COMPLEX NETWORKS

Diego Garlaschelli, Frank den Hollander, Aske Plaat
October 28, 2015

Preface
Transportation networks, communication networks and social networks form the
backbone of modern society. In recent years there has been a growing fascination
with the complex connectedness such networks provide. This connectedness
manifests itself in many ways: in the rapid growth of Internet and the WorldWide Web, in the ease with which global communication takes place, in the
speed at which news and information travel around the world, and in the fast
spread of an epidemic or a financial crisis. These phenomena involve networks,
incentives, and the aggregate behaviour of groups of people. They are based on
the links that connect people and their decisions, with global consequences.
A network can be viewed as a graph. The vertices in the graph represent the
nodes of the network, the edges connecting the vertices represent the links between the nodes. Neurons in the brain are connected by synapses, proteins
inside the cell by physical contacts, people within social groups by common interests, countries of the world by economic relationships and financial markets,
companies by trade, and computers in the Internet by cables transferring data.
Despite the very different origin and nature of these links, all systems share an
underlying networked structure representing their large-scale organisation. This
organisation in turn leads to self-organisation, and other emergent properties,
which can only be understood by analysing the overall architecture of the system
rather than its constituents elements alone.
The Science of Complex Networks constitutes a young and active area of research, inspired by the empirical study of real-world networks (either physical, chemical, biological, economic or social). Big Data are continuously being
recorded and stored into large data sets, from biological data resulting from
DNA sequencing and the investigation of protein interactions and function, via
financial data reporting the high-frequency behaviour of stock markets, to informatics data mapping the structure and dynamics of the Internet and the
World-Wide Web. Each such data set is the analogue of the outcome of a
large-scale experiment that is rather different from the experiments carried out
in a laboratory. We are therefore experiencing an unprecedented possibility of
analysing experimental data and using them to formulate and test theoretical
models of complex networks.
Most complex networks display non-trivial topological features, with patterns
of connection that are neither purely regular nor purely random. Such features
include a heavy tail in the empirical distribution of the number of edges incident to a vertex (“scale freeness”), insensitivity of this distribution to the size
of the network (“sparseness”), small distances between most vertices (“small
2

3

world”), likeliness that two neighbours of a vertex are also neighbours of each
other (“highly clustered”), positivity of the correlation coefficient between the
numbers of edges incident to two neighbouring vertices (“assortativity”), community structure and hierarchical structure. The challenge is to understand the
effect such features have on the “performance” of the network, via the study of
models that allow for computation, prediction and control.
The present document contains the notes for the course on Complex Networks
offered by the Departments of Mathematics, Physics and Computer Science of
the Faculty of Science at Leiden University, The Netherlands. This course is
intended for third-year bachelor students and first-year master students. Its aim
is to provide an introduction to the area, covering both theoretical principles
and practical applications from various different directions (see the course schedule and the table of contents below). Complex Networks is a multi-disciplinary
course: it exposes views on the area from mathematics, physics and computer
science, and is open to students from all programs in these three disciplines.
At the same time it assumes basic knowledge at the bachelor level in each of
these disciplines, including key concepts from calculus (differentiation, integration, limits), probability theory (probability distributions, random variables,
stochastic processes), statistical physics (ensembles, entropy), and computer
programming (C, Java or Python). The course is both challenging in terms of
panorama and rewarding in terms of insight.
In the course we highlight some of the many fruitful ways in which mathematics,
physics and computer science come together in the study of complex networks.
The course is divided into two parts:
(I) Theory of Networks. In Chapter 1 we provide a general introduction
to complex networks by reporting on some of the empirically observed
properties of real-world networks, highlighting the universal behaviour
observed across many of them. In Chapters 2–3 we introduce some of the
most important mathematical models of networks, from simple models
to more difficult models aimed at reproducing the empirical properties of
real-world networks. In Chapters 4–5 we offer an empirical characterization of real-world networks and a statistical-physics description of some
of the aforementioned models, with the aim of providing tools to identify
structural patterns in real-world networks. In Chapters 6–7, finally, we
review various key contributions of computer science, including algorithms
to generate random graphs and measure their properties, as well as the use
of visualization tools to gain insight into the structure of random graphs.
(II) Applications of Networks. We exploit the theoretical and methodological tools introduced in (I) to illustrate important applications in Percolation (Chapter 8), Epidemiology (Chapter 9), Pattern detection (Chapter 10), Self-organisation (Chapter 11), Network dynamics and Network
properties (Chapter 12) and Real Networks (Chapter 13). Much of (II)
deals with the interplay between structure and functionality of random
graphs, i.e., with the question how the topology of a network affects the
behaviour of a process taking place on it.
As a red thread through the course we use the so-called Configuration Model,
a random graph with a prescribed degree sequence. This allows us to link up
concepts and tools.

In this case. Diego Garlaschelli (Physics) Prof. The final grade is a weighted average of the grades for the homework (30%) and the written examination (70%). Frank den Hollander (Mathematics) Prof. Physics. 6 applications) and 6 exercise sessions. clearly listing all names of the students whose work it represents. It is allowed and may be helpful to form teams of up to three students to work on the homework exercises. Aske Plaat (Computer Science) . Dr. a single file per group should be submitted. Dr. At the end of the course there is a 3-hour open-book written examination.4 The course consists of 13 lectures (1 introduction. and Computer Science each. 6 theory. Each of lectures 2–13 contains two types of exercises: • Exercise: Routine exercise that supports the presentation in the text. Dr. preferably teams with a student from Mathematics. • Homework: In-depth exercise that needs to be handed in one week after the lecture.

adjacency lists Applications 5 . epidemic. sparseness.Course overview Chapter Teacher Topic Introduction 1 DG+FdH+AP real-world networks. examples. topological features. preferential attachment models 4 DG empirical topological properties of real-world networks 5 DG maximum-entropy. challenges Theory 2 FdH random graphs. invasion percolation. vulnerability 9 FdH contact process. degree distribution. Erd˝ os-R´enyi random graph 3 FdH configuration model. small world. visualization 7 AP implementing the Configuration Model 8 FdH ordinary percolation. network ensembles 6 AP implementation basics. adjacency matrix. spread of rumour 10 DG pattern detection in networks 11 DG self-organised networks 12 AP network dynamics and higher order properties 13 AP implementing real networks. scale freeness.

Contents
I

Theory of Networks

9

1 Real-world Networks
1.1 Complex networks . . . . . . . . . . . . . . . . . . . . . . .
1.2 Social Networks . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Acquaintance Networks . . . . . . . . . . . . . . . .
1.2.2 Collaboration Networks . . . . . . . . . . . . . . . .
1.2.3 World-Wide Web . . . . . . . . . . . . . . . . . . . .
1.3 Technological Networks . . . . . . . . . . . . . . . . . . . .
1.3.1 Internet . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.2 Transportation Networks . . . . . . . . . . . . . . .
1.3.3 Energy Networks . . . . . . . . . . . . . . . . . . . .
1.4 Economic Networks . . . . . . . . . . . . . . . . . . . . . . .
1.4.1 Financial Networks . . . . . . . . . . . . . . . . . . .
1.4.2 Shareholding Networks . . . . . . . . . . . . . . . . .
1.4.3 World Trade Web . . . . . . . . . . . . . . . . . . .
1.5 Biological Networks . . . . . . . . . . . . . . . . . . . . . .
1.5.1 Metabolic Networks . . . . . . . . . . . . . . . . . .
1.5.2 Protein Interaction Networks and Genetic Networks
1.5.3 Neural Networks and Vascular Networks . . . . . . .
1.5.4 Food Webs . . . . . . . . . . . . . . . . . . . . . . .
1.6 Still other types of networks . . . . . . . . . . . . . . . . . .
1.6.1 Semantic Networks . . . . . . . . . . . . . . . . . . .
1.6.2 Co-occurrence Networks . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

10
10
11
11
12
15
18
18
21
22
22
22
23
23
25
25
25
25
26
26
26
26

2 Random Graphs
2.1 Graphs, random graphs, four scaling features
2.2 Erd˝
os-R´enyi random graph . . . . . . . . . .
2.2.1 Percolation transition . . . . . . . . .
2.2.2 Scaling features . . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

31
31
34
35
38

3 Network Models
3.1 The configuration model . . . . . .
3.1.1 Motivation . . . . . . . . .
3.1.2 Construction . . . . . . . .
3.1.3 Graphical degree sequences
3.1.4 Percolation transition . . .
3.1.5 Scaling features . . . . . . .
3.2 Preferential attachment model . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

40
40
40
40
41
44
45
45

6

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

CONTENTS

3.2.1
3.2.2
3.2.3
3.2.4

7

Motivation . . . . .
Construction . . . .
Scaling features . . .
Dynamic robustness

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

45
46
49
50

4 Network Topology
4.1 Basic notions . . . . . . . . . .
4.2 Empirical topological properties
4.2.1 First-order properties .
4.2.2 Second-order properties
4.2.3 Third-order properties .
4.2.4 Global properties . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

53
53
56
57
61
66
70

5 Network Ensembles
5.1 Equiprobability in the Erd˝
os-R´enyi model . .
5.2 Implementations of the Configuration Model
5.2.1 Link stub reconnection . . . . . . . . .
5.2.2 The local rewiring algorithm . . . . .
5.2.3 The Chung-Lu model . . . . . . . . .
5.2.4 The Park-Newman model . . . . . . .
5.3 Maximum-entropy ensembles . . . . . . . . .
5.3.1 The Maximum Entropy Principle . . .
5.3.2 Simple undirected graphs . . . . . . .
5.3.3 Directed graphs . . . . . . . . . . . . .
5.3.4 Weighted graphs . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.

