PHYSICAL REVIEW X 13, 021038 (2023)

Emergence of Geometric Turing Patterns in Complex Networks

Jasper van der Kolk ,1,2,* Guillermo García-Pérez,3 Nikos E. Kouvaris ,4
M. Ángeles Serrano ,1,2,5,† and Marián Boguñá 1,2,‡
1
Departament de Física de la Matèria Condensada, Universitat de Barcelona,
Martí i Franquès 1, E-08028 Barcelona, Spain
2
Universitat de Barcelona Institute of Complex Systems (UBICS),
Martí i Franquès 1, E-08028 Barcelona, Spain
3
Algorithmiq Ltd, Kanavakatu 3C, FI-00160 Helsinki, Finland
4
Dribia Data Research Sociedad Limitada, Carrer Llacuna 162, 08018 Barcelona, Spain
5
Institució Catalana de Recerca i Estudis Avançats (ICREA), Passeig Lluís Companys 23,
E-08010 Barcelona, Spain

(Received 12 December 2022; revised 5 May 2023; accepted 22 May 2023; published 22 June 2023)

Turing patterns, arising from the interplay between competing species of diffusive particles, have long
been an important concept for describing nonequilibrium self-organization in nature and have been
extensively investigated in many chemical and biological systems. Historically, these patterns have been
studied in extended systems and lattices. Recently, the Turing instability was found to produce topological
patterns in networks with scale-free degree distributions and the small-world property, although with an
apparent absence of geometric organization. While hints of explicitly geometric patterns in simple network
models (e.g., Watts-Strogatz) have been found, the question of the exact nature and morphology of geometric
Turing patterns in heterogeneous complex networks remained unresolved. In this work, we study the Turing
instability in the framework of geometric random graph models, where the network topology is explained
using an underlying geometric space. We demonstrate that not only can geometric patterns be observed,
their wavelength can also be estimated by studying the eigenvectors of the annealed graph Laplacian.
Finally, we show that Turing patterns can be found in geometric embeddings of real networks. These results
indicate that there is a profound connection between the function of a network and its hidden geometry, even
when the associated dynamical processes are exclusively determined by the network topology.
DOI: 10.1103/PhysRevX.13.021038 Subject Areas: Complex Systems,
Interdisciplinary Physics,
Nonlinear Dynamics

I. INTRODUCTION structures. For instance, in the framework of embryonic

morphogenesis, which was Turing’s original motivation, it
In his 1952 seminal paper [1], Turing described how
has been argued that embryos should be thought of as a
competing species of diffusive particles can cause a wide
multicellular network rather than a continuous medium
variety of patterns, starting from a spatially homogeneous
[10]. However, complex network architecture makes the
state. Ever since, reaction-diffusion processes have been
investigation of dynamical processes difficult, and studies
shown to host a wide variety of self-organizational behavior
have often been restricted to small networks [11–13].
[2–9]. Historically, these behaviors have mostly been
In the context of network science, Nakao and Mikhailov
studied on regular lattices or in continuous media, but
made a step forward and demonstrated the existence of the
there has long been evidence for their existence also in
Turing instability in systems of activator-inhibitor species
systems of interconnected elements with complex
diffusing in large random graphs [11]. In that work, the
instability and the corresponding emerging patterns are
*
[email protected] intimately related to the degree heterogeneity usually
†
[email protected] present in real networks and are, thus, purely topological
‡
[email protected] in nature.
Similar results have since been found for directed [14],
Published by the American Physical Society under the terms of non-normal [15], and temporal networks [16,17], as well as
the Creative Commons Attribution 4.0 International license.
Further distribution of this work must maintain attribution to
for networks with higher order interactions such as hyper-
the author(s) and the published article’s title, journal citation, graphs [18] and simplicial complexes [19]. Another exam-
and DOI. ple of nongeometric pattern formation in networks can be

2160-3308=23=13(2)=021038(18) 021038-1 Published by the American Physical Society

JASPER VAN DER KOLK et al. PHYS. REV. X 13, 021038 (2023)

found in the context of multilayered networks, where particles, coexisting with regions of low concentration.
particles diffuse not only within network layers but also When the dynamics takes place on a metric space, these
between them [20–23]. Here, the dynamical parameters inhomogeneities give rise to geometric patterns. It is
can be chosen such that the particle concentration within important to note that the first assumption of this
a layer is homogeneous whereas the distribution between framework is that diffusion alone, without the activation-
layers is heterogeneous. Similar patterns can be found in inhibition mechanism, should lead to a homogeneous
networks displaying community structure [24], where distribution of particles.
concentrations can be chosen to be homogeneous within In continuous media, diffusion is modeled by
a community and heterogeneous between communities. It Brownian particles and, in degree regular lattices, by
has been shown that these types of patterns can also be standard random walks. Diffusion in complex networks
reproduced in empirical networks, provided their com- is, however, different. Indeed, a simple random walk
munity structure is sufficiently strong. process on a heterogeneous network leads to a steady
While hints of explicitly geometric patterns in simple state where the concentration of particles is proportional
network models were found in several works [14,25–27], to the degree of each node and, so, highly hetero-
no comprehensive study of this phenomenon has been geneous. Thus, before adding the activation-inhibition
undertaken. Most importantly, the network models dynamics, we have to carefully define diffusion on
described in these studies are not suitable to describe networks to ensure that a homogeneous concentration of
real systems. In fact, geometric domains in a classical species is a fixed point of the dynamics. In Appendix A 1
sense, typical of Turing patterns in lattices and continu- we show that this can be achieved by a continuous
ous media, have not been observed in real complex time random walk with jump rates proportional to the
networks. This is a consequence of the apparent lack current node’s degree. Then, the evolution equation for the
of geometric structure in small-world networks, causing average concentration of particles at a given node i, ui ðtÞ,
topological distances between nodes—measured by the can be written as
shortest path lengths on the graph—to collapse around
the average value, scaling logarithmically (or slower) X
dui ðtÞ N
with the system size. ¼ −ϵ Lij uj ðtÞ; ð1Þ
This paradigm has changed recently with the develop- dt j¼1
ment of network geometry [28]. Indeed, a latent metric
space with hyperbolic geometry underlying real complex where ϵ is the diffusion coefficient, N the number of nodes,
networks provides the simplest explanation to many of and Lij the Laplacian matrix defined as in Ref. [50]:
their observed topological properties, including hetero-
geneous degree distributions [29–31], clustering [30–34],
Lij ¼ kj δij − aij : ð2Þ
small-worldness [35–37], and percolation [38,39],
spectral [40], and self-similarity properties [29,41,42].
These models have been extended to growing networks In this last equation aij is the adjacency matrix of the
[43], weighted networks [44], multilayer networks network, that hereafter is assumed to describe a single
[45,46], networks with community structure [47–49], connected component.
and they are also the basis for defining a renormalization
group for complex networks [41,42]. A. Activation-inhibition dynamics
In this paper, we show that the Turing instability triggers
In the diffusion process described by Eq. (1), particles
the emergence of purely geometric patterns that become
do not interact when they meet in the same node.
evident in the latent space of real complex networks. We
However, in realistic settings, particles may undergo all
analyze this phenomenon within the framework of network
sorts of reaction processes. In particular, the Turing
geometry and find the relation between the parameters of
instability arises when two different species, U and V,
the dynamical model and the general properties of the
interact upon meeting at the same node, under an
network, such as the congruency of the network with the
activation-inhibition process. Within each node, species
underlying network space and the level of heterogeneity of
U undergoes an autocatalytic process and, simultaneously,
its degree distribution.
favors the increase of the population of species V. At the
same time, species V inhibits the growth of species U even
II. BACKGROUND
though it cannot survive without it. When the average
The Turing instability arises as a consequence of the number of particles per node is large, we can neglect
different diffusivity rates between activator and inhibitor fluctuations in the number of particles and work under the
species. Starting from a homogeneous concentration of mean field approximation. Let ui ðtÞ and vi ðtÞ be the
particles, upon small perturbations, the system evolves average number of particles at node i of species U and V,
towards a stable state with regions of high concentration of respectively. Their evolution equations are given by

021038-2
EMERGENCE OF GEOMETRIC TURING PATTERNS IN COMPLEX … PHYS. REV. X 13, 021038 (2023)

