Get the full ebook with Bonus Features for a Better Reading Experience on ebookgate.

com

Dynamical Processes on Complex Networks 1st

Edition Alain Barrat

https://ebookgate.com/product/dynamical-processes-on-
complex-networks-1st-edition-alain-barrat/

OR CLICK HERE

DOWLOAD NOW

Download more ebook instantly today at https://ebookgate.com

Instant digital products (PDF, ePub, MOBI) available
Download now and explore formats that suit you...

Channel Coding in Communication Networks From Theory to

Turbo Codes 1st edition Edition Alain Glavieux

https://ebookgate.com/product/channel-coding-in-communication-
networks-from-theory-to-turbo-codes-1st-edition-edition-alain-
glavieux/
ebookgate.com

Numerical Continuation Methods for Dynamical Systems Path

following and boundary value problems Understanding
Complex Systems 1st Edition Bernd Krauskopf
https://ebookgate.com/product/numerical-continuation-methods-for-
dynamical-systems-path-following-and-boundary-value-problems-
understanding-complex-systems-1st-edition-bernd-krauskopf/
ebookgate.com

Iterated Maps on the Interval as Dynamical Systems 1st

Edition Pierre Collet

https://ebookgate.com/product/iterated-maps-on-the-interval-as-
dynamical-systems-1st-edition-pierre-collet/

ebookgate.com

Lecture Notes on Complex Analysis Ivan Francis Wilde

https://ebookgate.com/product/lecture-notes-on-complex-analysis-ivan-
francis-wilde/

ebookgate.com
Probability Statistics and Random Processes with Queueing
Theory and Queueing Networks 4th Edition Veerara

https://ebookgate.com/product/probability-statistics-and-random-
processes-with-queueing-theory-and-queueing-networks-4th-edition-
veerara/
ebookgate.com

Recovery of Values from Low Grade and Complex Minerals

Development of Sustainable Processes 1st Edition Elvis
Fosso-Kankeu
https://ebookgate.com/product/recovery-of-values-from-low-grade-and-
complex-minerals-development-of-sustainable-processes-1st-edition-
elvis-fosso-kankeu/
ebookgate.com

Probability on Trees and Networks 1st Edition Russell

Lyons

https://ebookgate.com/product/probability-on-trees-and-networks-1st-
edition-russell-lyons/

ebookgate.com

Collective Dynamics in Complex Networks of Noisy Phase

Oscillators Towards Models of Neuronal Network Dynamics
1st Edition Bernard Sonnenschein
https://ebookgate.com/product/collective-dynamics-in-complex-networks-
of-noisy-phase-oscillators-towards-models-of-neuronal-network-
dynamics-1st-edition-bernard-sonnenschein/
ebookgate.com

Superhydrophobic Surfaces 1st Edition Alain Carré

https://ebookgate.com/product/superhydrophobic-surfaces-1st-edition-
alain-carre/

ebookgate.com
This page intentionally left blank
 DYNAMICAL PROCESSES ON
COMPLEX NETWORKS

The availability of large data sets has allowed researchers to uncover complex
properties such as large-scale fluctuations and heterogeneities in many networks,
leading to the breakdown of standard theoretical frameworks and models. Until
recently these systems were considered as haphazard sets of points and connec-
tions. Recent advances have generated a vigorous research effort in understanding
the effect of complex connectivity patterns on dynamical phenomena. This book
presents a comprehensive account of these effects.
A vast number of systems, from the brain to ecosystems, power grids and the
Internet, can be represented as large complex networks. This book will interest
graduate students and researchers in many disciplines, from physics and statis-
tical mechanics, to mathematical biology and information science. Its modular
approach allows readers to readily access the sections of most interest to them,
and complicated maths is avoided so the text can be easily followed by non-experts
in the subject.

ALAIN BARRAT is Senior Researcher at the Laboratoire de Physique Théorique

(CNRS and Université de Paris-Sud, France), and the Complex Networks Lagrange
Laboratory at the Institute for Scientific Interchange in Turin, Italy. His research
interests are in the field of out-of-equilibrium statistical mechanics.
MARC BARTH ÉLEMY is Senior Researcher at the Département de Physique
Théorique et Appliquée at the Commissariat à l’Énergie Atomique (CEA),
France. His research interests are in the application of statistical physics to com-
plex systems.
ALESSANDRO VESPIGNANI is Professor of Informatics and Adjunct Professor of
Physics and Statistics at Indiana University, USA, and Director of the Complex
Networks Lagrange Laboratory at the Institute for Scientific Interchange in Turin,
Italy. His research interests are in the theoretical and computational modeling of
complex systems and networks.
DYNAMICAL PROCESSES ON
COMPLEX NETWORKS

ALAIN BARRAT
Laboratoire de Physique Théorique, Université de Paris-Sud, France
Centre National de la Recherche Scientifique, France
Complex Networks Lagrange Laboratory, ISI Foundation, Italy

M A R C B A R T H É L E M Y
Commissariat à l’Energie Atomique – Département de Physique
Théorique et Appliquée, France

ALESSANDRO VESPIGNANI
School of Informatics, Department of Physics and Biocomplexity Institute,
Indiana University, USA
Complex Networks Lagrange Laboratory, ISI Foundation, Italy
Centre National de la Recherche Scientifique, France
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo

Cambridge University Press

The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York

www.cambridge.org
Information on this title: www.cambridge.org/9780521879507
© A. Barrat, M. Barthelemy and A. Vespignani 2008

This publication is in copyright. Subject to statutory exception and to the

provision of relevant collective licensing agreements, no reproduction of any part
may take place without the written permission of Cambridge University Press.
First published in print format 2008

ISBN-13 978-0-511-45558-2 eBook (EBL)

ISBN-13 978-0-521-87950-7 hardback

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.
 To Luisa;
To Loulou, Simon, Théo;
To Martina
 Contents

Preface page xi
Acknowledgements xv
List of abbreviations xvii
1 Preliminaries: networks and graphs 1
1.1 What is a network? 1
1.2 Basic concepts in graph theory 2
1.3 Statistical characterization of networks 11
1.4 Weighted networks 19
2 Networks and complexity 24
2.1 Real-world systems 24
2.2 Network classes 34
2.3 The complicated and the complex 47
3 Network models 50
3.1 Randomness and network models 50
3.2 Exponential random graphs 58
3.3 Evolving networks and the non-equilibrium approach 60
3.4 Modeling higher order statistics and other attributes 72
3.5 Modeling frameworks and model validation 74
4 Introduction to dynamical processes: theory and simulation 77
4.1 A microscopic approach to dynamical phenomena 77
4.2 Equilibrium and non-equilibrium systems 79
4.3 Approximate solutions of the Master Equation 82
4.4 Agent-based modeling and numerical simulations 85
5 Phase transitions on complex networks 92
5.1 Phase transitions and the Ising model 92
5.2 Equilibrium statistical physics of critical phenomena 96
5.3 The Ising model in complex networks 101

vii
viii Contents

5.4 Dynamics of ordering processes 108

5.5 Phenomenological theory of phase transitions 111
6 Resilience and robustness of networks 116
6.1 Damaging networks 116
6.2 Percolation phenomena as critical phase transitions 120
6.3 Percolation in complex networks 124
6.4 Damage and resilience in networks 126
6.5 Targeted attacks on large degree nodes 129
6.6 Damage in real-world networks 135
7 Synchronization phenomena in networks 136
7.1 General framework 136
7.2 Linearly coupled identical oscillators 138
7.3 Non-linear coupling: firing and pulse 148
7.4 Non-identical oscillators: the Kuramoto model 151
7.5 Synchronization paths in complex networks 156
7.6 Synchronization phenomena as a topology probing tool 158
8 Walking and searching on networks 160
8.1 Diffusion processes and random walks 160
8.2 Diffusion in directed networks and ranking algorithms 166
8.3 Searching strategies in complex networks 170
9 Epidemic spreading in population networks 180
9.1 Epidemic models 180
9.2 Epidemics in heterogeneous networks 189
9.3 The large time limit of epidemic outbreaks 197
9.4 Immunization of heterogeneous networks 207
9.5 Complex networks and epidemic forecast 212
10 Social networks and collective behavior 216
10.1 Social influence 216
10.2 Rumor and information spreading 218
10.3 Opinion formation and the Voter model 225
10.4 The Axelrod model 232
10.5 Prisoner’s dilemma 235
10.6 Coevolution of opinions and network 238
11 Traffic on complex networks 242
11.1 Traffic and congestion 242
11.2 Traffic and congestion in distributed routing 246
11.3 Avalanches 256
11.4 Stylized models and real-world infrastructures 264
12 Networks in biology: from the cell to ecosystems 267
12.1 Cell biology and networks 268
 Contents ix

12.2 Flux-balance approaches and the metabolic activity 271

12.3 Boolean networks and gene regulation 274
12.4 The brain as a network 279
12.5 Ecosystems and food webs 282
12.6 Future directions 293
13 Postface: critically examining complex networks science 294
Appendix 1 Random graphs 298
Appendix 2 Generating functions formalism 303
Appendix 3 Percolation in directed networks 306
Appendix 4 Laplacian matrix of a graph 310
Appendix 5 Return probability and spectral density 311
References 313
Index 344
 Preface

In the past few years, the study of large networked systems has received a boost
from the ever-increasing availability of large data sets and computer power for their
storage and manipulation. In particular, mapping projects of the World Wide Web
and the physical Internet offered the first chance to study the topology of large
complex networks. Gradually, other maps followed describing many networks of
practical interest in social science, critical infrastructures, and biology. Indeed,
large complex networks arise in a vast number of natural and artificial systems.
The brain consists of many interconnected neurons; ecosystems consist of species
whose interdependency can be mapped into intricate food webs. Social systems
may be represented by graphs describing various interactions among individuals.
Large networked infrastructures such as power grids and transportation networks
are critical to our modern society. Finally, the living cell is not an exception, its
organization and function being the outcome of a complex web of interactions
among genes, proteins, and other molecules. In the search for the underlying laws
governing the dynamics and evolution of these complex systems, researchers have
started the systematic analysis and characterization of their network representa-
tions. A central result of these activities is that large-scale networks are generally
characterized by complex topologies and very heterogeneous structures. These
features usually find their signature in connectivity patterns statistically charac-
terized by heavy tails and large fluctuations, scale-free properties, and non-trivial
correlations such as high clustering and hierarchical ordering.
The large size and dynamic nature of complex networks has attracted the atten-
tion of the statistical physics community. In this context the statistical physics
approach has been exploited as a very convenient strategy because of its deep con-
nection with statistical graph theory and the possibility of characterizing emergent
macroscopic phenomena in terms of the dynamical evolution of the basic elements
of the system. In addition, the large scale analysis of networks appearing in very