77
78
79
80
81
82
84
88
89
91
92
93

6 Random Graph Implementation
6.1 Random graph . . . . . . . . . . . .
6.1.1 Adjacency Matrix . . . . . .
6.1.2 Random Graph . . . . . . . .
6.1.3 Mean and Variance . . . . . .
6.1.4 Computing Graph Properties
6.2 Visualization . . . . . . . . . . . . .
6.2.1 Downloading Gephi . . . . .
6.2.2 Running Gephi . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

95
96
96
97
97
98
99
99
99

7 Configuration Model Implementation
7.1 Configuration Model Implementation .
7.1.1 Pre-specified degree sequence .
7.1.2 Visualization . . . . . . . . . .
7.2 Repeated Configuration Model . . . .
7.2.1 Self-Loops and Multi-Edges . .
7.2.2 Check Routines . . . . . . . . .
7.2.3 Repeated Configuration Model

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

102
102
102
103
104
104
104
105

II

Applications of Networks

106

8 Percolation
107
8.1 Ordinary percolation . . . . . . . . . . . . . . . . . . . . . . . . . 107

8

CONTENTS

8.2
8.3

Invasion percolation . . . . . . . . . . . . . . . . . . . . . . . . . 109
Vulnerability of the configuration model . . . . . . . . . . . . . . 112

9 Epidemiology
9.1 The contact process on infinite lattices . .
9.1.1 Construction . . . . . . . . . . . .
9.1.2 Shift-invariance and attractiveness
9.1.3 Convergence to equilibrium . . . .
9.1.4 Critical infection threshold . . . .
9.2 The contact process on large finite lattices
9.3 The contact process on random graphs . .
9.4 Spread of a rumour on random graphs . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

116
116
116
117
118
118
119
120
121

10 Pattern detection in networks
10.1 The maximum-likelihood principle . . . . . . . . . . .
10.1.1 Motivation . . . . . . . . . . . . . . . . . . . .
10.1.2 Generalities . . . . . . . . . . . . . . . . . . . .
10.1.3 Erd˝
os-R´enyi random graph . . . . . . . . . . .
10.1.4 More complicated models . . . . . . . . . . . .
10.2 Detecting structural patterns in networks . . . . . . .
10.2.1 Maximum likelihood in the configuration model
10.2.2 Directed graphs . . . . . . . . . . . . . . . . . .
10.2.3 General case . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

124
124
125
126
127
127
129
129
134
141

11 Self-Organized Networks
11.1 Introduction . . . . . . . . . . . . . . .
11.2 Scale invariance and self–organization
11.2.1 Geometric fractals . . . . . . .
11.2.2 Self–Organized Criticality . . .
11.3 The fitness model . . . . . . . . . . . .
11.3.1 Particular cases . . . . . . . . .
11.4 A self–organized network model . . . .
11.4.1 Motivation . . . . . . . . . . .
11.4.2 Definition . . . . . . . . . . . .
11.4.3 Analytical solution . . . . . . .
11.4.4 Particular cases . . . . . . . . .
11.5 Conclusions . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

145
145
146
146
147
151
152
153
154
155
155
159
163

12 Visualizing Dynamics and Higher Order Properties
12.1 Network Dynamics . . . . . . . . . . . . . . . . . . . .
12.1.1 Netlogo . . . . . . . . . . . . . . . . . . . . . .
12.1.2 Preferential Attachment . . . . . . . . . . . . .
12.1.3 Percolation Transition . . . . . . . . . . . . . .
12.2 Network Properties . . . . . . . . . . . . . . . . . . . .
12.2.1 Gnuplot . . . . . . . . . . . . . . . . . . . . . .
12.2.2 Empirical Network properties . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

170
170
170
173
174
175
175
176

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.

13 Analysing Real Networks
179
13.1 Adjacency Lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
13.2 SNAP real networks . . . . . . . . . . . . . . . . . . . . . . . . . 180

Part I Theory of Networks 9 .

For instance. More surprisingly. Examples include electrical power grids. which means that most nodes are separated from each other by relatively short chains of links. 10 . Facebook and Twitter. and what the qualitative and the quantitative aspects of these features are. a global description of their topology is impossible. See Barab´asi [4].. however. as well as collaboration and citation networks of scientists. the topology of social networks affects the spread of information and disease. Networks are modelled as graphs. i. Watts [39] and Newman. telephony networks. Watts and Barab´asi [32] for expository accounts on the discovery of network properties and the empirical measurement of these properties.Chapter 1 Real-world Networks 1. Complex networks plays an increasingly important role in science. which is why researchers have turned to local properties: How many vertices 1 In Chapters 2. while the topology of Internet affects its success as a means of communication. Internet and the World-Wide Web. 3 and 4 we provide rigorous definitions of the above structural properties. The structure of such networks affects their performance. A striking conclusion from this empirical work is that real-world networks share fascinating features.1 The above observations have had fundamental implications for scientific research on networks. a set of vertices connected by a set of edges. A common feature of real-world networks is that they are large and complex. Consequently.1 Complex networks The advent of the computer age has incited a mounting interest in the fundamental properties of real-world networks. i. Many are small worlds.e. they are scale-free. The aim of this research is to understand why networks share these features. which means that the empirical distribution of the degree (= number of links to other nodes) of the nodes is almost independent of the size of the network. this property was perhaps to be expected. In addition. many real-world networks appear to have power-law degree distributions. as well as extensive supporting empirical evidence. which means that the fraction of nodes with degree k is approximately proportional to k −τ for some τ > 1.. transportation and traffic networks. Due to the vast computational power that is presently available. Because networks tend to operate efficiently.e. large data sets can be easily stored and analysed. many networks are sparse. This has had a profound impact on the empirical study of large networks.

transportation. These models can be divided into two distinct types: (1) static.1 Social Networks Acquaintance Networks In 1967. Most models use random graphs as a way to model the uncertainty and the lack of regularity in real-world networks. Newman [30]. Most real-world networks can be classified into four broad classes: (I) Social Networks: WWW. The goal is to explain the universal behaviour exhibited by real-world networks. In Section 1. and this fact was dubbed with the phrase Six Degrees of Separation. In Sections 1. Kansas.6 we mention a few further examples that lie beyond. only 3 letters (roughly 5%) reached their destination. In this course we survey some of the proposals made for network models. Massachusetts. and by providing more clues about the recipient. USA. (II) Technological Networks: Internet.1. WhatsApp. interfirm. The exposition below borrows from Chapter 1 of the latter reference. in a short story called Chains. SOCIAL NETWORKS 11 does the network have? According to what rules are the vertices connected to one another by edges? What cluster sizes and cluster shapes are most common? What is the average distance between two vertices? What is the maximal distance between two vertices? These local properties are typically probabilistic.2 1.5 we describe examples drawn from each of these classes.2. The main conclusion from the work of Milgram was that most people are connected by a chain of at most 6 “friends of friends”. 95% by pretending that the value of the package was high. He sent 60 letters to various people in Wichita. respectively. Dorogovtsev and Mendes [18]. protein interaction. where the aim is to explain how networks came to be as they are. While 50 people responded to the challenge. Reviews on the subject can be found in Albert and Barab´asi [2]. 1. Twitter. input/output. psychologist Stanley Milgram performed the following experiment.2. Facebook. traffic. and van der Hofstad [22]. where new vertices are more likely to be attached to vertices that already have large degrees. The observation that many real-world networks share the properties mentioned above has incited a burst of activity in network modeling. neural. interbank. Later playwright . Dynamic explanations often focus on the growth of the network as a way to explain power-law degree distributions by means of preferential attachment growth. (IV) Biological Networks: metabolic. In later experiments.2–1. who were asked to forward the letter to a specific person in Cambridge. (2) dynamic. USA. The participants could only pass the letters (by hand) to personal acquaintances who they thought might be able to reach the target. where the aim is to describe networks and their topology at a given instant of time. either directly or via “friends of friends”. which leads to the study of random graphs. Milgram managed to increase the success rate to 35%. The idea of “close connectedness” was first proposed in 1929 by the Hungarian writer Frigyes Karinthy. (III) Economic Networks: trade. such as his/her occupation. power grids.

