (2002). 23. McFarland, D. D. & Brown, D. J. Social Distance as a 7426–7431 (2005).
5. Dorogovtsev, S. N., Goltsev, A. V. & Mendes, J. F. F. Metric: A Systematic Introduction to Smallest Space 40. Brockmann, D. & Helbing, D. The hidden geometry
Critical phenomena in complex networks. Rev. Mod. Analysis 213–252 (John Wiley, 1973). of complex, network-​driven contagion phenomena.
Phys. 80, 1275–1335 (2008). 24. Hoff, P. D., Raftery, A. E. & Handcock, M. S. Science 342, 1337–1342 (2013).
6. Gao, J., Buldyrev, S. V., Stanley, H. E. & Havlin, S. Latent space approaches to social network analysis. 41. Taylor, D. et al. Topological data analysis of contagion
Networks formed from interdependent networks. J. Am. Stat. Assoc. 97, 1090–1098 (2002). maps for examining spreading processes on networks.
Nat. Phys. 8, 40–48 (2012). 25. Boguñá, M., Krioukov, D. & Claffy, K. C. Navigability Nat. Commun. 6, 7723 (2015).
7. D’Souza, R. M. & Nagler, J. Anomalous critical and of complex networks. Nat. Phys. 5, 74–80 (2009). 42. De Domenico, M. Diffusion geometry unravels
supercritical phenomena in explosive percolation. 26. Gulyás, A., Bíró, J. J., Kőrösi, A., Rétvári, G. & the emergence of functional clusters in collective
Nat. Phys. 11, 531–538 (2015). Krioukov, D. Navigable networks as Nash equilibria phenomena. Phys. Rev. Lett. 118, 168301
8. Gao, J., Barzel, B. & Barabási, A.-L. Universal of navigation games. Nat. Commun. 6, 7651 (2015). (2017).
resilience patterns in complex networks. Nature 530, 27. Allard, A. & Serrano, M. Á. Navigable maps 43. Hens, C., Harush, U., Haber, S., Cohen, R. & Barzel, B.
307–312 (2016). of structural brain networks across species. Spatiotemporal signal propagation in complex
9. Bianconi, G. Interdisciplinary and physics challenges PLoS Comput. Biol. 16, e1007584 (2020). networks. Nat. Phys. 10, 403–412 (2019).
of network theory. Europhys. Lett. 111, 56001 (2015). 28. Muscoloni, A. & Cannistraci, C. V. Navigability 44. Feder, J. Fractals (Springer Science & Business Media,
10. Estrada, E. The Structure of Complex Networks: evaluation of complex networks by greedy routing 2013).
Theory and Applications (Oxford Univ. Press, 2012). efficiency. Proc. Natl Acad. Sci. USA 116, 1468–1469 45. Frisch, U. & Kolmogorov, A. N. Turbulence: The Legacy
11. Garlaschelli, D., Capocci, A. & Caldarelli, G. (2019). of AN Kolmogorov (Cambridge Univ. Press, 1995).
Self-​organized network evolution coupled to extremal 29. Zuev, K., Boguñá, M., Bianconi, G. & Krioukov, D. 46. Cardy, J. Scaling and Renormalization in Statistical
dynamics. Nat. Phys. 3, 813–817 (2007). Emergence of soft communities from geometric Physics Vol. 5 (Cambridge Univ. Press, 1996).
12. Garlaschelli, D. & Loffredo, M. Generalized Bose– preferential attachment. Sci. Rep. 5, 9421 (2015). 47. Lesne, A. & Laguës, M. Scale Invariance: From Phase
Fermi statistics and structural correlations in weighted 30. García-​Pérez, G., Boguñá, M., Allard, A. & Transitions to Turbulence (Springer Science & Business
networks. Phys. Rev. Lett. 102, 038701 (2009). Serrano, M. Á. The hidden hyperbolic geometry Media, 2011).
13. Kalinin, N. et al. Self-​organized criticality and pattern of international trade: World Trade Atlas 1870–2013. 48. Bak, P. How Nature Works: The Science of Self-​
emergence through the lens of tropical geometry. Sci. Rep. 6, 33441 (2016). organized Criticality (Springer Science & Business
Proc. Natl Acad. Sci. USA 115, E8135–E8142 31. Wang, Z., Li, Q., Jin, F., Xiong, W. & Wu, Y. Hyperbolic Media, 2013).
(2018). mapping of complex networks based on community 49. Carlson, J. M. & Doyle, J. Complexity and robustness.
14. Song, C., Havlin, S. & Makse, H. A. Self-​similarity information. Physica A 455, 104–119 (2016). Proc. Natl Acad. Sci. USA 99, 2538–2545 (2002).
of complex networks. Nature 433, 392–395 32. García-​Pérez, G., Serrano, M. Á. & Boguñá, M. Soft 50. Bollobás, B. Modern Graph Theory Vol. 184 (Springer
(2005). communities in similarity space. J. Stat. Phys. 173, Science & Business Media, 2013).
15. Gallos, L. K., Song, C., Havlin, S. & Makse, H. A. 775–782 (2018). 51. Chung, F. & Lu, L. The average distances in random
Scaling theory of transport in complex biological 33. Muscoloni, A. & Cannistraci, C. Leveraging the graphs with given expected degrees. Proc. Natl Acad.
networks. Proc. Natl Acad. Sci. USA 104, 7746–7751 nonuniform PSO network model as a benchmark for Sci. USA 99, 15879–82 (2002).
(2007). performance evaluation in community detection and 52. Cohen, R. & Havlin, S. Scale-​free networks are
16. Condamin, S., Bénichou, O., Tejedor, V., Voituriez, R. link prediction. New J. Phys. 20, 063022 (2018). ultrasmall. Phys. Rev. Lett. 90, 058701 (2003).