dui ðtÞ XN If we combine this with the fact that each component is
¼ f½ui ðtÞ; vi ðtÞ − ϵ Lij uj ðtÞ; ð3Þ associated to a given node of the network, we find that the
dt j¼1 dynamics will necessarily evolve toward a state with some
of the nodes containing a high concentration of species U
dvi ðtÞ XN
and the rest a low concentration.
¼ g½ui ðtÞ; vi ðtÞ − σϵ Lij vj ðtÞ; ð4Þ
dt In Ref. [11] it was shown that there exist unstable
j¼1
eigenvalues (i.e., λα > 0) whenever
where functions fðu; vÞ and gðu; vÞ represent the activa- 2 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3
tion-inhibition dynamics within the nodes and ϵ and σϵ are
f g 6f g f g 7
the activator and inhibitor diffusion coefficients, respec- σ ≥ σ c ¼ v 2 u 4 u v − 2 − 2 1 − u v 5: ð7Þ
tively. We further assume that the system has a stable f u f v gu f v gu
homogeneous stationary state, that is, usti ¼ ū and vsti ¼ v̄,
for which fðū; v̄Þ ¼ 0 and gðū; v̄Þ ¼ 0, so that ðū; v̄Þ is a
In this regime, all eigenvalues that lie within the interval
stable fixed point of the dynamics. The activation-
Λ ∈ ½Λ− ; Λþ , with
inhibition dynamics and its stability condition impose
constraints on functions fðu; vÞ and gðu; vÞ. In particular,  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
its first derivatives f u ≡ ∂u fðū; v̄Þ, f v ≡ ∂v fðū; v̄Þ, gu ≡ 1
Λ ¼ σf u þ gv  ðσf u − gv Þ2 þ 4σf v gu ; ð8Þ
∂u gðū; v̄Þ, and gv ≡ ∂v gðū; v̄Þ must satisfy f u > 0, f v < 0, 2σϵ
gu > 0, gv < 0, and f u þ gv < 0, f u gv − f v gu > 0 (see
Ref. [11] for further details). become unstable, with the most unstable one being
The Turing instability arises when, due to the differences
 rffiffiffiffiffiffiffiffiffiffiffiffiffi
in the diffusion coefficients of the two species, the fixed 1 −f v gu
point becomes unstable so that any small perturbation Λmax ¼ f u − gv − ð1 þ σÞ : ð9Þ
1−σ σ
drives the system away from it. Interestingly, the conditions
for this to take place depend only on f u , f v , gu , gv and the
The functional dependence of Eqs. (7) and (8) implies
ratio between the diffusion coefficients of the two species
that, for a given set of parameters f u , f v , gu , gv and for one
σ. Similarly to the case of continuous media, where the
particular eigenvalue Λα , it is always possible to select
perturbation around the fixed point can be expanded in
parameters ϵ and σ such that Λα ∈ ½Λ− ; Λþ . In the case that
Fourier modes—the eigenfunctions of the Laplacian ∇2 —
Λþ ≈ Λ− ≈ Λα , one can assume that the differences in the
and where the eigenvalues are related to the wave numbers,
concentration of species in the nodes are encoded in the
here the perturbation can be expanded in terms of the
eigenvector ϕ ⃗ α.
eigenvectors of the graph Laplacian Lij [11],
Interestingly, these results depend only on the first
X
N derivatives of functions fðu; vÞ and gðu; vÞ coupled to
ui ðtÞ − ū ≈ cα eλα t ϕαi ; ð5Þ the diffusion coefficients. Therefore, there exists a whole
α¼1 class of systems where such instability may emerge, from
chemical reactions [51–53] or biological morphogenesis
where cα and λα are some constants that depend on the [3,54] to ecosystems [55–58] and game theory applied to
parameters of the model Eqs. (3) and (4) and on the ecological systems [59,60]. In ecology, one such model is
eigenvalue Λα (similarly for species V). This expansion the Mimura-Murray model [56]. Here, u and v represent
is possible because the eigenvectors of the Laplacian prey and predator densities, respectively, and the dynamics
matrix fϕ ⃗ α ¼ ðϕα ; ϕα ; …; ϕα Þ; α ¼ 1; …; Ng form a com- is governed by the functions
1 2 N
plete orthonormal basis of RN . If λα < 0; ∀ Λα , then  
perturbations around the fixed point are absorbed expo- a þ bu − u2
nentially fast, so that the fixed point is stable. However, if fðu; vÞ ¼ − v u; ð10Þ
c
there exists at least one Λα such that λα > 0, then the fixed
point becomes unstable. In this case, those nodes with
cα ϕαi > 0 will increase the concentration of species U, gðu; vÞ ¼ ½u − ð1 þ dvÞv: ð11Þ
whereas those with negative value will decrease it. Note
that for all eigenvectors associated to nonvanishing eigen- From Eqs. (3), (4), (10), and (11), we see that in the absence
values we have of preys, predators go extinct and in the absence of
predators, preys attain a constant population. In general,
X
N there is a fixed point with positive densities of preys and
ϕαi ¼ 0: ð6Þ predators. In this paper, we use this dynamics as a case
i¼1 study (technical details are given in Appendix A 3).

021038-3
JASPER VAN DER KOLK et al. PHYS. REV. X 13, 021038 (2023)

B. Geometric description of complex networks: choice, the model becomes small-world whenever
The S1 =H2 model 1 < β ≤ 2, or 2 < γ ≤ 3, or both. Additionally, degree
As we have seen, the shape of emerging patterns is heterogeneity is modulated by the exponent γ and the limit
dictated by the arrangements of the signs in the components γ → ∞ is equivalent to a homogeneous distribution
of the eigenvectors of the Laplacian matrix. In continuous of hidden degrees ρðκÞ ¼ δðκ − hkiÞ and a Poisson degree
media or regular lattices, these signs are arranged in the distribution.
metric space in an alternating fashion, so that when the As it is defined, the S1 model represents the only
parameters of the dynamics are tuned to select these maximally random ensemble of geometric graphs that are
eigenvectors as unstable, the concentration of species simultaneously sparse, clustered, small-world, with
also follows the same geometric pattern. In random graphs heterogeneous degree distribution, and with only struc-
models, like the configuration model [61–64] or the tural degree-degree correlations [67]. In the regime β > 1,
Barábasi-Albert model [65], the situation is different. For the connection probability does not depend on the system
these type of models, there is no associated metric space size and, therefore, it generates networks with finite
and the most distinctive feature of nodes is the degree. In clustering in the thermodynamic limit. In this paper, we
this case, the Turing instability leads to patterns that are restrict ourselves to this regime of temperatures. However,
purely topological, with high or low concentration of the model has been extended to the case β < 1, showing
species depending on nodes’ degrees [11]. Real complex interesting quasigeometric properties in the neighborhood
networks, however, are better described by geometric of β ≲ 1 even though clustering vanishes for β ≤ 1 in the
random graph models. In such models, nodes are embedded thermodynamic limit [34]. It is worth mentioning that the
in a metric space and the connection probability depends model is defined in arbitrary dimensions. However, as
on the distances among nodes in this metric space. This shown in Ref. [70], most real networks are well described
approach has led to the development of the field of network by low dimensional spaces. Thus, to keep the analysis
geometry [28], giving rise to the most comprehensive simple, we restrict ourselves to the one-dimensional case.
description of real complex networks. The S1 model is isomorphic to a purely geometric model
To generate geometric networks, we use the S1 =H2 in the hyperbolic plane of constant negative curvature −1,
model [29,30,66], otherwise known as the geometric soft the so-called H2 [30] model. Upon mapping the hidden
configuration model [67]. The model combines a similarity degree κ to a radial coordinate r as
dimension, encoded in a one-dimensional sphere, and a
κ
popularity dimension, quantified by nodes’ degrees. In r ¼ RH2 − 2 ln ; ð13Þ
particular, each node is given a pair of hidden variables κ0
κi ∈ ½κ 0 ; ∞Þ and θi ∈ ½0; 2πÞ. The former accounts for the
ith node’s ensemble average degree, and can be generated the connection probability Eq. (12) becomes
from an arbitrary distribution ρðκÞ. The latter is the 1
angular position of node i on the circle, as an abstraction pij ¼ ðβ=2ÞðdH2 ;ij −RH2 Þ
; ð14Þ
of its position in the similarity space. In the simplest 1þe
version of the model, κ and θ are statistically independent
random variables and θ is homogeneously distributed. where RH2 ¼ 2 ln ðN=μπκ 20 Þ is the radius of a hyperbolic
This final assumption, however, is not critical to our disk and dH2 ;ij is the hyperbolic distance between nodes i
model and, in fact, in real systems the angular distribution and j given by the hyperbolic law of cosines:
is inhomogeneous, inducing the emergence of geometric
communities [47–49,68,69]. The model is fully deter- cosh ðdH2 ;ij Þ ¼ cosh ri cosh rj − sinh ri sinh rj cos Δθij ;
mined by the following connection probability between ð15Þ
two nodes i and j:
where Δθij is the angular separation between the nodes.
1
pij ¼  β ; ð12Þ Because of this isomorphism, we are free to work with one
1þ
xij version or the other indistinctly. In this paper, we use the S1
μκ i κj
model to perform the analytical calculations and the H2
model for visualization.
where xij is the distance between the nodes in the circle of
radius R ¼ N=2π, β > 1 is the inverse of the temperature
of the ensemble [30,34,67], and μ ¼ ðβ=2πhkiÞ sinðπ=βÞ a III. MIMURA-MURRAY DYNAMICS
parameter fixing the average degree hki of the network. ON GEOMETRIC NETWORKS
While the model is defined for arbitrary distributions of To illustrate the role of geometry on the Turing insta-
hidden degrees, here we choose a power law distribution bility, we run the Mimura-Murray dynamics on a real
of the form ρðκÞ ¼ ðγ − 1Þκ γ−1 −γ
0 κ , with κ ≥ κ 0 . With this network, namely that of the connectome of a mouse

021038-4
EMERGENCE OF GEOMETRIC TURING PATTERNS IN COMPLEX … PHYS. REV. X 13, 021038 (2023)

associated to the geometric character of the network. We

do observe, however, topological patterns induced by
degree, with high degree nodes holding a low concen-
tration of species U. These are the type of patterns
documented in Ref. [11]. Instead, in the case of the less
heterogeneous network, we find a clear geometric pattern
associated to the angular coordinates of nodes and very
weak degree-related patterns, similar to the results found
in the real network. These patterns emerge despite the fact
that networks have the small-world property, so that
diffusion induces flows of species between nodes that
are far apart in the underlying metric space. These results
suggest that some eigenvectors of the Laplacian of the
S1 =H2 model have a well-defined periodic structure in the
similarity space. Next, we develop a theoretical frame-
work allowing us to calculate an approximation for the
frequency associated with such eigenvectors.