1. REAL-WORLD NETWORKS John Guare popularised the phrase when he chose it as the title for his 1990 play. Between us and everybody else on this planet. A gondolier in Venice . disease transmission. The research of Watts and the advent of the computer age have opened up new areas of inquiry related to “Six Degrees of Separation” in diverse areas of network theory. who was extremely prolific and wrote around 1500 papers with 511 collaborators. I am bound to everyone on this planet by a trail of six people. The Erd˝ os number of a mathematician counts how many papers that mathematician is away from the legendary mathematician Paul Erd˝os. we should make precise what the latter means.2. questionaires take time to fill out and to collect.12 CHAPTER 1. says: “Everybody on this planet is separated only by six other people.) An Eskimo. you can find . Of all the mathematicians who are connected to Erd˝os by a trail of collaborators. In this play. which is popularized under the name “Erd˝osnumber project”. (. and we draw an edge between two people when they “know each other”. or address each other on a first-name basis. It’s anyone. for instance. A native in the rain forest.. It’s not just the big names. Watts again found that the average number of intermediaries was 6. etc. The idea of Milgram was taken up afresh in 2001. the maximal Erd˝ os-number is claimed to be 15. the vertices are mathematicians. Questionaires cannot always be trusted. To put the idea of a small world into network language.. It is a profound thought. or social networks such as Hyves and Facebook (see Fig.. On the above website. Duncan Watts.” The fact that. Surprisingly. Possibilities are various: it could mean that the two people involved have shaken hands at some point. and computer circuitry. 1. Of course. In this graph.1). researchers are interested in examples of social networks that can be more easily measured.. with the added possibilities of the computer era. because they are electronic. Also. we define the vertices of the social graph to be the inhabitants of the world (n ≈ 7 × 109 ).000 senders and 19 targets in 157 different countries. Six degrees of separation. Ousa. or meet regularly. and there is an edge between two mathematicians when they have co-authored a paper. recreated Milgram’s experiment using an e-mail message as the “package” that needed to be delivered. people can be reached by a chain of at most 6 intermediaries is rather striking.2 Collaboration Networks An interesting example of a complex network that has drawn attention is the collaboration graph in mathematics. The president of the United States. on average. One of the main difficulties with social networks is that they are notoriously hard to measure. It implies that any two people in remote areas such as Greenland and the Amazone can be linked by a sequence of on average 6 intermediaries. As a result. because people have different ideas about what a certain social relation is. The precise choice affects the connectivity of the social graph and hence the conclusions we may draw about its topology. one of the main characters. such as electrical power grids. This makes the phrase It is a small world we live in! very appropriate indeed. a professor at Columbia University. In 2001. corporate communication. after reviewing the data collected by 48. Examples are e-mail networks.

See Fig.1: A map of the network of all friendships formed on Facebook across the world (from https://www.2 for an artistic impression of the collaboration graph in mathematics taken from http://www. The distribution of the Erd˝ os numbers is given in the following table (based on data collected in July 2004): Erd˝ os number number of mathematicians 0 1 2 3 4 5 6 7 8 9 10 11 12 13 1 504 6593 33605 83642 87760 40014 11591 146 819 244 68 23 4 The median is 5. We note that the Erd˝ os number is finite if and only if the corresponding mathematician is in the largest connected component of the collaboration graph.com/Erdos.com/zuck). 1.1.65.2.orgnet. SOCIAL NETWORKS 13 Figure 1. it is possible to find the distance between any two mathematicians worldwide. out how far your own professors are away from Erd˝os. and the standard deviation is 1.facebook.html . the mean is 4.21. Also.

In July 2004. They found that Nobel prize laureates have Erd˝ os-numbers of at most 8 with an average of 4–5. the collaboration graph consisted of about 1. with a total of about 401.4% 27. See http://www.000 edges. The percentage of papers with a given number authors is: number of authors 1 2 3 4 5 ≥6 percentage 62. while Fields medal winners have Erd˝ os-numbers of at most 5 with an average of 3–4. starting out above 90% in the 1940’s and currently standing at under 50%. so that the average number of collaborators per person is 3.7% 0. De Castro and Grossman [15.4% 0.”.000 different authors. 1. Figure 1.4% 8. REAL-WORLD NETWORKS and Fig.edu/enp . in which case the number of co-authors is not known precisely.3 for the degree distribution in the collaboration graph. The entire graph has about 676.oakland.1% The largest number of authors shown for a single item lies in the 20’s.9 million authored papers in the Mathematical Reviews database.2: An artist impression of the collaboration graph in mathematics.14 CHAPTER 1.36.0% 1. Sometimes the author list includes “et al. 16] investigated the Erd˝os-numbers of Nobel prize laureates and Fields medal winners. The fraction of items authored by just one person has steadily decreased over time.

who has 511 co-authors. and the reviews cited at the end of Section 1. the clustering coefficient describes how often two neighbors of a vertex are adjacent to each other.1. the average number of collaborators for people who have collaborated is 4. In other words. Raghavan.1. Rajagopalan. The largest degree is for Erd˝ os. Broder. b. In the collaboration graph. The relatively high value of this number. together with the fact that average path lengths are small. The clustering coefficient of a graph is equal to the number of ordered triples of vertices a. and its inventors were also the founders of Google (see Brin . Rajagopalan and Tomkins [25]. SOCIAL NETWORKS 15 Number of vertices with given degree 1000000 100000 10000 1000 Series1 100 10 1 1 10 100 1000 Degree Figure 1.25. Maghoul. While Internet is physical.3: The degree sequence in the collaboration graph. With the rapid growth of the WWW. bc and ac are present. The Page-Rank algorithm is claimed to be the main reason behind the success of Google. Notorious is the Page-Rank problem: to rank web pages in such a way that the most important pages come up first. It is of great practical importance to know what the structure of the WWW is. to allow search engines to explore it efficiently. the edges are hyperlinks (or URLs) pointing from one web page to another. c in which the edges ab. since hyperlinks are not necessarily reciprocated.2. The properties of the WWW have been studied by a number of authors: see e. Jeong and Barab´ asi [3]. for example.14.g. Stata. the WWW is virtual.2. the interest in its properties is growing as well. Raghavan. indicates that the collaboration graph is a small-world graph.3 World-Wide Web The vertices of the WWW are electronic web pages. There are only 5 mathematicians with degree at least 200. The clustering coefficient of the collaboration graph is 1308045/9125801 = 0. 1. divided by the number in which ab and bc are present. Kleinberg. Kumar. Albert. Tomkins and Wiener [7]. Kumar. The WWW is therefore a directed network.

16 CHAPTER 1. Raghavan. They divide the WWW into four parts (see Fig. They found that the in-degrees obey a power-law distribution with exponent τin ≈ 2. Figure 1. the surprisingly small diameter of the web means that all information is just a few clicks away. Raghavan. 1.edu and whitehouse. Furthermore. Kumar. and found it to grow linearly with the logarithm of the size n of the domain. Albert.5. Maghoul. they investigated the average distance d between the vertices in these domains. 1. Their analysis was based on several Web domains.edu.1. The most extensive analysis of the WWW was performed by Broder.35 + 2. which prompted them to the following quote: “Fortunately. Extrapolating this relation to the estimated size of the WWW at the time (n = 8 × 108 ). they concluded that the diameter of the WWW was 19. while the out-degrees obey a power-law distribution with exponent τout ≈ 2. Massachusetts Institute of Technology and the White House). with an estimated dependence of the form d = 0.4).” Kumar. REAL-WORLD NETWORKS and Page [6] for the original reference). mit.4: The in-degree sequence in the WWW.06 log n.gov (the Web domains of Notre Dame University. Rajagopalan. Stata. such as nd. Tomkins and Wiener [7]. Jeong and Barab´asi [3] studied the degree distribution of the WWW.5): . Rajagopalan and Tomkins [27] were the first to observe that the WWW has a power-law degree distribution (see Fig.

(a) The Strongly Connected Component (SCC) – the central core consisting of those pages that can reach each other along the directed links (28% of the pages). the distances between most pairs of vertices within the SCC are at most 7. consisting of pages that can be reached from the SCC. The relatively high values of these numbers are due in part to the fact that the graph for the WWW is directed. 1. (b) The IN part. (d) The TENDRILS and other components. Furthermore.1. consisting of pages that can reach the SCC. When the WWW is considered as a directed graph. it was found that both the in-degrees and the out-degrees in the WWW follow a power-law distribution.5. with exponents τin ≈ 2. nor be reached from it (30% of the pages). but do not link back to it (21% of the pages). in accordance with the rough findings obtained earlier. When the WWW is considered as an undirected graph. but cannot be reached from it (21% of the pages).6 for a histogram of pairwise distances in the sample. See Fig. while the WWW as a whole has diameter at least 500. (c) The OUT part.5: The structure of the WWW.1 and τout ≈ 2. It was found that the SCC has diameter at least 28. consisting of pages that can neither reach the SCC. SOCIAL NETWORKS 17 Tendrils 44 Million nodes IN SCC OUT 44 Million nodes 56 Million nodes 44 Million nodes Tubes Disconnected components Figure 1. . similar to the “Six Degrees of Separation” found in social networks. the average distance between vertices decreases to around 7.2.

even though the Internet itself is rather large. connected by cables transferring data. Shenker and Willinger [10]. organisations. which is the number of routers traversed by an e-mail message between two uniformly chosen routers. treating as vertices whole groups of computers. It is an undirected network: information can travel both ways along the cables. but between which there are large-scale stable connections. Therefore the network is usually studied at a coarse-grained level. 1. These groups of computers are called autonomous systems. Fig. within which rearrangements may occur frequently due to local handling. Marchetti and Pietronero [9]. Govindan. e-mail service would break down. and the book by Pastor-Satorras and Vespignani [35]. Chen.e. Jamin.18 CHAPTER 1. 1. A snapshot of Internet is portrayed in Fig. Chang. which approximately correspond to domain names. and the AS-count (= the number of Autonomous Systems that are traversed by an e-mail data set). log k 7→ log Nk .1 Technological Networks Internet Internet is a physical network of computers. If distances in the Internet would be larger than this threshold. Caldarelli.. due to local rearrangements of networked computers within.g. The “fine structure” of Internet changes continuously. 1.3.9 plots the degree distribution on a log-log scale. PastorSatorras.3 1. In Internet.8 depicts the hopcount. Fig. Consequently.7. Vazquez and Vespignani [34]. IP-packets cannot use more than a certain threshold of physical links. for instance. i. REAL-WORLD NETWORKS Figure 1.6: Average distances in the Strongly Connected Component of the WWW (Adamic [1]). The properties of Internet have been studied in many references: see e. 1. which is typically bounded by 7. the graph of the Internet has evolved in such a way that typical distances are relatively small.

