Visit https://ebookultra.

com to download the full version and

explore more ebooks

The Shortest Path to Network Geometry A

Practical Guide to Basic Models and Applications
M. Ángeles Serrano

_____ Click the link below to download _____

https://ebookultra.com/download/the-shortest-path-to-
network-geometry-a-practical-guide-to-basic-models-and-
applications-m-angeles-serrano/

Explore and download more ebooks at ebookultra.com

Here are some suggested products you might be interested in.
Click the link to download

M A A Practical Guide to Doing the Deal Jeffrey C. Hooke

https://ebookultra.com/download/m-a-a-practical-guide-to-doing-the-
deal-jeffrey-c-hooke/

Liver regeneration basic mechanisms relevant models and

clinical applications 1st Edition Udayan M. Apte Ph.D.
Dabt
https://ebookultra.com/download/liver-regeneration-basic-mechanisms-
relevant-models-and-clinical-applications-1st-edition-udayan-m-apte-
ph-d-dabt/

Practical Guide for Policy Analysis The Eightfold Path to

More Effective Problem Solving 5th Edition Eugene Bardach

https://ebookultra.com/download/practical-guide-for-policy-analysis-
the-eightfold-path-to-more-effective-problem-solving-5th-edition-
eugene-bardach/

802 1aq Shortest Path Bridging Design and Evolution The

Architect s Perspective 1st Edition David Allan

https://ebookultra.com/download/802-1aq-shortest-path-bridging-design-
and-evolution-the-architect-s-perspective-1st-edition-david-allan/
Network and Discrete Location Models Algorithms and
Applications Second Edition Mark S. Daskin

https://ebookultra.com/download/network-and-discrete-location-models-
algorithms-and-applications-second-edition-mark-s-daskin/

Practical Guide to Female Pelvic Medicine 1st Edition

Gamal M. Ghoniem

https://ebookultra.com/download/practical-guide-to-female-pelvic-
medicine-1st-edition-gamal-m-ghoniem/

Oregon off the beaten path a guide to unique places 10th

ed Edition Oakley

https://ebookultra.com/download/oregon-off-the-beaten-path-a-guide-to-
unique-places-10th-ed-edition-oakley/

Practical Guide to Common Clinical Procedures and

Emergencies 1st Edition Chandra M. Kumar

https://ebookultra.com/download/practical-guide-to-common-clinical-
procedures-and-emergencies-1st-edition-chandra-m-kumar/

Telecommunication Network Economics From Theory to

Applications Maille P.

https://ebookultra.com/download/telecommunication-network-economics-
from-theory-to-applications-maille-p/
The Shortest Path to Network Geometry A Practical
Guide to Basic Models and Applications M. Ã​ngeles
Serrano Digital Instant Download
Author(s): M. Ã​ngeles Serrano, Marián Boguñá
ISBN(s): 9781108791083, 1108791085
File Details: PDF, 4.91 MB
Year: 2022
Language: english
 Serrano and Boguñá
Real networks comprise hundreds to millions of interacting
elements and permeate all contexts, from technology to
biology to society. All of them display non-trivial connectivity
patterns, including the small-world phenomenon, making
nodes to be separated by a small number of intermediate links.
As a consequence, networks present an apparent lack of metric The Structure and Dynamics
structure and are difficult to map. Yet, many networks have a
hidden geometry that enables meaningful maps in the two- of Complex Networks
dimensional hyperbolic plane. The discovery of such hidden
geometry and the understanding of its role have become
fundamental questions in network science, giving rise to the
field of network geometry. This Element reviews fundamental
The Shortest Path to

The Shortest Path to Network Geometry

models and methods for the geometric description of
real networks with a focus on applications of real network
maps, including decentralized routing protocols, geometric
community detection, and the self-similar multiscale unfolding
of networks by geometric renormalization.
Network Geometry
A Practical Guide to Basic
About the series Series editors
Models and Applications
This cutting-edge new series provides Guido Caldarelli
authoritative and detailed coverage Ca’ Foscari
University of
M. Ángeles Serrano
of the underlying theory of complex
networks, specifically their structure and Venice
dynamical properties. Each Element within
and Marián Boguñá
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press
the series will focus upon one of three
primary topics: static networks, dynamical
networks, and numerical/computing
network resources.

Cover image: jivacore/Shutterstock ISSN 2516-5763 (online)

ISSN 2516-5755 (print)
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press
 Elements in the Structure and Dynamics of Complex Networks
edited by
Guido Caldarelli
Ca’ Foscari University of Venice

THE SHORTEST PATH TO

NETWORK GEOMETRY
A Practical Guide to Basic Models
and Applications

M. Ángeles Serrano
University of Barcelona,
University of Barcelona
Institute of Complex
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

Systems (UBICS) and

Catalan Institution for
Research and Advanced
Studies (ICREA)
Marián Boguñá
University of Barcelona and
University of Barcelona
Institute of Complex
Systems (UBICS)
 University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre,
New Delhi – 110025, India
103 Penang Road, #05–06/07, Visioncrest Commercial, Singapore 238467

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781108791083
DOI: 10.1017/9781108865791
© M. Ángeles Serrano and Marián Boguñá 2021
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2021
A catalogue record for this publication is available from the British Library.
ISBN 978-1-108-79108-3 Paperback
ISSN 2516-5763 (online)
ISSN 2516-5755 (print)
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

Cambridge University Press has no responsibility for the persistence or accuracy of

URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
 The Shortest Path to Network Geometry

A Practical Guide to Basic Models and Applications

Elements in the Structure and Dynamics of Complex Networks

DOI: 10.1017/9781108865791
First published online: December 2021

M. Ángeles Serrano
University of Barcelona, University of Barcelona Institute of Complex Systems
(UBICS) and Catalan Institution for Research and Advanced Studies (ICREA)
Marián Boguñá
University of Barcelona and University of Barcelona Institute of Complex Systems
(UBICS)
Author for correspondence: M. Ángeles Serrano, marian.serrano@ub.edu

Abstract: Real networks comprise hundreds to millions of interacting

elements and permeate all contexts, from technology to biology to society.
All of them display non-trivial connectivity patterns, including the
small-world phenomenon, making nodes to be separated by a small
number of intermediate links. As a consequence, networks present an
apparent lack of metric structure and are difficult to map. Yet, many
networks have a hidden geometry that enables meaningful maps in the
two-dimensional hyperbolic plane. The discovery of such hidden geometry
and the understanding of its role have become fundamental questions in
network science, giving rise to the field of network geometry. This Element
reviews fundamental models and methods for the geometric description of
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

real networks with a focus on applications of real network maps, including

decentralized routing protocols, geometric community detection, and the
self-similar multiscale unfolding of networks by geometric renormalization.

Keywords: network geometry, mapping techniques, navigability, community

detection, renormalization

© M. Ángeles Serrano and Marián Boguñá 2021

ISBNs: 9781108791083 (PB), 9781108865791 (OC)
ISSNs: 2516-5763 (online), 2516-5755 (print)
 Contents

1 From Networks to Maps 1

2 Geometric Models for Static Topologies 5

3 Mapping Real Networks 15

4 Mesoscale Organisation and Community Detection 17

5 Self-Similarity and Renormalisation 19

6 Navigability 29

7 Geometry of Weighted, Multiplex, and Growing Networks 32

8 Conclusions 35

References 38
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press
 The Shortest Path to Network Geometry 1

1 From Networks to Maps

Throughout history, maps have been at the center of political, economic, and
geostrategic decisions and have become a critical piece of our everyday lives,
serving as a precise and relevant information source. Maps provide an accurate
way of visualizing, storing, and communicating information, help to recognize
locational distributions, spatial patterns, and relationships, and allow us to track
processes that operate through space at different length scales. Our work in the
last decade led us to prove that many real complex networks are natural geomet-
ric objects and can be mapped into hidden low-dimensional metric spaces with
hyperbolic geometry, where distances determine the likelihood of the interac-
tions and encode the different intrinsic attributes determining how similar the
elements of the system are (Allard et al. [2017]; Boguñá, Papadopoulos, and
Krioukov [2010]; García-Pérez et al. [2016]; García-Pérez, Boguñá, and Ser-
rano [2018]; Kleineberg et al. [2016]; Krioukov et al. [2012]; Krioukov et al.
[2010]; Papadopoulos et al. [2012]; Serrano, Boguñá, and Sagues [2012]; Ser-
rano, Krioukov, and Boguñá [2008]). We took advantage of the large amount
of empirical data available and the current explosion in computing power to
create meaningful geometric maps of large real networks by embedding them
in an underlying space that ought not to be geographical or spatially obvious.
In this Element, we review our most relevant research on this topic, with a spe-
cial focus on models and applications to real networks. These results triggered
the field of network geometry to become one of the fundamental areas within
network science devoted to the discovery and modeling of nontrivial geometric
properties of complex networks (Boguñá et al. [2020]).
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

Complex networks typically have been studied as topological objects (Doro-

govtsev and Mendes [2003]; M. E. J. Newman [2010]), graphs where elements
are represented as nodes and their interactions as links. Graphs of real net-
works are not regular lattices nor are they completely disordered or random,
and their structure is imprinted with universal features. One of the most par-
adigmatic examples is the small-world phenomenon, connecting every pair
of nodes in a network, on average, by a small number of intermediate links
(Amaral et al. [2000]; Watts and Strogatz [1998]). Other ubiquitous proper-
ties are scale-free, or heavy tailed, distributions of the number of connections
per node (degree) (Barabási and Albert [1999]), with a few nodes linked to an
enormous number of neighbors (hubs with very high degrees, while most other
nodes are poorly connected), modularity, and hierarchical structure (Amaral
[2008]). These and other prevalent features are not a mere curiosity but arise
as the outcome of evolutionary pressures or functional needs and affect the
dynamics that characterize or take place within and between networks (Barrat,
Barthélemy, and Vespignani [2008]).
 2 The Structure and Dynamics of Complex Networks

B
C
network topology A

hidden metric space

Figure 1 Hidden metric spaces help to understand the structure and

function of complex networks. The smaller the distance between two nodes
in the hidden metric space – the more similar they are – the more likely they
are connected in the observable network topology. If node A is close to node
B, and B is close to C, then A and C are necessarily close because of the
triangle inequality in the metric space. Therefore, triangle ABC exists in the
network topology with high probability, which explains the strong clustering
observed in real complex networks.

One of the main consequences of the small-world effect is the apparent

lack of a metric structure defined on the system. Certainly, in a small-world
network, the distribution of shortest path lengths among pairs of nodes is
sharply peaked around its average and, therefore, any pair of nodes is roughly
separated by the same minimal number of intermediate links. This is the rea-
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

son why complex networks are often considered as pure topological objects
and are difficult to map. Yet, many networks are embedded in metric spaces.
Some are explicit (Barthélemy [2011]) – like in airport networks (Barrat et al.
[2004]; Guimerà et al. [2005]), power grids, or urban networks – whereas
some are hidden yet shaping the network topological structure (Boguñá et al.
[2010]; Krioukov et al. [2010]; Krioukov et al. [2009]; Serrano et al. [2008]);
see Figure 1. This idea led to hidden metric space models of complex net-
works with an underlying effective hyperbolic geometry. These models are
able to explain universal features observed in real-world systems, includ-
ing the small-world property, scale-free degree distributions, clustering, and
also fundamental mechanisms like preferential attachment in growing net-
works (Papadopoulos et al. [2012]), the emergence of communities (Zuev
et al. [2015]), and multiscale self-similarity (García-Pérez, Boguñá, and Ser-
rano [2018]). The discovery of the hidden geometry of real complex networks
led to the emergence of the field of network geometry (Boguñá et al. [2020]),
a major research area within network science.
 The Shortest Path to Network Geometry 3

The hidden metric space network models of complex networks couple their
topology to an underlying geometry through a probabilistic connectivity law
depending on distances in the space, which combine popularity and simi-
larity dimensions in such a way that more popular and similar nodes have
more chance to interact (Krioukov et al. [2010]; Papadopoulos et al. [2012];
Serrano et al. [2008]). The basic assumptions in our model are that there
exists some similarity between nodes which, along degrees, plays an important
role in how connections are established and that, since similarity is transi-
tive, geometry is an appropriate mathematical formalism to encode it. The
clue for the connection between topology and geometry is then clustering –
transitive relationships, or triangles – which arises as a reflection in the topol-
ogy of the network of the triangle inequality in the underlying hidden metric
space. These models can be combined with statistical inference techniques to
find the coordinates of the nodes in the underlying metric space that maximize
the likelihood that the topology of the network is reproduced by the model
(Blasius et al. [2018]; Boguñá et al. [2010]; García-Pérez et al. [2019];
Papadopoulos, Aldecoa, and Krioukov [2015]). One of the key properties of
these maps is that the shortest paths in the topology of the networks follow
closely geodesic lines in the underlying space. This ensures that networks
highly congruent with the hidden metric space model are navigable, where nav-
igability is understood as efficient routing of information based on the metric
embedding (Allard and Serrano [2020]; Boguñá and Krioukov [2009]; Boguñá
Krioukov, and Claffy [2009]; Boguñá et al. [2010]; Gulyás et al. [2015];
Krioukov et al. [2010]; Papadopoulos et al. [2010]).
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

One example of the power of this geometric approach is the discovery of the
hyperbolic plane as the effective geometry of many real networks (see Fig. 2),
including such disparate systems as the Internet at the Autonomous Systems
level (Boguñá, Papadopoulos, and Krioukov [2010]), genome-scale reconstruc-
tions of metabolic networks in the cell (Serrano et al. [2012]), the World Trade
Web from 1870 to 2013 (García-Pérez et al. [2016]), and brains of different spe-
cies (Allard and Serrano [2020]). In the case of the Internet, the metric space
provides a solution to the scalability limitations of current inter-domain rout-
ing protocols. For metabolic networks, it allows us to redefine the concept of
biological pathways and to quantify their crosstalk. In international trade, the
maps provide information about the long-term evolution of the system, unrav-
eling the role of globalization, hierarchization, and localization forces. Finally,
the effective geometry of human and nonhuman brain structures is also better
described as hyperbolic than Euclidean, thus implying that hyperbolic embed-
dings are universal and meaningful maps of brain structure that allow for an
efficient routing of information.
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

Figure 2 Hidden metric space maps of real complex networks in the hyperbolic plane. (a) Embedding of the human connectome, (b)
Internet graph at the Autonomous Systems level, and (c) embedding of the World Trade Web as of year 2013. In all cases, the size of a
node is proportional to the logarithm of its degree, and the radial coordinate decreases with increasing degree so that higher degree nodes
are placed closer to the center of the disk. In (a) and (b) brain regions’ and countries’ names are located at the average angular position of
all nodes belonging to the same region/country.
Source: Panel (a) modified from Zheng et al. (2020). Panel (b) reprinted from Boguñá et al. (2010). Panel (c) reprinted from García-Pérez et al. (2016).
 The Shortest Path to Network Geometry 5

These results suggest that the geometric paradigm improves our knowledge
of the basic principles underlying the organization, function, and evolution of
complex systems. But, in the long run, it also will transform research on how to
model, predict, and control them, with potential implications for a large num-
ber of current challenges. These include efficient recommendation systems and
search engines, prediction of epidemic spreading, and drug design in cancer and
brain research.