& Klafter, J. First-​passage times in complex scale-​ 34. García-​Pérez, G., Boguñá, M. & Serrano, M. Á. 53. Dorogovtsev, S. N., Mendes, J. F. F. & Samukhin, A. N.
invariant media. Nature 450, 77–80 (2007). Multiscale unfolding of real networks by geometric Metric structure of random networks. Nucl. Phys. B
17. Radicchi, F., Ramasco, J., Barrat, A. & Fortunato, S. renormalization. Nat. Phys. 14, 583–589 (2018). 653, 307–422 (2003).
Complex networks renormalization: flows and fixed 35. Papadopoulos, F., Kitsak, M., Serrano, M. Á., 54. Mandelbrot, B. B. The Fractal Geometry of Nature
points. Phys. Rev. Lett. 101, 148701 (2008). Boguñá, M. & Krioukov, D. Popularity versus similarity Vol. 2 (WH Freeman, 1982).
18. Rozenfeld, H. D., Song, C. & Makse, H. A. in growing networks. Nature 489, 537–540 (2012). 55. Harary, F. Graph Theory (Addison-​Wesley, 1994).
Small-​world to fractal transition in complex networks: 36. Zuev, K., Papadopoulos, F. & Krioukov, D. Hamiltonian 56. Wiener, H. Structural determination of paraffin
a renormalization group approach. Phys. Rev. Lett. dynamics of preferential attachment. J. Phys. A 49, boiling points. J. Am. Chem. Soc. 69, 17–20
104, 025701 (2010). 105001 (2016). (1947).

132 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

57. Havlin, S., Trus, B. & Stanley, H. Cluster-​growth model 89. Nekrashevych, V. Hyperbolic spaces from self-​similar networks. Random Struct. Algorithms 49, 65–94
for branched polymers that are “chemically linear”. group actions. Algebra Discret. Math. 2, 77–86 (2016).
Phys. Rev. Lett. 53, 1288–1291 (1984). (2003). 121. Bringmann, K., Keusch, R., Lengler, J., Maus, Y.
58. Wilson, K. G. The renormalization group and critical 90. Nekrashevych, V. Self-​Similar Groups (Mathematical & Molla, A. R. Greedy routing and the algorithmic
phenomena. Rev. Mod. Phys. 55, 583–600 (1983). Surveys and Monographs Vol. 117, American small-​world phenomenon. In Proc. ACM Symp. Princ.
59. Efrati, E., Wang, Z., Kolan, A. & Kadanoff, L. P. Mathematical Society, 2005). Distrib. Comput. - PODC’17 371–380 (ACM Press,
Real-​space renormalization in statistical mechanics. 91. Furstenberg, H. Ergodic Theory and Fractal Geometry 2017).
Rev. Mod. Phys. 86, 647 (2014). Vol. 120 (American Mathematical Society, 2014). 122. Abdullah, M. A., Fountoulakis, N. & Bode, M. Typical
60. Gallos, L. K., Song, C. & Makse, H. A. Scaling of 92. Furstenberg, H. & Weiss, B. Markov processes and distances in a geometric model for complex networks.
degree correlations and its influence on diffusion in Ramsey theory for trees. Comb. Probab. Comput. 12, Internet Math. https://doi.org/10.24166/im.13.2017
scale-​free networks. Phys. Rev. Lett. 100, 248701 547–563 (2003). (2017).
(2008). 93. Dorogovtsev, S. N. & Mendes, J. F. F. Evolution 123. Bläsius, T., Friedrich, T. & Krohmer, A. Cliques in
61. Bunde, A. & Havlin, S. Fractals and Disordered of networks. Adv. Phys. 51, 1079–1187 (2002). hyperbolic random graphs. Algorithmica 80,
Systems (Springer Science & Business Media, 2012). 94. Buyalo, S. & Schroeder, V. Elements of Asymptotic 2324–2344 (2018).
62. Goh, K.-I., Salvi, G., Kahng, B. & Kim, D. Skeleton and Geometry (European Mathematical Society Publishing 124. Fountoulakis, N. & Müller, T. Law of large numbers
fractal scaling in complex networks. Phys. Rev. Lett. House, 2007). for the largest component in a hyperbolic model of
96, 018701 (2006). 95. Grigorchuk, R., Nekrashevych, V. & Šunić, Z. From complex networks. Ann. Appl. Probab. 28, 607–650
63. Kim, J. et al. Fractality in complex networks: critical Self-​Similar Groups to Self-​Similar Sets and Spectra (2018).
and supercritical skeletons. Phys. Rev. E 75, 016110 175–207 (Birkhäuser, 2015). 125. Friedrich, T. & Krohmer, A. On the diameter of
(2007). 96. Boguñá, M. & Krioukov, D. Navigating ultrasmall hyperbolic random graphs. SIAM J. Discret. Math. 32,
64. Yook, S.-H., Radicchi, F. & Meyer-​Ortmanns, H. worlds in ultrashort time. Phys. Rev. Lett. 102, 1314–1334 (2018).
Self-​similar scale-​free networks and disassortativity. 058701 (2009). 126. Kiwi, M. & Mitsche, D. Spectral gap of random
Phys. Rev. E 72, 045105 (2005). 97. Boguñá, M., Papadopoulos, F. & Krioukov, D. hyperbolic graphs and related parameters.
65. Galvão, V. et al. Modularity map of the network of Sustaining the Internet with hyperbolic mapping. Ann. Appl. Probab. 28, 941–989 (2018).
human cell differentiation. Proc. Natl Acad. Sci. USA Nat. Commun. 1, 62 (2010). 127. Bringmann, K., Keusch, R. & Lengler, J. Geometric
107, 5750–5755 (2010). 98. Penrose, M. Random Geometric Graphs (Oxford Univ. inhomogeneous random graphs. Theor. Comput. Sci.