0 1 2 3 4 5 6 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 1 2 3 4 5 6 7 8 9 10 11 12 13 Figure 1.2D.png).1 0.7: Portrait of a particular snapshot of Internet (from http://www.watblog. .3. Data courtesy Hongsuda Tangmunarunkit.00 0. 0.1.08 0.04 0. TECHNOLOGICAL NETWORKS 19 Figure 1.4 0.12 0. com/wp-content/uploads/2013/01/1069524880.3 0.2048x2048.2 0.8: Internet hopcount data and number of AS traversed in hopcount data.

Û/Ñ Ã Ï ÚSckÚ Ñ-−τ Ñ Ã Õ$Ñ(Í Âà Ï$Í Ã Í Ö ÂfÈ Ñ-Õ$Í ÂH×dà Õ$Ñ Â Ï$ê È Ñ Ð=ÖÜ.9: à ÖÜÜ Ð Ñ@ÝÏ$Ñ Â$in Þ ßJÌà@the Ú ç!Ï ÓÚ:Ôà Õ$Ñ%7Ö Ö Ï Ï Ã511/1997 ÌÍ Î Ï Ð Degree Ñ=ÒÓÔÕ$Ñ7Ö distributions Ï Ã5× Ñ-ÎÐ Ñ-Ñ7Ø$Ù Ö Ã Ú ÓÛ/Ö of Î Ä Ù Ö AS ÎXØ$Ù Ö domains Ã5× × Ñ-Ñ-ÎÎÐ Ð Ñ-Ñ-Ñ7Ñ7ãHØ$ä Ù Ö Ã and Ú ÓÛ/Ö Î 12/1998 Ä Ù Ö ÎXØ$Ù Ö Ã ÖÜÜ Ð Ñ@ÝÏ$Ñ Â$Þ ßJà@á=â Ñ Ð Ú Ï Ú:à Õ$Ñ%Ö Ï Ã5× Ñ-ÎÐ Í á=Î Ïâ Ð Ñ ÑÐ months on a log-log scale (Faloutos [20]): Power-law distribution with exponent ≈ 2.15632 ) 10000 "980410.68585) * x ** ( -2.52124) * x 1000 1000 1000 1000 100 100 100 100 10 10 10 10 1 1 ¾¿ÀÁÂÃÄ Å 10Å ÄÆÇ 100 11 11 ¾¾È[¿ÀÀÁÁÂÂÃÃÄÄ É10 ÅDÊ10å@ÄÄÆ Æ Ë Ë 100 100 1 1 ¾È[Àæ Ö Ï 10à ÄÆ Ò Figure 1.20288)) exp(8.15 − 2..out" exp(7.16356 -2.e.1000 1000 100 100 10 10 1 1 100 1 ¾ ¿ÀÁÂÃÄ Å 10Å ÄÆÇ ¾È[ÀÁÂÃÄÉ10ÊÄÆ Ë CHAPTER 1. i. ÚUÃ Ï ×!ß Ã Õ$Ñ Ú Í è-Ñ_ÖÜ Ã Õ$Ñ Â Ñ-Í Î Õ È ÖÐ Õ$ÖÖ ×¨é Í Ã Õ$Í ÂcÚ Ö êXÑ × Í ÚUÃ5¿Â$Þ Ñ ë ÜeÖÐ>Í ÂÃ Ñ Ð Ä× Ö ê ¿ Í Â ÎÐ ¿ Ø$Õ Ú>¿ÂH×cÉ!þ Æ ç! à Õ$Ñ × Í ÚUÃ5¿Â$Þ Ñ:Í Ã Ú Ñ-Ù ÜäZì ¿ êXÑ-Ù ß ë é ÑWÏ Ú Ñ Ã Õ$Ñ Ã Ö Ã5¿ Ù Â Ï$ê Ä Ú Ñ-Ñ(Í Âfÿ Ø$Ø[Ñ ÂH× Í õ ä  ÜeÖÐ Ã Ï ÂH¿Dà ÑÍ Â$ÚUÃ Ñ ¿ × ÖÜ 10000 È Ñ ÐÖÜ:Ø ¿ Í Ð Ú ÖÜ Â Ö × Ñ Ú=íO¾î[ÀWé exp(8.11393) **xx****(( -2. When N is proportional to an Ã Ö î Õ$Ö Ø Ú ë[1000 Ú Ñ â Ñ Ð ¿ Ù/Ï Ú Ñ ÜeÏ$Ù ¿ Ø$Ø$Ù Í Þ@¿DÃ Í Ö Â$Ú.out" "981205.Ø ¿ Í Ð Ú 1000 Ãé Í Þ Ñ ä inverse power of k.52124) Ú ê ¿ Ù Ù Â Ï$ê È Ñ Ð Ã Ö â Ñ Ð Í Ü ß ÖÐ × Í Ú ØÐ Ö â Ñ Í Ã Õ$Í Â î Õ$Ö *Ø xÚ **ë é ("routes.out" Õ$-2.89793) exp(8.¿Ú:é Ñ Ú Í Â$Þ Ù Ï × Í Â Î Ú Ñ-Ù Ü Ä Ø ¿ Í Ð Ú ë ¿kÂH×_Þ Ö Ï ÂÃ Í Â Î ¿ Ù Ù3Ö Ã Õ$Ñ Ð.Ø ¿ Í Ð Ú ÖÜ   .Ö é Ñ â Ñ Ð@ë/Ñ â Ñ Â_à Õ$Í Ú with Nk the number of vertices of degree k.20.out" exp(7.48626 Í Þ Õ é Ñ ) × Ñ ï Â Ñ  ¿ = Ú Ã Ã 5 à ¿ Â È ¿ Ú Â × = Ú é à   Ú Ú ¿ $ Õ Ñ Ö Ù $ Ï ê Ñ  Ð  Ö 1 Ü Ø Í Ð  Ö Ü Ö Ñ Í $ Õ Í Ù Ñ  Ö O Ð @ Ñ  Ý Ï Ù Ø[Ñ Ð Í êXÑ ÂÃ5¿ Ù Ù ß ä. REAL-WORLD NETWORKS ÌÍ Î ÏÐ Ñ=ÒÓÔÕ$Ñ7Ö Ï Ã5× Ñ-ÎÐ Ñ-Ñ7Ø$Ù Ö Ã Ú ÓÛ/Ö Î Ä Ù Ö ÎXØ$Ù Ö Ã ÖÜÜ Ð Ñ@ÝÏ$Ñ Â$Þ ßJà@á=â Ñ Ð Ú Ï Ú:à Õ$Ñ%Ö Ï Ã5× Ñ-ÎÐ 20 10000 1 10000 10000 "971108.