A. Eigenvectors of the S1 =H2 Laplacian matrix

in the annealed approximation
FIG. 1. Examples of Turing patterns in networks. The first, The goal of this section is to determine the dispersion
(a)–(c), is a real network representing the connectome of a mouse, relation ωðΛÞ between the eigenvalue Λ and the frequency
the second, (d)–(f), a strongly heterogeneous S1 =H2 network of its associated eigenvector. It needs to be stressed that this
with parameters ðN; hki; β; γÞ ¼ ð1000; 50; 1.1; 2.1Þ, and the
is a structural property of the network, and is not con-
third, (g)–(i), a weakly heterogeneous S1 =H2 network with
ditioned by the dynamics on top of the network.
parameters ðN; hki; β; γÞ ¼ ð1000; 50; 2.4; 2.86Þ. In all panels
the colors correspond to the density of the activators in the final Let us first note some important properties of the
stationary state. In panels (a), (d), and (g) nodes are located Laplacian matrix [50]. The Laplacian is a real symmetric
according to their coordinates in hyperbolic space. Panels (b), (e), matrix with real eigenvalues Λ1 ¼ 0 ≤ Λ2 ≤    ≤ ΛN ≤
and (h) show the density of activators as a function of nodes’ 2kc , where kc is the maximum degree of the network. The
angular coordinates, and in panels (c), (f), and (i) the activators’ average of all eigenvalues is equal to the network average
densities are plotted against nodes’ degree ranking. degree, that is,
hΛi ¼ hki; ð16Þ
described in Appendix A 5. Figure 1(a) shows the results on
the concentration of species U in the hyperbolic represen- and the second moment is
tation of the network (see Sec. V for technical details). As
can be seen, there is a clear pattern associated to the hΛ2 i ¼ hk2 i þ hki: ð17Þ
geometric organization of the network, with low concen- These results suggest that eigenvalues are strongly corre-
trations of particles localized in the same angular position lated to the sequence of degrees in the network.
and high concentrations in the rest of the network. In Additionally, as we show in Appendix A 4, in geometric
Figs. 1(b) and 1(c), we also represent the concentration as a random graph models, bounds on the isoperimetric (or
function of the angular coordinate θ and the node index i, Cheeger) constant [71,72] imply that the smallest non-null
respectively. In the latter case the nodes are ordered from eigenvalue—the spectral gap—of the Laplacian approaches
high to low degree. It is important to note that the Mimura- zero in the thermodynamic limit. In this limit, the spectrum
Murray dynamics does not use any information from the is necessarily continuous on the positive real line as long as
underlying metric space, so that the observed geometric the node density along the circle and the connection
pattern is a highly nontrivial result. probability are continuous functions and the degree dis-
To shed light on this problem, we also run the dynamics tribution is unbounded. These facts imply that any eigen-
on networks generated by the S1 =H2 model with a vector frequency given by ωðΛÞ can be accessed by the
heterogeneous degree distribution with exponent γ. system, as any real positive number will necessarily be
Figures 1(d)–1(i) show results of the dynamics on two arbitrarily close to an eigenvalue of the graph Laplacian for
different networks: one highly heterogeneous (γ ¼ 2.1) sufficiently large networks.
and weakly clustered and a second one less heterogeneous Species whose dynamics are governed by the
(γ ¼ 2.84) and highly clustered. In the strongly hetero- Mimura-Murray model diffuse within a network with a
geneous network shown in the figure, there are no patterns disordered but quenched structure, a single realization of

021038-5
JASPER VAN DER KOLK et al. PHYS. REV. X 13, 021038 (2023)

the network ensemble defined by the S1 =H2 model for a where F stands for Fourier transform. Note that μ−1 ¼
given sequence of hidden variables ðκ i ; θi Þ, i ¼ 1; …; N. hkiΨ̂ð0Þ. From Eq. (21), it is easy to see that the integral
Unfortunately, for quenched networks it is not possible to over the spatial coordinate of nontrivial eigenvectors must
get any analytical insight on the structure of the eigen- vanish, that is, ϕ̂ðκ; 0Þ ¼ 0, similarly to the case of the
vectors of the Laplacian matrix. To overcome this problem, quenched Laplacian matrix Eq. (6). Indeed, by setting
we use the annealed approximation, which has been
ω ¼ 0 in Eq. (21), we conclude that ϕ̂ðκ; 0Þ ¼ 0 is the only
extensively used in the literature to tackle a wide variety
solution when Λ ≠ 0. Thus, the annealed approximation
of problems, from neural networks and opinion dynamics
preserves the basic property of the Laplacian matrix given
to epidemic spreading [73–81]. In this approach, the
in Eq. (6).
network structure is resampled from the ensemble at a
rate faster than the diffusion dynamics. This allows us to
B. Homogeneous ensemble
replace the adjacency matrix aij by its ensemble average,
the connection probability pij, and consider the network The analytic solution of Eq. (21) can be found in
not as a quenched system but as a dynamic one. With this particular cases. One such case is a homogeneous ensemble
approximation, the eigenvalue problem of the annealed where all nodes have the same hidden degree, that is,
Laplacian matrix can be written as ρðκÞ ¼ δðκ − hkiÞ, leading to a Poisson degree distribution
X  of average hki. In this situation, the eigenvector is just a
X ϕðκ j ; θj Þ 1 function of the frequency ω and Eq. (21) becomes
 β ¼  β − Λ ϕðκi ; θi Þ; ð18Þ
xij xij
j≠i 1 þ μκ κ j≠i 1 þ μκ κ  
i j i j Ψ̂ðμhki2 ωÞ hki − Λ
− ϕ̂ðωÞ ¼ 0: ð23Þ
where Λ is the eigenvalue and ϕðκ i ; θi Þ the component of Ψ̂ð0Þ hki
the corresponding eigenvector of a node i with hidden
variables ðκi ; θi Þ. Note that, as in the case of the true Assuming that ϕ̂ðωÞ ≠ 0, this equation defines a dispersion
Laplacian matrix, Λ ¼ 0 is an eigenvalue with constant relation between the eigenvalue Λ and its characteristic
eigenvector. frequency ωc as a solution of the transcendent equation:
In the thermodynamic limit, the curvature of the circle  
goes to zero and, thus, nodes become distributed in R1 Λðωc Þ Ψ̂ðμhki2 ωc Þ
¼ 1− : ð24Þ
according to a Poisson point process of density one. In this hki Ψ̂ð0Þ
limit, the distance between two nodes can be evaluated as
xij ¼ jxi − xj j, where xi and xj are the positions of nodes i The characteristic frequency ωc in Eq. (24) is constrained
and j in R1 . Finally, we take the continuum limit in Eq. (18) by the boundary conditions and the discretization of nodes
by replacing the sum over index j by a double integral over in the space. Indeed, since nodes are distributed on the line
nodes coordinates ðκ; xÞ. In this way, Eq. (18) can be at density 1, frequencies above π (and so wavelengths of the
written as the following integral equation: order 1) cannot be detected. Additionally, if the system is
Z ∞ Z ∞ finite of length L ¼ N, frequencies below 2π=N will have
ϕðκ 0 ;x0 Þ
ρðκ 0 Þdκ0 dx0  β ¼ ðκ − ΛÞϕðκ; xÞ: ð19Þ associated wavelengths comparable to the system size and,
κ0 −∞ jx−x0 j
1 þ μκκ0 so, will not be detected either. Therefore, we look for
solutions of Eq. (24) in the domain ωc ∈ ½2π=N; π.
The spatial part of Eq. (19) has the form of a convolution It is illustrative to analyze in detail the homogeneous
integral. Thus, we can take advantage of the convolution case with β ¼ ∞, which corresponds to a connection
theorem for Fourier transforms. By defining the Fourier probability being a step function. In this case, function
transform of the eigenvector as Ψ̂ðzÞ takes the form
Z ∞
ϕ̂ðκ; ωÞ ≡ e−iωx ϕðκ; xÞdx; ð20Þ sin z
Ψ̂ðzÞ ¼ 2 ; ð25Þ
−∞ z
Eq. (19) can be written in Fourier space as and μ ¼ 1=ð2hkiÞ. Combining these results, the dispersion
Z relation defining ωc becomes
∞ κ 0 ρðκ 0 Þ Ψ̂ðμκκ 0 ωÞ κ−Λ
ϕ̂ðκ 0 ; ωÞdκ 0 ¼ ϕ̂ðκ; ωÞ; ð21Þ  
κ0 hki Ψ̂ð0Þ κ Λðωc Þ 2 sin hkiωc
2
¼ 1− : ð26Þ
hki hkiωc
where we define Ψ̂ as
  In the case of β ¼ ∞, nodes connect to nearest neighbors
1
Ψ̂ðzÞ ≡ F ðzÞ; ð22Þ on the line that are below a certain critical distance, so that
1 þ jxjβ the average number of such neighbors is hki. Therefore,

021038-6
EMERGENCE OF GEOMETRIC TURING PATTERNS IN COMPLEX … PHYS. REV. X 13, 021038 (2023)

and become identical as the connectivity increases, in

agreement with the fact that the continuum approximation
becomes exact at infinite average degree.
Despite the good agreement shown in Fig. 2, the
comparison with quenched networks generated by the
S1 =H2 model has to be made with care. Indeed, for
homogeneous networks, the annealed approximation com-
pletely neglects fluctuations of nodes’ degrees. In a
quenched network this is far from true, as the degree
distribution follows a Poisson distribution of average hki.
FIG. 2. Dispersion relation for the annealed approximation and
a degree-regular one-dimensional lattice. Solid orange lines show However, the dispersion relations Eqs. (26) and (27) are
exact results from Eq. (27) and black dashed lines for the strongly dependent on the network connectivity, especially
annealed approximation in Eq. (26) for homogeneous networks in the high frequency domain. Therefore, due to the
with β ¼ ∞. Results are shown for average degrees hki ¼ 4, 8, randomness of nodes’ degrees, it is not possible to
and 12. characterize each eigenvector of the quenched Laplacian
by a unique frequency but by a collection of frequencies
apart from the fluctuations in the number of neighbors, the around a given average.
dispersion relation in Eq. (26) should be equivalent to the We cannot talk then of a deterministic dispersion relation
dispersion relation of a one-dimensional lattice with hki but a fuzzy one where different frequencies may coexist
symmetric nearest neighbors, which reads in the same eigenvector. This will destroy any possible
high frequency periodic pattern, leaving only low fre-
2 X
hki=2
Λ quency eigenvectors (with Λ ≪ hki) visible. The reason
¼1− cos nωc : ð27Þ
hki hki n¼1 is that low frequency modes correspond to long wave-
lengths, so that local fluctuations of degrees become
Figure 2 shows the comparison between Eqs. (26) and (27) irrelevant. This is clearly visible in Figs. 3(a)–3(c), showing
for hki ¼ 4, 8, and 12. Both expressions are very similar three different eigenvectors of the annealed Laplacian

FIG. 3. (a)–(c),(g)–(i) Examples of three eigenvectors of the quenched Laplacian matrix with low, medium, and high associated
frequencies for a single network generated by the homogeneous ensemble with N ¼ 1000 and hki ¼ 20 and β ¼ ∞ (a)–(c) and β ¼ 1.5
(g)–(i). (d)–(f),(j)–(l) Concentration of species U in the Mimura-Murray dynamics for the eigenvalues in the column to the left of
each panel.