66. Newman, M. E. J. Modularity and community Press, 2003). 760, 35–54 (2019).
structure in networks. Proc. Natl Acad. Sci. USA 103, 99. Boguñá, M., Krioukov, D., Almagro, P. & Serrano, M. Á. 128. Müller, T. & Staps, M. The diameter of KPKVB random
8577–8582 (2006). Small worlds and clustering in spatial networks. graphs. Adv. Appl. Probab. 51, 358–377 (2019).
67. Fortunato, S. Community detection in graphs. Phys. Rev. Res. 2, 023040 (2020). 129. Papadopoulos, F., Psomas, C. & Krioukov, D. Network
Phys. Rep. 486, 75–174 (2010). 100. van der Hoorn, P., Lippner, G. & Krioukov, D. Sparse mapping by replaying hyperbolic growth. IEEE/ACM
68. Stanley, H. E. Scaling, universality, and maximum-​entropy random graphs with a given power-​ Trans. Netw. 23, 198–211 (2015).
renormalization: three pillars of modern critical law degree distribution. J. Stat. Phys. 173, 806–844 130. Papadopoulos, F., Aldecoa, R. & Krioukov, D.
phenomena. Rev. Mod. Phys. 71, S358–S366 (2018). Network geometry inference using common neighbors.
(1999). 101. Voitalov, I., van der Hoorn, P., van der Hofstad, R. Phys. Rev. E 92, 022807 (2015).
69. Radicchi, F., Barrat, A., Fortunato, S. & Ramasco, J. & Krioukov, D. Scale-​free networks well done. 131. Blasius, T., Friedrich, T., Krohmer, A. & Laue, S.
Renormalization flows in complex networks. Phys. Rev. Res. 1, 033034 (2019). Efficient embedding of scale-​free graphs in the
Phys. Rev. E 79, 026104 (2009). 102. Cannon, J., Floyd, W., Kenyon, R. & Parry, W. hyperbolic plane. IEEE/ACM Trans. Netw. 26,
70. Stanley, H. E. Introduction to Phase Transitions and Hyperbolic Geometry 59–116 (MSRI, 1997). 920–933 (2018).
Critical Phenomena (Oxford Univ. Press, 1971). 103. Matoušek, J. in Lectures on Discrete Geometry. 132. Muscoloni, A., Thomas, J. M., Ciucci, S., Bianconi, G.
71. Kleinberg, J. Navigation in a small world. Nature 406, Graduate Texts in Mathematics Vol 212. & Cannistraci, C. V. Machine learning meets complex
845 (2000). (ed. Matoušek J.) Ch. 15 (Springer, 2002). networks via coalescent embedding in the hyperbolic
72. Viswanathan, G. M. et al. Optimizing the success of 104. Munzner, T. Exploring large graphs in 3D hyperbolic space. Nat. Commun. 8, 1615 (2017).
random searches. Nature 401, 911–914 (1999). space. IEEE Comput. Graph. Appl. 18, 18–23 (1998). 133. Alanis-​Lobato, G., Mier, P. & Andrade-​Navarro, M. A.
73. Strogatz, S. H. Complex systems: Romanesque 105. Aste, T., Dimatteo, T. & Hyde, S. Complex networks Efficient embedding of complex networks to hyperbolic
networks. Nature 433, 365–366 (2005). on hyperbolic surfaces. Physica A 346, 20–26 space via their Laplacian. Sci. Rep. 6, 30108 (2016).
74. Dorogovtsev, S. N. & Mendes, J. F. F. Evolution of (2005). 134. Wang, Z., Wu, Y., Li, Q., Jin, F. & Xiong, W. Link
Networks: From Biological Nets to the Internet and 106. Lohsoonthorn, P. Hyperbolic Geometry of Networks. prediction based on hyperbolic mapping with
WWW (Oxford Univ. Press, 2003). PhD thesis, USC (2003). community structure for complex networks. Physica A
75. Rozenfeld, H. D., Gallos, L. K., Song, C. & Makse, H. A. 107. Jonckheere, E., Lohsoonthorn, P. & Bonahon, F. 450, 609–623 (2016).
in Encyclopedia of Complexity and Systems Science Scaled Gromov hyperbolic graphs. J. Graph. Theory 135. Muscoloni, A. & Cannistraci, C. V. Minimum curvilinear
(ed. Meyers R.) 3924–3943 (Springer, 2009). 57, 157–180 (2008). automata with similarity attachment for network
76. Kitano, H. Systems biology: a brief overview. Science 108. Kleinberg, R. Geographic routing using hyperbolic embedding and link prediction in the hyperbolic space.
295, 1662–1664 (2002). space. In IEEE INFOCOM 2007 - 26th IEEE Int. Preprint at https://arxiv.org/abs/1802.01183 (2018).
77. Jin, Y., Turaev, D., Weinmaier, T., Rattei, T. & Conf. Comput. Commun. 1902–1909 (IEEE, 136. Alanis-​Lobato, G., Mier, P. & Andrade-​Navarro, M. A.