4× Ñ Ú=é Í Ã Õ$Í Â Ù Ñ ÚÚ ÖÐOÑ@ÝÏ ¿ Ù Ø[Ñ Ð Í êXÑ ÂÃ5¿ Ù Ù ß  ä .15 Ö Ï Ã5× Ñ-Î− Ð Ñ-Ñ72.ï$Ð ÚUÃWÜeÊ ÖÐ> Ä× Ã Ö Õ$ê Ñ%¿ Í ÂÍ Â Ã Ñ ÎÐ ÄÐ × ¿ Ö Ø$ê Õ Ú>¿ Í ¿ ÂHÎ×cÐ ¿ É!Ø$þ Æ Õ ç!Ú ë:¿ÜeÖÂHÐ ×(à à Õ$Õ$Ñ Ñ.Í ÂKÿ Ø$Ø[Ñ ÂH× Õ$Ö Í ÂØ Ã Ú Ñ ÍÐ Â> ÚUÃ Ï ×!ß Ã Õ$Ñ Ú Í è-Ñ_ÖÜ Ã Õ$Ñ Â Ñ-Í Î Õ È ÖÐ Õ$ÖÖ ×¨é Í Ã Õ$Í ÂcÚ Ö êXÑ × Í ÚUÃ5¿Â$Þ Ã Ñ Õ$ë Ñ.Ø ¿ Í Ð Ú ÖÜ   . î ôëxÜeÖÐ î\û ó!ä7ü©ÑSÑ õ$¿ êXÍ Â Ñ é Õ$Ñ Ã Õ$Ñ Ð(1.K(BC0C.H  Ï$ê È Ñ Ðand Ä Ù Ö Î ÚÞ@¿ Ù Ñ äÔÕ$this ÜeÏ Â$Þ Ãlog-log Í Ö Â ÖÜ Ã Õ$Ñ plot. Ø$Ù Ö Ã Ú ÓÛ/Naturally.    !"$#  ÂN íO¾î[À 100 Ö × kÑ Ú. îaû ó!ë Ú Õ$Ö éWÂd¿Ú%¿(Ú Ö Ù Í × Ù Í Â ÑSÍ Â_à Õ$ÑØ$Ù Ö Ã 5Ú ä%ü©Ñ ¿ Ø$ØÐ Ö õ Í ê ¿DÃ Ñ Ù Ö Î Í Þ@¿ Ù/ÖÐ × Ñ Ð@ë ¿Úé Ñ Ú Ñ-Ñ.Í Â>à Õ$Í ÚWÚ Ñ Þ Ã Í Ö Â ä Õ$Ö Ø Ú ë[pdwfkhv Í Â$Þ Ù Ï × Í Â Î uhdolw| vlprqlrxv lq lwv sdudphwhuv dv srvvleoh Ãé Í Þ Ñ ä dv forvh dv srvvleoh1 Wr uvw rughu/ wkhÛ/Ñ sorwv Ã Ï ÚSÚ lq Ñ-Ñ ÃFkhq Õ$Ñ(Í Âà Ï$hwÍ Ã Í do1 Ö ÂfÈ Ñ-Õ$Í ÂH×dà Õ$Ñ Â Ï$ê È Ñ Ð=ÖÜ.=/=>N for some constant c and some exponent à Õ$Ñ7Ö Ã Õ$Ñ Ð./_79&:< Ø[Ö é Ñ Ð Ä Ù ¿-é ä 1 1 Here.20.=/=>O1 1 ≈ denotes an10 uncontrolled approximation. The τ can beÚ¿ esti.Ö é Ñ â Ñ Ð@ë/Ñ â Ñ Â_à Õ$Í Ú Ð Ö Ï$Î Õ ¿ Ø$ØÐ Ö õ Í ê ¿DÃ Í Ö Â Õ ¿Ú prgholqj frqvlvwv lq sursrvlqj d prgho Ã Ö î wkdw Ú Ñ-Ù Ü Ä Ø ¿ Í Ð Ú ë ¿ÂH×_Þ Ö Ï ÂÃ Í Â Î ¿ Ù Ù3Ö Ã Õ$Ñ Ð.Í ÂÃ Ñ Ð Ä× ÌÍ Î ÏÐ Ñç!τÓÔ≈Õ$Ñ72.HBC7 î .2) Í × ë log Ñ íOτ¾î[log à Õ$Ñ Â Ï$ê È Ñ Ð. 100 Á ïHÎ ÏÐ Ñ Ú:Ç7 1 ¿ÂH×XË ë é Ñ1 Ø$Ù Ö ¾ÃÈ[à ÀÕ$exponent ¾¿ÀÁÂÃÄ ÅDå@ÄÆ Ë æ Ñ Ö Â Ï Ï$10Ãê ÄÆ È Ò Ñ ÐBÖÜQØ ¿ Í Ð ÚíO¾î[ÀB¿100 î–û KS/eGH/Cf(7)/g0C.Ø ¿ Í Ð Ú Ú Ñ â Ñ Ð ¿ Ù/Ï Ú Ñ ÜeÏ$Ù ¿ Ø$Ø$Ù Í Þ@¿DÃ Í Ö Â$Ú.Ø[Ö ÚÚ Í È Ù Ñ=Ø ¿ Í Ð Ú Ó íO¾ ó ÀZð ò_ôë é Õ$Í Þ Õ_Í Ú Ø ¿ Í Ð Ú Ø$Ù Ï Ú1τ¿ .≈ äZÌ$ÖÐ î>ðñÉ ë é Ñ7Ö Â Ù ß Õ ¿@â Ñ Ã Õ$Ñ Ú Ñ-Ù Ü Ä Ø ¿ Í Ð Ú Ó íO¾ÉÀZð (1.ÖÜØ ¿ Í Ð Ú7íO¾î[À ÜeÖÐ Ã Õ$Ñ ÁÂÃ Ñ Ð Â Ñ Ã ÜeÖ Ù Ù Ö éWÚ=¿(Ú Í êXÍ Ù ¿ Ð Z %H[(   ]\^ /=09&D)@6 1+00C./FD=6 1=@/e1?0C. ÖÜBÕ$Ö for Ø ÚWî the Í Â Ù Ö Î AS-data Ñ ×$¿DÃ5¿givesÑ õ Ø[Ö Â ó Ñ I mated by the slope of the line in the Âà I Uc ÚÍ Ð Ñ-ØÐ Ñ Ú Ñ ÂÃ Ñ ×aÈßa× Í ¿ êXÖ ÂH×Ú ë ¿ÂH×_à Õ$Ñ × Ö P ÃÃ Ñ × Õ$ÖÐ Í è-Ö ÂÃ5¿ ÙBÙ Í Â Ñ the estimate ¿DÞ õ ßJ  РÏ$Ú Ï ê Ú:È Ã Ñ Õ$Ðthat Õ$Ñ@ÝÑ%Ï$ê Ñ Â$must Í êSà@Ï$á=ê â Ñ have Õ$ãHÍ Þä ÕKN Í Ú òk ô = ä3ü©n./79&:<./R79&:<.=/=>1?A@5+BC>=DE1?F71GH/=DJI 100 ò\äZÌ$ÖÐ Ã Õ$Ñ × Í ¿ êXÑ Ã Ñ ÐÖÜ Ã Õ$Ñ1ÎÐ ¿ Ø$Õ(ó!ë î>ð ó!ë é Ñ1Õ ¿@â Ñ Ã Õ$Ñ Ú Ñ-Ù Ü Ä BCDL@ >"1=@1+>=0BM1+7546021<0C.1) 02546879&:<.æï$Ð Ö ÚUÏ Ã7ÃÅDÄÆ å Òë ¿ÚXé Õ$Ö Ñ Ø Ä Ø$Ù Ö Ã Ñ õ Ø[Ö Â Ñ ÂÃ× Ñ ÚÞ Ð Í È Ñ Ú¿ÂX¿Ú Ö Ï ÿ ÃÄØ$Æ Ø[ÒÑ äZÂHÔ× Í Õ$õÑ Þ ÖäÐÐ  Ñ-Ù Â ¿DÜeà ÖÍ ÐÖ Ã Â_Ï ÂHÞ Ö¿DÑ Ã Ñ-ý Ù ß Þ ëQÍ Ñ ÜeÂÖ Ã ÏÚ1ÐS¿ Ø[Ð Ñ% Í Â$ÚUÃ Ñ ¿ × ÖÜ Ã Õ$Ñ × Í ÚUÃ5¿Â$Þ Ñ:Í Ã Ú Ñ-Ù ÜäZì ¿ êXÑ-Ù ß ë é ÑWÏ Ú Ñ Ã Õ$Ñ Ã Ö Ã5¿ Ù Â Ï$ê Õ$Ö Ä Ø Ú Í Â>Ã Ú Õ$Ñ-Ñ (æ Í Âf Ö Í Í ÂÚWà ÚÉ!þ Æ Í ÚSË ¿ Ð ¿Dà Õ$Ã Ñ Õ$Ð Ñ%ÎÐ ¿ Ø$ÕKÍ ÂK¿=Ú Í Â Î Ù Ñ Â Ï$ê È Ñ Ð@äBÔÕ$ dqdo|vlv ri wkh ghjuhh judsk prghoÈ Ñ +vhf1 Â Ö × Ñ Ú=duw íO¾î[riÀWé Í Ã Õ$Í Â î Õ$Ö Ø Ú ë é Õ$Í Þ Õ é Ñ × Ñ ï Â Ñ Ú ê ¿ Ù Ù Â Ï$ê È Ñ Ð Ã Ö â Ñ Ð Í Ü ß ÖÐ × Í Ú ØÐ Ö â Ñ ¿ Ù Í Â Ñ ¿ Ð Í Ãß Õ ß Ø[Ö Ã Õ$Ñ Ú Í Ú Ñ õ!Ä ÐÖÜ:Ø ¿ LLL.Ù Ù then 5TJ1+7DJ025+70I$UV à Õ$Ñê ¿Dõ Í êSÏ$êöØ[Ö ÚÚ Í È Ù Ñ Â Ï$ê È Ñ Ð:ÖÜZØ ¿ Í Ð Ú äWÌ$ÖÐ ¿ Õ ß Ø[Ö Ã Õ$Ñ Ã Í Þ@¿ Ù 10 Ð Í Â Î Ã Ö Ø[Ö Ù Ö Î 10ß ë é Ñ7Õ ¿@â Ñ íO¾î[ÀZ÷øî/ù ë ¿ÂH× ë$ÜeÖÐ ¿(å@Ä× Í êXÑ Â$Ú Í Ö ÂH¿ Ù íO¾î[ÀB÷ñîWYX log NkÎÐ ≈ é ÑSÕ c¿@â − À1÷úk.¿Ú:é Ñ Ú Õ$Ö é Ù ¿DÃ Ñ Ð.1=@ DR`4/=>=DJ9&DR0C. Ñ so i È$Ú Ñ Ð â Ñ Ã Õ ¿DÃà Õ$Ñ Ã ÕÐ Ñ-Ñ. Ö Î Ä Ù Ö Ð Ñ-ÎXØØ$Ð Ñ Ù Ö Ú Ñà ÂÖà ÜÚ:Ü Ã Ð we Ñ%ÖÖ ÜBÏ Ø Ã5¿ × Í Ñ-Ð ÎÚ Ð ë Ñ-é Ñ7k∈N é:¿Âà ÚÞ Ð Í È Ñ Ã Õ$ÑOØ$Ù Ö Ã=Èß\¿ Ù Í Â ÑXÍ Â Ù Ñ ¿ÚUÃÄ0Ú ÝÏ ¿ Ð Ñ Ú ï à ëQÜeÖÐ Ã Í Þ@¿ Ù Ù ß Ñ@ÝÏ ¿ Ù Õ$Ö Ø Ä Ø$Ù Ö Ã Ñ õ Ø[Ö Â Ñ Âà kÚ jÊ that it is reasonable to assume that τ à Ö>× Ñ 1.1 Í Ð Ú ÖÜ Wkh ¿Ú=à Õ$Ñ dvÃ Ö Ã5vlpsoh ¿ Ù Â Ï$ê È dqg Ñ ÐÖÜ1sdu0 Ø ¿ Í Ð Ú ÖÜ Â Ö0.