xi
xii Preface

different fields has provided evidence for the sharing of several asymptotic proper-
ties and raised the issue of the supposed emergence in complex networks of general
and common organizing principles that go beyond the particulars of the individual
system, and are thus amenable to general modeling principles.
The advances in understanding large complex networks have generated
increased attention towards the potential implications of their structure for the
most important questions concerning the various physical and dynamical processes
occurring on top of them. Questions of how pathogens spread in population net-
works, how blackouts can spread on a nationwide scale, or how efficiently we can
search and retrieve data on large information structures are all related to spreading
and diffusion phenomena. The resilience and robustness of large infrastructures
can be studied by percolation models in which we progressively damage the net-
work. Social behavior may be often modeled through simple dynamical processes
and agent-based models. Since the early studies of percolation and spreading pro-
cesses in complex networks, a theoretical picture has emerged in which many of
the standard results were radically altered by the topological fluctuations and the
complex features observed in most real-world networks. Indeed, complex prop-
erties often imply a virtual infinite heterogeneity of the system and large scale
fluctuations extending over several orders of magnitude, generally corresponding
to the breakdown of standard theoretical frameworks and models. Therefore, while
the definitions of the basic models are expected to remain unchanged, the scien-
tific community has become aware of the need to investigate systematically the
impact of the various network characteristics on the basic features of equilibrium
and non-equilibrium dynamical processes.
The study of models unfolding on complex networks has generated results of
conceptual and practical relevance. The resilience of networks, their vulnerability
to attacks, and their synchronization properties are all strongly affected by topolog-
ical heterogeneities. Consensus formation, disease spreading, and our accessibility
to information can benefit from or be impaired by the connectivity pattern of the
population or infrastructure we are looking at. Noticeably, all these results have
been obtained in different contexts and have relevance across several disciplines
including physics, biology, and computer and information science. The purpose of
this book is to provide a unified presentation of these results and their impact on
our understanding of a wide range of networked systems.
While several techniques and methods have been used in the field, the mas-
sive size and heterogeneity of the systems have favored the use of techniques akin
to the analysis of non-linear, equilibrium, and non-equilibrium physical systems.
As previously mentioned, the statistical physics approach has indeed revealed its
convenience and has proven very useful in this context, particularly regarding its
ability to shed light on the link between the microscopic dynamical evolution of
 Preface xiii

the basic elements of the system and the emergence of macroscopic phenomena.
For this reason we will follow this approach and structure the book according to the
general framework provided by statistical physics. In doing this, however, we will
make a special effort in defining for each phenomenon or system under study the
appropriate language used in the field and offer to the reader a mapping between
languages and techniques used in the different disciplines.
For the sake of clarity and readability we have used a modular approach to the
material that allows the reader interested in just one phenomenon to refer directly
to the corresponding specific chapter. The basic definitions and methods used
throughout the book are introduced in the first four chapters that set the stage for
the remaining chapters, each one clearly self-contained and addressing a specific
class of processes. The material is ordered in a path that goes from equilibrium
to non-equilibrium phenomena, in accordance with the general framework of sta-
tistical physics and in terms of the homogeneity of the techniques used in each
context.
More precisely, we start in Chapter 1 by defining the basics of graph theory
and some tools for the statistical characterization and classification of large net-
works. These tools are used in Chapter 2, where we discuss the complex features
which characterize a large number of real-world networks. Chapter 3 is devoted to
a brief presentation of the various network modeling techniques and the general
approaches used for their analysis. Approximations and computational techniques
are discussed as well. Chapter 4 provides an introduction to the basic theoretical
concepts and tools needed for the analysis of dynamical processes taking place on
networks. All these introductory chapters will allow researchers not familiar with
mathematical and computational modeling to get acquainted with the approaches
and techniques used in the book.
In Chapter 5, we report the analysis of equilibrium processes and we review the
behaviors of basic equilibrium physical models such as the Ising model in com-
plex networks. Chapter 6 addresses the analysis of damage and attack processes
in complex networks by mapping those processes into percolation phase transi-
tions. Chapter 7 is devoted to synchronization in coupled systems with complex
connectivity patterns. This is a transition chapter in that some of the phenomena
analyzed and their respective models can be considered in the equilibrium pro-
cesses domains while others fall into the non-equilibrium class. The following four
chapters of the book are devoted to non-equilibrium processes. One chapter con-
siders search, navigation, and exploration processes of complex networks; the three
other chapters concern epidemic spreading, the emergence of social behavior, and
traffic avalanche and congestion, respectively. These chapters, therefore, are con-
sidering far from equilibrium systems with absorbing states and driven-dissipative
xiv Preface

dynamics. Finally, Chapter 12 is devoted to a discussion of the recent applica-

tion of network science to biological systems. A final set of convenient appendices
are used to detail very technical calculations or discussions in order to make the
reading of the main chapters more accessible to non-expert readers. The postface
occupies a special place at the end of the book as it expresses our view on the value
of complex network science.
We hope that our work will result in a first coherent and comprehensive expo-
sition of the vast research activity concerning dynamical processes in complex
networks. The large number of references and research focus areas that find room
in this book make us believe that the present volume will be a convenient entry ref-
erence to all researchers and students who consider working in this exciting area of
interdisciplinary research. Although this book reviews or mentions more than 600
scientific works, it is impossible to discuss in detail all relevant contributions to the
field. We have therefore used our author privilege and made choices based on our
perception of what is more relevant to the focus and the structure of the present
book. This does not imply that work not reported or cited here is less valuable, and
we apologize in advance to all the colleagues who feel that their contributions have
been overlooked.
 Acknowledgements

There are many people we must thank for their help in the preparation and com-
pletion of this book. First of all, we wish to thank all the colleagues who helped us
shape our view and knowledge of complex networks through invaluable scientific
collaborations: J. I. Alvarez-Hamelin, F. P. Alvarez, L. A. N. Amaral, D. Balcan, A.
Baronchelli, M. Boguñá, K. Börner, G. Caldarelli, C. Caretta Cartozo, C. Castel-
lano, F. Cecconi, A. Chessa, E. Chow, V. Colizza, P. Crépey, A. De Montis, L.
Dall’Asta, T. Eliassi-Rad, A. Flammini, S. Fortunato, F. Gargiulo, D. Garlaschelli,
A. Gautreau, B. Gondran, E. Guichard, S. Havlin, H. Hu, C. Herrmann, W. Ke, E.
Kolaczyk, B. Kozma, M. Leone, V. Loreto, F. Menczer, Y. Moreno, M. Nekovee,
A. Maritan, M. Marsili, A. Maguitman, M. Meiss, S. Mossa, S. Myers, C. Nardini,
M. Nekovee, D. Parisi, R. Pastor-Satorras, R. Percacci, P. Provero, F. Radicchi, J.
Ramasco, C. Ricotta, A. Scala, M. Angeles Serrano, H. E. Stanley, A.-J. Valleron,
A. Vàzquez, M. Vergassola, F. Viger, D. Vilone, M. Weigt, R. Zecchina, and C.-H.
Zhang.
We also acknowledge with pleasure the many discussions with, encouragements,
and useful criticisms from: R. Albert, E. Almaas, L. Adamic, A. Arenas, A. L.
Barabási, J. Banavar, C. Cattuto, kc claffy, P. De Los Rios, S. Dorogovtsev, A.
Erzan, R. Goldstone, S. Havlin, C. Hidalgo, D. Krioukov, H. Orland, N. Mar-
tinez, R. May, J. F. Mendes, A. Motter, M. Mungan, M. Newman, A. Pagnani,
T. Petermann, D. Rittel, L. Rocha, G. Schehr, S. Schnell, O. Sporns, E. Trizac, Z.
Toroczkai, S. Wasserman, F. van Wijland, W. Willinger, L. Yaeger, and S. Zapperi.
Special thanks go to E. Almaas, C. Castellano, V. Colizza, J. Dunne, S. Fortu-
nato, H. Jeong, Y. Moreno, N. Martinez, M. E. J. Newman, M. Angeles Serrano,
and O. Sporns for their kind help in retrieving data and adapting figures from
their work. A. Baronchelli, C. Castellano, L. Dall’Asta, A. Flammini, S. Fortu-
nato, and R. Pastor-Satorras reviewed earlier versions of this book and made many
suggestions for improvement. B. Hook deserves a special acknowledgement for

xv
xvi Acknowledgements

proofreading and editing the first version of the manuscript. Simon Capelin and
his outstanding editorial staff at Cambridge University Press were a source of
continuous encouragement and help and greatly contributed to make the writing
of this book a truly exciting and enjoyable experience. We also thank all those
institutions that helped in making this book a reality. The Centre National de la
Recherche Scientifique, the Laboratoire de Physique Théorique at the University
of Paris-Sud, the Commissariat à l’Energie Atomique, the School of Informatics
at Indiana University, and the Institute for Scientific Interchange in Turin have
provided us with great environments and the necessary resources to complete this
work. We also acknowledge the generous funding support at various stages of our
research by the following awards: NSF-IIS0513650, EC-001907 Delis, and CRT-
foundation Lagrange laboratory. Finally, A.B. thanks the SNCF for providing fast
trains between Paris, Dijon, and Turin which are not always late, M.B. expresses his
gratitude to the Cannibale Café which is probably the best coffee-office in Paris,
and A.V. thanks Continental Airlines for having power plug-ins in all classes of
reservations.
 Abbreviations

AS Autonomous System
BA Barabási–Albert
BGP Border Gateway Protocol
CAIDA Cooperative Association for Internet Data Analysis
CPU central processing unit
ER Erdős–Rényi
IATA International Air Transport Association
DIMES Distributed Internet Measurements and Simulations
GIN giant in-component
GOUT giant out-component
GSCC giant strongly connected component
GWCC giant weakly connected component
HOT Heuristically Optimized Trade-off
ICT Information and communication technology
LAN Local Area Network
ME master equation
MF mean-field
P2P peer-to-peer
SCN scientific collaboration network
SI Susceptible-Infected model
SIR Susceptible-Infected-Removed model
SIS Susceptible-Infected-Susceptible model
TTL time-to-live
WAN worldwide airport network
WS Watts–Strogatz
WWW World Wide Web

xvii
 1
Preliminaries: networks and graphs

In this chapter we introduce the reader to the basic definitions of network and graph
theory. We define metrics such as the shortest path length, the clustering coefficient,
and the degree distribution, which provide a basic characterization of network sys-
tems. The large size of many networks makes statistical analysis the proper tool for
a useful mathematical characterization of these systems. We therefore describe the
many statistical quantities characterizing the structural and hierarchical ordering
of networks including multipoint degree correlation functions, clustering spec-
trum, and several other local and non-local quantities, hierarchical measures and
weighted properties.
This chapter will give the reader a crash course on the basic notions of net-
work analysis which are prerequisites for understanding later chapters of the book.
Needless to say the expert reader can freely skip this chapter and use it later as a
reference if needed.

1.1 What is a network?

In very general terms a network is any system that admits an abstract mathemat-
ical representation as a graph whose nodes (vertices) identify the elements of the
system and in which the set of connecting links (edges) represent the presence of a
relation or interaction among those elements. Clearly such a high level of abstrac-
tion generally applies to a wide array of systems. In this sense, networks provide a
theoretical framework that allows a convenient conceptual representation of inter-
relations in complex systems where the system level characterization implies the
mapping of interactions among a large number of individuals.
The study of networks has a long tradition in graph theory, discrete mathema-
tics, sociology, and communication research and has recently infiltrated physics
and biology. While each field concerned with networks has introduced, in many
cases, its own nomenclature, the rigorous language for the description of networks

1
2 Preliminaries: networks and graphs

is found in mathematical graph theory. On the other hand, the study of very large
networks has spurred the definitions of new metrics and statistical observables
specifically aimed at the study of large-scale systems. In the following we pro-
vide an introduction to the basic notions and notations used in network theory and
set the cross-disciplinary language that will be used throughout this book.

1.2 Basic concepts in graph theory

Graph theory – a vast field of mathematics – can be traced back to the pioneering
work of Leonhard Euler in solving the Königsberg bridges problem (Euler, 1736).
Our intention is to select those notions and notations which will be used throughout
the rest of this book. The interested reader can find excellent textbooks on graph
theory by Bergé (1976), Chartrand and Lesniak (1986), Bollobás (1985, 1998) and
Clark and Holton (1991).