021038-7
JASPER VAN DER KOLK et al. PHYS. REV. X 13, 021038 (2023)

The fact that ðβ − 1Þ−1 < 1=2 implies that, in this regime, a
certain eigenvalue leads to a much lower frequency in the
case of small β’s than it does for high β’s, an observation
that is corroborated by Fig. 4(b).
Figures 3(g)–3(l) show the same analysis as in
Figs. 3(a)–3(f) but for the case β ¼ 1.5, so that networks
are deep into the small-world regime, with many long-
range connections among distant nodes in the metric space.
We observe the same qualitative behavior as in the case
β ¼ ∞, except that, as expected, results have more noise.
FIG. 4. (a) Rescaled eigenvalue as a function of the character-
istic frequency in log-log scale. (b) Characteristic frequency as a
Yet, we can clearly identify geometric patterns both in the
function of the rescaled eigenvalue. In both cases hki ¼ 10 and eigenvectors and in the steady state of the Mimura-Murray
several β’s are investigated. dynamics when the dynamic parameters are properly tuned.
However, we notice that, in agreement with our theoretical
prediction in Eq. (29), eigenvalues with similar associated
matrix corresponding to low, medium, and high frequencies eigenvector frequencies as those in Figs. 3(a) and 3(d) are
of a graph generated by the homogeneous ensemble. The significantly larger.
periodic pattern is very clear in the case of low eigenvalue
with very low frequency. In the medium frequency case, we
can still identify a periodic behavior, although distorted by C. Heterogeneous ensemble
noise. In the case of high frequency, periodic behavior is In real networks, the degree distribution is typically
totally absent. We also run the Mimura-Murray dynamics heterogeneous. Therefore, we have to solve the eigenvalue
on the same network selecting parameters’ values such that problem in Eq. (21) for a heterogeneous distribution of
the same eigenvalues used in Figs. 3(a)–3(c) become hidden degrees ρðκÞ. In particular, we choose a scale-free
unstable (see Appendix A 3 for technical details of the distribution ρðκÞ ∝ κ −γ . For this distribution of hidden
dynamics). Figures 3(d)–3(f) show the concentration of degrees, Eq. (21) can be rewritten as
species U in the three cases. We observe that the patterns in Z  
the concentration of species U closely follow the patterns of 1 z0γ−3 ω̃
ðγ − 2Þ Ω̂ Θ̂ðz0 ; ω̃Þdz0 ¼ Θ̂ðz; ω̃Þ; ð30Þ
the corresponding eigenvectors, justifying the relevance 0 1 − Λ̃z 0 zz0
of the structure of eigenvectors for dynamical processes on
networks. where we redefine variables ðκ; Λ; ωÞ as
Based on these observations, we conclude that the
annealed approximation provides good results in quenched κ0 Λ
z≡ ; Λ̃ ≡ ; ω̃ ≡ μκ 20 ω; ð31Þ
networks for small eigenvalues such that Λ ≪ hki, which κ κ0
corresponds to low frequencies, that is, μhki2 ωc ≪ 1. For
arbitrary values of β, we can then take the low frequency and functions as
limit in Eq. (24) to derive the relation between ωc and Λ.  
Using the definition of the Fourier transform, it is easy to Ψ̂ðxÞ κ 0 ω̃
Ω̂ðxÞ ≡ ; Θ̂ðz; ω̃Þ ≡ ð1 − Λ̃zÞϕ̂ ; : ð32Þ
see that for small z, Ψ̂ð0Þ z μκ 20

Ψ̂ðzÞ zβ−1 β≤3 When Λ̃ < 1, the singularity of the kernel in the integral of
1− ∼ ð28Þ Eq. (30) falls outside the domain of integration. In this case,
Ψ̂ð0Þ z2 β > 3:
Eq. (30) can be solved numerically by Gaussian quadrature
as the solution of the system of equations,
Figure 4(a) shows the low frequency limit of the dispersion
relation computed numerically for different values of β,  
Xn
wj zγ−3
j ω̃
which corroborates this scaling behavior. Using this result, ðγ − 2Þ Ω̂ Θ̂ðzj ; ω̃Þ ¼ Θ̂ðzi ; ω̃Þ; ð33Þ
the characteristic frequency scales as j¼1 1 − Λ̃zj z i zj

8  1=ðβ−1Þ
>
> 1 Λ where zi , with i ¼ 1; …; n, are the zeros of the orthogonal
>
> β≤3
>
< hki hki polynomials used in the quadrature, wi their associated
ωc ∼ sffiffiffiffiffiffiffi ð29Þ weights, and n ≫ 1 the number of points within the domain
>
> 1 Λ of integration used to evaluate the integral.
>
> β > 3:
>
: hki hki Equation (33) defines a homogeneous system of linear
equations with kernel matrix KðΛ; ωÞ given by

021038-8
EMERGENCE OF GEOMETRIC TURING PATTERNS IN COMPLEX … PHYS. REV. X 13, 021038 (2023)

FIG. 5. (a) Dispersion relation obtained with Gaussian quad-

rature for β ¼ 2 for various γ’s and (b) the dispersion relation
for γ ¼ 2.5 and varying β. The average degree is hki ¼ 12 in
both panels.

 
wj zγ−3
j ω̃
K ij ≡ ðγ − 2Þ Ω̂ − δij : ð34Þ
1 − Λ̃zj zi zj

A nonzero solution is found when det½KðΛ; ωÞ ¼ 0,

defining thus the dispersion relation Λðωc Þ. We use this
condition to compute numerically the dispersion relation
for different values of β and γ. Figure 5(a) shows the
results for β ¼ 2 and different values of γ and Fig. 5(b)
shows the results for γ ¼ 2.5 and different values of β.
The dispersion relations are qualitatively similar to the FIG. 6. (a)–(c) Examples of three eigenvectors with low,
homogeneous case with β ≤ 2. The main difference medium, and high associated frequencies for a single network
generated by the heterogeneous ensemble with N ¼ 1000,
appears in the asymptotic behavior of Λðωc Þ, which
β ¼ 2.5, γ ¼ 2.5, and hki ¼ 20. (d)–(f) Concentration of species
approaches κ0 at high frequencies. Given the relation U in the Mimura-Murray dynamics for the eigenvalues of the left-
between the average degree and κ 0 , this result implies that hand column.
the ratio Λ=hki approaches ðγ − 2Þ=ðγ − 1Þ, a value that is
below 1. Therefore, the condition hΛi ¼ hki implies that
the number of eigenvalues with periodic behavior can only determining when an eigenvector is periodic and, if so,
account for a small fraction of all eigenvalues, and this what its dominant frequency is. To do so, we perform a
fraction vanishes when γ ≈ 2, as shown in Fig. S1 of discrete Fourier transform (DFT) on the eigenvectors and
Supplemental Material [82]. We conjecture that the apply a statistical p-value test with the null hypothesis
remaining eigenvalues, with Λ > κ0 , are those which being that the Fourier signal is the result of white noise (see
cannot be characterized by a single frequency. Appendix A 2 for further details). We apply this procedure
Figure 6 shows the same analysis performed in Fig. 3 but to all eigenvalues and select those with statistically sig-
for heterogeneous networks with γ ¼ 2.5 and β ¼ 2.5. The nificant periodic signals. For each set of parameters β and γ,
low frequency eigenvector is clearly periodic, although we we repeat this procedure for a large number of network
observe some deviations corresponding to low degree realizations and average the values of Λ obtained corre-
nodes. By increasing the eigenvalue, the periodicity is still sponding to a given frequency.
preserved although the sinusoidal behavior is strongly The result of this program is shown in Figs. 7(a)–7(d) for
modified. Finally, periodicity is destroyed at high eigen- the homogeneous ensemble with β ¼ ∞ and β ¼ 1.5, and
values. This behavior is, again, translated into the steady in Figs. 7(e)–7(h) for the heterogeneous ensemble with γ ¼
state of the Mimura-Murray dynamics so that, even in 2.5 and β ¼ ∞ and β ¼ 1.5. As can be seen, the agreement
the presence of strong heterogeneity in the degree distri- with the annealed approximation is very good in all cases,
bution, we are able to find geometric Turing patterns in the with small deviations in the case of β ¼ 1.5. These are
dynamics. understandable because networks generated by the S1 =H2
model lose their metric character when β approaches one
(as clustering goes to zero). Additionally, as mentioned in
IV. NUMERICAL RESULTS ON SYNTHETIC
our previous discussion, it is not possible to find periodic
S1 =H2 NETWORKS
behavior at high frequencies as the fluctuations of node
To confirm numerically our theoretical predictions for degrees induce incoherent signals in the eigenvector
the dispersion relations, we need a systematic way of structures. In the examples shown in Fig. 7, the null model

021038-9
JASPER VAN DER KOLK et al. PHYS. REV. X 13, 021038 (2023)

FIG. 7. (a),(c),(e),(g) Spectrum of an ensemble of networks generated by the homogeneous (a),(c) and heterogeneous (e),(g)
ensembles with β ¼ 1.5 (a),(e) and β ¼ 1.5 (c),(g). For the heterogeneous ensemble γ ¼ 2.5 is used. In all cases the average degree is
given by hki ¼ 12. (b),(d),(f),(h) Dispersion relation obtained by applying the fast Fourier transform to the eigenvectors and detecting
the frequency with highest contribution. The opacity of the points is proportional to total amount of eigenvectors with the frequency
corresponding to the point. The more transparent a point is, the fewer eigenvectors have that frequency. The black dashed lines are the
theoretical predictions given by the annealed approximation. The parameter r̂ gives the percentage of eigenvectors classified as periodic
by the procedure described in Appendix A 2. The notation  implies that the results are significant with p ¼ 0.01.

combined with the DFT—using a confidence level of 2σ— observable frequencies is small. Third, as shown in
is able to detect only between 8% and 12% of periodic Ref. [70], some real networks may be better represented
eigenvalues. in similarity spaces with dimension higher than D ¼ 1.
Thus, a geometric pattern in a high dimensional space can
V. REAL NETWORKS be distorted in a lower dimensional projection, making its
detection difficult. Finally, the angular distribution in real
To analyze the emergence of Turing patterns in real networks is not perfectly uniform, with significant fluctua-
networks, we first need to find an embedding of the tions defining geometric communities. As shown in
network under study in the hyperbolic plane. This amounts Ref. [24], these communities have an impact on the
to finding, for each node of the network, its hidden degree κ eigenvector structure of the Laplacian matrix, so that in
and angular coordinate θ. For this task, we use Mercator this case we should expect patterns with mixed effects from
[83], a tool which combines machine learning and maxi- both the geometry of the network and its community
mum likelihood methods to find optimal embeddings structure. Despite all these limitations, we expect to find
congruent with the S1 =H2 model. We then run the patterns in real networks associated to small nonvanishing
Mimura-Murray dynamics on the network and look for eigenvalues.
geometric patterns in the angular similarity space inferred Figure 8 shows results for four different real networks
by Mercator. from different domains. Detailed definitions of these
Of course, in the case of real networks, we expect less networks as well as those of several others which are
clean results as compared to synthetic graphs generated by analyzed in Supplemental Material [82] can be found in
the S1 =H2 model. Indeed, there are four main factors Appendix A 5. Figures 8(a)–8(d) show the structure of the
affecting the emergence of Turing patterns in real networks. eigenvector corresponding to a small eigenvalue (and
First, the embedding method is noisy, so that periodic lower frequency) for the four studied networks. The
patterns will also be noisy. Second, even if the embedding geometric patterns in the similarity space found by
is perfect, the degree distribution is generally hetero- Mercator are very clear, with a wavelength of the order
geneous so that, as in the S1 =H2 model, the range of of the system size. The same structure is translated into