2 Geometric Models for Static Topologies

Our first remarkable observation was to identify clustering – a measure of the
number of triangles in a graph – as the key connection between complex net-
works and an underlying hidden geometry. Indeed, the triangle inequality in
a metric space induces clustering in the structure of the graph, as illustrated
in Figure 1. In Serrano et al. (2008), we analyzed the clustering coefficient of
several real complex networks and found that their topological structure was
compatible with an underlying hidden metric space. This finding led us to intro-
duce the S1 class of network models (Serrano et al. [2008]). In these models,
nodes are embedded in a metric space and connections exist with a gravity-
law-like connection probability balancing the distance between nodes and their
degrees; see Figure 3a. The connection probability encodes, in a simple and
general way, the two major forces at play, namely, the effect of a similarity
distance and the effect of the importance of the nodes involved in the connec-
tions. In this way, the model is able to generate scale-free, small-world, and
clustered graphs very similar to those found in real complex networks, where
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

the heterogeneity in the distribution of the number of contacts per node can be
controlled independently of the level of clustering that measures the coupling
with the metric space.
The S1 model is a mixed model in the sense that it combines a metric
component and a topological component. Nodes are given coordinates in a
metric similarity space but are also given degrees, determining their number
of neighbors. At first glance, it seems difficult to combine, in a purely geo-
metric framework, the small-world and scale-free properties that we observe
in real networks. The major complication arises as a consequence of the small-
world effect. This effect implies an exponential expansion of space, that is, the
number of nodes within a disk of a certain radius grows exponentially with the
radius (up to the finite size of the system). This behavior is in stark contrast
to what happens in Euclidean spaces, where space grows as a power of the
radius, but it is similar to what happens in hyperbolic geometry. In Krioukov
et al. (2009, 2010), we developed the theory of random geometric graphs in
hyperbolic geometry; see Figure 3b. Interestingly, scale-free graphs are the
 6 The Structure and Dynamics of Complex Networks

Figure 3 (a) Model S1 . The similarity distances da between pairs of nodes

A1-A2, B1-B2, and C1-C2 have been highlighted. The size of a node is
proportional to its hidden degree . (b) Model H2 in the hyperbolic plane.
Nodes in the different pairs are separated by the same hyperbolic distance.
Nodes are equally sized, but nodes with higher hidden degree are positioned
closer to the center. The similarity distance is the same in the two
representations.
Source: Modified from the Supplementary Information in García-Pérez et al. [2016].

natural outcome of the formalism, indicating that this geometry is the most
appropriate to model complex networks. Indeed, it produces in a natural way
scale-free, small-world, and clustered graphs. However, the most surprising
result is that this class of models, which we call H2 , is isomorphic to our
previous S1 version (Krioukov et al. [2010]; Serrano et al. [2008]).
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

This duality allows us to use either model indistinctly, depending on the

particular application. The S1 version is especially convenient for theory
development, analytical calculations, and the implementation of embedding
techniques – which estimate the coordinates that maximize the likelihood of the
observed structure being produced by the model. Instead, the H2 version is well
suited for visualization purposes, to analyze navigation properties (Allard and
Serrano [2020]; Boguñá and Krioukov [2009]; Boguñá et al. [2009]; Boguñá
et al. [2010]; Gulyás et al. [2015]; Krioukov et al. [2010]; Papadopoulos et al.
[2010]), or to define hierarchies within the network (García-Pérez et al. [2016]).
An interesting aspect is that our geometric class of models corresponds
to an entropy-maximizing probabilistic mixture of grand canonical network
ensembles, where network links can be thought of as noninteracting fermions
whose energies depend on distances on the underlying geometry, with the
particular choice of the functional form of this dependency defining network
properties. At present, these models provide the simplest class of models
capturing sparsity, the small-world property, power-law degree distributions,
 The Shortest Path to Network Geometry 7

and nonvanishing clustering in a geometric framework with explicit symmetry

structure (Boguñá et al. [2020]).

2.1 The S1 Model

In the S1 model (Serrano et al. [2008]), a node i is assigned two hidden vari-
ables: a hidden degree i quantifying its popularity; and an angular position i
in a one-dimensional sphere (or circle), the similarity space, where distances
with the other nodes serve as a proxy for their similarity. The radius of the
circle is adjusted to R D N=2, where N is the number of nodes, so that the
density is set to 1 without loss of generality. The probability of connection
between any pair of nodes takes the form of a gravity law, whose magnitude
increases with the product of the hidden degrees (i.e., their combined popular-
ities) and decreases with the angular distance between the two nodes. In other
words, more-similar nodes are angularly closer and are, therefore, more likely
to be connected, whereas not-so-similar pairs of nodes have a high probabil-
ity of being connected only if they are popular. Specifically, nodes i and j are
connected with probability
1 1
pij D D  ˇ ; (2.1)
1C ˇij 1C
dij
i j

where  controls the average degree of the network, ˇ controls its level of clus-
tering, and dij D Rij , and ij D  j ji j jj is the angular distance
between nodes i and j. Notice that there are no constraints on the distribution of
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

hidden variables  and  . The angular distribution could be nonhomogeneous,

and both hidden variables could even be correlated. This is an important obser-
vation because such angular inhomogeneities or correlations can explain the
emergence of communities and other nontrivial topological patterns observed
in real networks (Allard and Serrano [2020]).
A priori, the functional form of the connection probability could be any
d
integrable function of argument f iji j . However, the Fermi–Dirac form of
the connection probability in Eq. (2.1) is the only possible choice that defines
maximally random ensembles of graphs that are simultaneously sparse,1 heter-
ogeneous, clustered, small-worlds, and maximally degree–degree uncorrelated
(Boguñá et al. [2020]).2 Besides, with this choice, parameter ˇ has full control

1 By sparse networks we mean ensembles of networks with size-independent average degree.

2 By maximally degree–degree uncorrelated we mean that the probability of a node with hidden
variable  having a neighbor with hidden variable  0 is independent of . Yet, for heteroge-
neous scale-free networks, some level of degree–degree correlation is unavoidable, as shown
in Boguñá, Pastor-Satorras, and Vespignani (2004).
 8 The Structure and Dynamics of Complex Networks

of the level of clustering without affecting the degree distribution. It can be

shown that the model undergoes a structural phase transition at ˇ D 1 so that,
for ˇ < 1, networks are unclustered, whereas for ˇ > 1, the ensemble gener-
ates networks with finite clustering in the thermodynamic limit (Serrano et al.
[2008]).

AS A G N S1 E
The algorithm below generates networks from the S1 ensemble in the limit
N  1, in the simple scenario of uncorrelated hidden variables  and  ,
and with the similarity coordinate homogeneously distributed.

1. Fix the number of nodes N, parameter ˇ > 1, and the target average
degree hki
2. Set  to
 
ˇ 
D sin
2hki ˇ
3. Assign a hidden degree  to every node from ./ so that hi D hki.
Assign also an angular position  to each node sampled uniformly at
random within the interval Œ0; 2.
4. Connect every pair of nodes with probability given by Eq. (2.1).

With this parametrization – and in the thermodynamic limit – the expected

degree of a node with hidden degree  is simply k./ N D , which justifies the
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

name of hidden degree. Indeed, the expected degree of any node i with hidden
P
variables .i ; i / can be evaluated as k. N i ; i / D
j pij , where the connection
probability is given in Eq. (2.1). If the network is homogeneous with respect
N i ; i / is independent of i . Thus, the expected degree
to the similarity space, k.
of any node with hidden degree , located without loss of generality at  D 0,
can be expressed as
Z "Z N #
Nk./ D 2  0 . 0 /
2 0 dt
d 0 D (2.2)
0 1 C tˇ

Z  ˇ !
0 1 1 N
DN . /2 F1 1; ; 1 C ; d 0 ;
ˇ ˇ 2 0

where 2 F1 .1; ˇ1 ; 1 C ˇ1 ; xˇ / is the hypergeometric function, whose asymp-

totic behavior when x ! 1 is 2 F1 .1; ˇ1 ; 1 C ˇ1 ; xˇ /   csc.=ˇ/=.ˇx/.
 The Shortest Path to Network Geometry 9

Using this result, we recover the proportionality between expected and hidden
degrees.
The degree distribution of the model can be evaluated as
Z
1
P.k/ D  k e  ./d; (2.3)
kŠ
that is, a mixture of Poisson distributions weighted by ./. Eq. (2.3) shows that
the model generates nodes with degree zero with probability P.0/ D he  i, so
that the expected number on nonzero degree nodes is Nobs D NŒ1 P.0/,
whereas the observable average degree (counting only nodes with nonzero
degree) is hkiobs D hki=Œ1 P.0/.
In the case of scale-free networks, we consider ./ to be a power-law
distribution of the form
2
./ D . 1/0 1  ;  > 0 D hki ; > 2: (2.4)
1
Plugging this expression into Eq. (2.3), the degree distribution reads
C1 1 €.k
; 0 /
P.k/ D . 1/0 ; (2.5)
kŠ
where €.kC1 ; 0 / is the incomplete gamma function, so that the asymptotic
behavior of the degree distribution is P.k/  k . To simulate sparse scale-free
networks with < 2 (as found, for instance, in airport networks) we need to
introduce a cutoff in the distribution of hidden degrees c . In particular, we
choose a hard cutoff of the form
1
. 1/0
./ D  1  with 0 <  < c ; (2.6)
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

c
1 0

where the lower cutoff is the solution of the equation

 2
c
1 1 0
hki D 0  1 : (2.7)
2 c
1 0

Equations (2.6) and (2.7) can also be used to compensate for finite size effects
in scale-free networks with & 2. Indeed, to prevent extreme fluctuations
arising when is very close to 2, instead of generating values of  from the
unbounded distribution Eq. (2.4), we introduce a hard cutoff whose value is the
same as the natural cutoff of the unbounded distribution, which can be approx-
imated by c D 0 N1=. 1/ (Boguñá, Pastor-Satorras, and Vespignani [2004]).
Then, we generate values of  from Eq. (2.6) with parameter 0 equal to
1
1 N 2
0 D 2
hki: (2.8)
1 N 1 1
 10 The Structure and Dynamics of Complex Networks

Notice that when is very close to 2, finite size effects can be very impor-
tant even for large networks. However, notice that this is not the only source
of finite size effects. To fully account for finite size effects, we must also con-
sider the effects coming from the upper limit in the integral in Eq. (2.2), as
done in García-Pérez et al. (2019). However, in many practical applications,
the correction in Eq. (2.8) is enough.
The S1 model can be used to produce synthetic ensembles with controllable
structural features or for high-fidelity replication of a specific real network. To
that end, observed degrees in the real network can be taken as good proxies of
hidden degrees, and parameters  and ˇ can be tuned to reproduce the average
degree and clustering of the real network. This procedure is not very accurate
for heterogeneous networks due to strong fluctuations. Actual hidden degrees
could be estimated from real data to avoid the mismatch between hidden and
observed degrees, but this operation can be demanding and, besides, there is
no guarantee that all nodes end up with the exact same degree they had in the
real network. An alternative is the implementation of the geometric randomiza-
tion model introduced in Starnini, Ortiz, and Serrano (2019), which preserves
exactly the degree sequence of the input network while producing a version of
the network maximally congruent with the S1 model.
The geometric randomization model assumes the same form of the connec-
tion probability as in the S1 model. Given a real network, nodes are given
angular coordinates in the similarity space uniformly at random. Then, the net-
work is rewired in order to maximize the likelihood that the new topology is
generated by the S1 model while preserving the observed degrees and, thus, the
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

total number of edges. After selecting a value of ˇ, for instance, the one that
replicates the level of clustering of the original network, the rewiring procedure
is conducted by executing a Metropolis–Hastings algorithm as follows.

G R M

1. Assign each node an angular coordinate uniformly at random.

2. Choose two links at random, say between nodes i and j and between
nodes l and m.
3. Compute the probability of rewiring (connecting i and l and j and m) as
     
Lnew ij lm ˇ
pr D min 1; D min 1; il jm ; (2.9)
Lold
where Lnew corresponds to the value of the likelihood function after the
swap and Lold before the swap, both evaluated using Eq. (3.1) (see next
section) and the probability of connection in Eq. (2.1). Notice that pr
only requires information about the angular coordinates of nodes.
 The Shortest Path to Network Geometry 11

4. Accept the link swap with probability pr , making sure not to create a
multiple link or a self-loop.
5. Stop the algorithm after the likelihood fluctuates steadily around a
plateau.

Notice that the resulting network might not have global connectivity even if the
degrees of the nodes have not changed. The geometric randomization model
can be used as a null model in the analysis of features which are particularly
sensitive to fluctuations in the degree cutoff, like the behavior of dynamical
processes such as epidemic spreading or synchronization.