1.2.1 Graphs and subgraphs

An undirected graph G is defined by a pair of sets G = (V, E), where V is a
non-empty countable set of elements, called vertices or nodes, and E is a set of
unordered pairs of different vertices, called edges or links. Throughout the book
we will refer to a vertex by its order i in the set V. The edge (i, j) joins the vertices
i and j, which are said to be adjacent or connected. It is also common to call
connected vertices neighbors or nearest neighbors. The total number of vertices in
the graph (the cardinality of the set V) is denoted as N and defines the order of
the graph. It is worth remarking that in many biological and physical contexts, N
defines the physical size of the network since it identifies the number of distinct
elements composing the system. However, in graph theory, the size of the graph
is identified by the total number of edges E. Unless specified in the following, we
will refer to N as the size of the network.  
For a graph of size N , the maximum number of edges is N2 . A graph with
E = N2 , i.e. in which all possible pairs of vertices are joined by edges, is called
a complete N -graph. Undirected graphs are depicted graphically as a set of dots,
representing the vertices, joined by lines between pairs of vertices, representing the
corresponding edges.
An interesting class of undirected graph is formed by hierarchical graphs where
each edge (known as a child) has exactly one parent (node from which it origi-
nates). Such a structure defines a tree and if there is a parent node, or root, from
which the whole structure arises, then it is known as a rooted tree. It is easy to
prove that the number of nodes in a tree equals the number of edges plus one, i.e.,
 1.2 Basic concepts in graph theory 3

N = E +1 and that the deletion of any edge will break a tree into two disconnected
trees.
A directed graph D, or digraph, is defined by a non-empty countable set of ver-
tices V and a set of ordered pairs of different vertices E that are called directed
edges. In a graphical representation, the directed nature of the edges is depicted
by means of an arrow, indicating the direction of each edge. The main difference
between directed and undirected graphs is represented in Figure 1.1. In an undi-
rected graph the presence of an edge between vertices i and j connects the vertices
in both directions. On the other hand, the presence of an edge from i and j in a
directed graph does not necessarily imply the presence of the reverse edge between
j and i. This fact has important consequences for the connectedness of a directed
graph, as will be discussed in more detail in Section 1.2.2.
From a mathematical point of view, it is convenient to define a graph by means
of the adjacency matrix X = {xi j }. This is a N × N matrix defined such that

1 if (i, j) ∈ E
xi j = . (1.1)
0 if (i, j) ∈
/E

For undirected graphs the adjacency matrix is symmetric, xi j = x ji , and therefore

contains redundant information. For directed graphs, the adjacency matrix is not

Undirected Original graph Subgraph of original graph

graphs 0 0
0 1 2 3 0 1 2
0 0 1 1 1 1 2 0 0 1 1 1 2
1 1 0 1 1 1 1 0 1
2 1 1 0 1 2 1 1 0
3 1 1 1 0 3
0 0
0 1 2 3 0 1 2
0 0 1 0 0 1 2 0 0 1 0 1 2
1 1 0 1 1 1 1 0 1
2 0 1 0 1 2 0 1 0
3 0 1 1 0 3

Directed 0 0
0 1 2 3 0 1 2
graphs 1 2 1 2
0 0 1 0 1 0 0 1 0
1 0 0 0 0 1 0 0 0
2 0 1 0 0 2 0 1 0
3 0 1 1 0 3

Fig. 1.1. Adjacency matrix and graphical representation of different networks.

In the graphical representation of an undirected graph, the dots represent the ver-
tices and pairs of adjacent vertices are connected by a line (edge). In directed
graphs, adjacent vertices are connected by arrows, indicating the direction of the
corresponding edge.
4 Preliminaries: networks and graphs

symmetric. In Figure 1.1 we show the graphical illustrations of different undirected

and directed graphs and their corresponding adjacency matrices.
An important feature of many graphs, which helps in dealing with their structure,
is their sparseness. The number of edges E for  a connected graph (i.e., with no
disconnected parts) ranges from N − 1 to 2 . There are different definitions of
N

sparseness, but we will adopt the convention that when the number of edges scales
as E ∼ N α with α < 2, the graph is said to be sparse. In the case where E ∼ N 2 ,
the corresponding graph is called dense. By defining the connectance or density of
a graph as the number of existing edges divided by the maximal possible number
of edges D = E/[N (N − 1)/2], a graph is then sparse if D  1. This feature
implies, in the case of large graphs, that the adjacency matrix is mostly defined by
zero elements and its complete representation, while costly, does not contain much
relevant information. With large graphs, it is customary to represent the graph in
the compact form defined by the adjacency lists (i, v ∈ V(i)), where the set of all
neighbors of a fixed vertex i is called the neighborhood (set) of i and is denoted by
V(i). The manipulation of these lists is obviously very convenient in computational
applications because they efficiently store large sparse networks.
In many cases, we are also interested in subsets of a graph. A graph G  = (V  , E  )
is said to be a subgraph of the graph G = (V, E) if all the vertices in V  belong to V
and all the edges in E  belong to E, i.e. E  ⊂ E and V  ⊂ V. A clique is a complete
n-subgraph of size n < N . In Figure 1.1 we provide the graphical and adjacency
matrix representations of subgraphs in the undirected and directed cases. The abun-
dance of given types of subgraphs and their properties are extremely relevant in the
characterization of real networks.1 Small, statistically significant, coherent sub-
graphs, called motifs, that contribute to the set-up of networks have been identified
as relevant building blocks of network architecture and evolution (see Milo et al.
[2002] and Chapter 12).
The characterization of local structures is also related to the identification of
communities. Loosely speaking, communities are identified by subgraphs where
nodes are highly interconnected among themselves and poorly connected with
nodes outside the subgraph. In this way, different communities can be traced back
with respect to varying levels of cohesiveness. In directed networks, edge direc-
tionality introduces the possibility of different types of local structures. A possible
mathematical way to account for these local cohesive groups consists in examin-
ing the number of bipartite cliques present in the graph. A bipartite clique K n,m
identifies a group of n nodes, all of which have a direct edge to the same m

1 Various approaches exist to determine the structural equivalence, the automorphic equivalence, or the regular
equivalence of subnetworks, and measures for structural similarity comprise correlation coefficients, Euclidean
distances, rates of exact matches, etc.
 1.2 Basic concepts in graph theory 5

nodes. The presence of subgraphs and communities raises the issue of modularity
in networks. Modularity in a network is determined by the existence of specific sub-
graphs, called modules (or communities). Clustering techniques can be employed
to determine major clusters. They comprise non-hierarchical methods (e.g., sin-
gle pass methods or reallocation methods), hierarchical methods (e.g., single-link,
complete-link, average-link, centroid-link, Ward), and linkage based methods (we
refer the interested reader to the books of Mirkin (1996) and Banks et al. (2004)
for detailed expositions of clustering methods). Non-hierarchical and hierarchical
clustering methods typically work on attribute value information. For example, the
similarity of social actors might be judged based on their hobbies, ages, etc. Non-
hierarchical clustering typically starts with information on the number of clusters
that a data set is expected to have and sorts the data items into clusters such that an
optimality criterion is satisfied. Hierarchical clustering algorithms create a hierar-
chy of clusters grouping similar data items. Connectivity-based approaches exploit
the topological information of a network to identify dense subgraphs. They com-
prise measures such as betweenness centrality of nodes and edges (Girvan and
Newman, 2002; Newman, 2006), superparamagnetic clustering (Blatt, Wiseman
and Domany, 1996; Domany, 1999), hubs and bridging edges (Jungnickel, 2004),
and others. Recently, a series of sophisticated overlapping and non-overlapping
clustering methods has been developed, aiming to uncover the modular structure
of real networks (Reichardt and Bornholdt, 2004; Palla et al., 2005).

1.2.2 Paths and connectivity

A central issue in the structure of graphs is the reachability of vertices, i.e. the
possibility of going from one vertex to another following the connections given by
the edges in the network. In a connected network every vertex is reachable from
any other vertex. The connected components of a graph thus define many properties
of its physical structure.
In order to analyze the connectivity properties let us define a path Pi0 ,in in a
graph G = (V, E) as an ordered collection of n + 1 vertices VP = {i 0 , i 1 , . . . , i n }
and n edges EP = {(i 0 , i 1 ), (i 1 , i 2 ), . . . , (i n−1 , i n )}, such that i α ∈ V and
(i α−1 , i α ) ∈ E, for all α. The path Pi0 ,in is said to connect the vertices i 0 and i n .
The length of the path Pi0 ,in is n. The number Ni j of paths of length n between two
nodes i and j is given by the i j element of the nth power of the adjacency matrix:
Ni j = (Xn )i j .
A cycle – sometimes called a loop – is a closed path (i 0 = i n ) in which all
vertices and all edges are distinct. A graph is called connected if there exists a path
connecting any two vertices in the graph. A component C of a graph is defined
6 Preliminaries: networks and graphs

as a connected subgraph. Two components C1 = (V1 , E1 ) and C2 = (V2 , E2 ) are

disconnected if it is impossible to construct a path Pi, j with i ∈ V1 and j ∈ V2 .
It is clear that for a given number of nodes the number of loops increases with
the number of edges. It can easily be shown (Bergé, 1976) that for any graph
with p disconnected components, the number of independent loops, or cyclomatic
number, is given by
 = E − N + p. (1.2)

It is easy to check that this relation gives  = 0 for a tree.

A most interesting property of random graphs (Section 3.1) is the distribution
of components, and in particular the existence of a giant component G, defined
as a component whose size scales with the number of vertices of the graph, and
therefore diverges in the limit N → ∞. The presence of a giant component implies
that a macroscopic fraction of the graph is connected.
The structure of the components of directed graphs is somewhat more complex
as the presence of a path from the node i to the node j does not necessarily guar-
antee the presence of a corresponding path from j to i. Therefore, the definition
of a giant component needs to be adapted to this case. In general, the component
structure of a directed network can be decomposed into a giant weakly connected
component (GWCC), corresponding to the giant component of the same graph in
which the edges are considered as undirected, plus a set of smaller disconnected
components, as sketched in Figure 1.2. The GWCC is itself composed of several
parts because of the directed nature of its edges: (1) the giant strongly connected

Giant weakly Disconnected

connected component components

Giant strongly Giant

Giant connected component out-component
in-component

Tendril

Tendril
Tube Tendril

Fig. 1.2. Component structure of a directed graph. Figure adapted from

Dorogovtsev et al. (2001a).
 1.2 Basic concepts in graph theory 7

component (GSCC), in which there is a directed path joining any pair of nodes;
(2) the giant in-component (GIN), formed by the nodes from which it is possible to
reach the GSCC by means of a directed path; (3) the giant out-component (GOUT),
formed by the nodes that can be reached from the GSCC by means of a directed
path; and (4) the tendrils containing nodes that cannot reach or be reached by the
GSCC (among them, the tubes that connect the GIN and GOUT), which form the
rest of the GWCC.
The concept of “path” lies at the basis of the definition of distance among ver-
tices. Indeed, while graphs usually lack a metric, the natural distance measure
between two vertices i and j is defined as the number of edges traversed by the
shortest connecting path (see Figure 1.3). This distance, equivalent to the chemi-
cal distance usually considered in percolation theory (Bunde and Havlin, 1991), is
called the shortest path length and denoted as i j . When two vertices belong to two
disconnected components of the graph, we define i j = ∞. While it is a symmet-
ric quantity for undirected graphs, the shortest path length i j does not coincide in
general with  ji for directed graphs.