Makse, H. A. The evolutionary dynamics of protein– 2007). Manifold learning and maximum likelihood estimation
protein interaction networks inferred from the 109. Ratcliffe, J. Foundations of Hyperbolic Manifolds for hyperbolic network embedding. Appl. Netw. Sci. 1,
reconstruction of ancient networks. PLoS ONE 8, (Springer, 2006). 10 (2016).
e58134 (2013). 110. Dall, J. & Christensen, M. Random geometric graphs. 137. Garcia-​Perez, G., Allard, A., Serrano, M. A. &
78. Villani, C. Optimal Transport: Old and New Vol. 338 Phys. Rev. E 66, 016121 (2002). Boguna, M. Mercator: uncovering faithful hyperbolic
(Springer Science & Business Media, 2008). 111. Krioukov, D. et al. Network cosmology. Sci. Rep. 2, embeddings of complex networks. New J. Phys. 21,
79. Ben-​Avraham, D. & Havlin, S. Diffusion and Reactions 793 (2012). 123033 (2019).
in Fractals and Disordered Systems (Cambridge Univ. 112. Allard, A., Serrano, M. Á., García-​Pérez, G. & 138. Kitsak, M., Voitalov, I. & Krioukov, D. Link prediction
Press, 2000). Boguñá, M. The geometric nature of weights in real with hyperbolic geometry. Phys. Rev. Res. 2, 043113
80. Haynes, C. P. & Roberts, A. P. Generalization of the complex networks. Nat. Commun. 8, 14103 (2017). (2020).
fractal Einstein law relating conduction and diffusion 113. Kleineberg, K.-K., Boguñá, M., Ángeles Serrano, M. 139. Serrano, M. Á., Boguñá, M. & Sagués, F. Uncovering
on networks. Phys. Rev. Lett. 103, 020601 (2009). & Papadopoulos, F. Hidden geometric correlations in the hidden geometry behind metabolic networks.
81. Masuda, N., Porter, M. A. & Lambiotte, R. Random real multiplex networks. Nat. Phys. 12, 1076–1081 Mol. Biosyst. 8, 843–850 (2012).
walks and diffusion on networks. Phys. Rep. 716, (2016). 140. Kleineberg, K.-K. Metric clusters in evolutionary
1–58 (2017). 114. Kleineberg, K.-K., Buzna, L., Papadopoulos, F., games on scale-​free networks. Nat. Commun. 8,
82. Gallos, L. K., Potiguar, F. Q., Andrade Jr, J. S. & Boguñá, M. & Serrano, M. A. Geometric correlations 1888 (2017).
Makse, H. A. IMDb network revisited: unveiling fractal mitigate the extreme vulnerability of multiplex 141. García-​Pérez, G., Aliakbarisani, R., Ghasemi, A. &
and modular properties from a typical small-​world networks against targeted attacks. Phys. Rev. Lett. Serrano, M. Á. Precision as a measure of predictability
network. PLoS ONE 8, e66443 (2013). 118, 218301 (2017). of missing links in real networks. Phys. Rev. E 101,
83. Gallos, L. K., Makse, H. A. & Sigman, M. A small world 115. Muscoloni, A. & Cannistraci, C. V. A nonuniform 052318 (2020).
of weak ties provides optimal global integration of popularity-​similarity optimization (nPSO) model to 142. Kerrache, S., Alharbi, R. & Benhidour, H. A scalable
self-​similar modules in functional brain networks. efficiently generate realistic complex networks with similarity–popularity link prediction method. Sci. Rep.
Proc. Natl Acad. Sci. USA 109, 2825–2830 (2012). communities. New J. Phys. 20, 052002 (2018). 10, 6394 (2020).
84. Avena-​Koenigsberger, A., Misic, B. & Sporns, O. 116. Maldacena, J. The large N limit of superconformal 143. Voitalov, I., Aldecoa, R., Wang, L. & Krioukov, D.
Communication dynamics in complex brain networks. field theories and supergravity. Adv. Theor. Math. Geohyperbolic routing and addressing schemes. ACM
Nat. Rev. Neurosci. 19, 17–33 (2017). Phys. 2, 231–252 (1998). SIGCOMM Comput. Commun. Rev 47, 11–18 (2017).
85. Gallos, L. K., Song, C. & Makse, H. A. A review of 117. Gugelmann, L., Panagiotou, K. & Peter, U. Random 144. Zheng, M., Allard, A., Hagmann, P., Alemán-​Gómez, Y.
fractality and self-​similarity in complex networks. Hyperbolic Graphs: Degree Sequence and Clustering & Serrano, M. Á. Geometric renormalization unravels
Physica A 386, 686–691 (2007). 573–585 (Springer, 2012). self-​similarity of the multiscale human connectome.
86. Bloch, A. On methods for the construction of networks 118. Fountoulakis, N. On a geometrization of the Chung–Lu Proc. Natl Acad. Sci. USA 117, 20244–20253
dual to non-​planar networks. Proc. Phys. Soc. 58, model for complex networks. J. Complex Netw. 3, (2020).
677–694 (1946). 361–387 (2015). 145. Cacciola, A. et al. Coalescent embedding in the
87. Estrada, E. The communicability distance in graphs. 119. Candellero, E. & Fountoulakis, N. Clustering and the hyperbolic space unsupervisedly discloses the hidden
Linear Algebra Appl. 436, 4317–4328 (2012). hyperbolic geometry of complex networks. Internet geometry of the brain. Preprint at https://arxiv.org/
88. Gromov, M. in Essays in Group Theory Math. 12, 2–53 (2016). abs/1705.04192 (2017).
(ed. Gersten S. M.) 75–263 (Mathematical Sciences 120. Bode, M., Fountoulakis, N. & Müller, T. The probability 146. Faqeeh, A., Osat, S. & Radicchi, F. Characterizing
Research Institute Publications Vol. 8, Springer, 1987). of connectivity in a hyperbolic model of complex the analogy between hyperbolic embedding and

NaTure RevIewS | PhySicS volume 3 | February 2021 | 133

Reviews

community structure of complex networks. In Proc. Twelfth Work Graph-​Based Methods Nat. 206. Estrada, E., Sánchez-​Lirola, M. & De La Peña, J. A.