13 2. between AS’s in ehkdyhv glhuhqwo|1 srwhqwldo wr pdwfk wkh uvw rughu fkdudfwhulvwlfv +H ^kQ ` wkhvh revhuydwlrqv/ glhuhqw are Lq quite Lqvsluhg close. iru prgholqj ï$Ð ÚUÃWDV0krsfrxqw Õ$Ö Ø Ä Ø$Ù Ö Ã Ñ õ Ø[Ö Â Ñ ÂÃ× Ñ ÚÞ Ð Í È Ñ Ú¿ÂX¿Ú Ø[Ñ Þ Ã ÖÜ Ã Õ$Ñ Þ Ö Â$Â Ñ Þ Ã Í â Í Ãß ÖÜ Ú >  à æ  à  Ä Æ Þ D ¿ à _ Â Þ Þ   à 1 Ú ¿ W Ú ! É þ Æ Ë $ Õ Ö Ø Í $ Õ Ñ Ö Ï  Ò Z ä  Ô $ Õ Ñ  Ö  Ð Ð Ñ Ù Í Ö  Ö Ñ ý Í Ñ Ð % Ñ Í Figure+vhf1 1.=/=>O1?_.06 0. and hints dqg ydu ^kQ `.=/=>1?A@5+BC>=DE1?F71GH/=DJI íO¾î[À IAK(BC0C.Í ÂÃ Ñ Ð Ä× Ö ê ¿ Í Â>×$¿DÃ5¿Ú Ñ Ã Ú Õ ¿@â Ñ.04 1.Í ÂKÿ Ø$Ø[Ñ ÂH× Í < à Õ$Ñ. respectively.ï$Ð ÚUÃ7ÅDå ghjuhh judsk +vhf1 LLL./ Õ$Ö Ø Ä Ø$Ù Ö Ã rughu uroh1 Khqfh/ zh eholhyh wkhuh Üelv Þ Ã Í Ö Â ydoxh Ï Â$vwloo ÖÜ Ã Õ$Ñ lq Ï$vwxg|lqj ê È Ñ ÐÖÜBÕ$Ö Ø ÚWî Í Â Ù Ö Î Ä Ù Ö Î ÚÞ@¿ Ù Ñ äÔÕ$Ñ ×$¿DÃ5¿ Ñ õ Ø[Ö Â Ñ Âà I Uc Í Ú Ð Ñ-ØÐ Ñ Ú Ñ ÂÃ Ñ ×aÈßa× Í ¿ êXÖ ÂH×Ú ë ¿ÂH×_à Õ$Ñ × Ö ÃÃ Ñ × Õ$ÖÐ Í è-Ö ÂÃ5¿ ÙBÙ Í Â Ñ wkh ghjuhh judsk1 Ú   à : Ú Ã D ¿ õ Â È ¿ Ú é Þ Ú Ð Ñ  Ø Ð Ñ Ñ $ Õ % Ñ ê Í S ê $ Ï ê $ Ï ê Ñ  Ð  Ö B Ü Ø Í Ð ë $ Õ Í K Õ Í ò 3 ä © ü Ñ ô i È$Ú Ñ Ð â Ñ Ã Õ ¿DÃà Õ$Ñ Ã ÕÐ Ñ-Ñ.98 2.20288 ) Pr[h AS = k] RIPE AMSIX LINX S E[hAs ] Var[hAs ] alfa 2.97 168398 # points .10./R@1+KS/=>N1? wulexwlrq/ dowkrxjk qrw d shuihfw rqh1 Ø ¿ Í Ð Ú Ø$Ù Ï Ú1¿ Ù Ù rià Õ$srzhu0olnh Ñ7Ö Ã Õ$Ñ Ð.=/F0C.1Ï$ê È Ñ ÐBÖÜQØ ¿ Í Ð ÚíO¾î[ÀB¿Ú¿ î–û ó IKS/eGH/Cf(7)/g0C.2¿ÂH× ë$ÜeÖÐ ¿(å@Ä× Í êXÑ Â$Ú Í Ö ÂH¿ Ù íO¾î[ÀB÷ñîWYX î>û ó é ÑSÕ ¿@â Ñ ghjuhh1 íO¾î[À1÷úî Wkh dwlrqv ri wkh srzhu0odz lq uhjlphvÎÐ riÍ × ë vpdoohu ôëxÜeÖÐ î\û ó!ä7ü©ÑSÑ õ$¿ êXÍ Â Ñ é Õ$Ñ Ã Õ$Ñ Ð Ã Â È ¿ 7 Ú O í  ¾ [ î À à  Á   à  à W é = Ú ( ¿ Ú ¿ $ Õ Ñ $ Ï ê Ñ . ri wkh DV dqg LS krsfrxqw/Europe uhvshfwlyho|1 zloo eh glvfxvvhg= wkh ghjuhh judsk lq vhf1 LLL dv d the dependence at uhvxowv the fact thatprghov the AS-graph is homogenous.81 3.1=@ DRBC76 1Ja4326 1JaDbTJ546 /c_d1+> yldwlrqv iurp dq h{dfw sro|qrpldoØ[odz Á ïHÎ ÏÐ Ñ Ú: Ç7¿ÂH×XË ë é Ñ1Ø$Ù Ö Ãà Õ$Ñ Â 0.HBC7 î .91 1./_79&:<. In other words. Ð  Ö  Ü Ø Í Ð e Ü  Ö Ð $ Õ Ñ Ñ Ð Ñ e Ü Ö Ù Ù Ö Í X ê Í Ù Ð ehdxw| ri dq dv|pswrwlf dqdo|vlv lv wkdw wkhvh vpdoo gh0 Z %H[(   ]\^ /=09&D)@6 1+00C.Ø[Ö ÚÚ Í È Ù Ñ=Ø ¿ Í Ð Ú Ó íO¾ ó ÀZð ò_ôë é Õ$Í Þ Õ_Í Ú 5TJ1+7DJ025+70I$UV frqvlghuhg khuh wkh prvw jhqhudo ghvfulswlrq à Õ$Ñê ¿Dõ Í êSÏ$êöØ[Ö ÚÚ Í È Ù Ñ Â Ï$ê È Ñ Ð:ÖÜZØ ¿ Í Ð Ú äWÌ$ÖÐ ¿ Õ ß Ø[Ö Ã Õ$Ñ Ã Í Þ@¿ Ù glvwulexwlrq ixqfwlrqv/ vshflhg lqÐ Í +5./79&:<./FD=6 1=@/e1?0C.=/=>R1?)@5+BC>=D_1?L71GH/=DJI íO¾î[À I Ö é Ñ Ð Ärqo| Ù ¿-é ä sod| d vhfrqg K(BC0C. surylghv lqiru0 frqvwuxfw iurp d jlyhq ghjuhh vhtxhqfh d frqqhfwhg judsk pdwlrq derxw EJS urxwlqj lq wkh Lqwhuqhw1 Wkh ULV lv d zlwkrxw vhoi0orrsv1 Pruhryhu/ wkh| ghprqvwudwh wkdw/ li surmhfw ri ULSH +vhh iru pruh ghwdlov ^54`./ zklfk doorzv ghyl0 Â Î Ã Ö Ø[Ö Ù Ö Î ß ë é Ñ7Õ ¿@â Ñ íO¾î[ÀZ÷øî/ù ë 0.95 366075 2. dgglwlrq/ zh suhvhqw qhz dqg pruh suhflvh rq wkh prgho iru wkh DV judsk dqg wkh udqgrp judsk zlwk sro|0 ghjuhh judsk wkdq suhylrxvo| rewdlqhg e|ofQhzpdq hw do1 the AS-count on the of YL thedv network is LS0krsfrxqw fairly weak. dqg wkh ULV fro0 wkh ghjuhh vhtxhqfh vdwlvhv fhuwdlq frqvwudlqwv vxfk dv 10000 "981205.HBC7 î ./R79&:<.. 1.íOlq ¾î[À wkh é Ñ7Ö Â Ù ß Õ ¿@0. AskDV shown in Fig.10: of AS traversed in à various ÂK¿=Ú Í Â Î sets. |hii _gihi?| hi4L|i hL|i UL**iU|Lh +W./02143 ^9` vwloo ghprqvwudwh d srzhu0olnh ehkdylru Â Ö × Ñ Ú.11393) * x ** ( -2.6Wj t 4i?|L?i_ ? |i *@t| UL*4? kLS 1 Vhfwlrq Y sorwv vlpxodwlrq uhvxowv ri wkh DV krs0 frxqw glvwulexwlrq ghulyhg iurp wkh ghjuhh judsk dqg frp0 Interestingly.1=@ DR`4/=>=DJ9&DR0C.    !"$# %& '( %)+*-.ØÐ ¿Þ Ä 0. even though a qrpldo olqngeometry zhljkwv lq vhf1 prgho iru wkh ^46` ru Uhlwwx dqg Qruurv ^47`/ exw suhvhqw wkh lqyroyhg lq dq DV1 pdwkhpdwlfdo surriv hovhzkhuh ^53`1 Rxu frqvwuxfwlrq ri wkh ghjuhh judsk doprvw dozd|v dyrlgv +xquhdolwlvwlf.HBCDY@6 1+0021h.=/=>N1?O.out" exp(8.3â Ñ Ã Õ$Ñ Ú Ñ-Ù Ü Ä Ø ¿ Í Ð Ú Ó íO¾ÉÀZð äZÌ$Öghjuhh Ð î>ðñÉ ë glv0 02546879&:<.0 Lq wklv sdshu/ zh irfxv sulpdulo|é:rq ¿ÂÃwkh Ã Ö × prgholqj Ã=Èß\¿ Ù Í Â ÑX0Í Â Ù Ñ ¿ÚU1ÃÄ0Ú ÝÏ ¿ Ð 2Ñ Ú ï à ëQ3ÜeÖÐ 4 Ã Í Þ@¿ 5Ù Ù ß Ñ@ÝÏ 6¿ Ù Õ$Ö Ø Ä 7Ø$Ù Ö Ã Ñ õ 8Ø[Ö Â Ñ Âà 9ÚkjÊþ ç X Êþ Ç ë ¿ÂH×%Êþ Ë çWÍ ÂÞ ÕÐ Ö Â Ö Ä Ñ ÚÞ Ð Í È Ñ Ã Õ$ÑOriØ$Ù Öwkh prghov= wkh DV0krsfrxqw kDV 1 Zh sursrvh wzr îaû glhuhqw Ú W é d   ¿ % Ú ( ¿ Ú × Â _ Â Ã Ã Ú ¿ õ D ¿ à ó!ë Õ$Ö Ö Ù Í Ù Í ÑSÍ Õ$ÑØ$Ù Ö ä%ü©Ñ Ø$ØÐ Ö Í ê Ñ # AS Ù Hops õ äZÔÕ$Í ÚÚ Õ$Ö éWÚÃ Õ ¿DÃà Õ$Ñ Ö Î Í Þ@k¿ Ù/ÖÐ × Ñ Ð@ë ¿Úé Ñ Ú Ñ-Ñ. e| This indicates thatwzr theedvlfdoo| AS-count is robust.70 1163687 2.Number Õ$Ñ%ÎÐ ¿ Ø$ÕKÍ data Ù Ñ Â Ï$ê Data Ñ æ Ö Ï ÃÄÆ Ò7Ø$Ù Piet Ö Ã ë$Í Â ïHvan Îä Ë ä È ë 6}  Ai ThLM@M*|) _i?t|) u?U|L? Lu |i 5 LTUL?| uLh dv d prgho iru wkh LS0krsfrxqw lq dq DV1 Hduolhu zrun Mieghem.wkh Ê Õ$Ö Ø Ú Í Â>à Õ$Ñ%Í Âdqg Ã Ñ Ð Ä× Ö ê ¿ Í Â ÎÐ ¿ Ø$Õ Ú ë ¿ÂH×(à Õ$Ñ. È Ñ Ð@äBÔÕ$courtesy wkh udqgrp judsk zlwk sro|qrpldo olqn zhljkwv YL.c 5Wj @?_ wWj +^48` wr ^4<`. vhf1 YL dqg YLL/the zkloh wkh DV krsfrxqw ri wzr ixqgdphqwdoo| glhuhqw prghov wkdw kdyh erwk wkh between the AS-count AS’s in North-America. vhoi0 YYY Cj aj}ijj }iBV orrsv/ zklfk duh wrohudwhg lq ^46` dqg ^47`1 Ilqdoo|/ zh sur0 Wkh uvw dssurdfk wr prgho wkh DV krsfrxqw vwduwv e| srvh dq lqwhjudwhg prgho iru wkh hqg0wr0hqg LS0krsfrxqw zklfk lv edvhg rq wkh wzr0ohyho urxwlqj klhudufk| lq Lqwhu0 frqvlghulqj d judsk zlwk Q qrghv frqvwuxfwhg iurp d jlyhq ghjuhh vhtxhqfh/ qhw1 G4> G5> = = = > GQ YY jBtij6jA|t Nu 7 NVWNA| A YA|jiAj| Kdyho dqg Kdnlpl ^8/ ss1 49` kdyh sursrvhg dq dojrulwkp wr Wkh Urxwlqj Lqirupdwlrq Vhuylfh +ULV. the AS-counts of various different data sets (focussing on differsduhv wkhvh zlwk wkh suhvhqwhg Lqwhuqhw phdvxuhphqwv ri d Srlvvrq udqgrp yduldeoh dv h{sodlqhg lq ^48` dqg ixuwkhu vhfwlrq LL1 Wkh qryhow| ri wkh sdshu olhv lq glvfxvvlrq entwkhparts of the hoderudwhg Internet)lqyield roughly same picture.1=@ DP021Q0C.1=@ DJI à Õ$Ñ1ÎÐ ¿ Ø$Õ(ó!ë î>ð ó!ë é Ñ1Õ ¿@â Ñ Ã Õ$Ñ Ú Ñ-Ù Ü Ä ò\äZÌ$ÖWkhuhiruh/ Ð Ã Õ$Ñ × Í ¿ êXÑ Ã zh Ñ ÐÖÜkdyh BCDL@ >"1=@1+>=0BM1+7546021<0C. zdv pruh eldvhg wr prgho wkh LS krsfrxqw *tL |i ?4Mih Lu _gihi?| 5AOt E| _gihi?| +.