A 2 B
3

C=0 C = 0.5 C=1

Fig. 1.3. Basic metrics characterizing a vertex i in the network. A, The degree
k quantifies the vertex connectivity. B, The shortest path length identifies the
minimum connecting path (dashed line) between two different vertices. C, The
clustering coefficient provides a measure of the interconnectivity in the vertex’s
neighborhood. As an example, the central vertex in the figure has a cluster-
ing coefficient C = 1 if all its neighbors are connected and C = 0 if no
interconnections are present.
8 Preliminaries: networks and graphs

By using the shortest path length as a measure of distance among vertices, it

is possible to define the diameter and the typical size of a graph. The diameter is
traditionally defined as
dG = max i j . (1.3)
i, j

Another effective definition of the linear size of the network is the average short-
est path length,2 defined as the average value of i j over all the possible pairs of
vertices in the network
1 
 = i j . (1.4)
N (N − 1) i j

By definition  ≤ dG , and in the case of a well-behaved and bounded shortest

path length distribution, it is possible to show heuristically that in many cases the
two definitions behave in the same way with the network size.
There are also other measures of interest which are related to the characterization
of the linear size of a graph. The eccentricity of a vertex i is defined by ec(i) =
max j =i i j , and the radius of a graph G by radG = mini ec(i). These different
quantities are not independent and one can prove (Clark and Holton, 1991) that the
following inequalities hold for any graph

radG ≤ dG ≤ 2 radG . (1.5)

Simple examples of distances in graphs include the complete graph where

 = 1 and the regular hypercubic lattice in D dimensions composed by N
vertices for which the average shortest path length scales as  ∼ N 1/D . In
most random graphs (Sections 2.2 and 3.1), the average shortest path length
grows logarithmically with the size N , as (  ∼ log N ) – a much slower
growth than that found in regular hypercubic lattices. The fact that any pair of
nodes is connected by a small shortest path constitutes the so-called small-world
effect.

1.2.3 Degree and centrality measures

When looking at networks, one of the main insights is provided by the impor-
tance of their basic elements (Freeman, 1977). The importance of a node or edge
is commonly defined as its centrality and this depends on the characteristics or
specific properties we are interested in. Various measurements exist to character-
ize the centrality of a node in a network. Among those characterizations, the most

2 It is worth stressing that the average shortest path length has also been referred to in the physics literature as
another definition for the diameter of the graph.
 1.2 Basic concepts in graph theory 9

commonly used are the degree centrality, the closeness centrality, or the between-
ness centrality of a vertex. Edges are frequently characterized by their betweenness
centrality.
Degree centrality
The degree ki of a vertex i is defined as the number of edges in the graph incident
on the vertex i. While this definition is clear for undirected graphs, it needs some
refinement for the case of directed graphs. Thus, we define the in-degree kin,i of the
vertex i as the number of edges arriving at i, while its out-degree kout,i is defined as
the number of edges departing from i. The degree of a vertex in a directed graph is
defined by the sum of the in-degree and the out-degree, ki = kin,i + kout,i . In terms
of the adjacency matrix, we can write
 
kin,i = x ji , kout,i = xi j . (1.6)
j j

For an undirected graph with a symmetric adjacency matrix, kin,i = kout,i . The
degree of a vertex has an immediate interpretation in terms of centrality quantifying
how well an element is connected to other elements in the graph. The Bonacich
power index takes into account not only the degree of a node but also the degrees
of its neighbors.
Closeness centrality
The closeness centrality expresses the average distance of a vertex to all others as
1
gi =  . (1.7)
j =i i j
This measure gives a large centrality to nodes which have small shortest path
distances to the other nodes.

Betweenness centrality
While the previous measures consider nodes which are topologically better con-
nected to the rest of the network, they overlook vertices which may be crucial
for connecting different regions of the network by acting as bridges. In order to
account quantitatively for the role of such nodes, the concept of betweenness cen-
trality has been introduced (Freeman, 1977; Newman, 2001a): it is defined as the
number of shortest paths between pairs of vertices that pass through a given vertex.
More precisely, if σh j is the total number of shortest paths from h to j and σh j (i) is
the number of these shortest paths that pass through the vertex i, the betweenness
of i is defined as
 σh j (i)
bi = . (1.8)
h = j =i
σh j
10 Preliminaries: networks and graphs

A similar quantity, the load or stress centrality, does not discount the multiplic-

ity of equivalent paths and reads as L i = h = j =i σh j (i). The above definitions
may include a factor 1/2 to avoid counting each path twice in undirected networks.
The calculation of this measure is computationally very expensive. The basic algo-
rithm for its computation would lead to a complexity of order O(N 2 E), which
is prohibitive for large networks. An efficient algorithm to compute betweenness
centrality is reported by Brandes (2001) and reduces the complexity to O(N E) for
unweighted networks.
According to these definitions, central nodes are therefore part of more shortest
paths within the network than less important nodes. Moreover, the betweenness
centrality of a node is often used in transport networks to provide an estimate of
the traffic handled by the vertices, assuming that the number of shortest paths is a
zero-th order approximation to the frequency of use of a given node. Analogously
to the vertex betweenness, the betweenness centrality of edges can be calculated as
the number of shortest paths among all possible vertex couples that pass through
the given edge. Edges with the maximum score are assumed to be important for the
graph to stay interconnected. These high-scoring edges are the “bridges” that inter-
connect clusters of nodes. Removing them frequently leads to unconnected clusters
of nodes. The “bridges” are particularly important for decreasing the average path
length among nodes in a network, for speeding up the diffusion of information, or
for increasing the size of the part of the network at a given distance from a node.
However, networks with many such bridges are more fragile and less clustered.

1.2.4 Clustering
Along with centrality measures, vertices are characterized by the structure of their
local neighborhood. The concept of clustering3 of a graph refers to the tendency
observed in many natural networks to form cliques in the neighborhood of any
given vertex. In this sense, clustering implies the property that, if the vertex i is
connected to the vertex j, and at the same time j is connected to l, then with a
high probability i is also connected to l. The clustering of an undirected graph can
be quantitatively measured by means of the clustering coefficient which measures
the local group cohesiveness (Watts and Strogatz, 1998). Given a vertex i, the clus-
tering C(i) of a node i is defined as the ratio of the number of links between the
neighbors of i and the maximum number of such links. If the degree of node i is ki
and if these nodes have ei edges between them, we have
ei
C(i) = , (1.9)
ki (ki − 1)/2

3 Also called transitivity in the context of sociology (Wasserman and Faust, 1994).
 1.3 Statistical characterization of networks 11

where it is worth remarking that this measure of clustering only has a meaning for
ki > 1. For ki ≤ 1 we define C(i) ≡ 0. Given the definition of ei , it is easy to check
that the number of edges among the neighbors of i can be computed in terms of the
adjacency matrix X as

1
ei = xi j x jl xli . (1.10)
2 jl

In Figure 1.3, we provide an illustration of some simple examples of the clustering

of vertices with a given neighborhood. The average clustering coefficient of a graph
is simply given by

1 
C = C(i). (1.11)
N i

It is worth noting that the clustering coefficient has been defined in a number of
similar ways, for instance as a function of triples in the network (triples are defined
as subgraphs which contain exactly three nodes) and reversing the order of average
and division in Equation (1.11)

3 × number of fully connected triples

C = , (1.12)
number of triples

where the factor 3 is due to the fact that each triangle is associated with three nodes.
This definition corresponds to the concept of the fraction of transitive triples intro-
duced in sociology (Wasserman and Faust, 1994). Different definitions give rise
to different values of the clustering coefficient for a given graph. Hence, the com-
parison of clustering coefficients among different graphs must use the very same
measure. In any case, both measures are normalized and bounded to be between 0
and 1.

1.3 Statistical characterization of networks

One of the elements that has fostered the recent development of network science
can be found in the recent possibility of systematic gathering and handling of data
sets on several large-scale networks. Indeed, in large systems, asymptotic regular-
ities cannot be detected by looking at local elements or properties. In other words,
one has to shift the attention to statistical measures that take into account the global
behavior of these quantities.
12 Preliminaries: networks and graphs

1.3.1 Degree distribution

The degree distribution P(k) of undirected graphs is defined as the probability that
any randomly chosen vertex has degree k. It is obtained by constructing the nor-
malized histogram of the degree of the nodes in a network. In the case of directed
graphs, one has to consider instead two distributions, the in-degree P(kin ) and
out-degree P(kout ) distributions, defined as the probability that a randomly cho-
sen vertex has in-degree kin and out-degree kout , respectively. The average degree
of an undirected graph is defined as the average value of k over all the vertices in
the network,
1   2E
k = ki = k P(k) ≡ , (1.13)
N i k
N

since each edge end contributes to the degree of a vertex. For a directed graph, the
average in-degree and out-degree must be equal,
  k
kin = kin P(kin ) = kout = kout P(kout ) ≡ , (1.14)
kin kout
2

since an edge departing from any vertex must arrive at another vertex. Analo-
gously to the average degree, it is possible to define the nth moment of the degree
distribution,

kn = k n P(k). (1.15)
k

A sparse graph has an average degree that is much smaller than the size of the
graph, k  N . In the next chapters we will see that the properties of the degree
distribution will be crucial to identify different classes of networks.

1.3.2 Betweenness distribution

Analogously to the degree, it is possible to introduce the probability distribution
Pb (b) that a vertex has betweenness b, and the average betweenness b defined as
 1 
b = b Pb (b) ≡ bi . (1.16)
b
N i

For this quantity it is worth showing its relation with the average shortest path
length  . By simply reordering the sums in the betweenness definition we have
   σh j (i)  1 
bi = = σh j (i). (1.17)
i i h, j =i
σh j σ
h = j h j i =h, j
 1.3 Statistical characterization of networks 13

A simple topological reasoning gives i =h, j σh j (i) = σh j (h j − 1). Plugged into

the previous equation, this yields i bi = N (N − 1)(  − 1), and therefore
b = (N − 1)(  − 1). (1.18)
As this formula shows, it is easy to realize that the betweenness usually takes values
of the order O(N ) or larger. For instance, in a star graph, formed by N − 1 vertices
with a single edge connected to a central vertex, the betweenness takes a maximum
value (N − 1)(N − 2) at the central vertex (the other peripheral vertices having 0
betweenness). For this reason, in the case of very large graphs (N → ∞) it is
sometimes convenient to define a rescaled betweenness b̃ ≡ N −1 b.
Finally, it is clear that in most cases the larger the degree ki of a node, the larger
its betweenness centrality bi will be. More quantitatively, it has been observed
that for large networks and for large degrees there is a direct association of the
form (Goh, Kahng and Kim, 2001, 2003; Barthélemy, 2004)
η
bi ∼ ki , (1.19)
where η is a positive exponent depending on the network. Such associations are
extremely relevant as they correspond to the fact that a large number of short-
est paths go through the nodes with large degree (the hubs). These nodes will
therefore be visited and discovered easily in processes of network exploration (see
Chapter 8). They will also typically see high traffic which may result in conges-
tion (see Chapter 11). Of course, fluctuations are also observed in real and model
networks, and small-degree nodes may also have large values of the betweenness
centrality if they connect different regions of the network, acting as bridges. Such
low-degree nodes with high centrality appear, for instance, in networks where
spatial constraints limit the ability of hubs to deploy long links (Guimerà and
Amaral, 2004; Barrat, Barthélemy and Vespignani, 2005).