Phys. Rev. Lett. 121, 098301 (2018). Lang. Process 59–69 (Association for Computational Hyperspherical embedding of graphs and networks
147. Boguñá, M., Pastor-​Satorras, R. & Vespignani, A. Linguistics, 2018). in communicability spaces. Discret. Appl. Math. 176,
Cut-​offs and finite size effects in scale-​free networks. 176. Ganea, O.-E., Bécigneul, G. & Hofmann, T. Hyperbolic 53–77 (2014).
Eur Phys. J. B 38, 205–209 (2004). entailment cones for learning hierarchical embeddings. 207. Estrada, E. & Hatano, N. Communicability angle
148. Catanzaro, M., Boguñá, M. & Pastor-​Satorras, R. In Proc 35th Int. Conf. Mach. Learn. (eds Dy, J., & and the spatial efficiency of networks. SIAM Rev. 58,
Generation of uncorrelated random scale-​free Krause, A.) 1646–1655 (PMLR, 2018). 692–715 (2016).
networks. Phys. Rev. E 71, 027103 (2005). 177. Ganea, O.-E., Bécigneul, G. & Hofmann, T. Hyperbolic 208. Akbarzadeh, M. & Estrada, E. Communicability
149. Korman, A. & Peleg, D. Dynamic Routing Schemes for neural networks. Adv. Neural Inf. Process. Syst. (2018). geometry captures traffic flows in cities. Nat. Hum.
General Graphs 619–630 (Lecture Notes in Computer 178. Nickel, M. & Kiela, D. Learning continuous hierarchies Behav. 2, 645–652 (2018).
Science Vol. 4051, Springer, 2006). in the Lorentz model of hyperbolic geometry. In 209. Estrada, E. Communicability geometry of multiplexes.
150. Serrano, M. Á., Krioukov, D. & Boguñá, M. Percolation Int. Conf. Mach. Learn. (eds Dy, J., & Krause, A.) New J. Phys. 21, 015004 (2018).
in self-​similar networks. Phys. Rev. Lett. 106, 048701 3776–3785 (PMLR, 2018). 210. De Domenico, M., Granell, C., Porter, M. A. &
(2011). 179. Ovinnikov, I. Poincaré Wasserstein autoencoder. Arenas, A. The physics of spreading processes
151. Cohen, R., Erez, K., Ben-​Avraham, D. & Havlin, S. Preprint at https://arxiv.org/abs/1901.01427 (2019). in multilayer networks. Nat. Phys. 12, 901–906
Resilience of the internet to random breakdowns. 180. Sala, F. et al. Representation tradeoffs for hyperbolic (2016).
Phys. Rev. Lett. 85, 4626–4628 (2000). embeddings. In Int. Conf. Mach. Learn. (eds Dy, J., 211. Estrada, E., Hatano, N. & Benzi, M. The physics of
152. Pastor-​Satorras, R. & Vespignani, A. Epidemic & Krause, A.) 4460–4469 (PMLR, 2018). communicability in complex networks. Phys. Rep.
spreading in scale-​free networks. Phys. Rev. Lett. 86, 181. Gulcehre, C. et al. Hyperbolic attention networks. 514, 89–119 (2012).
3200–3203 (2001). In Int. Conf. Learn. Represent ICLR 2019 (OpenReview. 212. Chebotarev, P. A class of graph–geodetic distances
153. Bianconi, G. Mean field solution of the Ising model net, 2019). generalizing the shortest-​path and the resistance
on a Barabási–Albert network. Phys. Lett. A 303, 182. Chami, I., Ying, R., Ré, C. & Leskovec, J. Hyperbolic distances. Discret. Appl. Math. 159, 295–302
166–168 (2002). graph convolutional neural networks. In Adv. Neural (2011).
154. Dorogovtsev, S. N., Goltsev, A. V. & Mendes, J. F. F. Inf. Process. Syst. (Wallach, H. et al) 4868–4879 213. Chebotarev, P. The walk distances in graphs.
Ising model on networks with an arbitrary distribution (2019). Discret. Appl. Math. 160, 1484–1500 (2012).
of connections. Phys. Rev. E 66, 016104 (2002). 183. Liu, Q., Nickel, M. & Kiela, D. Hyperbolic graph 214. Iannelli, F., Koher, A., Brockmann, D., Hoevel, P.
155. Leone, M., Vázquez, A., Vespignani, A. & Zecchina, R. neural networks. In Adv. Neural Inf. Process. Syst. & Sokolov, I. M. Effective distances for epidemics
Ferromagnetic ordering in graphs with arbitrary (Wallach, H. et al) 8230–8241 (2019). spreading on complex networks. Phys. Rev. E 95,
degree distribution. Eur. Phys. J. B 28, 191–197 184. Suzuki, A., Wang, J., Tian, F., Nitanda, A. & 012313 (2016).
(2002). Yamanishi, K. Hyperbolic ordinal embedding. 215. Barthélemy, M. Spatial networks. Phys. Rep. 499,
156. Kadanoff, L. P. Statistical Physics: Statics, Dynamics In Proc. Elev. Asian Conf. Mach. Learn (Lee, W. S., 1–101 (2011).
and Renormalization (World Scientific, 2000). & Suzuki, T.) 1065–1080 (PMLR, 2019). 216. Bonamassa, I., Gross, B., Danziger, M. M. & Havlin, S.
157. Arenas, A., Díaz-​Guilera, A., Kurths, J., Moreno, Y. 185. Tifrea, A., Bécigneul, G. & Ganea, O.-E. Poincaré Critical stretching of mean-​field regimes in spatial
& Zhou, C. Synchronization in complex networks. GloVe: hyperbolic word embeddings. In Int. Conf. networks. Phys. Rev.Lett. 123, 088301 (2019).
Phys. Rep. 469, 93–153 (2008). Learn. Represent (OpenReview.net, 2019). 217. Bertagnolli, G., Agostinelli, C. & Domenico, M. D.