2.2 The H2 Model

The H2 model (Krioukov et al. [2010, 2009]) is a quasi-isomorphic purely
geometric version of the S1 model, in which the popularity and similarity
dimensions are combined into a single distance in the hyperbolic plane, such
that hyperbolically closer nodes are more likely to be connected; see Figure 3b.
There are many representations of the hyperbolic plane H2 . In the two-
dimensional hyperboloid model that we use in our work, points in H2 (of
constant curvature K D 1) are characterized by two coordinates .r;  / and
the metric tensor reads

ds2 D dr2 C sinh2 rd 2 : (2.10)

https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

By comparing it to the metric of the familiar Euclidean two-dimensional plane

in polar coordinates, ds2 D dr2 Cr2 d 2 , we immediately notice that the perime-
ters of hyperbolic circles and the areas of disks of radius r grow much faster
with r than do those of Euclidean circles. In particular, they grow exponentially
fast when r  1, akin to growth behavior in trees. This is the first indication
that networks in hyperbolic geometry may be endowed with the small-world
property while still being purely geometric.
The hyperbolic distance between two points at radial coordinates ri and rj ,
separated by an angular distance ij , can be computed with the hyperbolic
law of cosines and reads

cosh xij D cosh ri cosh rj sinh ri sinh rj cos ij : (2.11)

This expression can be rewritten in a more convenient way:

 
ij ij
cosh xij D cosh.ri rj / 1 sin2 C cosh.ri C rj / sin2 : (2.12)
2 2
 12 The Structure and Dynamics of Complex Networks

From the previous equation, it is easy to see that when

ij cosh.ri rj /
sin2  ; (2.13)
2 cosh.ri rj / C cosh.ri C rj /
then the first term in the right side of Eq. (2.12) can be neglected. Besides, if
ri > rj  1, then the term in the right hand side of Eq. (2.13) goes as e 2rj and
the inequality Eq. (2.13) is fulfilled even for very small angular separations.
Therefore, in this limit, a very good approximation for the hyperbolic distance
is given by
ij ij
xij  ri C rj C 2 ln sin  ri C rj C 2 ln : (2.14)
2 2
This expression will show its importance when analyzing the equivalence
between the hyperbolic network model described next and the S1 model.
Networks in the H2 model are generated by distributing points within a disk
of radius RH2 in the hyperbolic plane and connecting pairs of nodes i and j with
probability pij D f .xij RH2 /, where xij is the hyperbolic distance between the
two nodes and f .
/ is a decreasing function of its argument. The distribution of
points within the disk is, a priori, arbitrary and, as in the case of the S1 model,
the radial and angular coordinates could be correlated. In the simplest case,
the angular distribution is uniform and the radial distribution is proportional to
the volume element of H2 , so that points are then homogeneously distributed
within the disk. This corresponds to distributing points radially according to
the probability density
sinh r
.r/ D ; with r 2 Œ0; RH2 : (2.15)
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

cosh RH2 1
R 2
H
The model generates sparse graphs when N / e 2 , in which case the degree
distribution has a power-law tail with exponent D 3 (Krioukov et al. [2010,
2009]). If, instead, nodes are distributed (quasi-uniformly) with the probability
density
sinh ˛r
.r/ D ˛ ; with r 2 Œ0; RH2  (2.16)
cosh ˛RH2 1
and ˛  1=2, then the degree distribution has a power-law tail with expo-
nent D 2˛ C 1. In both cases, the underlying metric structure induces the
emergence of clustering in the thermodynamic limit, modulated by the specific
functional form of the connection probability pij . As in the S1 model, maxi-
mum entropy ensembles are obtained when this connection probability takes
the Fermi–Dirac form,
1
pij D ˇ
; (2.17)
1 C e 2 .xij RH2 /
 The Shortest Path to Network Geometry 13

with ˇ > 1. As in the S1 model, parameter ˇ controls the level of clustering of

the ensemble, approaching zero when ˇ ! 1C and converging to a constant
value when ˇ ! 1. Interestingly, Eq. (2.17) suggests that we can interpret
the network ensemble as a system of noninteracting fermions – the edges –
that can occupy different available states – all possible pairs among nodes –
with energies defined by the corresponding hyperbolic distances, ˇ being the
inverse of the system’s temperature, and RH2 the chemical potential controlling
the expected number of fermions (Krioukov et al. [2010]).
To establish a mapping between the S1 model and the H2 model, the angular
coordinates remain as in the S1 model, but the hidden degrees are transformed
into radial coordinates according to
i
ri D RH2 2 ln ; (2.18)
0
where the radius of the two-dimensional hyperbolic disk containing all nodes
is
N
RH2 D 2 ln : (2.19)
02

Higher-degree nodes are therefore located closer to the center of the H2

disk, whereas low-degree nodes are placed near its boundary. In fact, if hid-
den degrees are power-law distributed according to the probability density
in Eq. (2.4), then the mapping Eq. (2.18) causes the radial coordinates to be
distributed as in Eq. (2.16) for large RH2 .
Substituting Eqs. (2.18) and (2.19) into the connection probability of the S1
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

model Eq. (2.1) yields

1
pij D ˇ
; (2.20)
1Ce 2 .Q
xij RH2 /


where xQ ij D ri C rj C 2 ln 2 ij is, as discussed above, a very good approxi-
mation of the hyperbolic distance between two points with coordinates .ri ; i /
and .rj ; j / in the hyperbolic disk with curvature K D 1.3 Besides, the dis-
crepancy between xQ ij and the true hyperbolic distance xij is not relevant in the
case of networks. Indeed, the expected value of the smallest radial coordinate
corresponds to the expected value of the largest hidden degree, which scales

3 Note that the Fermi–Dirac connection probability is not a requirement for the mapping to hold.
In fact, if the connection probability in the S1 model is an integrable but otherwise arbitrary
d
function of iji j , then in the hyperbolic representation the connection probability is a function
1
of the argument e 2 .Qxij RH 2 /
.
 14 The Structure and Dynamics of Complex Networks
1 2
as c  0 N 1 , and so, it scales as rmin  2 1
ln N. For such nodes, the
inequality in Eq. (2.13) becomes
2
  2N2 1 : (2.21)

However, in a homogeneous angular distribution, the average distance between

two consecutive nodes – and so the minimal distance – is of the order N 1 ,
which is larger than the value in Eq. (2.21) for > 3 and, thus, the approxi-
mation xij  xQ ij holds for all pairs of nodes. When 2 < < 3, the number of
closest nodes to the hub for which the approximation does not hold exactly
3
scales as N 1 , which is a vanishing fraction of its total number of neigh-
bors. Consequently, the number of pairs of nodes for which the approximation
xij  xQ ij does not hold is a vanishing fraction in the thermodynamic limit.
Besides, for all these pairs of nodes, the connection probability is very close
to 1 both in the S1 and H2 models, so that those pairs of nodes are almost
surely connected in both models. After these considerations, we conclude that
the mapping between S1 and H2 models becomes exact in the thermodynamic
limit. This mapping allows us to use both models indifferently depending on
the application at hand.

AS A G N H2 E
The algorithm below generates scale-free networks with exponent from
the H2 ensemble in the limit N  1.

1. Fix the number of nodes N, parameter ˇ > 1, and the target average
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

degree hki.
2. Set the radius of the hyperbolic disk to
"  #
2N 1 2
RH2 D 2 ln :
ˇ sin .=ˇ/hki 2

3. Assign a radial coordinate r to every node from Eq. (2.16). Assign also
an angular position  to each node sampled uniformly at random within
the interval Œ0; 2.
4. Connect every pair of nodes with probability given by Eq. (2.17).

Alternatively, synthetic H2 networks can be obtained by generating syn-

thetic S1 networks and transforming the obtained hidden degrees into radial
coordinates using Eq. (2.18).

The two quasi-isomorphic models S1 and H2 generalize to spheres SD of any

dimension D and to hyperbolic spaces HDC1 , with the connection probability
 The Shortest Path to Network Geometry 15

d
being a function of the argument . ij/1=D in the case of the SD model (Serrano
i j
et al. [2008]). Interestingly, since the group of symmetries of hyperbolic spaces
is the Lorentz group SO(1,D+1), the equivalence between the two models is a
reflection of the isomorphisms between the Lorentz group SO(1,D+1) and the
Möbius group acting on the sphere SD as the group of its conformal transfor-
mations. This isomorphism is a starting point of the anti-de Sitter/conformal
field theory (AdS/CFT) correspondence in string theory (Maldacena [1998]).
Apart from the information in the previous paragraph, in this Element we
present our hidden metric space network models in a two-dimensional under-
lying geometry, where a one-dimensional similarity coordinate is combined
with a second popularity coordinate related to the degrees of the nodes. Even
if our models in similarity dimension D = 1 are very good at reproducing the
structure of real complex networks, there is no fundamental reason to believe
that the similarity space should be one-dimensional for them all. For instance,
the likelihood to trade between countries in the world can be dictated by cul-
tural, political, geographical, administrative, and economic dimensions, and
two countries can be at the same time close along one of them and far apart
along another.
In fact, if networks are metric and related to an underlying space, the dif-
ferences in the structure of that space due to changes in its dimensionality
should be reflected in the topology of the resulting networks. In particular, it
has been shown that the maximum clustering coefficient that can be obtained
from a geometric model decreases as the dimension of the space increases;
see supplementary material in García-Pérez, Boguñá, and Serrano (2018). This
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

suggests that real-world networks, which typically exhibit strong clustering,

must belong to the low-dimensionality regime. However, an excessive dimen-
sionality reduction applied when mapping real networks into the underlying
space would distort embedding distances, such that in general it would be dif-
ficult to tally the probability of connection in the networks with its metricity as
measured by the clustering coefficient. Taken together, this indicates that the
dimension of the similarity subspace of real complex networks will be typically
small. However, in general, we will need more resolution than that provided
by a single similarity coordinate to understand the different relationships that
shape the structure of the network and to be able to reproduce its structural
features with high fidelity.

3 Mapping Real Networks

Maps of real networks can be found by reverse-engineering our geometric mod-
els with statistical inference techniques to find the coordinates of the nodes in
 16 The Structure and Dynamics of Complex Networks

the underlying space that maximize the likelihood that the topology of the net-
work is generated by the model, an approach that was implemented in different
embedding tools (Blasius et al. [2018]; Boguñá et al. [2010]; Papadopoulos,
Aldecoa, and Krioukov [2015]; Papadopoulos, Psomas, and Krioukov [2015]).
More precisely, the maps are inferred by finding the hidden degree and angular
position of each node, f.i ; i /; i D 1;


; Ng, that maximize the likelihood
that the structure of the network was generated by the S1 /H2 model, where the
likelihood L is evaluated as
Y  aij  1 aij
LD pij 1 pij : (3.1)
i<j

Here pij is the connection probability, in our case given by Eq. (2.1), and faij g
are the entries of the adjacency matrix of the network (aij D aji D 1 if nodes
i and j are connected, aij D aji D 0 otherwise). This maximization, however,
is computationally expensive and cannot be performed using a brute-force
approach. To find meaningful results, we use heuristic optimization techniques
that explore the fundamental properties of the model, for instance, the fact
that, for scale-free networks, subgraphs of high-degree nodes have a higher
internal average degree as compared to the complete network (Serrano et al.
[2008]; Serrano, Krioukov, and Boguñá [2011]). This allows us to define an
onion-like structure that helps us to guide the maximization process (Boguñá
et al. [2010]; García-Pérez et al. [2019]; Papadopoulos, Aldecoa, and Krioukov
[2015]; Papadopoulos, Psomas, and Krioukov [2015]).
Recently, unsupervised machine learning state-of-the-art techniques have
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

also been used to embed complex networks into the hyperbolic plane, hav-
ing a very competitive computational complexity (Alanis-Lobato, Mier, and
Andrade-Navarro [2016a, 2016b]; Muscoloni et al. [2017]). These techniques
project high-dimensional data to a much lower-dimensional space (Sarveniazi
[2014]), leveraging on the redundancy of high-dimensional datasets where the
original features are correlated, such that only a small subspace of the original
representation space is eventually populated by the underlying process. Low-
dimensional representations can then be produced with minimal information
loss using dimensional reduction methods. However, the inherent randomness
of real complex systems can degrade the quality of unsupervised methods
that, on the other hand, will never inform about fundamental principles that
explain the structure of the observed data. In our recently published embedding
tool Mercator (García-Pérez et al. [2019]), we used the best of both machine
learning data-driven and maximum likelihood model-driven approaches to find
high-quality two-dimensional maps at the cost of an acceptable computational
complexity.
 The Shortest Path to Network Geometry 17

Using these mapping tools, meaningful and reliable hyperbolic maps have
been obtained for many real networks (Allard and Serrano [2020]; Boguñá
et al. [2010]; García-Pérez et al. [2019]; García-Pérez et al. [2016]; Serrano
et al. [2012]); see examples in Figure 2. The maps are highly congruent with
metadata not contained in the topology of the network or provided in the
embedding, like geopolitical information in the Internet or the World Trade
Web (Boguñá et al. [2010]; García-Pérez et al. [2016]), biochemical pathways
in metabolic networks (Serrano et al. [2012]), or neuroanatomical modules in
the brain (Allard and Serrano [2020]). One of the key properties of these maps is
that the shortest paths in the topology of real networks follow closely geodesic
lines in hyperbolic space. In other words, real network maps are navigable.
Navigability is understood as efficient transport of information, energy, matter,
or other media based on the metric embedding without the global knowledge
of the network structure and without finding shortest paths in the network, a
computationally intensive combinatorial problem. Instead, latent space guides
navigation in the network based on distances between nodes in the latent space.
However, not all networks are intrinsically navigable; a combination of degree
heterogeneity and clustering is needed to guarantee that geometric navigation
is able to discover long-range links to approach the target node in few hops and,
at the same time, to be able to find the target when reaching the local neigh-
borhood of the target. Interestingly, the vast majority of real complex networks
fulfill such constraints due to their scale-freeness and metric structure, which
suggests the interesting possibility that some real complex networks evolved
to optimize navigability (Boguñá et al. [2009]).
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

4 Mesoscale Organisation and Community Detection

One of the interesting empirical observations from these maps is that, in real
networks, nodes are not distributed homogeneously in the hyperbolic plane.
They display a radial gradient related to the degree distribution, so that higher-
degree nodes tend to appear closer to the center of the disk. More importantly,
nodes clusterize in the similarity subspace into densely populated angular
regions separated by voids zones, which correlates well with metadata not
explicitly contained in the topology of the network. For instance, in the case
of the Internet, Autonomous Systems in the same country tend to be close
in angular distance (Boguñá et al. [2010]). In the World Trade Web (García-
Pérez et al. [2016]), our mapping is highly congruent with geopolitical aspects,
placing in the same angular sectors countries that are either geographically or
politically close. The same type of results can be observed for metabolic net-
works in relation to biological pathways (Serrano et al. [2012]) and for the brain
and neuroanatomical modules (Allard and Serrano [2020]).
 18 The Structure and Dynamics of Complex Networks

This observation poses fundamental questions about the origin of such com-
munities and, at the same time, it offers the opportunity to design completely
new methods to detect them in real networks. While communities can be set by
hand in our models by prescribing a given angular distribution of nodes (Mus-
coloni and Cannistraci [2018]), in García-Pérez, Serrano, and Boguñá (2018)
and Zuev et al. (2015) we introduced a new mechanism that explains their
emergence. The mechanism is a generalization of the concept of preferential
attachment called geometric preferential attachment and tries to mimic the fact
that newborn nodes have a preference to appear in similarity areas that are
already highly populated. A single parameter, ƒ, modulates the strength of
the mechanism interpolating between a purely random angular choice for new-
born nodes and a strong preference to emerge close to already highly populated
regions. In this way, it is possible to generate random geometric networks with
the same local properties as the original models but with a complex mesoscale
organization, as found in real networks.
Beyond these theoretic considerations, the angular inhomogeneities
observed in maps of real networks suggest new methods to detect their com-
munity organization. The main idea is to use the angular distribution of nodes
in the similarity subspace to cluster nodes into what we call soft communi-
ties. This clusterization can be performed in different ways. For instance, in
our first method, named Geometric Critical Gap Method (G-CGM) (Serrano
et al. [2012]; Zuev et al. [2015]), angular gaps,  , between consecutive nodes
are measured, and values that exceed a certain critical value, c , are consid-
ered to separate adjacent soft communities of nodes in the similarity circle,
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

which determines a unique partition into nonoverlapping communities. The

critical gap is defined as the expected value of the largest gap that would be
obtained if the N nodes were distributed uniformly at random in Œ0; 2, so
that no communities are expected in this case. Following this idea, the crit-
ical gap can be calculated assuming a Poisson point process on the circle
of unit radius with density ı D N=.2/. In this case, the distribution of the
angular gaps is approximately exponential with rate ı, and the critical gap is
given by

ln N C
c D 2 ; (4.1)
N

where is the Euler–Mascheroni constant (Zuev et al. [2015]).

Our second method, the Topological Critical Gap Method (T-CGM) (García-
Pérez et al. [2016]), is a hybrid method combining geometry and topology
to select the partition of nodes in similarity space that maximizes modular-
ity Q, where modularity is the standard metric quantifying the quality of the
 The Shortest Path to Network Geometry 19

division of networks into clusters (Newman and Girvan [2004]). The method
is implemented as follows.