1.3.3 Mixing patterns and degree correlations

As a discriminator of structural ordering of large-scale networks, the attention of the
research community has initially been focused on the degree distribution, but it is
clear that this function is only one of the many statistics characterizing the struc-
tural and hierarchical ordering of a network. In particular, it is likely that nodes do
not connect to each other irrespective of their property or type. On the contrary, in
many cases it is possible to collect empirical evidence of specific mixing patterns in
networks. A typical pattern known in ecology, epidemiology, and social science as
“assortative mixing” refers to the tendency of nodes to connect to other nodes with
similar properties. This is common to observe in the social context where people pre-
fer to associate with others who share their interests. Interesting observations about
14 Preliminaries: networks and graphs

assortative mixing by language or race are abundant in the literature. Likewise, it

is possible to define a “disassortative mixing” pattern whenever the elements of the
network prefer to share edges with those who have a different property or attribute.
Mixing patterns have a profound effect on the topological properties of a network as
they affect the community formation and the detailed structural arrangements of the
connections among nodes.
While mixing patterns can be defined with respect to any type or property of
the nodes (Newman, 2003a), in the case of large-scale networks the research com-
munity’s interest has focused on the mixing by vertex degree. This type of mixing
refers to the likelihood that nodes with a given degree connect with nodes with
similar degree, and is investigated through the detailed study of multipoint degree
correlation functions. Most real networks do exhibit the presence of non-trivial
correlations in their degree connectivity patterns. Empirical measurements provide
evidence that high or low degree vertices of the network tend, in many cases, to
preferentially connect to other vertices with similar degree. In this situation, corre-
lations are named assortative. In contrast, connections in many technological and
biological networks are more likely to attach vertices of very different degree. Cor-
relations are then referred to as disassortative. The correlations, although other
possibilities could be considered, are characterized through the conditional proba-
bilities P(k  , k  , . . ., k (n) | k) that a vertex of degree k is simultaneously connected
to a number n of other vertices with corresponding degrees k  , k  , . . ., k (n) . Such
quantities might be the simplest theoretical functions that encode degree correla-
tion information from a local perspective. A network is said to be uncorrelated
when the conditional probability is structureless, in which case the only relevant
function is just the degree distribution P(k).
In order to characterize correlations, a more compact quantity is given by the
Pearson assortativity coefficient r (Newman, 2002a)
  2
e je ke /E − e ( je
+ ke )/(2E)
r =    2 , (1.20)
(
e ej 2 + k 2 )/(2E) −
e e ( je + k e )/(2E)

where je and ke denote the degree of the extremities of edge e and E is the total
number of edges. This quantity varies from −1 (disassortative network) up to 1
(perfectly assortative network). However, such a measure can be misleading when
a complicated behavior of the correlation functions (non-monotonous behavior) is
observed. In this case the Pearson coefficient gives a larger weight to the more
abundant degree classes, which in many cases might not express the variations of
the correlation function behavior.
More details on the degree correlations are provided by the two-point conditional
probability P(k  | k) that any edge emitted by a vertex with degree k is connected
to a vertex with degree k  . Even in this simple case, however, the direct evaluation
 1.3 Statistical characterization of networks 15

of the function from empirical data is a rather cumbersome task. In general, two
nodes’ degree correlations can be represented as the three-dimensional histograms
of P(k  | k) or related quantities4 (Maslov and Sneppen, 2002). On the other hand,
such a histogram is highly affected by statistical fluctuations and, thus, it is not
a good candidate when the data set is not extremely large and accurate. A more
practical quantity in the study of the network structure is given by the average
nearest neighbors degree of a vertex i
1 
knn,i = kj, (1.21)
ki
j∈V (i)

where the sum is over the nearest neighbors vertices of i. From this quan-
tity a convenient measure to investigate the behavior of the degree correlation
function is obtained by the average degree of the nearest neighbors, knn (k), for
vertices of degree k (Pastor-Satorras, Vázquez and Vespignani, 2001; Vázquez,
Pastor-Satorras and Vespignani, 2002)
1 
knn (k) = knn,i , (1.22)
Nk i/k =k
i

where Nk is the number of nodes of degree k. This last quantity is related to the
correlations between the degrees of connected vertices since on average it can be
expressed as

knn (k) = k  P(k  |k). (1.23)
k

If degrees of neighboring vertices are uncorrelated, P(k  |k) is only a function of k 

and thus knn (k) is a constant. In the presence of correlations, the behavior of knn (k)
identifies two general classes of networks (see Figure 1.4). If knn (k) is an increasing
function of k, vertices with high degree have a larger probability of being connected

knn(k )
Assortative

Disassortative

Fig. 1.4. Pictorial representation of the assortative (disassortative) mixing prop-

erty of networks as indicated by the behavior of the average degree of the nearest
neighbors, knn (k), for vertices of degree k.

4 For instance, the joint degree distribution P(k  , k) that defines the probability that a randomly chosen edge
connects two vertices of degrees k and k  .
16 Preliminaries: networks and graphs

with large degree vertices. This corresponds to an assortative mixing (Newman,

2002a). On the contrary, a decreasing behavior of knn (k) defines a disassortative
mixing, in the sense that high degree vertices have a majority of neighbors with
low degree, while the opposite holds for low degree vertices (Newman, 2002a).
It is important to stress, however, that given a certain degree distribution, a com-
pletely degree–degree uncorrelated network with finite size is not always realizable
owing to structural constraints. Indeed, any finite-size random network presents a
structural cut-off value kc over which the requirement of the lack of dangling edges
introduces the presence of multiple and self-connections and/or degree–degree cor-
relations (Boguñá, Pastor-Satorras and Vespignani, 2004; Moreira, Andrade and
Amaral, 2002). Networks with bounded degree distributions and finite second
moments k 2 present a maximal degree kmax that is below the structural one kc .
However, in networks with heavy-tailed degree distribution (see Section 2.1), this
is not always the case and kc is generally smaller than kmax . In this instance, struc-
tural degree–degree correlations and higher order effects, such as the emergence
of large cliques (Bianconi and Marsili, 2006a; Bianconi and Marsili, 2006b), are
present even in maximally random networks.5 Structural correlations are genuine
correlations, which is not surprising since they are just imposed by topological
constraints and not by a special ordering or evolutionary principle shaping the net-
work. A more detailed discussion on topological constraints and the properties of
random networks can be found in Appendix 1.
In the case of random uncorrelated networks, it is possible to obtain an explicit
form for the conditional probability P(k  | k). In this case P(k  | k) does not depend
on k and its functional form in terms of k  can be easily obtained by calculating the
probability that any given edge is pointing to a node of degree k  . The probability
that one edge is wired to one node of degree k  is just the total number of edges
departing from nodes of degree k  divided by the number of all edges departing
from nodes of any degree. Since each one of the Nk  nodes emanates k  edges, we
obtain
k  Nk 
Punc (k  | k) =   , (1.24)
k  k Nk 
where considering that P(k) = Nk /N , finally yields
1 
Punc (k  | k) = k P(k  ). (1.25)
k
This expression states that even in an uncorrelated network, the probability that
any edge points to a node of a given degree k  is not uniform but proportional to

5 Operatively, the maximally random network can be thought of as the stationary ensemble of networks visited
by a process that, at any time step, randomly selects a couple of links of the original network and exchanges
two of their ending points (automatically preserving the degree distribution).
 1.3 Statistical characterization of networks 17

the degree itself. In other words, by following any edge at random it is more likely
you will end up in a node with large degree: the more connected you are, the easier
it is to find you. This result is particularly relevant in the case of networks where
degree values may be very different, and will prove to be particularly useful in
several analytical calculations where the heterogeneous nature of the network is
affecting equilibrium and non-equilibrium processes even in the uncorrelated case.
The behavior of knn (k) in the uncorrelated case can be easily derived from Equation
(1.25) and reads as
k2
unc
knn (k) = . (1.26)
k
As we stated previously, in a random uncorrelated network the average nearest
neighbor degree does not depend on k and has a constant value determined by the
two first moments of the degree distribution.

1.3.4 Clustering spectrum

Correlations among three vertices can be measured by means of the probability
P(k  , k  | k) that a vertex of degree k is simultaneously connected to two ver-
tices with degrees k  and k  . As previously indicated, the conditional probabilities
P(k  , k  | k) or P(k  | k) are difficult to estimate directly from real data, so other
assessments have been proposed. All of them are based on the concept of cluster-
ing coefficient C(i) as expressed in Equation (1.9), which refers to the tendency
to form triangles in the network (see Section 1.2.4). A measure customarily used

in graph characterization is the average clustering coefficient C = N −1 i C(i)
which expresses the statistical level of cohesiveness measuring the global density
of interconnected vertex triplets in the network. Although statistical scalar mea-
sures are helpful as a first indication of clustering, it is always more informative
to work with quantities that explicitly depend on the degree. As in the case of
two vertices’ correlations, a uniparametric function is defined by the average clus-
tering coefficient C(k) restricted to classes of vertices with degree k (Vázquez
et al., 2002; Ravasz et al., 2002)
1 
C(k) = C(i) (1.27)
Nk i/k =k
i

where Nk is the number of vertices with degree k. A functional dependence of the

local clustering on the degree can be attributed to the presence of a complex struc-
ture in the three-vertex correlation pattern. Indeed, it has been observed that C(k)
exhibits, in many cases, a non-trivial dependence on k that is supposed to partly
18 Preliminaries: networks and graphs

encode the hierarchical structure of the network (Vázquez et al., 2002; Ravasz
et al., 2002).

1.3.5 Rich-club phenomenon

Several other statistical measures have been defined in the case of large-scale
networks as simple proxies for their architectures and many of them are specifi-
cally devised for certain types of graphs, as seen in the density of bipartite cliques
K n,m in directed graphs. Analogously, the “rich-club” phenomenon has been dis-
cussed in both social and computer sciences (de Solla Price, 1986; Wasserman and
Faust, 1994; Zhou and Mondragon, 2004; Pastor-Satorras and Vespignani, 2004),
and refers to the tendency of high degree nodes, the hubs of the network (the rich
nodes), to be very well connected to each other, forming well-interconnected sub-
graphs (clubs) more easily than low degree nodes. Zhou and Mondragon (2004)
have introduced a quantitative measure of this tendency through the rich-club
coefficient, expressed as
2E >k
φ(k) = , (1.28)
N>k (N>k − 1)
where E >k is the number of edges among the N>k nodes with degree larger than
k, and N>k (N>k − 1)/2 represents the maximum possible number of edges among
these N>k nodes, so that φ(k) measures the fraction of edges actually connect-
ing those nodes out of the maximum number of edges they might possibly share.
A growing behavior as a function of k indicates that high degree nodes tend to
be increasingly connected among themselves.6 On the other hand, a monotonic
increase of φ(k) does not necessarily imply the presence of a rich-club organizing
principle. Indeed, even in the case of the Erdős–Rényi graph – a completely random
network – there is an increasing rich-club coefficient. This implies that the increase
of φ(k) is a natural consequence of the fact that vertices with large degree have a
larger probability of sharing edges than low degree vertices. This feature is there-
fore imposed by construction and does not represent a signature of any particular
organizing principle or structure. For this reason, it is appropriate to measure the
rich-club phenomenon as ρran (k) = φ(k)/φran (k), where φran (k) is the rich-club
coefficient of the maximally random network with the same degree distribution
P(k) as the network under study (Colizza et al., 2006b). In this case an actual
rich-club ordering is denoted by a ratio ρran (k) > 1. In other words, ρran (k) is a

6 It is also worth stressing that the rich-club phenomenon is not trivially related to the mixing properties of
networks described in Section 1.3.3, which permit the distinction between assortative networks, where large
degree nodes preferentially attach to large degree nodes, and disassortative networks, showing the opposite
tendency (Colizza et al., 2006b).
 1.4 Weighted networks 19

normalized measure which discounts the structural correlations due to unavoidable

connectivity constraints, providing a better discrimination of the actual presence
of the rich-club phenomenon due to the ordering principles shaping the network.
This example makes explicit the need to consider the appropriate null hypotheses
when measuring correlations or statistical properties. Such properties might either
be simply inherent to the connectivity constraints present in networks, or be the
signature of real ordering principles and structural properties due to other reasons.