021038-10
EMERGENCE OF GEOMETRIC TURING PATTERNS IN COMPLEX … PHYS. REV. X 13, 021038 (2023)

FIG. 8. (a)–(d) Eigenvectors ϕðθÞ corresponding to the eigenvalue Λ as denoted in the figure for the real networks analyzed in
this study, namely, FriendsOFF, an off-line friendship network; WTW 2013, the World-Trade-Web; CElegans-G, a network of
genetic interaction of the nematode C. elegans; and Malaria-G, a network of highly variable genes of the human malaria parasite.
(e)–(h) Concentration of species U for the Mimura-Murray dynamics setting the parameters to make these eigenvalues unstable. The red
dotted line represents the trend obtained using the Savitzky-Golay filter with window size π. Panels (i)–(l) show the same as (e)–(h) but
in the hyperbolic representation of the four networks.

the dynamics, as shown in Figs. 8(e)–8(h), where the same

pattern can be recognized. We also plot a trendline,
obtained by applying the Savitzky-Golay filter to the
original signal. Finally, Figs. 8(i)–8(l) show the concen-
tration of species U in the hyperbolic representation of the
different networks inferred by Mercator. It is worth
stressing how highly nontrivial this result is. Indeed,
the Mimura-Murray dynamics generates steady states
with correlation lengths of the order of the system size
N in small-world networks where any pair of nodes is
separated by a very small number of intermediate steps,
scaling as ln N. Additionally, this is achieved despite the
fact that neither Mercator nor the Mimura-Murray dynam-
ics use geometric information from the latent space as they
take only the bare network topology as input. FIG. 9. Analysis of a human connectome network [84].
Finally, we analyze in detail a human connectome (a) Dispersion relation between the wave numbers of the dominant
network from Ref. [84] (see Appendix A 5 for details). mode of the Laplacian eigenvectors and their rescaled eigenvalues.
This is an interesting example for which a significant The size of the points is proportional to the height of the largest peak
number of patterns of different wavelengths can be in the Fourier spectrum. (b) Hyperbolic representation of the network,
detected. Figure 9(a) shows the dispersion relation between where the node size is proportional to the hidden degree κ. The
eigenvalues showing significant periodic signals versus colored outer ring shows the partition in communities according to
the Louvain algorithm [85]. (c),(d) The eigenvector with wave
their wave number. Despite the caveats discussed earlier,
number one (c) and two (d) for which the pattern is most significant.
the dispersion relation agrees qualitatively well with the The color follows ϕðθÞ but rescaled such that the values lie between
ones from the S1 model. In Figs. 9(c)–9(f) we show two of the 5th and 95th percentile of their original range. The background
these patterns, corresponding to wave numbers one and colors again represent the communities. The red dotted line represents
two, where the patterns are clearly visible. Although the the trend using the Savitzky-Golay filter with window size π.

021038-11
JASPER VAN DER KOLK et al. PHYS. REV. X 13, 021038 (2023)

network displays a modular or community organization, as ACKNOWLEDGMENTS

revealed by a Louvain algorithm analysis, this alone cannot
We acknowledge support from Grant No. PID2019–
account for the observed patterns. Rather, geometry is
106290 GB-C22 funded by MCIN/AEI/10.13039/
responsible for this phenomenon.
501100011033, and Generalitat de Catalunya Grant
No. 2021SGR00856. M. B. acknowledges the ICREA
VI. CONCLUSIONS Academia award, funded by the Generalitat de Catalunya.
The small-world property present in complex networks J. v d K. acknowledges support from the Ministry of
suggests, a priori, that networks in the real world are Universities of Spain in the form of the FPU predoctoral
infinite dimensional structures, so that geometry plays a contract.
marginal role. Yet, network geometry stands today as the
most parsimonious and comprehensive description of this APPENDIX
class of systems. However, beyond providing an explan-
ation of the observed topological structure of real net-
1. Diffusion dynamics on graphs
works, it is not clear how network geometry influences
the dynamics taking place on them. In this work, we show Let us focus on the diffusion of a single species on a
that geometric patterns may emerge in complex networks simple undirected and globally connected network with
in a wide class of reaction-diffusion dynamics, as adjacency matrix A ≡ faij g, i; j ¼ 1; …; N, and let ni ðtÞ
observed also in other types of dynamical processes such be the number of particles at node i at time t. We
as opinion dynamics [86] and games [87]. Such patterns assume particles do not interact with each other and that
are already encoded in the structure of the Laplacian nodes can accumulate an arbitrary number of particles
matrix as a result of the (hidden) metric structure of real on them. Once at node i, a particle jumps away from it
networks. Interestingly, reaction-diffusion processes tak- at constant rate ζ i . That is, each particle defines a
ing place on such networks are able to sense their continuous-time random walk [90,91] with dwell time
geometric nature even when they show extreme disorder probability density ψ i ðτÞ ¼ ζ i e−ζi τ. When a particle
induced by heterogeneous degree distributions and a jumps from node i, it chooses one of i’s neighbors at
large fraction of long-range links making them small random. We can write the following equation for the
worlds. stochastic evolution of ni ðtÞ from t to t þ dt:
In this work, we provide a theoretical framework to
nj ðtÞ ni ðtÞ
understand this phenomenon and to quantify the role that X
N
aij X X
the different topological properties of networks have on ni ðt þ dtÞ ¼ ni ðtÞ þ ηlj ðt; dtÞ − ηli ðt; dtÞ;
j¼1
kj l¼1 l¼1
the emerging patterns. This is done in the framework of
undirected networks. Of course, real networks often ðA1Þ
contain directionality [88], and it has previously been
shown that directed networks also support the Turing where ηlj ðt; dtÞ is a random variable controlling the event
instability. In fact, the existence of directed links in a
that one of the nj ðtÞ particles sitting at node j jumps to
network has even been shown to amplify the region in
parameter space where the homogeneous steady state is a different node in the interval ðt; t þ dtÞ. That is,
unstable, allowing for more easily accessible pattern ηlj ðt; dtÞ ¼ 1 with probability ζj dt and ηlj ðt; dtÞ ¼ 0
formation. It is therefore important to expand the results otherwise. By taking first averages over the random
we present in this paper to the directed geometric net- variables η and then over nðtÞ, we obtain the equation
works [89]. for the average number of particles at node i,
In addition to developing the theoretical tools for hni ðtÞi ≡ ui ðtÞ:
studying geometric patterns formation, significant insight
 
is gained from the fact that Turing patterns can also be dui ðtÞ X N
aij XN
ζj
detected in real complex networks when the hidden ¼ ζj − δij uj ðtÞ ¼ − Lij uj ðtÞ;
dt j¼1
k j k
j¼1 j
similarity space is properly highlighted. This strongly
suggests that there must exists a deep connection between ðA2Þ
networks’ functions and the underlying geometry.
Finding such yet unknown connection may shed light where Lij is the Laplacian matrix [50],
on one of the most fundamental questions in network
science, namely, the connection between the topology Lij ¼ kj δij − aij : ðA3Þ
and function of complex systems. This is particularly
relevant in biological networks, where morphogenesis The steady state solution of Eq. (A2) is given by the
stands as a fundamental principle behind the organization eigenvector of zero eigenvalue of the Laplacian matrix.
of living systems. This implies that

021038-12
EMERGENCE OF GEOMETRIC TURING PATTERNS IN COMPLEX … PHYS. REV. X 13, 021038 (2023)