158. van den Heuvel, M. P., Kahn, R. S., Goni, J. & 186. Orsini, C. et al. Quantifying randomness in real Network depth: identifying median and contours
Sporns, O. High-​cost, high-​capacity backbone for networks. Nat. Commun. 6, 8627 (2015). in complex networks. J. Complex Netw. 8, cnz041
global brain communication. Proc. Natl Acad. Sci. USA 187. Gu, W., Tandon, A., Ahn, Y.-Y. & Radicchi, F. Defining (2019).
109, 11372–11377 (2012). and identifying the optimal embedding dimension 218. Bertagnolli, G. & De Domenico, M. Diffusion geometry
159. Goñi, J. et al. Exploring the morphospace of of networks. Preprint at https://arxiv.org/abs/ of multiplex and interdependent systems. Preprint at
communication efficiency in complex networks. 2004.09928 (2020). https://arxiv.org/abs/2006.13032 (2020).
PLoS ONE 8, e58070 (2013). 188. Chaudhari, P. et al. Entropy-​SGD: biasing gradient 219. De Domenico, M. et al. Mathematical formulation of
160. Fornito, A., Zalesky, A. & Breakspear, M. Graph descent into wide valleys. J. Stat. Mech. Theor. Exp. multi-​layer networks. Phys. Rev. X 3, 041022 (2013).
analysis of the human connectome: promise, progress, 2019, 124018 (2019). 220. De Domenico, M., Solé-​Ribalta, A., Gómez, S. &
and pitfalls. Neuroimage 80, 426–444 (2013). 189. Krioukov, D. Clustering implies geometry in networks. Arenas, A. Navigability of interconnected networks
161. Mišić, B., Sporns, O. & McIntosh, A. R. Phys. Rev. Lett. 116, 208302 (2016). under random failures. Proc. Natl Acad. Sci. USA 111,
Communication efficiency and congestion of signal 190. Holme, P. & Saramäki, J. Temporal networks. 8351–8356 (2014).
traffic in large-​scale brain networks. PLoS Comput. Phy. Rep. 519, 97–125 (2012). 221. Ravasz, E. & Barabási, A.-L. Hierarchical organization
Biol. 10, e1003427 (2014). 191. Kallenberg, O. Probabilistic Symmetries and in complex networks. Phys. Rev. E 67, 026112
162. Roberts, J. A. et al. The contribution of geometry to Invariance Principles (Springer, 2005). (2003).
the human connectome. Neuroimage 124, 379–393 192. Lovász, L. Large Networks and Graph Limits 222. Corominas-​Murtra, B., Goñi, J., Solé, R. V. &
(2016). (American Mathematical Society, 2012). Rodríguez-​Caso, C. On the origins of hierarchy in
163. Avena-​Koenigsberger, A. et al. Path ensembles and 193. Orbanz, P. & Roy, D. M. Bayesian models of graphs, complex networks. Proc. Natl Acad. Sci. USA 110,
a tradeoff between communication efficiency and arrays and other exchangeable random structures. 13316–13321 (2013).
resilience in the human connectome. Brain Struct. IEEE Trans. Pattern Anal. Mach. Intell. 37, 437–461 223. Newman, M. E. J., Strogatz, S. H. & Watts, D. J.
Funct. 222, 603–618 (2017). (2015). Random graphs with arbitrary degree distributions
164. Seguin, C., van den Heuvel, M. P. & Zalesky, A. 194. Aldous, D. J. Representations for partially and their applications. Phys. Rev. E 64, 26118 (2001).
Navigation of brain networks. Proc. Natl Acad. Sci. exchangeable arrays of random variables. J. Multivar. 224. Dorogovtsev, S. N., Mendes, J. F. F. & Samukhin, A. N.
USA 115, 6297–6302 (2018). Anal. 11, 581–598 (1981). Giant strongly connected component of directed
165. Tadić, B., Andjelković, M. & Šuvakov, M. Origin of 195. Hoover, D. N. Relations on Probability Spaces and networks. Phys. Rev. E 64, 025101 (2001).
hyperbolicity in brain-​to-brain coordination networks. Arrays of Random Variables Technical Report 225. Seidman, S. B. Network structure and minimum
Front. Phys. 6, 7 (2018). (Institute for Adanced Study, 1979). degree. Soc. Netw. 5, 269–287 (1983).
166. Avena-​Koenigsberger, A. et al. A spectrum of routing 196. Caron, F. & Fox, E. B. Sparse graphs using 226. Dorogovtsev, S. N., Goltsev, A. V. & Mendes, J. F. F.
strategies for brain networks. PLoS Comput. Biol. 15, exchangeable random measures. J. R. Stat. Soc. Ser. B k-​core organization of complex networks. Phys. Rev.
e1006833 (2019). 79, 1295–1366 (2017). Lett. 96, 040601 (2006).
167. Wang, X. et al. Synchronization lag in post stroke: 197. Veitch, V. & Roy, D. M. Sampling and estimation 227. Holme, P. Core–periphery organization of complex
relation to motor function and structural connectivity. for (sparse) exchangeable graphs. Ann. Stat. 47, networks. Phys. Rev. E 72, 046111 (2005).
Netw. Neurosci. 3, 1121–1140 (2019). 3274–3299 (2019). 228. Rombach, M. P., Porter, M. A., Fowler, J. H. &
168. Betzel, R. F. et al. Generative models of the human 198. Borgs, C., Chayes, J. T., Cohn, H. & Zhao, Y. An Lp Mucha, P. J. Core–periphery structure in networks.
connectome. Neuroimage, 124 1054–1064 (2015). theory of sparse graph convergence I: Limits, sparse SIAM J. Appl. Math. 74, 167–190 (2014).