1.4 Weighted networks

Along with a complex topological structure, real networks display a large hetero-
geneity in the capacity and intensity of their connections: the weights of the edges.
While the topological properties of a graph are encoded in the adjacency matrix xi j ,
weighted networks are similarly described by a matrix wi j specifying the weight
on the edge connecting the vertices i and j (wi j = 0 if the nodes i and j are not
connected). The weight wi j may assume any value and usually represents a phys-
ical property of the edge: capacity, bandwidth, traffic. A very significant measure
of a network’s properties in terms of the actual weights is also obtained by looking
at the vertex strength si defined as (Yook et al., 2001; Barrat et al. 2004a)

si = wi j , (1.29)
j∈V (i)

where the sum runs over the set V(i) of neighbors of i. The strength of a node
integrates the information about its degree and the importance of the weights of its
links and can be considered as the natural generalization of the degree. When the
weights are independent of the topology, the strength typically grows linearly with
the degree, i.e. with the number of terms in the sum (1.29): s  w k where w is
the average weight. In the presence of correlations we obtain in general s  Ak β
with β = 1 and A = w , or β > 1. Statistical measures for weighted networks
are readily provided by the probability distributions P(w) and P(s) that any given
edge and node have weight w and strength s, respectively.
In general, topological measures do not take into account that some edges are
more important than others. This can easily be understood with the simple exam-
ple of a network in which the weights of all edges forming triples of interconnected
vertices are extremely small. Even for a large clustering coefficient, it is clear that
these triples have a minor role in the network’s dynamics and organization, and
the network’s clustering properties are definitely overestimated by a simple topo-
logical analysis. Similarly, high degree vertices could be connected to a majority
of low degree vertices while concentrating the largest fraction of their strength
only on the vertices with high degree. In this case the topological information
20 Preliminaries: networks and graphs

would point to disassortative properties while the network could be considered

as effectively assortative, since the more relevant edges in terms of weight are
linking high degree vertices. In order to solve these incongruities, it is possi-
ble to introduce specific definitions that explicitly consider the weights of the
links and combine the topological information with the weight distribution of the
network. Many different clustering coefficient definitions have been introduced
in the literature (Barrat et al., 2004; Onnela et al., 2005; Serrano, Boguñá and
Pastor-Satorras, 2006; Saramäki et al., 2007). A convenient weighted clustering
coefficient is defined as

1  (wi j + wi h )
C w (i) = xi j xi h x j h . (1.30)
si (ki − 1) j,h 2

The quantity C w (i) is a count of the weight of the two participating edges of the
vertex i for each triple formed in the neighborhood of i. This definition not only
considers the number of closed triangles in the neighborhood of a vertex but also
considers their total relative edge weights with respect to the vertex’s strength. The
factor si (ki − 1) is a normalization factor that ensures that 0 ≤ C w (i) ≤ 1. Con-
sistently, the C w (i) definition recovers the topological clustering coefficient in the
case that wi j = const. It is customary to define C w and C w (k) as the weighted clus-
tering coefficient averaged over all vertices of the network and over all vertices with
degree k, respectively. For a large randomized network (without any correlations
between weights and topology), it is easy to see that C w = C and C w (k) = C(k).
In real weighted networks, however, we can face two opposite cases. If C w > C,
we observe a network in which the interconnected triples are more likely formed
by edges with larger weights (see Figure 1.5). In contrast, C w < C signals a net-
work in which the topological clustering is generated by edges with low weight.
In the latter, it is explicit that the clustering has a minor effect in the organization
of the network since the largest part of the interactions (traffic, frequency of the
relations, etc. . .) occurs on edges not belonging to interconnected triples. In order
to obtain a more detailed knowledge of the structure of the network, the variations
of C w (k) with respect to the degree class k can be analyzed and compared with
those of C(k).
Similarly, the weighted average nearest neighbors degree is defined as (Barrat
et al., 2004a)

1
N
w
knn,i = xi j wi j k j . (1.31)
si j=1
 1.4 Weighted networks 21

A B

i
i

w=5
w=1 w
knn, i < k nn, i
C(i ) < C w(i )

Fig. 1.5. A, The weighted clustering of a node i is larger than its topological coun-
terpart if the weights of the links (i, j) are concentrated on the cliques in which i
participates. B, Example of a node i with small average nearest neighbors degree
but large weighted average nearest neighbors degree: i is mostly connected to
low-degree nodes but the link with largest weight points towards a well-connected
hub.

This quantity (see also Serrano et al., 2006) is a simple generalization of the
average nearest neighbors degree (Pastor-Satorras et al., 2001)

1 
N
knn,i = xi j k j (1.32)
ki j=1

and performs a local weighted average of the nearest neighbor degree according to
the normalized weight of the connecting edges, wi j /si . This definition implies that
w
knn,i > knn,i if the edges with the larger weights point to the neighbors with larger
w
degree and knn,i < knn,i for the opposite (see Figure 1.5). Thus, knn,i
w
measures
the effective affinity to connect with high or low degree neighbors according to
w
the magnitude of the actual interactions. Also, the behavior of the function knn (k)
w
(defined as the average of knn,i over all vertices with degree k) marks the weighted
assortative or disassortative properties considering the actual interactions among
the system’s elements.
A final note should be made concerning the local heterogeneities introduced
by weights. The strength of a node i is the sum of the weights of all links in
which i participates. The same strength can, however, be obtained with very dif-
ferent configurations: the weights wi j may be either of the same order si /ki or
heterogeneously distributed among the links. For example, the most heteroge-
neous situation is obtained when one weight dominates over all the others. A
simple way to measure this “disparity” is given by the quantity Y2 introduced in
other contexts (Herfindal, 1959; Hirschman, 1964; Derrida and Flyvbjerg, 1987;
Barthélemy, Gondran and Guichard, 2003).
22 Preliminaries: networks and graphs

w = 10
w=1

Y2 = 0.46 Y2 = 0.17

Fig. 1.6. If there is a small number of important connections around a node, the
quantity Y2 is of order 1/m with m of order unity. In contrast, if all the connections
are of the same order, Y2 is small and of order 1/k where k is the node’s degree.

 wi j 2
Y2 (i) = . (1.33)
si
j∈V (i)

If all weights are of the same order then Y2 ∼ 1/ki (for ki  1) and if a small
number of weights dominate then Y2 is of the order 1/m with m of order unity (see
Figure 1.6).
A similar finding is encoded in the local entropy, defined for nodes of degree
larger than 2 as
1  wi j wi j
f (i) = − ln . (1.34)
ln ki si si
j∈V (i)

This quantity goes from 0 if the strength of i is fully concentrated on one link to the
maximal value 1 for homogeneous weights: it can thus be used as an alternative or
complement to the disparity Y2 to investigate the local heterogeneity of the weights.
It is important to stress regarding weighted networks that we have emphasized
some of the measures that are customarily used to analyze large-scale complex
systems. In particular, most of them have been adopted in the context of the
recent analysis of infrastructure and communication networks, and will be used
to evaluate the effect of heterogeneities and complexity in dynamical phenomena
affected by the weighted nature of the connections. Several other quantities related
to weighted properties are, however, defined and used in graph theory. In particular,
weighted distances between two nodes are defined by assigning an effective length
(depending on the weight) to each edge and summing these lengths along the path
followed. The minimum spanning tree corresponds to the way of connecting all the
vertices together with the smallest possible global weight. In standard graph the-
ory, mostly flows have been studied, which represent a particular case of weighted
networks where the weights must satisfy the conservation equation at each node,
stating that the total amount of flow into a node is equal to the amount of flow
 1.4 Weighted networks 23

going out of it. Although this problem is very important in many instances (such as
electrical networks, fluids in pipes, etc.), the conservation equation is not satisfied
in all real-world networks and we refer the reader interested in flow problems (such
as the maximum flow) to graph theory books (Clark and Holton, 1991; Cormen
et al., 2003; Jungnickel, 2004).
 2
Networks and complexity

Undeniably, the visualizations of the Internet or the airport network convey the
notion of intricate, in some cases haphazard, systems of a very complicated nature.
Complexity, however, is not the same as the addition of complicated features.
Despite the fact that there is no unique and commonly accepted definition of
complexity – it is indeed very unlikely to find two scientists sharing the same def-
inition of complex system – we discuss from what perspectives many real-world
networks can be considered as complex systems, and what are the peculiar fea-
tures signaling this occurrence. In this chapter we review the basic topological
and dynamical features that characterize real-world networks and we attempt to
categorize networks into a few broad classes according to their observed statisti-
cal properties. In particular, self-organized dynamical evolution and the emergence
of the small-world and scale-free properties of many networks are discussed as
prominent concepts which have led to a paradigm shift in which the dynamics
of networks have become a central issue in their characterization as well as in
their modeling (which will be discussed in the next chapter). We do not aim, how-
ever, at an exhaustive exposition of the theory and modeling of complex networked
structures since, as of today, there are reference textbooks on these topics such as
those by Dorogovtsev and Mendes (2003), Pastor-Satorras and Vespignani (2004),
and Caldarelli (2007), along with journal reviews by Albert and Barabási (2002),
Dorogovtsev and Mendes (2002), Newman (2003c), and Boccaletti et al. ([2006]).

2.1 Real-world systems

In recent times, the increased power of computers and the informatics revolution
have made possible the systematic gathering and handling of data sets on several
large-scale networks, allowing detailed analysis of their structural and functional
properties. As a first guide to the classification of the data obtained, we can pro-
vide a rudimentary taxonomy of real-world networks. Two main different classes

24
 2.1 Real-world systems 25

are infrastructure systems, and natural or living systems. Each of these classes can
be further divided into different subgroups. Natural systems networks can be differ-
entiated into the subgroups of biology, social systems, food webs and ecosystems,
among others. For instance, biological networks refer to the complicated sets of
interactions among genes, proteins and molecular processes which regulate bio-
logical life, while social networks describe the relations between individuals such
as family links, friendships, and work relations. In particular, leading sociologists
refer to our societies as networked societies and even if their analysis has been
largely focused on small and medium-scale networks, social network studies play a
key role in introducing basic definitions and quantities of network science. Indeed,
biological and social networks are prominent research topics and much of network
science is influenced by what has happened in these fields recently.
In turning our attention to infrastructure networks we can readily separate two
main subcategories. The first contains virtual or cyber networks. These networks
exist and operate in the digital world of cyberspace. The second subcategory
includes physical systems such as energy and transportation networks. This is
a rough classification since there are many interrelations and interdependencies
existing among physical infrastructure networks, as well as between physical and
digital networks. In the Internet, for instance, the cyber features are mixed with the
physical features. The physical Internet is composed of physical objects such as
routers – the main computers which allow us to communicate – and transmission
lines, the cables which connect the various computers. On top of this physical layer
lies a virtual world made of software that may define different networks, such as the
World Wide Web (WWW), email networks, and Peer-to-Peer networks. These net-
works are the information transfer channels for hundreds of millions of users and,
like the physical Internet, have grown to become enormous and intricate networks
as the result of a self-organized growing process. Their dynamics are the outcomes
of the interactions among the many individuals forming the various communities,
and therefore are mixtures of complex socio-technical aspects. Infrastructure net-
works represent a combination of social, economic, and technological processes.
Further examples can be found in the worldwide airport network (WAN) and power
distribution networks where physical and technological constraints cooperate with
social, demographic, and economic factors.