ζ i usti For each network, we repeat this procedure for all its
¼ const: ðA4Þ eigenvectors, leaving us with a set of pairs ðΛ; ωc Þ of
ki
statistically significant characteristic frequencies, from
Therefore, if particles diffuse at the same rate in all where we can derive the dispersion relation as an average
nodes—so with ζ i ¼ ζ; ∀ i—then the average concen- over many network realizations.
tration at node i is proportional to ki and only when
ζi ∝ ki the average concentration of particles is the same 3. Dynamics on the network
everywhere. We then make the choice ζ i ¼ ϵki and the
diffusion equation finally reads In this paper we use the Mimura-Murray model to perform
numerical simulations. In particular, as in Ref. [11], we
dui ðtÞ XN choose the parameters as a ¼ 35, b ¼ 16, c ¼ 9, and
¼ −ϵ Lij uj ðtÞ: ðA5Þ d ¼ 2=5, which gives the fixed point ðū; v̄Þ ¼ ð5; 10Þ. In
dt j¼1 this case, the critical value of σ above which the Turing
instability arises is σ c ¼ 15.5071. We then set the value to
Similarly to the case of diffusion in continuous media, σ ¼ 15.6, slightly above the critical point. Unstable eigen-
we call ϵ the diffusion coefficient of the species in the values will then lie within the interval ½Λ− ; Λþ  with
network, even though, unlike diffusion in the con-
tinuum, it has dimensions of inverse of time.
1.410 26 1.666 67
Λ− ¼ and Λþ ¼ ; ðA9Þ
ϵ ϵ
2. Dispersion relation
To compute the dispersion relation in quenched net- and the most unstable one given by
works, we generate a large number of them from the S1 =H2
model for a given set of parameters β, γ, and hki. For each
such network, we compute its eigenvectors and eigenvalues 1.533 25
Λmax ¼ : ðA10Þ
sorted in increasing order. We then compute the discrete ϵ
Fourier transform of the eigenvectors and measure the
characteristic frequency as the one corresponding to the To select a given eigenvalue of interest Λ , we set it to
highest peak in the Fourier spectrum. To asses whether a Λ ¼ Λmax and use Eq. (A10) to fix the value of ϵ. With
peak in the spectrum is statistically significant, we define a all the parameters of the dynamics fixed, we set every
null model where the entries of the components of the node’s initial condition to be the stationary one, that is,
eigenvector are independently drawn from a normal dis- ðui ð0Þ; vi ð0ÞÞ ¼ ðū; v̄Þ for all i except for a randomly chosen
tribution N ðμ; σ 2 Þ with μ ¼ 0 and σ 2 ¼ N −1 . The first one, j, for which ðuj ð0Þ; vj ð0ÞÞ ¼ ðū þ δ; v̄ þ δÞ, where
condition comes from Eq. (6) and the second from the fact δ ¼ 10−5 is a small perturbation. We then let the system
that eigenvectors are normalized. Denoting ϕ̂rand
i as one of evolve toward its steady state.
the N entries of the DFT of a random eigenvector from our
null model, it can be shown that it satisfies the following
distribution function: 4. Cheeger constant for the S1 =H2 model
It is worth mentioning an interesting property of the
2 −q
Probðjϕ̂rand
i j < qÞ ¼ 1 − e : ðA6Þ S1 =H2 model concerning its spectral gap, defined as the
smallest non-null eigenvalue of the Laplacian Λ2 . Given a
With this in hand, we can find the value q for which there is graph generated by the S1 model, GS1 , Cheeger’s inequal-
a probability p that at least one peak in the white noise ities state that
Fourier spectrum lies above q. We choose a probability
p ¼ 10−2 and take the corresponding value of q to be the ½hðGS1 Þ2
minimal value a peak in the Fourier spectrum of one of the ≤ Λ2 < 2hðGS1 Þ; ðA11Þ
2kc
eigenvectors of a network generated by the model needs to
cross to be considered periodic. Thus,
where hðGS1 Þ is the isoperimetric (or Cheeger) constant
Y
N of the graph [71,72]. This constant defines the optimal
p¼1− Probðjϕ̂i j2 < qÞ; ðA7Þ cut of the graph in two disjoint sets of nodes. For any
i¼1 partition A and B in two disjoint sets of sizes N A and N B ,
the Cheeger constant is defined as the ratio between the
so that number of edges connecting the two sets EAB and the size
of the smallest set, minimized over all possible bipartitions
q ¼ − ln½1 − ð1 − pÞ1=N : ðA8Þ of the graph; that is,

021038-13
JASPER VAN DER KOLK et al. PHYS. REV. X 13, 021038 (2023)

EAB visible and clear patterns. This derivation is valid only in

hðGS1 Þ ¼ min : ðA12Þ
A minðN A ; N B Þ the case of a homogeneous hidden degree distribution.
However, numerical analysis indicates that the Cheeger
While computing the Cheeger constant is, in general, not constant decays to zero in the case of a heterogeneous
possible, in the S1 =H2 model we can calculate its scaling distribution as well.
behavior with the system size as follows. Assuming that the
optimal cut minimizing the Cheeger constant is made by 5. Real network properties
splitting the unit circle in two continuous regions, in the
In this section we elaborate on the properties of the real
case of the S1 model Eq. (A12) can be written as follows:
networks considered in this paper as well as in the
 Z Z  2  Supplemental Material [82].
2π x 2π N dθ2 dθ1
hðGS1 Þ ¼ min ; (i) Airports, Ref. [92]: The data for the network were
x xN 0 x 2π 1 þ ð NΔθ122 Þβ
2π μ̂hki obtained from the Preferred RSocio-Commes Data-
ðA13Þ base of the National Flight Data Center (NFDC),
part of the U.S. Federal Aviation Administration.
where Δθ12 ¼ π − jπ − jθ1 − θ2 jj. Here we define the two Nodes represent airports or service centers and edges
disjoint regions as A ¼ fij0 < θi < xg and B ¼ fijx < are created from strings of preferred routes recom-
θi < 2πg ¼ GS1 nA. As we have chosen 0 < x < π, in the mended by the NFDC. The original network was
model the least amount of nodes reside in A and thus directed.
minðN A ; N B Þ ¼ N A ¼ xN=ð2πÞ, as shown in Eq. (A13). (ii) CElegans-C, Refs. [93,94]: The nervous system
The second part of this equation represents the amount of network of the nematode Caenorhabditis elegans.
links between regions A and B, i.e., EAB . It can be shown Here, the nodes are the neurons and the edges
that Eq. (A13) can be written as represent the connections, specifically gap junctions
and chemical synapses, between them. The original
 Z Z  network was directed.
1N x π dθdΔ
hðGS1 Þ ¼ min  β ; ðA14Þ (iii) CElegans-G, Ref. [95]: A genetic interaction net-
x xπ 0 θ
1þ NΔ
2π μ̂hki2
work for the nematode C. elegans where the nodes
represent genes and the links interactions on the
genetic level.
which in turn can be solved exactly, giving (iv) Commodities, Ref. [96]: A network where the nodes
  represent industrial sectors and the links the flows
N
hðGS1 Þ ¼ min fðN; xÞ ; ðA15Þ of goods and services between them. The data used
x 2π are from 2007 and were prepared by the U.S. Bureau
of Economic Analysis. The original network was
where directed.
 1; 1  β  (v) FriendsOFF, Ref. [97]: An off-line friendship net-
β N work among students, created from a survey where
fðN; xÞ ¼ 2π 2 F1 ;− each student was asked to list five male and five
1 þ β1 2μ̂hki
female friends. The nodes are in this case the
  1; 1  β  students and the edges represent these friendship
β Nx
þ x −22 F1 ;− relations. The original network was directed.
1 þ β1 2π μ̂hki
(vi) Human-C, Ref. [84]: A connectome of the human
 1; 2  β  brain where nodes represent small brain regions
β Nx
þ 2 F1 ; − : ðA16Þ and the edges represent fibers connecting these
1 þ β2 2π μ̂hki regions based on white matter tractography, which
is a method used to estimate axonal bundle trajec-
Here, 2 F1 ½a;b
c ; z is the ordinary hypergeometric
tories. In this network both hemispheres are
function. Note that we use a slightly different form than incorporated.
the standard 2 F1 ½a; b; c; z for aesthetic purposes. It can be (vii) List contacts, Ref. [98]: A network generated from
shown that this scales as hðGS1 Þ ≃ c1 N 1−β þ c2 N −1 , irre- the daily dynamic contact networks collected during
spective of the choice of x. Since β > 1, this scaling the Infectious SocioPatterns event that took place
relation implies that hðGS1 Þ approaches zero in the at the Science Gallery in Dublin, Ireland, during the
thermodynamic limit and so does the spectral gap. art science exhibition Infections: Stay Away. Nodes
This implies that for very large networks we expect to represent individuals attending the gallery, and two
find eigenvalues arbitrarily close to zero. This is impor- nodes are connected if they interacted with each
tant because small eigenvalues are the ones with more other during the visit.

021038-14
EMERGENCE OF GEOMETRIC TURING PATTERNS IN COMPLEX … PHYS. REV. X 13, 021038 (2023)

TABLE I. Properties of the five real networks studied in (xiii) WTW-2013, Refs. [105–107]: The world trade web
Figs. 1 and 8. The parameter r̂ gives the percentage of as of 2013, where nodes represent countries and
eigenvectors classified as periodic by the procedure described edges trade relations.
in Appendix A 2. The notation  implies that the results are Table I shows the network properties of these
significant with p ¼ 0.01. networks.
Network N hki β γ r̂
Airports 1226 3.9 1.0 4 7.1%
CElegans-C 279 16.4 1.5 3.3 3.2% [1] A. M. Turing, The Chemical Basis of Morphogenesis, Phil.
CElegans-G 878 7.2 2.6 2.7 5.9% Trans. R. Soc. B 237, 37 (1952).
Commodities 374 5.8 1.1 2.8 2.4% [2] A. Gierer and H. Meinhardt, A Theory of Biological
FriendsOFF 2539 8.2 1.3 9 4.4% Pattern Formation, Kybernetik 12, 30 (1972).
[3] H. Meinhardt, Models of Biological Pattern Formation
Human-C 989 36.1 2.3 8 7.5% (Academic Press, New York, 1982), 10.1016/S0070-2153
List contacts 410 13.5 2.2 7.5 6.1% (07)81001-5.
Malaria-G 307 18.3 2.9 9 3.9% [4] M. C. Cross and P. C. Hohenberg, Pattern Formation
Mouse-C 213 27.9 2.0 3.5 2.3% Outside of Equilibrium, Rev. Mod. Phys. 65, 851 (1993).
Power 4941 2.7 1.3 10 11.0% [5] A. Koch and H. Meinhardt, Biological Pattern Formation:
From Basic Mechanisms to Complex Structures, Rev.
URV-Email 1133 9.6 1.4 7 3.0%
Mod. Phys. 66, 1481 (1994).
Wikipedia 4589 46.4 1.2 2.8 1.0% [6] A. M. Zhabotinsky, M. Dolnik, and I. R. Epstein, Pattern
WTW 2013 189 5.8 1.9 2.4 5.8% Formation Arising from Wave Instability in a Simple
Reaction-Diffusion System, J. Chem. Phys. 103, 10306
(1995).
[7] M. Cross and H. Greenside, Pattern Formation and
(viii) Malaria-G, Ref. [99]: The network of recombinant Dynamics in Nonequilibrium Systems (Cambridge Uni-
antigen genes from the human malaria parasite versity Press, Cambridge, England, 2009).
Plasmodium falciparum, where nodes are var genes [8] S. Kondo and T. Miura, Reaction-Diffusion Model as a
encoding for an antigen protein expressed on the Framework for Understanding Biological Pattern For-
surface of the infected red blood cell. These var mation, Science 329, 1616 (2010).
genes are capable of interchanging bits of genetic [9] D. Walgraef, Spatio-Temporal Pattern Formation: With
Examples from Physics, Chemistry, and Materials Science
information, thus creating a vast amount of slightly
(Springer Science & Business Media, New York, 2012).
different genes that in turn create highly varying [10] H. Othmer and L. Scriven, Instability and Dynamic Pattern
antigens that are hard for the human immune system in Cellular Networks, J. Theor. Biol. 32, 507 (1971).
to detect. Two nodes are connected if they share a [11] H. Nakao and A. S. Mikhailov, Turing Patterns in
substring of nucleotides of significant length, in- Network-Organized Activator–Inhibitor Systems, Nat.
dicating that an interchange of genetic material Phys. 6, 544 (2010).
between the two nodes has taken place. [12] W. Horsthemke, K. Lam, and P. K. Moore, Network
(ix) Mouse-C, Ref. [100]: A connectome of the Topology and Turing Instabilities in Small Arrays of
mouse brain where nodes represent anatomical Diffusively Coupled Reactors, Phys. Lett. A 328, 444
(2004).
subregions of the brain and the edges the physical
[13] P. K. Moore and W. Horsthemke, Localized Patterns in
connections between them. The original network
Homogeneous Networks of Diffusively Coupled Reactors,
was directed. Physica (Amsterdam) 206D, 121 (2005).
(x) Power, Ref. [101]: A network which represents the [14] M. Asllani, J. D. Challenger, F. S. Pavone, L. Sacconi, and
power grid of the western U.S. Here, nodes are D. Fanelli, The Theory of Pattern Formation on Directed
transforms or power relay points and edges represent Networks, Nat. Commun. 5, 4517 (2014).
power lines running between these locations. [15] R. Muolo, M. Asllani, D. Fanelli, P. K. Maini, and T.
(xi) URV-Email, Ref. [102]: A network representing the Carletti, Patterns of Non-normality in Networked Systems,
exchange of emails between members of the Rovira i J. Theor. Biol. 480, 81 (2019).
Virgili University in Spain, in 2003. The nodes [16] J. Petit, B. Lauwens, D. Fanelli, and T. Carletti, Theory of
represent the members and the edges the email Turing Patterns on Time Varying Networks, Phys. Rev.
Lett. 119, 148301 (2017).
traffic between them. The original version of the
[17] R. A. V. Gorder, A Theory of Pattern Formation for
network was directed.
Reaction-Diffusion Systems on Temporal Networks, Proc.
(xii) Wikipedia, Refs. [103,104]: A network where nodes R. Soc. A 477, 20200753 (2021).10.1098/rspa.2020.0753
represent Wikipedia articles and where two nodes [18] R. Muolo, L. Gallo, V. Latora, M. Frasca, and T. Carletti,
are connected if there is a link between them. The Turing Patterns in Systems with High-Order Interactions,
original network was directed. Chaos Solitons Fractals 166, 112912 (2023).