T C G M
Given a map of a real network:

1. Measure and sort all angular gaps from the smallest to the largest,
fi I i D 1; Ng with 1 < 2 <


< N .
2. Set the critical gap c D 1 so that each node is in its own commu-
nity. Compute the modularity Q1 and set Qmax D Q1 .
3. Set the critical gap c D 2 , find the corresponding partition and
compute its modularity Q2 . If Q2 > Qmax set Qmax D Q2 .
4. Repeat the previous step for all values of i and update, if needed,
Qmax .
5. Select the partition with the final Qmax .

At the end of the process, the partition corresponding to Qmax is the optimal
one delivered by the algorithm. Interestingly, the mutual information between
partitions found by our method and by the Louvain algorithm (Blondel et al.
[2008]) show that both partitions have a very high overlap, making the T-CGM
an alternative method to detect communities (García-Pérez et al. [2016]). How-
ever, the number of communities discovered by T-CGM and the corresponding
modularity is typically higher than that of the Louvain method, as, in general,
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

Louvain modules integrate smaller T-CGM communities. From this point of

view, we could say that the T-CGM has a better resolution than the Louvain
method.

5 Self-Similarity and Renormalisation

Self-similarity has been studied in real complex networks from different per-
spectives (Alvarez-Hamelin et al. [2008]; García-Pérez, Boguñá, and Serrano
[2018]; Kim et al. [2007]; Serrano et al. [2008, 2011]; Song, Havlin, and
Makse [2005]). The first obvious observation is that in many networks, the
degree distribution can be approximated by a power law, a first indication of
a self-similar organization. A first step to unravel self-similarity was made
in Song et al. (2005), where a renormalization scheme was defined based
on topological shortest paths. However, shortest paths are a poor source of
metric distances due to the small-world effect, and results were limited. The
geometric approach provides an alternative that has unraveled self-similarity
 20 The Structure and Dynamics of Complex Networks

in networks from the analysis of the topological properties of nested hier-

archies of subgraphs (Serrano et al. [2008, 2011]), and from a geometric
renormalization technique applied to network maps discovered by our algo-
rithms (García-Pérez, Boguñá, and Serrano [2018]). Next, we discuss the two
symmetries.

5.1 Self-Similarity of Nested Subgraphs

Given a real network, there is an arbitrary number of ways one can define a
hierarchy of nested subgraphs. A priori, given one such hierarchy, we cannot
expect their subgraphs to have the same topological properties as in the original
network. However, in real-world scale-free networks, the hierarchy defined by
selecting nodes with degree above a given threshold kT defines a sequence of
graphs that, except for a larger internal average degree, have the very same
topological properties when rescaled by the internal average degree. Figure 4
shows a scheme of the filtering procedure and examples of the behavior of the
clustering coefficient for the Internet and PGP graphs (Serrano et al. [2008]).
Even though the internal average degree increases almost a decade from kT D 0
to kT D 100 in both networks, the behavior of the clustering coefficient remains
invariant, except for finite size fluctuations. The figure also shows the same
procedure applied to randomized versions of the same networks preserving the
degree sequence. In this case, the self-similarity of the corresponding subgraphs
is lost; see Serrano et al. (2008) for further details.
Interestingly, the S1 =H2 model has this self-similarity property built in when
hidden degrees are power law distributed and subgraphs are defined by select-
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

ing nodes with hidden degrees above a certain threshold T , which, given the
proportionality between  and expected degree, is equivalent to the procedure
described above. In the thermodynamic limit, any (infinite size) such subgraph
of a graph generated by the S1 model with average degree hki, exponent and
inverse temperature ˇ is a realization of the same S1 model with the same
and ˇ and an average degree given by

hk.T /i D T3 hki: (5.1)

This result predicts an increasing internal average degree of subgraphs as we go

deep in the hierarchy, exactly as we observe in real networks. Thus, the S1 =H2
model explains the observed self-similarity of real networks and the behavior
of the average degrees of self-similar subgraphs.4

4 The same property is also present in the configuration model and in growing models of networks.
However, in the former case, clustering vanishes in the thermodynamic limit, and in the latter
 The Shortest Path to Network Geometry 21

Figure 4 Hierarchy of nested subgraphs. Top row, a hierarchy of subgraphs

https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

is induced by removing nodes of degree below kT D 1; 2; and 3 in the original

graph. Bottom plots show the behavior of the clustering coefficient as a
function of the rescaled degree for the Internet and PGP networks (left) and
their randomized counterparts (right) for different values of the threshold
degree kT .
Source: Reprinted from Serrano et al. (2008).

This observation has a very important implication. Indeed, networks with

a self-similar nested hierarchy of subgraphs and increasing internal average
degree contain subgraphs that are macroscopic in the thermodynamic limit and
have an arbitrary large average degree. This immediately affects the critical

the internal average degree is kept constant by construction, at odds with what is observed in
real networks (Serrano et al. [2011]).
 22 The Structure and Dynamics of Complex Networks

behavior of any process featuring a phase transition taking place on this type of
networks. For instance, it predicts zero percolation and epidemic thresholds or
an infinite critical temperature in the Ising model. In general, this self-similarity
property sets either to zero or infinity the critical point for any phase transition
where the critical point is a monotonic function of the average degree (Serrano
et al. [2011]). Notice that this result is independent of the fact that the degree
distribution may or may not be power law distributed.

5.2 Renormalization and Self-Similar Multiscale Unfoldings

Our geometric models enable a rigorous definition of self-similarity and
scale invariance (Mandelbrot [1961]; Stanley [1971]) in real complex net-
works by affording a valid source of geometric length scales that can be
used to design renormalization techniques (García-Pérez, Boguñá, and Ser-
rano [2018]). In statistical physics, renormalization methods help to explore
rigorously the properties of physical systems at different length scales by
recursive averaging over short-distance degrees of freedom. This approach
successfully explained, for instance, the universality properties of critical
behavior in phase transitions (Wilson [1975, 1983]). Inspired by a precursor
of the renormalization group, the block spin renormalization method of Leo
Kadanoff (2000), previous efforts to understand the scaling behavior of net-
works took a purely topological approach and were based on shortest path
lengths between nodes (Goh et al. [2006]; Kim et al. [2007]; Radicchi et al.
[2008]; Rozenfeld, Song, and Makse [2010]; Song et al.[2005]; Havlin, and
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

Makse [2006]). However, the collection of shortest paths is a poor source of

length-based scaling factors in networks due to the small-world (Watts and
Strogatz [1998]) or even ultrasmall-world (Cohen and Havlin [2003]) prop-
erty. Other studies have faced the multiscale structure of network models in
a somewhat more geometric way (Boettcher and Brunson [2011]; Newman
and Watts [1999]), but their findings cannot be directly applied to real-world
networks.
In our geometric maps, the similarity dimension offers a reservoir of dis-
tance scales that can be used to define a geometric renormalization group for
real networks (RGN) (García-Pérez, Boguñá, and Serrano [2018]). It is more
convenient to work in the formulation of S1 model that makes explicit the sim-
ilarity dimension and is mathematically more tractable. The method defines a
new map by coarse-graining nodes and rescaling interactions so that longer-
range connections are progressively selected at lower resolution scales. The
process can be summarized as follows; see also Figure 5.
 The Shortest Path to Network Geometry 23

Figure 5 Renormalization scheme. Layer l D 0 represents the embedding

of the original network in the S1 representation. Layer l D 1; 2 are obtained
by merging pairs of consecutive nodes. Layer l D 2 can be obtained directly
from l D 1 by merging groups of four nodes.
Source: Reprinted from García-Pérez, Boguñá, and Serrano (2018).

G R G
Given a real network, first obtain its geometric map using the embed-
ding tool Mercator (García-Pérez et al. [2019]) (or an alternative method).
The embedding of the network topology will assign hidden degrees  and
angular coordinates in the similarity circle  to every node. Then:
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

1. Define nonoverlapping blocks of r consecutive nodes along the simi-

larity circle and apply a coarse-graining by merging the nodes to form
supernodes.
2. In the new map, place each supernode within the angular region defined
by the corresponding block so that the order of nodes is preserved.
3. Connect two supernodes i and j in the new layer if and only if some node
in block i is connected to some node in block j in the original network.

The operation can be iterated by taking the new layer in place of the origi-
nal network to produce a multiscale unfolding. In the limit N ! 1, where
N is the number of nodes, the RGN can be applied up to any desired scale
of observation, whereas it is bounded to O.log N/ iterations in finite sys-
tems. In addition, notice that the transformation has the abelian semigroup
structure in the sense that it is equivalent to performing two consecutive
renormalization steps with r D 2 than one with r D 4.
 24 The Structure and Dynamics of Complex Networks

Using this transformation, geometric scaling was found in several real

scale-free networks from very different domains: the Internet at the Auton-
omous Systems level (Claffy et al. [2009]), the airports network (Kunegis
[2013]; OpenFlights network dataset – KONECT [2016]), the human meta-
bolic network at the cell level (Serrano et al. [2012]), a human protein-protein
interaction network (Rolland et al. [2014]), and Drosophila Melanogaster con-
nectome (Takemura et al. [2013]), the network of Enron emails (Klimt and
Yang [2004]; Leskovec et al. [2009]), Music (Serrà et al. [2012]), and Words
co-ocurrences (Milo et al. [2004]). The resulting topological features of three
of the renormalized networks are shown in Figure 6. We observe that the
degree distributions, degree-degree correlations – as measured by the average
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

Figure 6 Self-similarity of real networks along the RGN flow. Each

column shows the RGN flow with r D 2 of different topological features of
the Internet AS network (left), the Human Metabolic network (middle), and
the Music network (right). Top: Complementary cumulative distribution of
rescaled degrees k.l/ .l/ .l/
res D k =hk i. Middle: Degree-dependent clustering
coefficient over rescaled-degree classes. The inset shows the normalized
average nearest neighbor degree kN nn;n .k.l/ N .l/ .l/ .l/ 2
res / D knn .kres /hk i=h.k / i.
Bottom: RGN flow of the community structure; Q.l/ stands for the
modularity in layer l, Q.l;0/ is the modularity that the community structure of
layer l induces in the original network, and nMI.l;0/ is the normalized mutual
information between the latter and the community structure detected directly
in the original network. The number of layers in each system is determined by
their original size.
Source: Reprinted from García-Pérez, Boguñá, and Serrano (2018).
 The Shortest Path to Network Geometry 25

nearest neighbors degree, clustering spectra, and community structures show

self-similar behavior.
The self-similarity exhibited by real-world networks can be understood in
terms of their congruency with the hidden metric space network model. As we
showed analytically, the S1 and H2 models are renormalizable in a geometric
sense (García-Pérez, Boguñá, and Serrano [2018]). This means that if a real
scale-free network is compatible with the model and admits a good embedding,
the model will be able to predict its self-similarity and geometric scaling. In
the S1 model, the transformation that ensures that the renormalized networks
remain maximally congruent with the model assigns a new hidden degree i0 to
supernode i in layer l C 1 as a function of the hidden degrees of the nodes it
contains in the previous layer l according to
0 11=ˇ
Xr
i0 D @ jˇ A ; (5.2)
jD1

whereas the angular coordinate of supernode i, i0 can be taken as any

(weighted) average of the angular coordinates of the nodes within the supern-
ode. Global parameters need to be rescaled as 0 D =r, ˇ 0 D ˇ, and R0 D R=r.
With these transformations, the probability p0ij for two supernodes i and j to be
connected in the new layer maintains its original form given by Eq. (2.1); see
Figure 7a.
However, the preservation of the form of the connection probability shows
that the S1 model is renormalizable but not that renormalized graphs are
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

self-similar. For that, it is also necessary that the statistical properties of

the transformed hidden degrees  0 and angular coordinates  0 be preserved
under renormalization. If the original distribution of hidden degrees is asymp-
totically a power law with exponent and ˇ > . 1/=2, then the hid-
den degree distribution in the renormalized layers is, asymptotically, also
a power law with the same exponent , with the only difference being
the average degree. Interestingly, the global parameter controlling the clus-
tering coefficient, ˇ, does not change along the flow, which explains the
self-similarity of the clustering spectra. Finally, the transformation for the
angles preserves the ordering of nodes and the heterogeneity in their angu-
lar density, and, as a consequence, the community structure is preserved in
the flow (Boguñá et al. [2010]; García-Pérez et al. [2016]; Serrano et al.
[2012]; Zuev et al. [2015]); see Figure 7b. The model is, therefore, renormal-
izable, and RGN realizations at any scale belong to the same ensemble with
a different average degree, which should be rescaled to produce self-similar
replicas.
 26 The Structure and Dynamics of Complex Networks

Figure 7 RGN of the S1 model. a. Empirical connection probability in a

synthetic S1 network. Fraction of connected pairs of nodes as a function of
ij D R.l/ ij.l/ =..l/ i.l/ j.l/ / in the renormalized layers, from l D 0 to l D 8,
.l/

and r D 2. The original synthetic network has N  225;000 nodes, D 2:5,

and ˇ D 1:5. The black dashed line shows the theoretic curve (Eq. (2.1)). The
inset shows the invariance of the mean local clustering along the RGN flow.
b. Hyperbolic embedding of the metabolic network (top) and its renormalized
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

layer l D 2 (bottom). The colors of the nodes correspond to the community

structure detected by the Louvain algorithm. Notice how the renormalized
network preserves the original community structure despite being four times
smaller. c. Real networks in the connectivity phase diagram. The synthetic
network above also is shown. Darker blue (green) in the shaded areas
represents higher values of the exponent . The dashed line separates the
-dominated region from the ˇ-dominated region. In phase I,  > 0 and the
network flows toward a fully connected graph. In phase II,  < 0 and the
network flows toward a one-dimensional ring. The thick red line indicates
 D 0 and, hence, the transition between the small-world and
non–small-world phases. In region III, the degree distribution loses its
scale-freeness along the flow. The inset shows the exponential increase of the
average degree of the renormalized real networks hk.l/ i with respect to l.
Source: Reprinted from García-Pérez, Boguñá, and Serrano (2018).
 The Shortest Path to Network Geometry 27

Small-World/Large-World Transition. A good approximation of the

behavior of the average degree for very large networks can be calculated by
taking into account the transformation of hidden degrees in the RGN flow
(Eq. (5.2)). The average degree of the renormalized layer l C 1 as a function of
the average degree in the previous layer reads

hki.lC1/ D r hki.l/ ; (5.3)

with a scaling factor  depending on the connectivity structure of the original

network. If 0 < ˇ 1  1, the flow is dominated by the exponent of the degree
distribution , and the scaling factor is given by

2
D 1; (5.4)
1
1
whereas the flow is dominated by the strength of clustering if 1  ˇ
< 2,
and
2
D 1: (5.5)
ˇ

Therefore, if < 3 or ˇ < 2 (phase I in Figure 7c), then  > 0 and the
model flows toward a highly connected graph; the average degree is preserved
if D 3 and ˇ  2 or ˇ D 2 and  3, which indicates that the network
is at the edge of the transition between the small-world and non–small-world
phases; and  < 0 if > 3 and ˇ > 2, causing the RGN flow to produce
sparser networks approaching a unidimensional ring structure as a fixed point
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