2.1.1 Networks everywhere

The various systems mentioned so far are characterized by the very different nature
of their constitutive elements. It is therefore appropriate to offer a list of some
network data sets prototypically considered in the literature, making clear to which
property and elements their graph representation refers.
26 Networks and complexity

Social networks
The science of social networks is one of the pillars upon which the whole field of
network science has been built. Since the early works of Moreno (1934) and the
definition of the sociogram, social networks have been the object of constant analy-
sis and study. Social networks represent the individuals of the population as nodes
and the social ties or relations among individuals as links between these nodes.
The links therefore may refer to very different attributes such as friendship among
classmates, sexual relations among adults, or just the belonging to common institu-
tions or work teams (collaborative interactions). The importance of these networks
goes beyond social sciences and affects our understanding of a variety of processes
ranging from the spreading of sexually transmitted diseases to the emergence of
consensus and knowledge diffusion in different kinds of organizations and social
structures.
Recently, the recording of social interactions and data in electronic format has
made available data sets of unprecedented size. The e-mail exchanges in large
corporate organizations and academic institutions make tracking social interac-
tions among thousands of individuals possible in a precise and quantitative way
(Ebel, Mielsch and Bornholdt, 2002; Newman, Forrest and Balthrop, 2002). Habits
and shared interests may be inferred from web visits and file sharing. Profes-
sional communities have been analyzed from wide databases such as complex
collaboration networks. Examples are the already classic collaboration network
of film actors (Watts and Strogatz, 1998; Barabási and Albert 1999; Amaral
et al., 2000; Ramasco, Dorogovtsev and Pastor-Satorras, 2004), the company direc-
tors network (Newman, Strogatz and Watts, 2001; Davis, Yoo and Baker, 2003;
Battiston and Catanzaro, 2004) and the network of co-authorship among scientists
(Newman, 2001a; 2001b; 2001c). In these last examples we encounter bipartite
networks (Wasserman and Faust, 1994), although the one mode projection is often
used so that members are tied through common participation in one or more films,
boards of directors, or academic papers. As an example, we report in Figure 2.1 an
illustration of a construction of the scientific collaboration network (SCN). Accord-
ing to this construction, the authors are the nodes of the network and share an edge
if they co-authored a paper. Similarly, for the actors’ collaboration network, two
actors are linked if they have co-starred in a movie. More information can be pro-
jected in the graph by weighting the intensity of the collaboration. A convenient
definition of the weight is introduced by Newman (2001b), who considers that the
intensity wi j of the interaction between two collaborators i and j is given by

 δip δ jp
wi j = , (2.1)
p
np − 1
 2.1 Real-world systems 27
Authors Papers

A
1
A B
B
2

C
C D
3

Fig. 2.1. A bipartite graph is defined as a graph whose nodes can be divided into
two separate sets or classes such that every edge links a node of one set to a node
of the other set. For a co-authorship network, authors and papers form two distinct
classes and the edges representing the authorship just go from authors to papers.
The one mode projection defines a unipartite graph, where the nodes of one class
share an edge if they are connected to a common node of the other class. The
one mode projection on the author set defines the co-authorship network in which
authors share an edge if they co-authored at least one paper.

where the index p runs over all co-authored papers, n p is the number of authors
p
of the paper p, and δi is 1 if author i has contributed to paper p, and 0 otherwise.
The strength of the interactions is therefore large for collaborators having many
papers in common but the contribution to the weight introduced by any given paper
is inversely proportional to the number of authors. Similarly, other definitions of
weight may be introduced in order to measure the impact of collaborations in the
scientific community. For instance, Börner et al. (2004) consider that the weight of
each edge is also a function of the number of citations of each paper, providing a
quantitative assessment of the impact of co-authorship teams. This is an example
of social science merging with bibliometrics, another discipline which has recently
benefited from the impact of the e-revolution and text digitalization on gathering
large-scale network data sets. Just to cite a paramount example, the networks of
citations (there is a directed link between two papers if one cites the other) among
scientific papers of several databases for journals such as the Proceedings of the
National Academy of Sciences (PNAS) or Physical Review contain thousands of
nodes (Redner, 2005).

Transportation networks
Transportation systems such as roads, highways, rails, or airlines are crucial in our
modern societies and will become even more important in a world where more
than 50% of the population lives in urban areas.1 The importance of this subject

1 Source: United Nations, population division http://www.unpopulation.org

28 Networks and complexity

justifies the huge literature published on these systems for at least the past 70 years.
Studies cover a broad range of approaches from applied engineering to mathematic
works on users’ equilibrium (for an introductory book on the subject, see Sheffi
[1985]). Geographers, regional science experts, and economists have long observed
the existence of structures such as hierarchies or particular subgraphs present in the
topology (see for example Yerra and Levinson [2005] and references therein). In
the air traffic network, for instance, O’Kelly (1998) has discussed the existence of
hubs and of particular motifs. Hierarchies present in a system of cities were also
noticed a long time ago, and Christaller (1966) proposed his famous central place
theory in which heterogeneous distributions of facilities and transportation costs
are the cause of the emergence of hierarchies. More recently, Fujita, Krugman and
Venables (1999) have proposed a model which explains the hierarchical formation
of cities in terms of decentralized market processes. These different models are,
however, very theoretical, and more thorough comparison with empirical data is
needed at this stage. Only recently, large-scale data and extensive statistics have
opened the path to better characterization of these networks and their interaction
with economic and geographic features. In this context, the network representa-
tion operates at different scales and different representative granularity levels are
considered.
The TRANSIMS project characterizes human flows at the urban level (Chowell
et al., 2003; Eubank et al., 2004). This study concerns the network of locations
in the city of Portland, Oregon, where nodes are city locations including homes,
offices, shops, and recreational areas. The edges among locations represent the
flow of individuals going at a certain time from a location to another. In Figure 2.2
we illustrate the network construction and we report an example of the type of
data set used by Chowell, Hyman, Eubank and Castillo-Chavez (2003). Other rele-
vant networks studied at this scale are subways (Latora and Marchiori, 2002), and
roads and city streets (Cardillo et al., 2006; Buhl et al., 2006; Crucitti, Latora and
Porta, 2006; Scellato et al., 2006; Kalapala et al., 2006).
At a slightly coarser scale, it is possible to map into a weighted network
the commuting patterns among urban areas, municipalities, and counties (Montis
et al., 2007), as well as railway systems (Sen et al., 2003). Finally, on the global
scale, several studies have provided extensive analysis of the complete worldwide
airport network and the relative traffic flows (Barrat et al., 2004a; Guimerà and
Amaral, 2004; Guimerà et al., 2005): this transportation system can be repre-
sented as a weighted graph where vertices denote airports, and weighted edges
account for the passenger flows between the airports. The graphs resulting from
large-scale transportation networks define the basic structure of metapopulation
or patch models used in studies of epidemic spread. In this representation the
 2.1 Real-world systems 29

Person Location Location Arrival Departure

ID ID type time time
wij
Location i 116 4356 home 00:00 07:00
wji 116 98135 work 08:00 11:00
116 71457 work 11:20 13:00
116 98135 work 13:20 17:00
Location j 116 4356 home 18:00 19:15
Location k 116 21343 social 19:30 21:00
116 4356 home 21:00 07:00
324 12679 home 00:00 07:00
324 431 school 08:00 14:00
324 12679 home 14:30 19:00
Location l
. . . . .
.. .. .. .. ..
.. .. .. .. ..

Fig. 2.2. Construction of the network between the various locations in the city
of Portland as reported by Chowell et al. (2003). The table is a small sample
of the data set, obtained through a TRANSIMS simulation, which records for
each individual the successive locations visited during a day, with the arrival and
departure time. The network between locations is constructed from this data set:
an edge exists between location i and j whenever an individual has gone directly
from i to j, and the weight wi j gives the daily traffic between i and j.

network’s nodes refer to different populations such as urban areas or cities, and an
edge between two nodes denotes an exchange of individuals between the two popu-
lations (see also Chapter 9). These networks are usually directed and weighted in
that the number of individuals going from one population to the other is not neces-
sarily symmetric and the weight of each edge quantifies the number of exchanged
individuals. In general, transportation networks naturally define such networked
structures, with the edges’ weights denoting the traffic flows (commuters, passen-
gers etc.) between the different locations, and network theory appears as the natural
tool for the study of these systems.

Internet
Characterizing how routers, computers, and physical links interconnect with each
other in the global Internet is very difficult because of several key features of the
network. One main problem is related to its exponential growth rate. The Internet,
in fact, has already increased by five orders of magnitude since its birth. A sec-
ond difficulty is the intrinsic heterogeneity of the Internet which is composed of
networks engineered with considerable technical and administrative diversity. The
different networking technologies are merged together by the TCP/IP architecture
which provides connectivity but does not imply a uniform behavior. Also very
important is the fact that the Internet is a self-organizing system whose properties
30 Networks and complexity

cannot be traced back to any global blueprint or chart. This means that routers and
links are added by competing entities according to local economic and technical
constraints, leading to a very intricate physical structure that does not comply with
any globally optimized plan (Pastor-Satorras and Vespignani, 2004; Crovella and
Krishnamurthy, 2006).
The combination of all these factors results in a general lack of understanding
about the large-scale topology of the Internet, but in recent years, several research
groups have started to deploy technologies and infrastructures in order to obtain a
more global picture of this network. Efforts to obtain such maps have been focused
essentially on two levels. First, the inference of router adjacencies amounts to a
measure of the internet router (IR) level graph. The second mapping level concerns
the Autonomous System (AS) graph of the Internet, referring to autonomously
administered domains which to a first approximation correspond to internet ser-
vice providers and organizations. Therefore, internet maps are usually viewed as
undirected graphs in which vertices represent routers or Autonomous Systems and
edges (links) represent the physical connections between them. Although these two
graph representations are related, it is clear that they describe the Internet at rather
different scales (see Figure 2.3). In fact, the collection of Autonomous Systems
and inter-domain routing systems defines a coarse-grained picture of the Internet
in which each AS groups many routers together, and links are the aggregations of
all the individual connections between the routers of the corresponding AS.
Internet connectivity information at the level of Autonomous Systems can be
retrieved by the inspection of routing tables and paths stored in each router (pas-
sive measurements) or by direct exploration of the network with a software probe
(active measurements). Based on the first strategy, the Oregon Route Views project
(RV) provides maps of the AS graph obtained from the knowledge of the routing

AS 2 AS 3
AS 3
AS 2

AS 1 AS 4
AS 4
AS 1
LAN

Fig. 2.3. Different granularity representations of the Internet. The hosts (squares)
are included in the Local Area Networks (LAN) that connect to routers (shaded
circles) and are excluded from the maps. Each Autonomous System is composed
of many routers.
Another Random Scribd Document
with Unrelated Content
 PLAN No. 472. GETTING A START WITH INK POWDERS
A young man whose ambition was to build up a permanent
business from a small beginning, as he was practically without
capital, concluded to start on one item at first, and gradually add
others as he could afford it, so he chose inks—not one, but several
kinds of inks. These inks he purposed to put up in the form of
powders, leaving only the hot water to be added by the customer.
But the different formulas were all so good that from anyone of
them an enterprising man could work into a good-paying business,
and they are therefore submitted herewith as separate plans.
Here is the formula he used for producing a powder for a writing
fluid that is equal to the best inks on the market and better than
most of them. And the beauty of it was that he could sell enough of
the powder for 10 cents to make a full pint of the very best ink, and
realize a very good profit on it over that price:
Nigrosin, 1 ounce; soluble blue or water blue anilin, 2 ounces;
salicylic acid, 15 grains; dextrin, 11⁄2 ounces.
This will make from one to two gallons of ink, when dissolved in
hot water, according to the shade desired—the more powder the
darker the ink. Fine for either ordinary or fountain pen, and sold
well.
 PLAN No. 473. BLUE INK POWDER
Many people prefer blue ink, and for them he made powders of an
excellent quality as follows:
Water-blue anilin, 1 dram; dextrin, 5 drams; or according to the
following formula:
Soluble Prussian blue, 1 dram; dextrin, 2 drams. Dissolve the
powder in hot water, varying the intensity of the blue shade as
desired, by using more or less powder.
This was a popular and profitable seller.
 PLAN No. 474. GREEN INK POWDER
Green ink is a novelty, and for that reason many people like to use
it. He made the powders for green ink as follows:
Green anilin, 1 dram; dextrin, 4 drams. To use, dissolve in hot
water, using more or less of the powder as darker or lighter shades
of the green are desired.
Very easy and cheap to make; very easy and profitable to sell.
 PLAN No. 475. RED INK POWDER
Red ink is always in demand, but many of the red inks on sale at
stationery and other stores are of a very inferior quality. Red ink
made from the following formula, as this man made it, gives
universal satisfaction in all cases where red ink is required:
Red anilin, 1 dram; dextrin, 1 dram. To use, dissolve the powder in
hot water.
These various ink powders are usually put up in packages of a
sufficient quantity to make a pint of ink, and this requires from a
teaspoonful to a tablespoonful of the powder.
Having no capital, the young man began with the direct selling
plan, canvassing from house to house and from store to store, and
selling to his acquaintances whenever possible. From the profits
these sales brought him, he was soon able to take up the trust
scheme, sending twenty-four packages of the powder, put up in
small envelopes, to boys and girls whose names he obtained in
various ways, offering them a premium of a watch, a camera, roller
skates, silver spoons, or other articles he could buy cheap in
quantities, when each one had sold and remitted for the twenty-four
packages.
Later he inserted 25-word ads. in various papers, and made a
large number of sales direct by mail from that source. Today he
owns the largest and best patronized stationery store in his town.
 HOW SEVEN BOYS EARNED MONEY
Seven boys, from 12 to 15 years of age, all pupils at the same city
school, and all close chums, adopted seven different ways of earning
a little money during vacation, and it is pleasing to know that all
seven succeeded. Here are the plans they followed, one boy to each
plan:

PLAN No. 476. CANCELED POSTAGE STAMPS

One boy went to the large business houses and collected all the
canceled stamps he could find on envelopes received through the
mails. Many of these were from foreign countries and brought good
prices when offered to dealers or boys making stamp collections,
while the domestic stamps he sold for 25 cents per thousand. During
the vacation period that year he made over $50.