021038-15
JASPER VAN DER KOLK et al. PHYS. REV. X 13, 021038 (2023)

[19] S. Gao, L. Chang, M. Perc, and Z. Wang, Turing Patterns in [39] N. Fountoulakis and T. Müller, Law of Large Numbers for
Simplicial Complexes, Phys. Rev. E 107, 014216 (2023). the Largest Component in a Hyperbolic Model of Complex
[20] M. Asllani, D. M. Busiello, T. Carletti, D. Fanelli, and G. Networks, Ann. Appl. Probab. 28, 607 (2018).
Planchon, Turing Patterns in Multiplex Networks, Phys. [40] M. Kiwi and D. Mitsche, Spectral Gap of Random
Rev. E 90, 042814 (2014). Hyperbolic Graphs and Related Parameters, Ann. Appl.
[21] M. Asllani, D. M. Busiello, T. Carletti, D. Fanelli, and G. Probab. 28, 941 (2018).
Planchon, Turing Instabilities on Cartesian Product Net- [41] G. García-Pérez, M. Boguñá, and M. Á. Serrano, Multi-
works, Sci. Rep. 5, 12927 (2015). scale Unfolding of Real Networks by Geometric Renorm-
[22] N. E. Kouvaris, S. Hata, and A. Díaz-Guilera, Pattern alization, Nat. Phys. 14, 583 (2018).
Formation in Multiplex Networks, Sci. Rep. 5, 10840 [42] M. Zheng, G. García-Pérez, M. Boguñá, and M. Á.
(2015). Serrano, Scaling Up Real Networks by Geometric
[23] D. M. Busiello, T. Carletti, and D. Fanelli, Homogeneous- Branching Growth, Proc. Natl. Acad. Sci. U.S.A. 118,
per-Layer Patterns in Multiplex Networks, Europhys. Lett. e2018994118 (2021).
121, 48006 (2018). [43] F. Papadopoulos, M. Kitsak, M. Á. Serrano, M. Boguñá,
[24] B. A. Siebert, C. L. Hall, J. P. Gleeson, and M. Asllani, and D. Krioukov, Popularity versus Similarity in Growing
Role of Modularity in Self-Organization Dynamics in Networks, Nature (London) 489, 537 (2012).
Biological Networks, Phys. Rev. E 102, 052306 (2020). [44] A. Allard, M. Á. Serrano, G. García-Pérez, and M.
[25] M. Asllani, T. Carletti, and D. Fanelli, Tune the Topology to Boguñá, The Geometric Nature of Weights in Real Com-
Create or Destroy Patterns, Eur. Phys. J. B 89, 260 (2016). plex Networks, Nat. Commun. 8, 14103 (2017).
[26] G. Cencetti, F. Battiston, T. Carletti, and D. Fanelli, [45] K.-K. Kleineberg, M. Boguñá, M. Á. Serrano, and F.
Generalized Patterns from Local and Non Local Reac- Papadopoulos, Hidden Geometric Correlations in Real
tions, Chaos Solitons Fractals 134, 109707 (2020). Multiplex Networks, Nat. Phys. 12, 1076 (2016).
[27] M.-T. Hütt, D. Armbruster, and A. Lesne, Predictable [46] K.-K. Kleineberg, L. Buzna, F. Papadopoulos, M. Boguñá,
Topological Sensitivity of Turing Patterns on Graphs, and M. Á. Serrano, Geometric Correlations Mitigate the
Phys. Rev. E 105, 014304 (2022). Extreme Vulnerability of Multiplex Networks against
[28] M. Boguñá, I. Bonamassa, M. D. Domenico, S. Havlin, D. Targeted Attacks, Phys. Rev. Lett. 118, 218301 (2017).
Krioukov, and M. Á. Serrano, Network Geometry, Nat. [47] K. Zuev, M. Boguñá, G. Bianconi, and D. Krioukov,
Rev. Phys. 3, 114 (2021). Emergence of Soft Communities from Geometric Prefer-
[29] M. Á. Serrano, D. Krioukov, and M. Boguñá, Self- ential Attachment, Sci. Rep. 5, 9421 (2015).
Similarity of Complex Networks and Hidden Metric [48] G. García-Pérez, M. Á. Serrano, and M. Boguñá, Soft
Spaces, Phys. Rev. Lett. 100, 078701 (2008). Communities in Similarity Space, J. Stat. Phys. 173, 775
[30] D. Krioukov, F. Papadopoulos, M. Kitsak, A. Vahdat, and (2018).
M. Boguñá, Hyperbolic Geometry of Complex Networks, [49] A. Muscoloni and C. V. Cannistraci, A Nonuniform
Phys. Rev. E 82, 036106 (2010). Popularity-Similarity Optimization (nPSO) Model to Effi-
[31] L. Gugelmann, K. Panagiotou, and U. Peter, Random ciently Generate Realistic Complex Networks with Com-
Hyperbolic Graphs: Degree Sequence and Clustering, in munities, New. J. Phys. 20, 052002 (2018).
Automata, Languages, and Programming. ICALP 2012,
[50] P. van Mieghem, Graph Spectra for Complex Networks
edited by A. Czumaj, K. Mehlhorn, A. Pitts, and R.
Wattenhofer, Lecture Notes in Computer Science Vol. 7392 (Cambridge University Press, Cambridge, England, 2010),
(Springer, Berlin, Heidelberg, 2012), 10.1007/978-3-642- 10.1017/CBO9780511921681.
31585-5_51. [51] I. Prigogine and R. Lefever, Symmetry Breaking Instabil-
[32] E. Candellero and N. Fountoulakis, Clustering and the ities in Dissipative Systems. II, J. Chem. Phys. 48, 1695
Hyperbolic Geometry of Complex Networks, Internet (1968).
Math. 12, 2 (2016). [52] V. Castets, E. Dulos, J. Boissonade, and P. De Kepper,
[33] N. Fountoulakis, P. van der Hoorn, T. Müller, and M. Experimental Evidence of a Sustained Standing Turing-
Schepers, Clustering in a Hyperbolic Model of Complex Type Nonequilibrium Chemical Pattern, Phys. Rev. Lett.
Networks, Electron. J. Pro 26, 1 (2021). 64, 2953 (1990).
[34] J. van der Kolk, M. Á. Serrano, and M. Boguñá, An [53] Q. Ouyang and H. L. Swinney, Transition from a Uniform
Anomalous Topological Phase Transition in Spatial Ran- State to Hexagonal and Striped Turing Patterns, Nature
dom Graphs, Commun. Phys. 5, 245 (2022). (London) 352, 610 (1991).
[35] M. A. Abdullah, N. Fountoulakis, and M. Bode, Typical [54] M. P. Harris, S. Williamson, J. F. Fallon, H. Meinhardt,
Distances in a Geometric Model for Complex Networks, and R. O. Prum, Molecular Evidence for an Activator–
Internet Math. 1 (2017).10.24166/im.13.2017 Inhibitor Mechanism in Development of Embryonic Feather
[36] T. Friedrich and A. Krohmer, On the Diameter of Hyper- Branching, Proc. Natl. Acad. Sci. U.S.A. 102, 11734 (2005).
bolic Random Graphs, SIAM J. Discrete Math. 32, 1314 [55] L. A. Segel and S. A. Levin, Application of Nonlinear
(2018). Stability Theory to the Study of the Effects of Diffusion on
[37] T. Müller and M. Staps, The Diameter of KPKVB Random Predator-Prey Interactions, AIP Conf. Proc. 27, 123
Graphs, Adv. Appl. Probab. 51, 358 (2019). (1976).
[38] M. Á. Serrano, D. Krioukov, and M. Boguñá, Percolation [56] M. Mimura and J. Murray, On a Diffusive Prey-Predator
in Self-Similar Networks, Phys. Rev. Lett. 106, 048701 Model which Exhibits Patchiness, J. Theor. Biol. 75, 249
(2011). (1978).