169. Horvát, S. et al. Spatial embedding and wiring cost random graph models, and power law distributions. 229. Papadopoulos, F., Krioukov, D., Boguna, M. &
constrain the functional layout of the cortical network Trans. Am. Math. Soc. 372, 3019–3062 (2019). Vahdat, A. Greedy forwarding in dynamic scale-​free
of rodents and primates. PLoS Biol. 14, e1002512 199. Janson, S. On edge exchangeable random graphs. networks embedded in hyperbolic metric spaces.
(2016). J. Stat. Phys. 173, 448–484 (2018). In 2010 Proc IEEE INFOCOM 1–9 (IEEE, 2010).
170. Stiso, J. & Bassett, D. S. Spatial embedding imposes 200. Muscoloni, A. & Cannistraci, C. V. Local-​ring network 230. Cannistraci, C. V. & Muscoloni, A. Geometrical
constraints on neuronal network architectures. automata and the impact of hyperbolic geometry in congruence and efficient greedy navigability of
Trends Cogn. Sci. 22, 1127–1142 (2018). complex network link-​prediction. Preprint at https:// complex networks. Preprint at https://arxiv.org/
171. Allen, Q. Y. et al. The intrinsic geometry of the human arxiv.org/abs/1707.09496 (2017). abs/2005.13255 (2020).
brain connectome. Brain Inform. 2, 197–210 (2015). 201. Klein, D. J. & Randić, M. Resistance distance. 231. Wu, Z., Menichetti, G., Rahmede, C. & Bianconi, G.
172. Farooq, H., Chen, Y., Georgiou, T. T., Tannenbaum, A. J. Math. Chem. 12, 81–95 (1993). Emergent complex network geometry. Sci. Rep. 5,
& Lenglet, C. Network curvature as a hallmark of 202. Babić, D., Klein, D., Lukovits, I., Nikolić, S. 10073 (2015).
brain structural connectivity. Nat. Commun. 10, 4937 & Trinajstić, N. Resistance-​distance matrix: 232. Mulder, D. & Bianconi, G. Network geometry and
(2019). a computational algorithm and its application. complexity. J. Stat. Phys. 173, 783–805 (2018).
173. Assaf, Y., Bouznach, A., Zomet, O., Marom, A. & Int. J. Quantum Chem. 90, 166–176 (2002). 233. Rideout, D. & Wallden, P. Spacelike distance from
Yovel, Y. Conservation of brain connectivity and wiring 203. McRae, B. H. Isolation by resistance. Evolution 60, discrete causal order. Class. Quantum Gravity 26,
across the mammalian class. Nat. Neurosci. 23, 1551–1561 (2006). 155013 (2009).
805–808 (2020). 204. Luxburg, U. V., Radl, A. & Hein, M. Getting lost in 234. Dowker, F. Spacetime discreteness, Lorentz invariance
174. Nickel, M. & Kiela, D. Poincaré embeddings for space: large sample analysis of the resistance distance. and locality. J. Phys. Conf. Ser. 306, 012016
learning hierarchical representations. Adv. Neural Adv. Neural Inf. Process. Syst. 23, 2622–2630 (2010). (2011).
Inf. Process. Syst. (2017). 205. Estrada, E. Complex networks in the euclidean space 235. Konopka, T., Markopoulou, F. & Severini, S. Quantum
175. Dhingra, B., Shallue, C. J., Norouzi, M., Dai, A. M. of communicability distances. Phys. Rev. E 85, graphity: a model of emergent locality. Phys. Rev. D
& Dahl, G. E. Embedding text in hyperbolic spaces. 066122 (2012). 77, 104029 (2008).

134 | February 2021 | volume 3 www.nature.com/natrevphys

 Reviews

236. Ambjørn, J., Jurkiewicz, J. & Loll, R. The spectral 253. van der Hoorn, P., Cunningham, W. J., Lippner, G., 270. Otter, N., Porter, M. A., Tillmann, U., Grindrod, P.
dimension of the universe is scale dependent. Trugenberger, C. & Krioukov, D. Ollivier–Ricci & Harrington, H. A. A roadmap for the computation
Phys. Rev. Lett. 95, 171301 (2005). curvature convergence in random geometric graphs. of persistent homology. EPJ Data Sci. 6, 17
237. Rovelli, C. & Speziale, S. On the geometry of loop Preprint at https://arxiv.org/abs/2008.01209 (2020). (2017).
quantum gravity on a graph. Phys. Rev. D 82, 44018 254. Sandhu, R. et al. Graph curvature for differentiating 271. Battiston, F. et al. Networks beyond pairwise
(2010). cancer networks. Sci. Rep. 5, 12323 (2015). interactions: structure and dynamics. Phys. Rep. 874,
238. Grüber, D., Sahlmann, H. & Zilker, T. Geometry and 255. Sia, J., Jonckheere, E. & Bogdan, P. Ollivier–ricci 1–92 (2020).
entanglement entropy of surfaces in loop quantum curvature-​based method to community detection in 272. Song, C., Gallos, L. K., Havlin, S. & Makse, H. A.
gravity. Phys. Rev. D 98, 066009 (2018). complex networks. Sci. Rep. 9, 9800 (2019). How to calculate the fractal dimension of a complex
239. Ambjørn, J., Jordan, S., Jurkiewicz, J. & Loll, R. 256. Ni, C.-C., Lin, Y.-Y., Luo, F. & Gao, J. Community network: the box covering algorithm. J. Stat. Mech.