as well as for the AS-graph.2 Transportation Networks Transportation networks such as road. the timetable is not sufficiently robust with respect to modifications in the circumstances caused by accidents.11: Flights worldwide on a single day. See Cohen.3. . The conclusion is that the topology of the Internet is critical for its vulnerability. 1. 13].1. ben Avraham and Havlin [12. the random graph models for the Internet cease to have the necessary connectivity properties. so that they can adapt their travel plans to changing circumstances. The most important issue in the Dutch railway network is that it is tight. railway or airline networks (see Fig. To allow these networks to function efficiently. Another objective is to be able to provide on-the-fly information to travelers. ignore geometry altogether. Figure 1. weather conditions. 1. or signaling breakdown.3. traffic controllers need to deal with disruptions (for instance. One objective is to develop robust scheduling algorithms that take the random nature of traffic into account and can properly cope with disturbances. due to bad weather conditions). A topic of research that is receiving considerable attention is how the Internet behaves under random breakdown or malicious attacks. most models for the Internet. due to the scarce space that is available for extending the network where needed. Also randomized algorithms can be very useful. Most of the scheduling and planning problems in railway and airline traffic are very hard (so-called “NP-complete”). but in practice good approximation algorithms may do a great deal. As a result. TECHNOLOGICAL NETWORKS 21 priori we might expect geometry to play a role. Erez. When vertices with high degrees are taken out.11) tend to become increasingly complex. As a result.

e. Battiston. Examples are reviewed for instance in Caldarelli. 1..3.3 Energy Networks Energy networks transport energy from providers to users. At the same time. technical failures or calamities. It is critical that the aggregate production rate of the energy sources is sufficient to meet the consumption rate of the users with extremely high probability. 1.g. These phenomena give rise to very distinct characteristics. the generation of energy increasingly exhibits random fluctuations over time. as an aggregation of all the transactions that occurred during a given time window. the production rate of conventional energy resources and power plants is subject to uncertainty and variability. In turn. . Therefore the time series of financial prices are at the same time the input and the output of a collective process involving the actions of a large number of people. rendering centralised operation impractical. these grids need to be designed to achieve consistently high levels of performance and reliability. With the rising deployment of renewable resources such as wind farms and solar panels. mechanisms must be put in place to ensure that it is highly unlikely for the aggregate arrival rate to exceed the service rate for any length of time. transactions modify the price of such assets.e. where nodes are individual traders in a financial market and links represent transactions among these traders. or statically.4. The availability of different sources of data allows us to represent various projections of thisnetwork. Whether market agents are willing to buy or sell depends on the prices of the financial assets that are being traded. and yet need to be cost-effective to operate. industries and whole countries. complex and networked system operating at different scales and with extremely heterogeneous components. Such links can be defined either dynamically. In addition. The main empirical ‘signals’ associated with such transactions are the (highly fluctuating) time series of prices of the financial entities being traded. i. In order to prevent overflow of buffers. and creating a strong need for distributed control mechanisms.4 Economic Networks The economy is a large. 1. stock markets. a large number of people (individual investors as well as companies) interact through financial transactions. once they occur. the rapid advance of smart-grid technology offers growing opportunities for actively controlling energy supply and demand. Networks of market traders can in principle be defined.1 Financial Networks In financial markets. Garlaschelli and Catanzaro [8]. the economic system is in fact an intricately connected network with many layers and degrees of complexity. However. organizations. Examples are electricity grids. From the level of individuals (which can be thought of as the basic ‘agents’ of the economy) up to the level of firms. due to supply disruptions. REAL-WORLD NETWORKS as they may yield an optimal profit on the average.22 CHAPTER 1. in such a way that they ‘appear and disappear’ over time. Because of their vital interest.