021038-16
EMERGENCE OF GEOMETRIC TURING PATTERNS IN COMPLEX … PHYS. REV. X 13, 021038 (2023)

[57] J. L. Maron and S. Harrison, Spatial Pattern Formation in [75] U. Bastolla and G. Parisi, Closing Probabilities in the
an Insect Host-Parasitoid System, Science 278, 1619 Kauffman Model: An Annealed Computation, Physica
(1997). (Amsterdam) 98D, 1 (1996).
[58] J. P. Gibert and J. D. Yeakel, Laplacian Matrices and [76] B. Luque and R. V. Solé, Phase Transitions in Random
Turing Bifurcations: Revisiting Levin 1974 and the Networks: Simple Analytic Determination of Critical
Consequences of Spatial Structure and Movement for Points, Phys. Rev. E 55, 257 (1997).
Ecological Dynamics, Theor. Ecol. 12, 265 (2019). [77] T. Rohlf and S. Bornholdt, Criticality in Random Thresh-
[59] A. K. Fahimipour, F. Zeng, M. Homer, A. Traulsen, old Networks: Annealed Approximation and Beyond,
S. A. Levin, and T. Gross, Sharp Thresholds Limit the Physica (Amsterdam) 310A, 245 (2002).
Benefit of Defector Avoidance in Cooperation on Net- [78] D. Vilone and C. Castellano, Solution of Voter Model
works, Proc. Natl. Acad. Sci. U.S.A. 119, e2120120119 Dynamics on Annealed Small-World Networks, Phys. Rev.
(2022). E 69, 016109 (2004).
[60] A. Brechtel, P. Gramlich, D. Ritterskamp, B. Drossel, and [79] M. Boguná, C. Castellano, and R. Pastor-Satorras, Lange-
T. Gross, Master Stability Functions Reveal Diffusion- vin Approach for the Dynamics of the Contact Process on
Driven Pattern Formation in Networks, Phys. Rev. E 97, Annealed Scale-Free Networks, Phys. Rev. E 79, 036110
032307 (2018). (2009).
[61] E. A. Bender and E. R. Canfield, The Asymptotic Number [80] B. Guerra and J. Gómez-Gardenes, Annealed and Mean-
of Labeled Graphs with Given Degree Sequences, J. Field Formulations of Disease Dynamics on Static and
Comb. Theory Ser. A 24, 296 (1978). Adaptive Networks, Phys. Rev. E 82, 035101(R) (2010).
[62] B. Bollobás, A Probabilistic Proof of an Asymptotic [81] S. C. Ferreira, R. S. Ferreira, and R. Pastor-Satorras,
Formula for the Number of Labelled Regular Graphs, Quasistationary Analysis of the Contact Process on
Eur. J. Combinatorics 1, 311 (1980). Annealed Scale-Free Networks, Phys. Rev. E 83, 066113
[63] M. Molloy and B. Reed, A Critical Point for Random (2011).
Graphs with a Given Degree Sequence, Random Struct. [82] See Supplemental Material at http://link.aps.org/
Algorithms 6, 161 (1995). supplemental/10.1103/PhysRevX.13.021038 for an analy-
[64] M. Molloy and B. Reed, The Size of the Giant Component sis of the periodicity of the eigenvectors of several real
of a Random Graph with a Given Degree Sequence, Comb. networks.
Probab. Comput. 7, 295 (1998). [83] G. García-Pérez, A. Allard, M. Á. Serrano, and M.
[65] A.-L. Barabási and R. Albert, Emergence of Scaling in Boguñá, Mercator: Uncovering Faithful Hyperbolic Em-
Random Networks, Science 286, 509 (1999). beddings of Complex Networks, New J. Phys. 21, 123033
[66] M. A. Serrano and M. Boguñá, The Shortest Path to (2019).
Network Geometry: A Practical Guide to Basic Models [84] P. Hagmann, L. Cammoun, X. Gigandet, R. Meuli, C. J.
and Applications, Elements in Structure and Dynamics Honey, V. J. Wedeen, and O. Sporns, Mapping the Struc-
of Complex Networks (Cambridge University Press, tural Core of Human Cerebral Cortex, PLoS Biol. 6, e159
Cambridge, England, 2022), 10.1017/9781108865791. (2008).
[67] M. Boguñá, D. Krioukov, P. Almagro, and M. Á. Serrano, [85] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E.
Small Worlds and Clustering in Spatial Networks, Phys. Lefebvre, Fast Unfolding of Communities in Large Net-
Rev. Res. 2, 023040 (2020). works, J. Stat. Mech. (2008) P10008.
[68] M. Boguñá, F. Papadopoulos, and D. Krioukov, Sustaining [86] E. Ortiz and M. Á. Serrano, Multiscale Voter Model on Real
the Internet with Hyperbolic Mapping, Nat. Commun. 1, 1 Networks, Chaos Solitons Fractals 165, 112847 (2022).
(2010). [87] K.-K. Kleineberg, Metric Clusters in Evolutionary Games
[69] M. Á. Serrano, M. Boguñá, and F. Sagués, Uncovering the on Scale-Free Networks, Nat. Commun. 8, 1888 (2017).
Hidden Geometry Behind Metabolic Networks, Mol. Bio- [88] J. D. O’Brien, K. A. Oliveira, J. P. Gleeson, and M. Asllani,
syst. 8, 843 (2012). Hierarchical Route to the Emergence of Leader Nodes in
[70] P. Almagro, M. Boguñá, and M. Á. Serrano, Detecting the Real-World Networks, Phys. Rev. Res. 3, 023117 (2021).
Ultra Low Dimensionality of Real Networks, Nat. Com- [89] A. Allard, M. Ángeles Serrano, and M. Boguñá, Geometric
mun. 13, 6096 (2022). Description of Clustering in Directed Networks, arXiv:
[71] G. Pólya and G. Szegö, Isoperimetric Inequalities in 2302.09055.
Mathematical Physics (AM-27) (Princeton University [90] E. W. Montroll and G. H. Weiss, Random Walks on
Press, Princeton, NJ, 1951), http://www.jstor.org/stable/j Lattices. II, J. Math. Phys. 6, 167 (1965).
.ctt1b9rzzn. [91] G. H. Weiss, Aspects and Applications of the Random
[72] J. Cheeger, A Lower Bound for the Smallest Eigenvalue Walk (North-Holland, Amsterdam, 1994), 10.1007/
of the Laplacian, in Problems in Analysis, edited by BF02179402.
R. C. Gunning (Princeton University Press, Princeton, [92] J. Kunegis, KONECT: The Koblenz Network Collection, in
NJ, 1969), pp. 195–199, 10.1515/9781400869312-013. Proceedings of the 22nd International Conference on World
[73] B. Derrida and Y. Pomeau, Random Networks of Wide Web (WWW '13 Companion) (ACM, New York,
Automata: A Simple Annealed Approximation, Europhys. 2013), pp. 1343–1350, 10.1145/2487788.2488173.
Lett. 1, 45 (1986). [93] Y.-Y. Ahn, H. Jeong, and B. J. Kim, Wiring Cost in the
[74] S.-i. Amari, N. Fujita, and S. Shinomoto, Four Types of Organization of a Biological Neuronal Network, Physica
Learning Curves, Neural Comput. 4, 605 (1992). (Amsterdam) 367A, 531 (2006).

021038-17
JASPER VAN DER KOLK et al. PHYS. REV. X 13, 021038 (2023)

[94] J. S. Kim and M. Kaiser, From Caenorhabditis elegans to [101] D. J. Watts and S. H. Strogatz, Collective Dynamics of
the Human Connectome: A Specific Modular Organization ‘Small-World’ Networks, Nature (London) 393, 440 (1998).
Increases Metabolic, Functional and Developmental Effi- [102] R. Guimerà, L. Danon, A. Díaz-Guilera, F. Giralt, and
ciency, Phil. Trans. R. Soc. B 369, 20130529 (2014). A. Arenas, Self-Similar Community Structure in a Network
[95] A. Cho, J. Shin, S. Hwang, C. Kim, H. Shim, H. Kim, H. of Human Interactions, Phys. Rev. E 68, 065103(R)
Kim, and I. Lee, WormNet v3: A Network-Assisted (2003).
Hypothesis-Generating Server for, Caenorhabditis ele- [103] R. West, J. Pineau, and D. Precup, Wikispeedia: An Online
gans, Nucleic Acids Res. 42, W76 (2014). Game for Inferring Semantic Distances between Concepts,
[96] D. Grady, C. Thiemann, and D. Brockmann, Robust in Proceedings of the International Joint Conference on
Classification of Salient Links in Complex Networks, Artificial Intelligence, California, 2009.
Nat. Commun. 3, 864 (2012). [104] R. West and J. Leskovec, Human Waynding in Information
[97] J. Moody, Peer Influence Groups: Identifying Dense Networks, in Proceedings of the 21st International
Clusters in Large Networks, Soc. Networks 23, 261 confErence on World Wide Web (WWW '12) (ACM,
(2001). 2012), pp. 619–628, 10.1145/2187836.2187920.
[98] L. Isella, J. Stehlé, A. Barrat, C. Cattuto, J.-F. Pinton, and [105] M. Á. Serrano and M. Boguñá, Topology of the World
W. V. den Broeck, What’s in a Crowd? Analysis of Face- Trade Web, Phys. Rev. E 68, 015101(R) (2003).
to-Face Behavioral Networks, J. Theor. Biol. 271, 166 [106] M. Á. Serrano, M. Boguñá, and A. Vespignani, Patterns of
(2011). Dominant Flows in the World Trade Web, J. Econ.
[99] D. B. Larremore, A. Clauset, and C. O. Buckee, A Network Interact. Coord. 2, 111 (2007).
Approach to Analyzing Highly Recombinant Malaria [107] G. García-Pérez, M. Boguñá, A. Allard, and M. A.
Parasite Genes, PLoS Comput. Biol. 9, e1003268 (2013). Serrano, The Hidden Hyperbolic Geometry of Interna-
[100] S. W. Oh et al., A Mesoscale Connectome of the Mouse tional Trade: World Trade Atlas 1870–2013, Sci. Rep. 6,
Brain, Nature (London) 508, 207 (2014). 33441 (2016).

021038-18