Second-​order phase transition in causal dynamical detection on networks with Ricci flow. Sci. Rep. 9, Theor. Exp. 2007, P03006 (2007).
triangulations. Phys. Rev. Lett. 107, 211303 (2011). 1–12 (2019). 273. Blondel, V. D., Guillaume, J.-L., Lambiotte, R. &
240. Sotiriou, T. P., Visser, M. & Weinfurtner, S. Spectral 257. Robles-​Kelly, A. & Hancock, E. R. A Riemannian Lefebvre, E. Fast unfolding of communities in large
dimension as a probe of the ultraviolet continuum approach to graph embedding. Pattern Recognit. 40, networks. J. Stat. Mech. Theor. Exp. 2008, P10008
regime of causal dynamical triangulations. 1042–1056 (2007). (2008).
Phys. Rev. Lett. 107, 131303 (2011). 258. Majid, S. Noncommutative Riemannian geometry
241. Loll, R. Quantum gravity from causal dynamical on graphs. J. Geom. Phys. 69, 74–93 (2013). Acknowledgements
triangulations: a review. Class. Quantum Grav. 37, 259. Franzosi, R., Felice, D., Mancini, S. & Pettini, M. S.H. thanks the Israel Science Foundation, ONR, the BIU
013002 (2020). Riemannian-​geometric entropy for measuring network Center for Research in Applied Cryptography and Cyber
242. Kelly, C., Trugenberger, C. A. & Biancalana, F. complexity. Phys. Rev. E 93, 062317 (2016). Security, NSF-​BSF grant number 2019740, and DTRA grant
Self-​assembly of geometric space from random 260. Narayan, O. & Saniee, I. Large-​scale curvature number HDTRA-1-19-1-0016 for financial support. M.B. and
graphs. Class. Quantum Gravity 36, 125012 (2019). of networks. Phys. Rev. E 84, 066108 (2011). M.A.S. acknowledge support from: a James S. McDonnell
243. Bianconi, G. & Rahmede, C. Complex quantum 261. Albert, R., DasGupta, B. & Mobasheri, N. Foundation Scholar Award in Complex Systems; the ICREA
network manifolds in dimension d > 2 are scale-​free. Topological implications of negative curvature for Academia award, funded by the Generalitat de Catalunya;
Sci. Rep. 5, 13979 (2015). biological and social networks. Phys. Rev. E 89, Agencia estatal de investigación project number PID2019-
244. Bianconi, G. & Rahmede, C. Network geometry 032811 (2014). 106290GB-​C22/AEI/10.13039/501100011033; the Spanish
with flavor: from complexity to quantum geometry. 262. Borassi, M., Chessa, A. & Caldarelli, G. Hyperbolicity Ministerio de Ciencia, Innovación y Universidades project
Phys. Rev. E 93, 032315 (2016). measures democracy in real-​world networks. number FIS2016-76830-​C 2-2-​P (AEI/FEDER, UE);
245. Bianconi, G. & Rahmede, C. Emergent hyperbolic Phys. Rev. E 92, 032812 (2015). project Mapping Big Data Systems: embedding large
network geometry. Sci. Rep. 7, 41974 (2017). 263. Tadić, B., Andjelković, M. & Melnik, R. Functional complex networks in low-​d imensional hidden metric
246. Higuchi, Y. Combinatorial curvature for planar graphs. geometry of human connectomes. Sci. Rep. 9, 12060 spaces, Ayudas Fundación BBVA a Equipos de Investigación
J. Graph. Theor. 38, 220–229 (2001). (2019). Científica 2017, and Generalitat de Catalunya grant number
247. Ollivier, Y. Ricci curvature of Markov chains on metric 264. Chatterjee, S. & Sloman, L. Average Gromov 2017SGR1064. D.K. acknowledges support from the NSF
spaces. J. Funct. Anal. 256, 810–864 (2009). hyperbolicity and the Parisi ansatz. Adv. Math 376, grant number IIS-1741355, and the ARO grant numbers
248. Lin, Y., Lu, L. & Yau, S.-T. Ricci curvature of graphs. 107417 (2019). W911NF-16-1-0391 and W911NF-17-1-0491.
Tohoku Math. J. Sec. Ser. 63, 605–627 (2011). 265. Shang, Y. Lack of Gromov-​hyperbolicity in small-​world
249. Sreejith, R., Mohanraj, K., Jost, J., Saucan, E. & networks. Cent. Eur. J. Math. 10, 1152–1158 (2012).
Author contributions
Samal, A. Forman curvature for complex networks. 266. Shang, Y. Non-​Hyperbolicity of random graphs with
The authors contributed equally to all aspects of the Review
J. Stat. Mech. Theor. Exp. 2016, 063206 (2016). given expected degrees. Stoch. Model. 29, 451–462
Article.
250. Sreejith, R., Jost, J., Saucan, E. & Samal, A. (2013).
Systematic evaluation of a new combinatorial 267. Aktas, M. E., Akbas, E. & El Fatmaoui, A. Persistence
curvature for complex networks. Chaos Solitons homology of networks: methods and applications. Competing interests
Fractals 101, 50–67 (2017). Appl. Netw. Sci. 4, 61 (2019). The authors declare no competing interests.
251. Samal, A. et al. Comparative analysis of two 268. Patania, A., Vaccarino, F. & Petri, G. Topological
discretizations of Ricci curvature for complex analysis of data. EPJ Data Sci. 6, 7 (2017). Publisher’s note
networks. Sci. Rep. 8, 8650 (2018). 269. Giusti, C., Pastalkova, E., Curto, C. & Itskov, V. Springer Nature remains neutral with regard to jurisdictional
252. Prokhorenkova, L., Samosvat, E. & van der Hoorn, P. Clique topology reveals intrinsic geometric structure claims in published maps and institutional affiliations.
Global Graph Curvature Vol. 2, 16–35 (Springer in neural correlations. Proc. Natl Acad. Sci. USA 112,
International Publishing, 2020). 13455–13460 (2015). © Springer Nature Limited 2021

NaTure RevIewS | PhySicS volume 3 | February 2021 | 135