(phase II in Figure 7c). In this case, the renormalized layers eventually lose
the small-world property. Finally, if ˇ < . 1/=2, the degree distribution
becomes increasingly homogeneous as r ! 1 (phase III in Fig. 7c), revealing
that degree heterogeneity is present only at short scales.
In Figure 7c, several real networks are displayed in the connectivity phase
diagram. All of them lie in the region of small-world networks having the
fully connected network as the fixed point, which indicates that long-range
connections are progressively selected by the RGN. Furthermore, all of them,
except the Internet, the Airports, and the Drosophila networks, belong to the ˇ-
dominated region. The inset shows the behavior of the average degree of each
layer l, hk.l/ i; as predicted, it grows exponentially in all cases.
Renormalization Explains Self-Similarity in the Brain. Beyond the
results for real networks presented above, the brain stands out as a paradigmatic
system where geometric renormalization can be contrasted with multiscale
empirical data. The architecture of the human brain underlies human behav-
ior and is extremely complex with multiple scales interacting with one another.
 28 The Structure and Dynamics of Complex Networks

In particular, multiscale human connectomes can be reconstructed at differ-

ent length scales by hierarchical coarse-graining of anatomical regions. The
multiscale organization of human connectomes at five different resolutions
was explored in Zheng et al. (2020). Strikingly, the structure of the human
brain remains self-similar when the resolution of observation is progressively
decreased, and geometric renormalization predicts the observed multiscale
properties. These results suggest that simple organizing principles underlie the
multiscale architecture of human structural brain networks, where the same
connectivity law dictates short- and long-range connections between different
brain regions over many resolutions. The implications are varied and can be
substantial for fundamental debates, such as whether the brain is working near
a critical point, as well as for applications including advanced tools to simplify
the digital reconstruction and simulation of the brain.
Scaled-down Network Replicas. The observed self-similarity of renormal-
ized real networks and their congruency with our model can be exploited
to produced scaled-down, high-fidelity replicas for useful applications; for
instance, as an alternative or guidance to sampling methods in large-scale simu-
lations and, in networked communication systems like the Internet, as a reduced
test bed to analyze the performance of new routing protocols (Papadopoulos
and Psounis [2007]; Papadopoulos, Psounis, and Govindan [2006]; Yao and
Fahmy [2008, 2011]). Scaled-down network replicas can also be used to per-
form finite-size scaling of critical phenomena in real networks, so that critical
exponents could be evaluated starting from a single-size instance network. Our
scaled-down networks can be produced at any scale in the range in which
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

self-similarity is preserved. The idea is to single out a specific scale after a

certain number of renormalization steps and to prune the extra links to adjust
its average degree to the level of the original network.

S -D R
Given a real network, a scaled-down network replica can be produced by
applying the following algorithm:

1. Obtain a S1 map of the real network, including its parameters, in

particular its value 0 .
2. Obtain a renormalized network layer using GRN up to the desired
size and evaluate the corresponding value of r . Typically, r > 0 so
that the average degree of the renormalized network is larger than the
original one. If this is the case, then proceed.
3. Set the auxiliary parameter  D 1.
 The Shortest Path to Network Geometry 29

4. Set new D  hk 0i
.
hkr i r
5. Go over all links in the renormalized network and keep each exist-
p .new /
ing link with probability qij D pij .r / , where pij ./ is the connection
ij
probability Eq. (2.1) with parameter . By doing so, the probability of
existence of a link in the pruned version is given by pij .new /.

Finite-size effects may play an important role in real networks, so that the
obtained average degree may not yet the target one. In this case, we readjust
the value of new as follows:

6. Set the tolerance for the difference between the obtained average
degree after pruning and the target average degree of the original
network, .
7. Compute the average degree hknew i of the pruned layer.
8. If hknew i hk0 i > 0, set  D  0:1u, where u is a random variable
uniformly distributed between .0; 1/. Go to step 4.
9. If hknew i hk0 i < 0, set  D  C 0:1u. Go to step 4.
10. The process ends when jhknew i hk0 ij < .

The high fidelity of scaled-down network replicas can be tested by repro-

ducing the behavior of dynamical processes in real networks. Three different
dynamical processes – the classic ferromagnetic Ising model, the susceptible-
infected-susceptible (SIS) epidemic spreading model, and the Kuramoto model
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

of synchronization – were tested in self-similar network layers of the different

real networks mentioned above. Results are shown in Figure 8. Quite remarka-
bly, for all dynamics and all networks, we observe very similar results between
the original and scaled-down replicas at all scales. This is particularly interest-
ing, as these dynamics have a strong dependence on the mesoscale structure of
the underlying networks.

6 Navigability
Interestingly, our geometric network model offers an explanation of the effi-
ciency of targeted transport in real networks (Boguñá et al. [2009]). Transport
of information, energy, or other media through networks is a universal phe-
nomenon in both natural and man-made complex systems. Examples include
the Internet, brain, or signaling, regulatory, and metabolic networks. The infor-
mation transport in these networks is not akin to diffusion. Instead, information
must be delivered to specific destinations, such as specific hosts in the Internet,
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

30
The Structure and Dynamics of Complex Networks
Figure 8 Dynamics on the scaled-down replicas. Each column shows the order parameters versus the control parameters of different
dynamical processes on the original and scaled-down replicas of the Internet AS network (left), the Human Metabolic network (middle),
and the Music network (right) with r D 2, that is, every value of l identifies a network 2l times smaller than the original one. All points
show the results averaged over 100 simulations. Error bars indicate the standard deviations of the order parameters.
a. Magnetization hjmji.l/ of the Ising model as a function of the inverse temperature 1=T. b. Prevalence hi.l/ of the SIS model as a function
of the infection rate . c. Coherence hri.l/ of the Kuramoto model as a function of the coupling strength  . In all cases, the curves of the
smaller-scale replicas are extremely similar to the results obtained on the original networks.
Source: Reprinted from García-Pérez, Boguñá, and Serrano (2018).
 The Shortest Path to Network Geometry 31

groups of neurons in the brain, or genes and proteins in regulatory networks.

At the same time the nodes in the network are not aware of the global network
structure, so that the questions we face are whether paths to specific destina-
tions in the network can be found without such global topology knowledge,
and how optimal these paths can be. In other words, we wonder whether real
complex networks are geometrically navigable.
Indeed, an underlying metric space can guide navigation in the network
based on distances between nodes (Boguñá et al. [2009]). That is, instead
of finding the shortest paths in the network – a computationally intensive
combinatorial problem in a network that changes dynamically, such as the
Internet (Korman and Peleg [2006]) – a transport process can be geometric,
relying only on geodesic distances in the space. Such processes are the more
efficient and robust (Muscoloni and Cannistraci [2019]) the higher the het-
erogeneity of the degree distribution – and so the smaller exponent – and
the larger the clustering coefficient – and so the larger the inverse tempera-
ture ˇ in our models. This defines a parameter range of navigable topologies,
where most real-world networks are found (Boguñá et al. [2009]). Interest-
ingly, networks from the S1 =H2 model are nearly maximally efficient for such
geometric navigation when the hyperbolic distance is used to define geodesic
paths (Bringmann et al. [2017]; Krioukov et al. [2010]).
The explanation of this phenomenon stems from the existence, for any pair
of nodes in hyperbolic networks, of topological shortest paths close to the cor-
responding geodesics in the underlying hyperbolic geometry. This is possible
due to the existence of superhubs interconnecting all parts of the network, pres-
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

ent as soon as < 3, in which case the networks are known to be ultrasmall
worlds (Cohen and Havlin [2003]). As shown in Boguñá and Krioukov (2009),
geometric navigation in hyperbolic networks with < 3 can always find these
ultrashort paths, and thus navigation in these networks is asymptotically opti-
mal. The other way around, networks that are maximally navigable by design
turned out to be similar to hyperbolic networks, and many real-world net-
works were found to contain large fractions of their maximum-navigability
skeletons (Gulyás et al. [2015]). Assuming that real-world networks evolve
to have a structure efficient for their functions, these findings provide an evo-
lutionary perspective on the emergence of latent geometry leading to structural
commonalities observed in many different real-world networks.
Finally, stochastic activation–inactivation dynamics of nodes enhances the
navigability of real networks with respect to the static case (Ortiz, Starnini,
and Serrano [2017]). The activation dynamics may represent temporal failures
of nodes due to random unknown events, or noise. Interestingly, there exists an
optimal intermediate activation value, which ensures the best trade-off between
 32 The Structure and Dynamics of Complex Networks

the increase in the number of successful paths and a limited growth of their
length. Contrary to expectations, the enhanced navigability is robust even when
the most connected nodes inactivate with very high probability. In fact, it is
even possible to improve the routing performance by switching on and off the
hubs of the network more often than the rest of the nodes. The results indicate
that some real networks are ultranavigable and remain highly navigable even if
the network structure is extremely unsteady, which may have important impli-
cations for the design and evaluation of efficient routing protocols that account
for the temporal or noisy nature of real complex networks.

7 Geometry of Weighted, Multiplex, and Growing Networks

In the previous sections, we have shown how to construct a geometric frame-
work for complex topologies that grants practical applications rooted on the
inference of hyperbolic maps. These applications include protocols for decen-
tralized navigability that suggest a new recipe for sustainable Internet routing
protocols, methods for the detection of hierarchical and community organiza-
tion, and a renormalization technique that unfolds networks into a multiscale
shell of self-similar layers revealing unexpected symmetries. The geometric
approach to networks does not stop here, but it has been extended beyond static
topologies to reach other network properties, such as weights and multiplex-
ity, and the emergence of fundamental mechanisms from local optimization
principles, such as preferential attachment.
Typically, weights in complex networks are coupled in a nontrivial way to
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

their binary topology, playing a central role in their structural organization,

function, and dynamic behavior. This is manifested, for instance, in a superlin-
ear relation between the strength s of a node (the sum of the weights attached
to it) and its degree k (Barrat et al. [2004]) Now, there is also empirical evi-
dence that weights in real networks present metric properties – links involved
in triangles tend to have larger normalized weights than the average link, where
the normalization is given by the average weight as a function of the product
of degrees of their endpoint nodes – and that the same underlying metric space
ruling the network topology also shapes its weighted organization (Allard et al.
[2017]).
These empirical findings are predicted by a very general model capable of
reproducing the coupling with the metric space in a very simple way (Allard
et al. [2017]). The model allows to fix the local properties of the nodes – their
joint degree–strength distribution – while varying the coupling of the topology
and, independently, of the weights with the hidden metric space. In real net-
works, the latter can be measured using a practical procedure that calibrates the
 The Shortest Path to Network Geometry 33

violation of the triangle inequality under the presence of noise in the weights.
Strong coupling with the metric space turns out to be a highly plausible expla-
nation for the observed weighted organization of many real networks, including
metabolic networks, human brain connectomes, international trade, cargo ship
movements, and more, although in some systems – for example the US air-
ports network – the coupling of weights with the underlying geometry is weak
in contrast with topology, which suggests that the formation of connections and
the assignment of their magnitude might be ruled by different processes.
Apart from weights, interactions between pairs of nodes in real networks
can be of different types. This leads to multiplex representations (Bianconi
[2018]), where links of different nature coexist and form different layers. These
structures are not random combinations of single networks but, in contrast,
exhibit significant hidden geometric correlations (Kleineberg et al. [2016]).
In real multiplexes, coordinates of nodes in hyperbolic maps of each separate
layer are significantly correlated, and so distances between nodes in the cor-
responding underlying hyperbolic spaces are significantly correlated. These
correlations are found in real multiplexes – the Internet, protein interaction
networks, collaboration networks, and more – at the level of both radial and
angular coordinates.
Radial correlations measured in real multiplexes are equivalent to corre-
lations among node degrees (Min et al. [2014]; Nicosia and Latora [2015];
Serrano, Buzna, and Boguñá [2015]). On the other hand, the observed cor-
relation among the angular similarity coordinates is a genuine geometric
feature with important theoretical and practical implications. Specifically,
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

hidden interlayer geometric correlations mitigate the vulnerability of multi-

plexes to targeted attacks on high-degree nodes, making multiplex networks
unexpectedly robust (Kleineberg et al. [2017]). Without geometric correla-
tions, multiplexes exhibit an abrupt breakdown of mutual connectivity, even
with interlayer degree correlations. With geometric correlations, a multistep
cascading process is instead observed, which does not destroy the system com-
pletely but leads into an eventually smooth percolation transition, with results
suggesting that it can be fully continuous in the thermodynamic limit.
On the practical side, geometric correlations can facilitate the definition and
detection of multidimensional communities, which are sets of nodes that are
simultaneously similar in multiple layers, and enable accurate trans-layer link
prediction, meaning that connections in one layer can be predicted by observing
the hidden geometric space of another layer. Also, sufficiently strong geomet-
ric correlations allow efficient targeted navigation in the multilayer system
using only local knowledge, outperforming navigation in the single layers.
These effects of geometric correlations can be assessed using a geometric
 34 The Structure and Dynamics of Complex Networks

multiplex network model that generates multiplex layers with realistic synthetic
topologies where correlations – both radial and angular – can be tuned with-
out altering the topological characteristics of each individual layer (Kleineberg
et al. [2017]).
Finally, beyond the structure of static networks, the geometric approach can
also explain mechanisms that underlie network growth in terms of local optimi-
zation processes such as preferential attachment, a common explanation for the
emergence of scaling in growing networks (Barabási and Albert [1999]). If new
connections are made preferentially to more popular nodes, then the resulting
distribution of the number of connections possessed by nodes follows power
laws, as observed in many real networks (Newman [2010]). Preferential attach-
ment can be a consequence of different underlying processes based on node
fitness, ranking, optimization, random walks, or duplication (Caldarelli et al.
[2002]; Dorogovtsev, Mendes, and Samukhin [2001]; D’Souza et al. [2007];
Fortunato, Flammini, and Menczer [2006]; Pastor-Satorras, Smith, and Sole
[2003]), but it can also emerge from a geometric description in which new
connections optimize the product between popularity and similarity. This idea
has been formalized in the PSO model (Papadopoulos et al. [2012]) that can
be thought of as a generalization of our geometric S1 =H2 model to growing
networks where the latent space is not hyperbolic but de Sitter space dS1;D ,
which has the same Lorentz group SO.1; D C 1/ of symmetries (Krioukov
et al. [2012]). As opposed to preferential attachment, PSO describes with good
accuracy the large-scale evolution of technological (the Internet), social (trust
relationships between people), and biological (metabolic) networks that grow
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