PLAN No. 477. BOUGHT A PRINTING PRESS

Another boy induced his father to help him buy a small printing
press, and cards of various sizes. He then took orders for the
printing of these cards for other boys and for men needing the
cheaper grade, charging 75 cents per hundred and cleared up nearly
$40 above expenses, besides paying for his printing press.

PLAN No. 478. PARLOR MAGIC

The next boy with a taste for entertaining, and being clever at
sleight-of-hand tricks, bought a book on parlor magic, and gave
entertainments at his own home and the homes of other boys,
charging 10 cents admission. He performed these tricks so well that
everyone felt that he or she had received full value for the dime paid
at the door, and the youthful entertainer realized a net profit of
almost $60 during the three months of his summer vacation.

PLAN No. 479. DID SCROLL-SAW WORK

The fourth boy, being of a mechanical turn of mind, bought a
scroll-saw, with which he made a great variety of very pretty things,
and for these the neighbors were glad to pay good prices, especially
where he made any special design to order. He was very skilful in his
work, and was kept busy most of the time, so that his net earnings
during vacation were $37.

PLAN No. 480. A LEMONADE STAND

The fifth boy had a taste for merchandising, and set up a
lemonade stand in the front yard of his home, where many people
passed every day. He had various-sized glasses in which he put his
lemonade, properly made and tastefully displayed, and sold his
product at 1 to 5 cents a glass, according to size. He also had some
very good ice cream which he sold in small dishes at 2 to 5 cents a
dish. Children were his principal customers, but even at these low
prices, he made a good profit on his sales, and the business netted
him a little more than $30 altogether.

PLAN No. 481. DOING ODD JOBS

The sixth boy did odd jobs wherever he could find them, such as
carrying satchels or parcels from stores, or to and from trains,
pushing baby carriages in the parks, running errands for neighbors,
and anything else that came handy. He was always on the lookout
for work and was very seldom idle. His earnings were $23.75, and
he was very well satisfied with that.
 PLAN No. 482. COLLECTING OLD MAGAZINES FOR SALE
The seventh boy went from house to house, collecting all the old
magazines that people were willing to give away, and sold these to
dealers at a good price per pound, as anything made of paper was in
good demand. This boy was more successful and his earnings were
$70 during that three months of vacation.
 SUGGESTIONS FOR THE FOLLOWING PLANS
A few of the following plans, are mere outlines containing
suggestions which may be worked out in more detail by those who
wish to make use of them. New features may be added as they
suggest themselves to each person adopting one or more of the
plans as a means of making a living. In giving so many under one
heading, space will not permit a separate method for handling each
plan.
In order to determine the best selling plan, or the best method of
profitably handling any of the ways outlined, it would be well for a
person to read as many of the plans set forth as possible, and
become familiar with the various means employed by others to
obtain the best possible results.
Selling plans for produce named in this book are of various kinds,
and include personal solicitation by a house-to-house canvass, the
employment of agents to sell on a commission basis, placing the
article on sale with druggists and dealers, mail order, advertising in
suitable mediums, giving away of coupons to dealers, who in turn
give them to their customers; the trust plan, or sending a certain
number of articles or packages to children, to be sold by them at a
certain low price, and paying a premium either in merchandise or
cash; filling orders by parcel post; placing of general advertising
through a reputable advertising agency, that will not only help to
prepare the proper kind of advertising, but also be able to select the
best mediums for that particular product; selling of certain items of
information direct to the customer, telling him how to make practical
use of certain ideas of which he had no previous knowledge.
All the above selling plans are set forth in various parts of this
book, in connection with the statement of how certain plans were
successfully worked by individuals who adopted them as a livelihood,
and the testimony of these persons should prove a valuable guide to
others seeking similar results.
PLAN No. 483. CADET OFFICER FOR U. S. SEE PLAN No. 217
 PLAN No. 484. LITTLE “TINKERING” JOBS
Replenishing and replacing batteries for doorbells, mending
kitchen-ware, and replacing various articles about the house will
often give a very good income in a small place where experts from
large establishments are not within reach. Many an elderly man, who
could not do anything else, has made a comfortable living by doing
these little “odd jobs.”
PLAN No. 485. CARPENTER FOR U. S. SEE PLAN No. 217
PLAN No. 486. DESIGNER-LANDSCAPE. SEE PLAN No. 217
 PLAN No. 487. THE “HOKEY-POKEY” SUMMER SELLER
One of the most delicious confections, and one that scores the
largest number of sales during the summer season, is made as
follows:
One can condensed milk; 2 tablespoonfuls cornstarch; a little cold
milk. Put the remainder of the milk in a double boiler, and when hot
add the cornstarch. Cook five minutes, then add the condensed milk,
and set aside to cool; then add the vanilla, and freeze. Cut into
squares or sticks and pack closely in a wooden pail, and it will sell
readily for 5 or 10 cents a stick.
A splendid seller at fairs, picnics, parties, etc., and a popular
delicacy in the city at soft-drink stands and confectionery stores.
Yields an unusually large profit.
 PLAN No. 488. A SHOE POLISH IN POWDER FORM
Shoe polishes always sell, and it is only a question as to which is
the best one. The following is not excelled:
Take powdered gum arabic, 5 pounds; sugar, 11⁄4 pounds; analine
black, 3 ounces. Powder these and mix well. Then divide into ten
packets, each of which will produce a pint of polish, or into twenty
packets that will make a half-pint each, though more may be made
from, a packet, as it is rather thick, especially for kid or glaze
leathers. It can be used with either water or vinegar, or these
combined, in which to dissolve the powder. Apply with a brush, and
continue the friction until the superfluous fluid dries and the polish
appears.
To make this a tan polish, use 1 ounce of chrysodine, instead of
the analine black.
A fine polish and a good profit in this preparation.
PLAN No. 489. LETTER CARRIERS FOR U. S. SEE PLAN No.
217
 PLAN No. 490. METAL POLISHING BLOCKS
These are made of precipitated chalk, 2 pounds; powdered tripoli,
1
⁄2 pound; jewelers’ rouge, 1 ounce. Mix into a stiff paste, with 1
ounce of glycerine and a pint of water, previously mixed, and pour
on just enough of the liquid to work the powders to the consistency
of fresh dough. Then place in little wooden butter molds to shape
them and set aside to dry, then force out and fill again. The blocks
are used with a soft cloth and a few drops of water, which will give
metal articles a fine polish. You can sell all you can make of these,
and realize a profit on them that will surprise you.
PLAN No. 491. CEMENT WORKER FOR U. S. SEE PLAN No.
217
PLAN No. 492. CERAMICS FOR U. S. SEE PLAN No. 217
 PLAN No. 493. SOAP LEAVES FOR TRAVELERS’ USE
These are made by passing sheets of paper over rollers and
through a hot solution of liquid soap, then passing it over drying
cylinders, and cutting it into sheets of the desired size.
They are so convenient and cheap that travelers will buy them and
there is a good profit in making and selling them.
 PLAN No. 494. HAVING THE BUTTER YELLOW IN WINTER
Just a little secret, but it is worth a good deal to buttermakers and
housewives who pride themselves upon the color of their butter, and
will pay something to know just how to obtain it:
Just before you finish churning, put the yolk of one or more eggs
into the churn, and you’ll have just the color you desire—a rich
yellow.
 PLAN No. 495. REMOVING FOUL AIR FROM WELLS
To determine whether or not the air at the bottom of the well is
foul, place a lighted torch or lamp in a bucket and lower it into the
well. If it continues to burn when the bucket rests on the water, it is
safe to descend. If it is extinguished, the air is foul. To remove this,
lower a pail filled with burning straw, or by dropping two or three
quarts of freshly slaked lime down the well. But test with the light
again before descending.
Plenty of people who have wells would gladly pay a small sum to
have this information mailed to them.
 PLAN No. 496. A QUICK FATTENING FOR FOWLS
Fowls will quickly fatten if given a mixture of ground rice, well
scalded with milk, to which some coarse sugar has been added,
making it rather thick. Feed several times a day, but not too much at
a time.
An ad. in poultry journals, offering to tell how this is done, for 25
cents, should bring excellent results.
 PLAN No. 497. ARM AND BUST DEVELOPER
Regarding it as every woman’s duty to look her best at all times, a
young lady in Denver prepared a most effective arm and bust
developer from the following formula:
Lanolin, 2 ounces; cocoa butter, 2 ounces; olive oil, 2 ounces.
These she melted in a double boiler, and heat until cold, when it was
ready to put up in 2-ounce jars that sold for 40 cents each, and
proved so satisfactory that she received hundreds of orders each
month, through a few ads. judiciously placed, besides having a good
sale through drug stores.
The directions she gave were to first bathe the parts with hot
water, to open the pores, and then rub in the cream very thoroughly
at bedtime for a number of nights.
 PLAN No. 498. REMEDY FOR BRITTLE NAILS
Women who are annoyed by having brittle nails are always glad to
learn of some effective way to make and keep them soft. This
prompted a young lady in St. Paul to utilize the following formula:
White petroleum, 1 ounce; powdered castile soap, 1 dram; oil of
bergamot, a few drops. This softens the nails, cures hang-nails and
renders the cuticle around the nails soft and pliable, so that it can be
easily removed with a towel or orange stick.
One small ad. in a leading magazine brought a great many orders,
and by repeating the ad. in other periodicals, the young lady earned
$1,500 clear profit the first year.
 PLAN No. 499. BATH POWDER
The delights of the bath are greatly multiplied by adding a well
prepared bath powder, and one of the very best of these was put up
by this lady, as follows:
Borax, 10 ounces; tartaric acid, 10 ounces; starch, 5 ounces. Mix
the ingredients together, and perfume with lavender water. Two
teaspoonfuls of the powder to a tub of water will soften and perfume
the same making it at the same time more cleansing and delightful.
She put this powder up in 8-ounce paper boxes, and sold it for 25
cents a box. It proved a good seller all the year round and the
profits were exceptionally large. The drug stores carried it in stock,
as it assisted greatly in making other sales, owing to the demand for
it.
Welcome to Our Bookstore - The Ultimate Destination for Book Lovers
Are you passionate about books and eager to explore new worlds of
knowledge? At our website, we offer a vast collection of books that
cater to every interest and age group. From classic literature to
specialized publications, self-help books, and children’s stories, we
have it all! Each book is a gateway to new adventures, helping you
expand your knowledge and nourish your soul
Experience Convenient and Enjoyable Book Shopping Our website is more
than just an online bookstore—it’s a bridge connecting readers to the
timeless values of culture and wisdom. With a sleek and user-friendly
interface and a smart search system, you can find your favorite books
quickly and easily. Enjoy special promotions, fast home delivery, and
a seamless shopping experience that saves you time and enhances your
love for reading.
Let us accompany you on the journey of exploring knowledge and
personal growth!

ebookgate.com