through the sequential addition of new nodes that connect to older ones in the
graph, predicting the probability of new links with high precision.
However, many real systems evolve in a self-similar way that preserves their
topology throughout the growth process over a long time span that is better
explained by branching of fundamental units – whether those be scientific fields
or countries (Zheng et al. [2021]). The Geometric Branching Growth model
predicts this evolution and explains the symmetries observed (Zheng et al.
[2021]). The model produces multiscale unfolding of a network in a sequence of
scaled-up replicas preserving network features, including clustering and com-
munity structure. When combined with scaled-down network replicas produced
by geometric renormalization (García-Pérez, Boguñá, and Serrano [2018]), the
model provides a full up-and-down self-similar multiscale unfolding of com-
plex networks that covers both large and small scales. Practical applications
in real instances that require optimization or control of system size in com-
plex networks are countless. They include the tuning of network size for best
response to external influence and finite-size scaling to assess critical behavior
 The Shortest Path to Network Geometry 35

under random link failures, assessment of scalability issues in dynamic pro-

cesses in core functions of real networks, such as in Internet routing protocols,
and more.
8 Conclusions
It is a very well-established empirical fact that most real complex networks
share a very special set of universal features. Among the most relevant ones,
networks have heterogeneous degree distributions, are small-worlds, and have
strong clustering. Our hidden metric space network model (Krioukov et al.
[2010]; Papadopoulos et al. [2012]; Serrano et al. [2008]), independently of
its S1 or H2 formulation, provides a very natural explanation of these prop-
erties with a limited number of parameters and using an effective hyperbolic
geometry of dimension 2. Even if the model and the corresponding renor-
malization group can be formulated in D dimensions, the high clustering
coefficient observed in real networks poses an upper limit on the potential
dimension of the similarity space (Dall and Christensen [2002]; García-Pérez,
Boguñá, and Serrano [2018]). Besides, in hyperbolic geometry space expands
exponentially fast,5 even in D D 2, so that networks can be faithfully embed-
ded in the one-dimensional S1 model even if the original network has a
higher-dimensional similarity subspace. This is in line with the accumulated
empirical evidence, which unambiguously supports the one-dimensional sim-
ilarity plus degrees as an extremely good proxy for the geometry of real
networks.
As we have discussed in Section 3, our models can be used both as topol-
ogy generators and as a mapping tool to obtain geometric representations of
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

real-world networks in the hyperbolic plane. One of the obvious advantages

offered by geometric maps of real networks is that many methods that have
been designed for extended systems in physics or for high-dimensional data in
computer science can be adapted to networks. Here we have introduced three
such applications: community detection, geometric navigation, and geometric
renormalization of complex networks. Concerning the latter, the existence of
a metric space underlying complex networks allows us to define a geometric
renormalization group that reveals their multiscale nature. Quite strikingly, our
geometric models of scale-free networks are shown to be self-similar under
such renormalization. Even more important is the finding that self-similarity
under geometric renormalization is a ubiquitous symmetry of real-world scale-
free networks, which provides new evidence supporting the hypothesis that
hidden metric spaces do underlie real networks.

5 Asymptotically, the volume of a ball of radius r grows as V.r/  e.D 1/r .

 36 The Structure and Dynamics of Complex Networks

The geometric renormalization group presented in this work is similar in

spirit to topological renormalization (studied in Goh et al. [2006]; Kim et al.
[2007]; Radicchi et al. [2008]; Rozenfeld et al. [2010]; Song et al. [2005,
2006]). However, instead of using shortest paths as a source of length scales to
explore the fractality of networks, geometric renormalization uses a continuum
geometric framework that allows us to unveil the role of degree heterogene-
ity and clustering in the self-similarity properties of networks. In our model,
a crucial point is the explicit contribution of degrees to the probability of
connection, giving the clue by which we can produce both short-range and
long-range connections using a single mechanism captured in a universal
connectivity law. The combination of similarity with degrees is a necessary
condition to make the model predictive of the multiscale properties of real
networks.
From a fundamental point of view, the geometric renormalization group
introduced here has proven to be an exceptional tool to unravel the global organ-
ization of complex networks across scales and promises to become a standard
methodology to analyze real complex networks. It can also help in areas like the
study of metapopulation models, in which transportation fluxes or population
movements happen on both local and global scales (Colizza, Pastor-Satorras,
and Vespignani [2007]). From a practical point of view, we envision many
applications. In large-scale simulations, scaled-down network replicas could
serve as an alternative or guidance to sampling methods, or for fast-track explo-
ration of rough parameter spaces in the search of relevant regions. Scaled-down
versions of real networks could also be applied to perform finite-size scal-
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

ing, which would allow for the determination of critical exponents from single
snapshots of their topology. Other possibilities include the development of a
new multilevel community detection method (Abou-Rjeili and Karypis [2006];
Karypis and Kumar [1999]) that would exploit the mesoscopic information
encoded in the different observation scales.
Our models have been extended to weighted networks, multiplexes, and
growing networks, but they still have to be extended to systems with asymmet-
ric interactions, represented as directed networks. Examples are found in many
different domains, from cellular networks, like in metabolic, gene-regulatory,
or neural networks; technological systems, like the World Wide Web or the
Internet, to social systems, as friendship interaction between two persons can
be perceived as different from person A to person B or from B to A. Such asym-
metry implies that the geometric paradigm must be adapted to this particular
type of system. The main caveat in this case stems from the breakdown of the
triangle inequality in directed networks, a fundamental property in any metric
space. Yet, given the ubiquitous presence of asymmetric relationships in real
 The Shortest Path to Network Geometry 37

systems, such extension promises to represent an important step forward in our

understanding of complex systems.
Network geometry today represents the best description of real complex net-
works, providing new insights into the fundamental principles that shape their
structure. Geometric models are able to encode in simple connectivity laws
most of the complex topological patterns of interactions in real complex sys-
tems. Future advances need to amplify the geometric framework to include the
study of dynamical processes that act as the necessary bridge between network
structure and function.
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press
 References
Abou-Rjeili, A., & Karypis, G. (2006). Multilevel algorithms for partition-
ing power-law graphs. In Proceedings 20th IEEE International Parallel &
Distributed Processing Symposium. doi: 10.1109/IPDPS.2006.1639360.
Alanis-Lobato, G., Mier, P., & Andrade-Navarro, M. A. (2016a). Efficient
embedding of complex networks to hyperbolic space via their Laplacian.
Sci. Rep., 6, 30108.
Alanis-Lobato, G., Mier, P., & Andrade-Navarro, M. A. (2016b, Nov. 15).
Manifold learning and maximum likelihood estimation for hyperbolic
network embedding. Applied Network Science, 1(1), 10. doi: https://doi
.org/10.1007/s41109-016-0013-0
Allard, A., & Serrano, M. Á. (2020). Navigable maps of structural brain
networks across species. PLOS Computational Biology, 16(2), e1007584.
Allard, A., Serrano, M. Á., García-Pérez, G., & Boguñá, M. (2017). The
geometric nature of weights in real complex networks. Nat. Commun., 8,
14103.
Alvarez-Hamelin, J. I., Dall’Asta, L., Barrat, A., & Vespignani, A.
(2008). K-core decomposition of internet graphs: hierarchies, selfsimilar-
ity and measurement biases. Networks and Heterogeneous Media, 3(2),
371–393.
Amaral, L. A. N. (2008). A truer measure of our ignorance. Proc. Natl. Acad.
Sci. USA, 105(19), 6795–6796.
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

Amaral, L. A. N., Scala, A., Barthélemy, M., & Stanley, H. E. (2000).

Classes of small-world networks. Proc. Natl. Acad. Sci. USA, 97(21), 11149–
11152.
Barabási, A. L., & Albert, R. (1999). Emergence of scaling in random networks.
Science, 286(5439), 509–512.
Barrat, A., Barthélemy, M., Pastor-Satorras, R., & Vespignani, A. (2004). The
architecture of complex weighted networks. Proc. Natl. Acad. Sci. USA,
101(11), 3747–3752.
Barrat, A., Barthélemy, M., & Vespignani, A. (2008). Dynamical processes on
complex networks. Cambridge: Cambridge University Press.
Barthélemy, M. (2011). Spatial networks. Phys. Rep., 499(1–3), 1–101.
Bianconi, G. (2018). Multilayer networks: structure and function. Oxford:
Oxford University Press.
Blasius, T., Friedrich, T., Krohmer, A. et al. (2018, Apr.). Efficient embedding
of scale-free graphs in the hyperbolic plane. IEEE/ACM Trans. Netw., 26(2),
920–933. doi: https://doi.org/10.1109/TNET.2018.2810186
 References 39

Blondel, V. D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, É. (2008). Fast
unfolding of communities in large networks. J. Stat. Mech., 2008(10),
P10008.
Boettcher, S., & Brunson, C. (2011). Renormalization group for critical
phenomena in complex networks. Frontiers in Physiology, 2, 102. doi:
https://doi.org/10.3389/fphys.2011.00102
Boguñá, M., Bonamassa, I., Domenico, M. D. et al. (2020). Network geometry.
Nat Rev Phys 3, 114–135. https://doi.org/10.1038/s42254-020-00264-4
Boguñá, M., & Krioukov, D. (2009). Navigating ultrasmall worlds in ultrashort
time. Phys. Rev. Lett., 102(058701). (arXiv:0809.2995v1)
Boguñá, M., Krioukov, D., Almagro, P., & Serrano, M. Á. (2020, Apr.). Small
worlds and clustering in spatial networks. Phys. Rev. Research, 2, 023040.
doi: https://doi.org/10.1103/PhysRevResearch.2.023040
Boguñá, M., Krioukov, D., & Claffy, K. (2009). Navigability of complex
networks. Nat. Phys., 5(1), 74–80.
Boguñá, M., Papadopoulos, F., & Krioukov, D. (2010). Sustaining the
Internet with hyperbolic mapping. Nat. Commun., 1, 62. doi: https://doi
.org/10.1038/ncomms1063
Boguñá, M., Pastor-Satorras, R., & Vespignani, A. (2004). Cut-offs and finite
size effects in scale-free networks. Eur. Phys. J. B, 38(2), 205–209.
Bringmann, K., Keusch, R., Lengler, J., Maus, Y., & Molla, A. R. (2017).
Greedy routing and the algorithmic small-world phenomenon. In PODC ’17:
Proceedings of the ACM Symposium on Principles of Distributed Comput-
ing. doi: https://doi.org/10.1145/3087801.3087829
https://doi.org/10.1017/9781108865791 Published online by Cambridge University Press

Caldarelli, G., Capocci, A., De Los Rios, P., & Muñoz, M. A. (2002, Decem-
ber). Scale-free networks from varying vertex intrinsic fitness. Phys. Rev.
Lett., 89(25), 258702. doi: https://doi.org/10.1103/PhysRevLett.89.258702
Claffy, K., Hyun, Y., Keys, K., Fomenkov, M., & Krioukov, D. (2009). Internet
mapping: From art to science. In 2009 Cybersecurity Applications Technol-
ogy Conference for Homeland Security (pp. 205–211). New York: IEEE. doi:
https://doi.org/10.1109/CATCH.2009.38
Cohen, R., & Havlin, S. (2003). Scale-free networks are ultrasmall. Phys. Rev.
Lett., 90(5), 058701.
Colizza, V., Pastor-Satorras, R., & Vespignani, A. (2007). Reaction-diffusion
processes and metapopulation models in heterogeneous networks. Nat.
Phys., 3, 276–282.
Dall, J., & Christensen, M. (2002). Random geometric graphs. Phys. Rev. E,
66(1), 016121. doi: https://doi.org/10.1103/PhysRevE.66.016121
Dorogovtsev, S. N., & Mendes, J. F. F. (2003). Evolution of networks: From
biological nets to the Internet and WWW. Oxford: Oxford University Press.
Another Random Scribd Document
with Unrelated Content
hoped it would, Mrs. Dingle’s objections and grievances, woven of
gossamer at first, began to grow tougher. She guessed that she
would catch more than herself in these elaborate reticulations, and
she persisted until she found another was becoming entangled also.
At first, to do her justice, Medora hesitated here. But she could not
pour her woes into Kellock’s ears without a reaction from him, and
his attitude towards her confession naturally influenced her. For,
while some of her elders suspected, according to the measure of
their wits, that Medora was acting, one man saw no shadow of
deception. Every word rang true on his ear, for circumstances
combined hopelessly to hoodwink him. His own serious nature, from
which any powers of illusion or sleight were excluded, read nothing
but the face value into Medora’s woeful countenance and the word
value into her hopeless speeches. Not for him to answer mock
heroics with banter, or reply to burlesque with irony. Had he been
made of different stuff, he might have saved Medora from herself at
this season; but being himself, the admirable man was terribly
perturbed and indeed found himself beset with sore questions and
problems from which both his character and personal attitude to the
girl precluded escape. For he loved her, and the fact that she was an
unhappy woman did not lessen his love; while, beyond that, his
altruistic instincts must have brought him into a delicate
complication in any case when once invited to participate. And now
he did enter, with motives that could not honestly be considered
mixed, for he was thus far influenced only by a conviction that it
might be possible to help both sufferers to a better understanding.
He knew that he enjoyed a far larger measure of intellect than Ned,
and he felt that to shirk an effort for Medora’s sake would be
cowardly. He had indeed convinced himself that it was his duty to
act.
He proceeded to tackle Ned, but he approached the task without the
attitude of mind vital to success. For success in such a ticklish matter
demanded in Kellock a standpoint of absolute impartiality. He must,
if he were to do any good whatever, come to Dingle with a mind as
open and unprejudiced as possible; whereas, though he knew it not,
Jordan’s mind by no means stood in that relation to the pair. Had it
done so, he had probably not interfered; for in truth it could not be
altruism alone that prompted him to the step he was now about to
take, but a very active and sincere sympathy for Medora in her
alleged griefs. He believed her with all his heart and he had a great
deal more concern for Mrs. Dingle’s point of view, which he
accepted, than for her husband’s, which he had neither heard nor
considered.
The men had eaten their dinner, and Ned, out of a cheerful
demeanour, which he brought from his work, presently sank into
taciturnity. From no will to do so, but powerlessness to prevent it, he
showed those about him that his thoughts were not pleasant.
Indeed the most casual had noticed that he was of late only himself
in the engine house, and that nothing but work sufficed to take him
out of himself. Away from it, he brooded and did not chatter and jest
as of old.
To-day he was more than usually abstracted and Kellock seized the
opportunity. Ned’s meal was finished in ten minutes and when he
began to stuff his pipe, the other asked him to come for a stroll up
the valley.
“Let’s go up to the ponds and see if there are any birds about, Ned,”
he said.
A little surprised, since the bird that interested Kellock was unknown,
Ned nevertheless agreed to take a walk.
“Certainly,” he answered. “Me and Trood flushed a woodcock there
yesterday, and I dare say on Saturday Trood will bring him down.
He’s a mark on a woodcock—never misses ’em.”
They strolled together up the valley where it fell gently to the Mill.
A quarter of a mile above the works the coomb narrowed to a bottle-
neck, through which a water-fall came down. The road wound
through this gap and on one side of it rose old, blue limestone
quarries, their jagged scarps and ridges fledged with gorse and oak
scrub; while on the other side of the water a limestone bluff
ascended, weathered to fine colour, and above it towered Scotch firs
and ivy-clad beeches that followed the foot of the hill and flung their
arms around a little mere, lying in the hollow of the undulating land.
In spring this cup shone emerald green; but now the place was grey
and silver. Alders and sallows towered black against the bright
water; sedges and reed mace had huddled into tangle of russet and
amber. They brightened where the sun touched them and burned
over the placid lake, while the highest colour note was a spindle
tree, whereon hung its harvest of pink and orange fruit, though all
the leaves were fled. The flame of it cast a brilliant reflection into the
face of the mirror below; and as Ned and Jordan approached by a
winding way, that skirted the mere, coot and moorhen scuttled off
leaving double trains behind them, widening out upon the waters.
Here it was that Kellock broached the great matter at his heart; and
because it was at his heart, whereas he imagined it solely in his
head, he found within the space of two minutes that he had made a
very grievous mistake.
Beside the lake spoke Jordan, while Ned had his eyes in the sedges
and distant mud flats for a woodcock.
“It’s about your wife I wanted to say a word, and I know we’re too
good friends for you to object. You see, Ned, when you look at the
past—”
“To hell with the past,” answered Dingle shortly. “It’s the future I
look at. You take my tip and keep out of this—specially seeing you
wanted her yourself once.”
“I must speak,” answered the vatman mildly, “and just for that
reason, Ned. When she took you, you’ll remember I followed a very
self-respecting line about it. But at your wish—at your wish, Ned—I
kept my friendship for Medora and you; and it’s out of that
friendship I want to say I think things might be bettered.”
“She’s been washing our dirty linen for your pleasure then?”
“Not at all. But—”
“God damn it!” burst out the other. “Ain’t there to be any peace left
in the world? You get out of this and keep out of it, or—”
“Don’t be silly, Ned,—listen.”
“To you? Not much. There’s some hooken-snivey going on here by
the looks of it. Blast you—there—that’s my answer to you!”
Dingle, in a white-hot passion, swung his arm, hit Kellock on the side
of his head with a tremendous blow and knocked him down. They
were on the edge of the lake and Medora’s champion rolled over and
fell into water ten feet deep. He was stunned and sank, then came
to the surface again.
Ned’s rage vanished with the blow, for now he saw in a moment the
gravity of the situation. Kellock appeared to be unconscious and
would certainly drown if left in the water.
The man on the bank flung himself upon his stomach, leant over,
gripped his victim by the collar and dragged him breast high under
the bank. In this position Kellock came at once to his senses.
“I’m sorry—I’m cruel sorry,” said Dingle. “Lift up your hands and put
’em round my neck—then I’ll heave you out.”
Kellock opened his eyes and panted, but did nothing for a moment.
“For God’s sake make an effort—I can’t help you else. Get your arms
round my neck, Jordan.”
The other obeyed and in a few moments he was safe. Ned fished his
cap out of the water, wrung it and handed it to him.
“I’m bitter sorry—my cursed temper.”
Kellock sat down for a moment and pressed the water out of his
clothes. He was quite calm.
“I dare say it was natural,” he answered. “If you’d but listened—”
“You can’t listen to things if you’re in hell. Take my arm. No good
biding here. I’ll see you to your house. You can have the law of me.
I deserve it. I’m no bloody good to anybody in the world now-a-
days. Better I was locked up, I reckon.”
“Don’t talk rot. We’re all learners. You’ve learned me something
anyway. See me home. I’m dazed, but I shall be all right in a
minute. And don’t let on about this. I shall say I slipped on the edge
of the water and fell in and bruised my head—just an accident and
my fault. And so it was my fault.”
“I won’t have that. You rub it in. I’ve earned it. I shall tell the people
what I am, if you don’t.”
“That won’t do,” answered the other. “Think of me as well as
yourself in that matter. You’re popular; I’m not; and if they hear
you’ve knocked me into the water, they’ll say there was a reason for
it.”
Dingle did not answer, but he knew this to be true.
“Least said soonest mended then.”
“For your wife’s sake, Ned.”
“Leave her out, please. I’m in your debt and I shan’t forget it.”
They met some women returning to the works and lied to them. All
expressed great concern. Then Ned brought Kellock to his rooms
and begged him to drink some spirits which he refused to do.
“Mind we tell the same tale about this,” said Jordan. “I fell in and
you grabbed me from the bank and brought me ashore. After all it’s
the truth, so far as it goes.”
Dingle agreed and then returned to his work; while the injured man,
though in considerable pain, only waited to change his clothes and
then hastened back to the Mill, to explain his accident and be
chaffed for his carelessness.
 CHAPTER IX
THE OLD PRIORY

There was none to drag up the melancholy blossoms of Medora’s

woe and display the fact that they had no roots; but she kept them
alive nevertheless; and since she was tickled to persist in folly by the
increasing interest created from her alleged sufferings, she woke up
to find those sufferings real at last. She had now earned a great deal
of pity and won a reputation for patience and endurance. She had
also awakened a certain measure of feeling against Ned, which was
inevitable, and now conditions which she had implied, knowing at
the bottom of her heart they did not exist, began to develop in
reality. The man was not built to watch Medora’s histrionics in
patience for ever, and she found him growing harsh and rough.
Then there was no more play-acting for Medora. Outraged in every
instinct, her sense of humour dead and her self-consciousness
morbidly hypertrophied, she began to hate the man she had
married. The cause of his changed attitude she forgot; and the bad
usage for which she had deliberately played, when it came she
resented with all her soul. Now she ceased to be a wife to him and
daily threatened to leave him.
A series of incidents more or less painful led to the threshold of
complete estrangement and Medora was always ahead of her
husband and always a good stage farther advanced to the final
rupture than was he. Indeed he never knew until the climax burst
upon him that it was so near. He did wrong things at this season,
was hard when he should have been gentle, and allowed himself
brutalities of speech and action. But again and again after such
ebullitions, he was contrite, abased himself and implored Medora to
help him to a better comradeship and understanding.
Each sought to confide, and Ned confided in Medora herself, while
she went elsewhere. Her interest was rapidly shifting and her
husband’s efforts at reconciliation meant nothing now. For the time
being she heartily loathed him, and the sound of his voice in the
house, and the fall of his foot. Yet between his furies he had
struggled hard to restore their friendship. He had confessed the
incident with Kellock and described to Medora how, in his passion
that anybody should presume to come between them, even with
good advice, he had turned on the vatman, knocked him into the
water and then pulled him out again.
“He meant well; but it shows what a state I’m in that I could do it.
He forgave me quickly enough, but I couldn’t forgive myself. And I
only tell you, Medora, to show what a perilous and unnatural frame
of mind I’ve got to. It’s all so properly cruel—as if some unseen devil
had poked his claws into our affairs and was trying to tear ’em apart.
And God knows I’ll do any mortal thing that man can do to right it.”
She was, however, much more interested in the disaster to Kellock.
“What did he say that made you try to murder him?” she asked.
“I didn’t try to murder him—I only shut his mouth. So I don’t know
what he was going to say. He admitted I was right anyway, and that
it was not his place to interfere.”
“Nobody’s got the right to talk sense to you seemingly.”
“I’m not telling you this for you to begin on me again,” he said. “I’m
telling you to show you what you’re doing and what you’ve done to
my temper. If anybody had told me a year ago I’d forget myself and
knock a man down for trying to do me a good turn, I’d never have
believed it. Yet such is my state that I did so. And since then I’ve
asked Jordan to speak about the thing and give me any advice he
could; but he’s told me frankly the time has passed for that. He
won’t speak now. He forgave me for knocking him into the water;
but I can see with half an eye he don’t want any more to do with
me.”
Medora, well knowing why this was, yet pretended not to know.
“You must ask yourself for a reason then and no doubt your
conscience will find it, Ned. We must cut a loss before long—you and
me—for I don’t want to die under this. I can’t stand very much more
and I dare say you feel the same.”
“What d’you mean by ‘cut a loss’?” he asked.
But after any pregnant remark of this description, Medora
temporised for a time and preferred to be indefinite.
“I don’t know what I mean,” she answered. “There’s times when I
wish I was out of it, young as I am. I can suffer and suffer of
course. I’m strong and there’s no limit to my endurance. But I’m
beginning to ask myself ‘why?’ And for that matter there are one or
two others asking me the same question.”
“No doubt,” he said. “The woman’s always right if her face is pretty
enough. You’ve got the art always to be in the right, and there’s only
one on God’s earth, and that’s me, who knows you’re wickedly in the
wrong quite as often as I am. It’s your wrongs in other people’s
mouths that made me do wrong; and when you saw me setting out
with all my heart to be patient and win you back again, you set
yourself wickedly to work to break down my patience and egg me
on. Again and again you’ve kept at me till I’ve gone too far and done
evil; and then you’ve run about everywhere and let everybody know
what a coward and brute I am.”
“That’s the way you talk,” she said, “and I can only listen with my
heart broken. You say these things for no reason but to make me
angry, and as to patience, even you will grant, if there’s any justice
left in you, that my patience has never broke down from the first.
And when the people have talked, I’ve laughed it off and put a
bright face on it.”
“Yes, I know that bright face—as though you were saying, ‘you see
I’m an angel already and only want the wings.’”
“Oh, your tongue!” she answered. “To think that ever you could
scourge a good wife with such bitter, biting words.”
Then she wept and he cursed and went out. It was a scene typical
of others; but from the moment that Medora heard of Kellock’s
immersion she could not rest until she had let him know she knew it.
They were meeting now unknown to Dingle, for though Jordan at
first protested against any private conference, Medora quickly over-
ruled him. For a month she had made it clear that only the wisdom
of Mr. Kellock was keeping her sane; and he believed it. Nor was this
altogether untrue, for Medora, now genuinely miserable, began to
seek increasing sustenance and support from her old lover.
As in the case of all her other schemes for entertainment and
exaltation, she crept to this and let it develop slowly. As the rift
between her and Ned grew wider, the gap narrowed between her
and Jordan Kellock. At each meeting she decreased the distance
between them, yet never by definite word or deed appeared to be
doing so. Kellock himself did not realise it. He knew the fact and
taxed his own conscience with it at first; but then for a time his
conscience left him in doubt as to his duty, until in the light of
Medora’s increasing sufferings, it spoke more distinctly and chimed
dangerously with his inclination.
His whole life was dominated by this great matter. It had become
personal and he wrestled with his difficulties by day and night.
Medora was one of those women who have a marvellous power of
influencing other judgments. She had a fatal gift to waken dislike
and distrust of another person in the mind of a third. She had
already created aversion for Ned in the minds of several women;
now Jordan, despite his own reason, felt himself beginning to hate
Dingle as heartily as Medora appeared to do. He fought this emotion
for a time; but found it impossible any longer to maintain an
impartial attitude. He told himself that it was only false sentiment to
pretend farther impartiality. Justice demanded antagonism to Ned in
the future—not because Medora had once been Jordan’s whole hope
and desire and was now herself unhappy and friendless; but
because, as an honest man, Kellock could not longer be impartial.
His views of life were changing; his orderly mind was beginning to
suspect that strong action might be necessary. Justice was the word
most often on his lips; and yet knowing that he loved Medora, he
was intelligent enough to perceive that inclination might be deluding
him and making apparently simple what, in reality, was complex. For
a time he hesitated; then came a day when he met Medora by
appointment and felt it impossible to stand outside her life any
longer. She, indeed, forced his hand and made it clear that she was
going to take definite steps for her own salvation.
Medora, on her way to Priory Farm one Sunday afternoon, had
arranged to meet Kellock at the ruins of the building that gave the
farm its name. Here they would be safe from any interruption.
The fragment of masonry crowned Mr. Dolbear’s orchard on the
summit of the hill that fell into Cornworthy. Here, heaved up against
the sky in its ivy mantle, stood the meagre remains of an old priory,
one of the smaller houses of the Austin nuns, founded by the
Norman lords of Totnes.
It consisted of a great gateway with a roof vaulted, ribbed and
bossed, and a lesser entrance that stood to the north of the first.
They pierced the mass and bore above them a chamber, of which
only the floor and ruined walls remained. It was reached by a stair,
where stone steps wound in the thickness of the wall and opened on
to the crown of the ruin fifty feet above. The space aloft was hung
with polypody and spleenwort in the chinks of its crumbling mortar,
and ivy knots seemed to hold the mass together. A whitethorn had
found foothold and rose above the central block of stone. Through a
ruined aperture facing east, one might see the orchard sloping to
the valley bottom and Cornworthy’s scattered dwellings, ascending
on the farther hill. The picture, set in the grey granite frame of the
priory window, revealed thatched houses grouped closely, with land
sweeping upwards on either side, so that the hamlet lay in a dingle
between the breasts of the red earth. The land climbed on beyond
the village and threw a hogged back across the sky. Here were
broad fallows and hedgerows where the leafless elms broke the line
with their grey skeletons. To this exalted but secret place, Medora
and Kellock were come. He had indeed been there some time when
she arrived.
“If you sit here,” he said, “you’re out of the wind.”
“We’re safe now,” she answered. “And ’twas like you to put yourself
about and tramp all this way. But I’ve got to be terrible careful,
Jordan, for if my husband thought I’d any friends working for me
and thinking for me, I don’t know what awful thing he’d do against
me. Nuns used to live here in past ages,” she continued. “Oh, my
God! I wish I’d been one of them. Then I should have spent my days
in peace and be at rest now.”
“Sit down and let’s use our time as best we can,” he advised.
“Time—time—I want for time to end. For two pins I’d jump out of
that window and end all time so far as I’m concerned.”
“You mustn’t talk or think like that, or else I shall fear I can’t be any
use. I tell you, before God, that my life’s all centred in you and your
troubles now. I shan’t have no peace till you have peace.”
“I’ll live for you then; and that’s about all I want to live for any
longer,” declared Medora. She felt in a theatrical mood and Ned’s
recent confession enabled her to speak with a great oncoming of
warmth and emotion. Her perception had fastened upon it from the
first and measured its value.
And now in the Priory ruin, she made the most of the matter. She
had worked it up and found it a tower of strength.
“I know what happened,” she said. “You hid it, Jordan, like the man
you are; but he told me how he knocked you into the water—cruel
devil.”
“I’m sorry he told you.—I asked him not to.”
“He wanted me to see what he could do, and would do again, and
will do again. He properly hates me now, and I shall soon be going
in fear of my life—I know that well enough. Not that I care much for
my life; but it’s awful to live with a tiger.”
“You don’t mean that, Medora?”
“I do then. He’s far ways different from what he was, or what
anybody thinks. He may pretend in the works; but he’s got the
temper of a devil; and sometimes I wish he’d strike and finish me;
and sometimes—I’m young and I don’t like to think of dying—
sometimes I say to myself I’ll make a bolt for it and go out into the
world and chance it. The world would be kinder than him and
anyway it couldn’t be crueller.”
“This is fearful—fearful,” he exclaimed. “I can’t stand you saying
these things, Medora.”
“I wouldn’t if they weren’t true. It can’t go on. I hate to distress you,
but there’s not a soul in the world cares a button what becomes of
me but you. I’m punished for the past I suppose. I deserve it. I took
that cruel tyrant when I might have took you—there, don’t listen to
me. I’m mad to-day.” She worked herself into tears and wept
convulsively, while he stared helplessly out at the world. His mind
moved. He could not stand her continued suffering, and the
confession and assurance of danger inspired him to thoughts of
action. Something must be done. She was in evident peril now. Any
day might bring the awful news of a disaster beyond repair. Such
things were in every newspaper. Not for an instant did he doubt the
critical nature of the situation. He hated to think Medora must
presently return home to sleep under the same roof as her husband.
To his order of mind the situation appealed with the uttermost
gravity, for not an inkling of the true Medora tinctured his impression
and he was as ignorant of the true Ned. He trusted the woman
absolutely and he loved her. He steadfastly believed now that the
most precious life in the world to him was in torment and in danger.
She had, under dreadful stress of emotion as it appeared, more than
once expressed her regret at the fatal step in the past. She had
mourned frankly and explicitly at taking Dingle, when she might
have married Kellock himself.
Here then was the tremendous problem for him; and so pressing
and immediate did it appear, that the young man was driven out of
his usual level attitude of mind and customary deliberation before
the demands of life. For the moment his future ambitions and
purposes were lost: he was only urged by the instant necessity to
decide what might best be done for Medora’s sake. Immense
prospects opened before him—knightly deeds, and unconventional
achievements calling for great efforts and an indifference to all
commonplace, social standards.
He was prepared at a future time to make war upon society for the
sake of his class, if the occasion demanded it. He fully intended
presently to stand forth with the protagonists of labour and fight for
socialism. He anticipated that battle and was educating and priming
himself for it. As yet the great revolt belonged to the future and
there his ultimate ambition lay; but now an immediate personal
appeal confronted him—a matter in which he himself and his own
happiness were deeply involved. And more than himself, for he felt
that Medora’s future now hung in the balance. Her destiny waited on
him.
But he did not tell Medora the result of his reflections. For the
moment he bade her be of good cheer and trust him.
While she sobbed, he considered and then, feeling it was time to
speak, comforted her.
“I’m glad you’ve told me all this,” he said. “It shows you know where
you can put your faith. And since you come to me with it, Medora,
I’ll make it my business. I’m only a human man and I loved you with
all my heart, and I do love you with all my heart still, and now the
case is altered. I should never have thought of you again—not in
that way—if your married life had turned out all right; but as it’s
turned out all wrong, then it’s up to me to come into your life again.
May I do so?”
“You’re the only thing in my life,” she said, drying her eyes.
“Everything else makes me want to end it—yes, I’ve thought often of
that, Jordan. But I’ll thankfully put myself in your hands and be
patient a bit longer if you tell me to.”
“It ain’t a case for waiting,” he said. “It’s a case for doing. I don’t
know what fear is myself, and more did you till he made you. It
looks very much to me as if you’d have to come to me, Medora.”
“Oh, my God—could you?”
“Yes, I could, and I will.”
“Think of yourself—it’s like your bravery to put me first and I’d be
your slave and live for you and thank Heaven for its blessings; but I
don’t want to ruin your life, you good, brave man.”
“Nobody can ruin your life but yourself,” he answered, “and if I save
your life, it won’t be to ruin my own. Say you’d like it to be so and
leave the rest to me. I mean it, Medora.”
A dream that had often filled the girl’s waking thoughts suddenly
promised to come true and for a moment she was frightened. But
only for a moment. She hardly hesitated. Here was romance, fame,
the centre of the stage—everything. She knew very well that she
could trust him, and if ever she loved and adored the impassive
vatman it was at this moment.
She took his hand and pressed her lips to it.
“Like it!” she cried. “It would be heaven on earth—heaven on earth.
And God’s my judge you shan’t repent it. I’ll live for you and die for
you.”
“So be it, Medora. It’s done.”
He put his arms round her and kissed her. Then both felt a secret
desire to be alone and consider the magnitude of the decision. He
voiced this wish.
“We’ll part now,” he said. “You go down to your mother and I’ll go
home. Be quite easy in your mind and cheerful and content. Leave
the rest to me. I’ll write to you to-night after I’ve gone all through it.
It ain’t so difficult as it sounds if we back each other up properly. I’ll
see you get the letter to-morrow out of sight of everybody at the
works. Be round by the vat house half after eleven. You’ve got a
man to deal with—remember that.”
“God bless you,” she answered very earnestly. “I’m yours now, and
never, never shall you repent of it, Jordan. You can trust me same as
I trust you in everything.”
They descended the winding stair of the ruin and then parted.
Medora went down through the orchard to her mother’s home at
Priory Farm, while Kellock, climbing through the hedge, presently set
his face to Dene and strolled down the Corkscrew Lane with his
mind full of the future. He found that thought persisted in drifting
away from Medora to her husband. He had just told her that she had
a man to deal with; and now it was impressed on Kellock that he,
also, had to deal with a man.
Meantime Ned’s wife reached the farm, and before she did so, she
bathed her eyes at a little stream under the orchard hedge.
She appeared in an unusually contented frame of mind and Lydia
was glad to see her so. Another guest had arrived, for Philander
Knox, at Mrs. Trivett’s invitation, visited Priory Farm. A friendship had
sprung up between him and the widow, for modest though Lydia
might be, she could not fail to perceive her company was agreeable
to Mr. Knox. He would listen to her opinions in a flattering manner
and often expressed surprise to mark how her sense chimed with his
experience. His own philosophy and general outlook on life were
approved by Mrs. Trivett and on this occasion she had invited him to
drink tea at Priory Farm and meet her brother and his family.
He had come and, as all who first penetrated into the life of the
farm, found himself bewildered by its complications. The children,
the mother, and the helpless father appeared to revolve as a system
of greater and lesser planets around the steadfast sun of Lydia. She
moved in the chaos as though it were her proper environment—“like
a ship in a storm,” as Mr. Knox afterwards told her.
Philander had designed to enliven the tea with humorous chatter. He
wished to impress Mr. Dolbear and his wife favourably, for he was a
sociable person and anxious to increase the number of friends in his
new home; but he found a meal at Priory Farm no occasion for much
intercourse or advancement of amenities. It proved a strenuous and
rather exasperating affair. The children dominated the tea and the
tea table. They chattered until they had eaten all they could and
departed; then, when the visitor hoped that his opportunity had
come, he found, instead, that their mother took up the conversation
and discussed the vanished youngsters one by one. She lingered
over each as a gardener over his treasures, or a connoisseur over his
collection. They were an incomparable group of children, it
appeared; and what puzzled Philander was to find that Lydia enjoyed
the subject as much as Mary herself. She also knew the children by
heart and was evidently devoted to each and all of them.
Tom Dolbear said very little, but enjoyed listening. His brood rejoiced
him and he lived now in hope of another boy.
It was Medora who strove to change the subject and allow Bobby
and Milly and Clara and Jenny and the rest to drop out of the
conversation.
“Mr. Knox will be sick to death of your babies, Aunt Polly,” she said.
“Far from it,” he declared. “A finer, hopefuller family I never wish to
see.”
Mr. Dolbear then invited Philander to come into the garden and
smoke, but finding the ladies were not prepared to accompany
them, he declined.
“If it’s all the same to you, I’ll rest here until I must get going,” he
answered. “I’m not used to your hills yet and they weary my legs a
lot. Never a great walker—after the way of town birds that have
lived all their lives by a tram line.”
So he sat and smoked, while Lydia cleared the tea things and
Medora helped her.
With Mrs. Trivett there were few opportunities for speech. She came
and went and worked. Then the dusk fell and the younger Dolbears
were brought in to go to bed. Medora nursed the baby for a time
and her mother noticed that she was more than usually cheerful.
Knox then declared that he must be going home and offered to
escort Medora. She agreed and having thanked Tom for his
hospitality and hoped that he might be privileged to accept it again
at some future time, he took his leave. On the way home he spoke
to his companion.
“Your mother’s a wonderful woman, Mrs. Dingle,” he said. “I see
these things from the outside and I’m properly astonished at her
cleverness.”
“So she is,” admitted Medora. “But I wish she wouldn’t work so hard
all the same. She does her day at the Mill and then comes back
home and instead of getting her proper rest—well, you see what it
is.”
“She’s like the mainspring of a watch,” declared Philander. “’Tis a
most delicate contrivance, yet all depends upon it; and if I may say
so, as an outsider, you can see with half an eye that her relations
depend upon her for everything.”
“They do—they do. If anything happened to mother, I don’t know
what would become of Aunt and Uncle—let alone all the children.”
“They don’t know their luck,” he said, and Medora agreed with him.
“I’m glad you see it. I’ve often thought that—so have other people.
My mother at Priory Farm is like a cheese-cake in a pigstye.”
“Strong, but not too strong. She must have great affection for them
to stand it.”
“Once a man offered for mother,” said Medora; “and, at the first
whisper of it, Uncle Tom and Aunt Polly pretty well went on their
knees to her not to leave them.”
“I can well believe it. It didn’t come to anything, however?”
“No, no—mother’s not for another husband.”
“If anything might make her think upon such a change, it would be
that household surely.”
“No,” answered Medora. “It’s just that helpless household that would
make her sacrifice herself. Duty’s her God. She’s mother to all those
children—more their mother than Aunt Polly in a way—for my aunt is
so busy bringing them into the world, that she’s got to leave all the
rest of the work to other people.”
Mr. Knox shook his head.
“It’s contrary to nature that such a fine woman as Mrs. Trivett should
hide her light under that bushel,” he asserted. “It’s a very selfish
thing to let her slave and wear her fingers to the bone like that; but
it often happens so. A husband and wife with a long family always
seem to fasten on some good-natured, kindly creature and drag her
in their house to be a slave to their children. There’s no selfishness
like the selfishness of a pair with a long quiver. They’ll fairly batter
the life out of anybody who’s fool enough to lend a hand; and the
more such a person does for the other woman’s children, the more
she may do. But I should hope your mother was too proud to let
herself be used as a nursemaid to her own nieces.”
“She’s never proud where children are concerned,” answered
Medora. “She’ll stop there till she’s worn out.”
“A very gloomy picture and I hope you’re wrong, Mrs. Dingle,” he
answered.
 CHAPTER X
THE LETTER

In the vat house there took place the transformation from liquid to
solid, from pulp to paper, from a gruel-like, tenuous compound to a
substance strong enough to stand strain of many pounds and last for
centuries.
Here was the largest building in the Mill—a very lofty, brightly
lighted, airy hall, from whose open roof descended electric lights
hanging above each vat. A steady whirr and throb of noisy engines
made a din here, but the vatmen and their couchers were used to it
and could hear themselves speak through the familiar riot.
To the right, elevated under the roof, stood the range of chests—
huge, round vessels, like little gasometers, into which the pulp
descended from Ned Dingle when he had perfected it. There were
eight of these fat monsters ranged in a row, and from them flowed
the material to the vats as it was needed. The vats stood on the
floor of the chamber—large, wide-mouthed troughs heated by steam
from within. For the pulp is warm for the vatman, and some of the
finest and most enduring papers demand such a high temperature
that an operative’s hands are blistered and boiled at his work. Beside
each vat is a hand-box of cold water, to dip and refresh the vatman’s
fingers when the need arises.
Within the vat revolves the “hog,” a toothed roller, which keeps the
heavy pulp mixed and moving, and prevents any settlement of the
fibre.
On stages before the breasts of the vats stood the paper makers,
and the wooden bands against which they leaned were polished with
the friction of their aprons. Their tools were two—the mould—a flat,
rectangular tray, or sieve, of copper wire as fine as gauze, with the
water-mark let in upon it to tell the story of the future paper, and
the deckle—a light wood and metal frame of four sides which fitted
exactly over the mould and lifted an edge all round it to hold the
pulp. The moulds varied from the size of two open sheets of
notepaper, to great squares of “double elephant,” the noblest stuff
the Mill produced. Moulds for these immense pieces once immersed
in the pulp, called for great physical power to draw them cleanly and
steadily back from the clinging fluid with their weight of material
spread upon them.
Kellock was making “double elephant” in a mighty mould. With his
thumbs firmly set on the deckle edge, he lowered the tray into the
snow-white pulp, sloping it towards him as he did so. He put it in,
sank it flat under the pulp and drew it out again with one beautiful,
rhythmic movement.
The pulp sucked hard at the great mould, to drag it to the depths,
but the man’s strength brought it steadily forth; and then he made
his “stroke”—a complicated gesture, which levelled and settled the
pulp on the mould and let the liquid escape through the gauze.
Kellock gave a little jog to the right and to the left and ended with
an indescribable, subtle, quivering movement which completed the
task. It was the work of two seconds, and in his case a beautiful
accomplishment full of grace and charm. He stood easily and firmly
while every muscle of breast and arm, back and loins played its
appointed part in the “stroke.”
Mr. Trood often stood and watched Jordan for the pleasure of the
sight. It was the most perfect style he had ever seen. He was a
theorist and calculated that Kellock produced the very greatest
amount of physical power for the least possible expenditure of
muscular loss; while others, who made as good paper as he,
squandered thousands of pounds of dynamical energy by a stroke
full of superfluous gesture. But the stroke is never the same in any
two vatmen. It develops, with each artificer’s knowledge of the craft,
to produce that highly co-ordinated effort embraced in the operation
of making a sheet of paper.
Mr. Knox operated at the next vat and offered an object lesson. He
did the same things that Kellock did; dipped his mould, drew it to
him, brought it squarely out, jogged to right and left and gave that
subtle, complex touch of completion; yet in his achievement a wholly
different display met the observer. It seemed that he performed a
piece of elaborate ritual before the altar of the vat.
He bowed his head to right and left; he moved his tongue and his
knees; he jerked his elbows and bent his back over the trough as a
priest consecrating the elements of some sacramental mass. Then
he bowed and nodded once more and the created sheet emerged
from his mould. The effect was grotesque, and seen at a little
distance a stranger had supposed that Mr. Knox was simply playing
the fool for the amusement of his coucher and layer; but in reality
he was working hard and making as fine and perfect paper as
Kellock himself. His muscles were tuned to his task; he had lifted his
sheer weight of forty tons or more by the end of the day and was
none the worse for it. Nor could he have omitted one gesture from
his elaborate style without upsetting everything and losing his
stroke.
So the transformation became accomplished and the millions of linen
and cotton fibres scooped on to the mould ran into a thin mat or
wad, which was a piece of paper. Why all these fragile and
microscopic atoms should become so inter-twisted and mingled that
they produce an integral fabric, it is difficult to understand; but this
was the result of the former processes; and those to come would
change the slab of wet, newly created stuff—now no more than a
piece of soaked blotting-paper—to the perfected sheet.
His stroke accomplished and the sediment levelled on the mould,
Kellock brought his mould to the “stay”—a brass-bound ledge on his
left hand. He lifted the deckle from it as he did so and the full mould
was drawn up the stay to the “asp,” where his coucher stood. Then
Kellock clasped the deckle on to his second mould, now returned
from the coucher, and dipped again, while his assistant, taking the
full mould from the asp, turned it over on to the accumulating pile of
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

ebookultra.com