2 Shareholding Networks Another possibility is to define firm ownership and shareholding networks (Kogut and Walker [26]).4. See Fig. .12(a) for an example. networks of financial correlations can be defined by representing the financial entities (e.g. However. See Fig. these networks are in some sense also social networks. Since the vertices often represent individual persons.1. ECONOMIC NETWORKS 23 these networks are extremely difficult to observe and analyse. describing the trade relationships among the world countries (Serrano and Bogu˜ n´a [38]).4. The ‘strongest correlations’ are defined either as the set {ρij } of correlations exceeding a given global threshold ρ. . 1. The correlation coefficient ρij between two time series xi (t) and xj (t) (where xi (t) denotes the increment of the i-th time series at time t) is defined as xi xj − xi xj (1. it is possible to calculate the n × n matrix of pairwise correlation coefficients between each pair of stocks. . 1.4. where boards are connected when they have at least one director in common (Newman. while avoiding the creation of loops. the time average f is dePT fined as f ≡ T1 t=1 f (t). for technical reasons that we do not discuss here. After the empirical correlation matrix is calculated.12(b) for an example. because of the high confidentiality of the data that are required as input. 1. an alternative and frequently used definition of increment is the difference between the logarithms of the prices. where the vertices are companies and/or shareholders and the edges represent the ownership relations between the corresponding vertices. Yoo and Baker [14]).3 World Trade Web Yet another important networked economic system is the World Trade Web. We note that corporate board and director networks can be re-defined in order to obtain economic networks.Networks of financial correlations have been used to study the returns of assets in a stock market [28. . What is much easier to obtain are the (publicly available) time series of price increments2 of stocks. 1. the t-th increment of a financial time series is defined as the difference between the price at time t and the price at time t − 1. Strogatz and Watts [31]. 33] and interest rates [17]. From a set of n synchronous time series.3) ρij ≡ q   2 xi − xi 2 x2j − xj 2 where. or as the minimum set of correlations (taken in decreasing value) that ensure some global connectivity property in the output network (such as the existence of paths connecting all pairs of stocks. or finally as the set of correlations that exceed a reference value calculated under some null hypothesis. 2 In the simplest case. if f (t) is a time series defined for t = 1. Davis. 5. T . stocks) as vertices and the strongest correlations as links. .

12: (a) Network formed by the strongest correlations among the stocks of the S&P500 index. (b) Snapshot of the shareholding network in Italy in 2001. 2005). the ownership relations among companies (from Garlaschelli et al. based on correlations between the log-returns of daily closing prices from 2001 to 2011.e. REAL-WORLD NETWORKS Figure 1. i.24 CHAPTER 1. Vertices are companies and edges represent ‘who owns whom’. Stocks are coloured according to their industrial classification (from MacMahon and Garlaschelli 2014).. .

1 25 Biological Networks Metabolic Networks Biological networks are shaped by natural evolution. BIOLOGICAL NETWORKS 1.1. Oltvai and Barab´ asi [24]). describing the (directed) transportation of nutrients between the various regions and tissues of an organism. Similarly. Tombor. and vascular networks (West. Southgate. Figure 1. [41]). where proteins are connected by an undirected edge when they interact by physical contact. Therefore their structure can shed light on how a specific function selects a particular topology. Mason.5. Barabasi and Oltvai [23]). describing the directed synaptic connections among neurons in the brain.3 Neural Networks and Vascular Networks Examples at the organism level include neural networks (White. . Genes sharing similar genetic interaction profiles are proximal to one another.13 for an example on a yeast cell).13: A functional network for a yeast cell of correlated genetic interaction profiles. Albert. such as blood vessels in animals and vessels in plants. and protein interaction networks (Jeong. genetic networks represent the correlations among the expression profiles of different genes in a cell (see Fig. 1. where metabolic substrates are linked by directed edges when a known biochemical reaction exists between them. Thomson and Brenner [42]).5.5.5.2 Protein Interaction Networks and Genetic Networks Examples at the cellular level include metabolic networks (Jeong. 2010). Colored genes are enriched for GO biological processes as indicated (from Costanzo et al. 1. Less similar genes are positioned further apart. 1.5 1. Brown and Enquist [40].

com.4 CHAPTER 1.6 Still other types of networks The four classes (I)–(IV) in Section 1. where nodes represent events and edges are established between events that co-occur together (possibly with a weight that quantifies the frequency of co-occurrence). 1. Two examples of undirected networks are word synonymy networks (Ravasz and Barab´asi [37]). Such networks are at the basis of the automatic recommendation systems routinely used e.2 Co-occurrence Networks Co-occurrence networks.g. form a huge class of networks.14 we show The Political Books Network compiled by Valdis Krebs [43]. are not exhaustive. where two biological species are connected by a directed edge when a predator-prey relation exists between them. REAL-WORLD NETWORKS Food Webs Examples at the community level include food webs (Elton [19]. by online shops. Yet another example is given by networks of co-purchased products.2–1. Briand and Newman [11]).26 1. of which examples were listed in Sections 1. where nodes are scientific articles and edges indicate that two articles have been co-authored by the same author. We give two further examples.2. as well as the collaboration networks discussed in Section 1. reporting the outcomes of psychological experiments where people are asked to associate “input” words to freely chosen “output” words. viewed as examples of social networks. 1. where words are connected when they appear one or two words apart from each other in the sentences of a given text. In Fig. For instance. as i ndicated by the “customers who bought this book also bought these other books” feature on Amazon.1 Semantic Networks In word networks. and of free associations.5. Examples of directed networks are given by networks of dictionary terms. where two products are linked when they have been frequently purchased together. co-cited by the same paper.6. 1. Pimm [36]. Examples include the aforementioned word co-occurrence networks. .5.1. 1. where words are connected when they are listed as synonyms in a dictionary. words are represented by vertices and edges are placed between words when some linguistic relation exists between them.2. examples in the field of scientometrics are co-authorship and cocitation networks. This network represents books about US politics sold by Amazon. Cohen. and word co-occurrence networks (Ferrer i Cancho and Sol´e [21]). respectively. where words are connected when a (directed) link between them is reported in a given dictionary. Edges represent frequent co-purchasing of books by the same buyers.6.

com) books about US politics. Modified from [43]. . “neutral” (triangles) and “conservative” (squares). respectively.1.2. The colour of vertices is assigned by a so-called community detection algorithm that finds groups of vertices that are more densely connected among themselves than with the rest of the network (see Chapter 4. The political viewpoints of these books are given by “liberal” (circles).14: The network of frequently co-purchased (on Amazon. Section 4.4).6. The more central vertices in the “liberal” and “conservative” communities are surrounded by black boxes. STILL OTHER TYPES OF NETWORKS 27 Figure 1.

[4] A. 2004. [5] G. H. book chapter in: Complex Networks (eds. 1990. Quantitative Finance 1 (2001) 96. Springer. 2002. Shenker and W. Cohen. K. E. Cohen. 85 (2000) 4626. 74 (2002) 47. Willinger. H. Rev. Yoo and W. 52 (2000) 386. Europhys. [2] R.N. Computer Networks and ISDN Systems 33 (1998) 107–117. Kumar. Davis.Bibliography [1] L.-L. Albert and A. Lett. Ben-Naim. Proceedings of the 21st Annual Joint Conference of the IEEE Computer and Communications Societies. [6] S. Lett. Community Food Webs: Data and Theory. [12] R. Mantegna. Perseus Publishing. Maghoul. Lillo and R. Lecture Notes in Physics 650. Graph structure in the web. Springer. The small world web. F. Resilience of the internet to random breakdowns. Wiener. 86 (2001) 3682. Marchetti and L. Erez. Catanzaro. S. Govindan. Barab´asi. Cambridge. R. Rev. in: Lecture Notes in Computer Science 1696.-L.J. Caldarelli. 399–423. Phys. S. Chang.M. Toroczkai). IEEE Computer Society. Pietronero. R. R. Phys. [3] R.E. M. Phys. Chen. 443–454. R. 1999. Nature 401 (1999) 130–131. [8] G.E. [11] J. Massachusetts. Linked: The New Science of Networks. ben Avraham and S. Newman. D. [14] G. Z. Baker. Jeong and A. Berlin. [10] Q. Rajagopalan. [13] R. Lett. D. Internet: Diameter of the worldwide web. ben Avraham and S. P. Breakdown of the internet under intentional attack. pp. Page. Computer Networks 33 (2000) 309–320. S. Garlaschelli and M. Frauenfelder. Stata. Erez. Bonanno.A. Rev. Barab´ asi. Springer. Brin and L. Caldarelli. pp. D. Adamic. [7] A. Havlin. Mod.F.-L. The anatomy of a large-scale hypertextual web search engine. Barab´asi. 2002. Cohen. Tomkins and J. 28 . F. K. [9] G. Jamin. Briand and C. Battiston. A. H. Raghavan. Havlin. Albert. S. Strategic Organization 1 (2003) 301. Broder. F.

Faloutsos and M. 2006.M.BIBLIOGRAPHY 29 [15] R.J.T. [32] M.W. Kleinberg. [25] J. Kumar. Proceedings of the Royal Society of London B268 (2001) 2261. Barab´asi and Z. S. Sneppen and A. Lecture Notes in Computer Science 1627. Kogut and G. With a sidebar by P. De Castro and J. 1999. B 30 (2002) 285. Volume I. Famous trails to Paul Erd˝os. A. Famous trails to Paul Erd˝os. Math. [24] H.J. J. Springer. S. Elton. Z.N. [31] M. Ramsden. S. P.J. Eur. [33] J.J. S. Phys.win. Tombor. Grossman.F.F. Barab´asi. Oltvai. Rajagopalan and A. [18] S. Jeong. Newman.E. K. Watts and A. J. SIAM Review 45 (2003) 167. [21] R. [28] R. Proceedings of the First Bonzenfreies Colloquium on Market Dynamics and Quantitative Economics.M. Aste. P. Dorogovtsev and J. [16] R.J. A. Tomkins.B.H. Intellingencer 21(1999) 51–63. The Structure and Dynamics of Networks. Mantegna. van der Hofstad. Sol´e. [22] R. Phys. Phys. Ferrer i Cancho and R. Rev. Sidgwick & Jackson. B 25 (1999) 193. Newman. Barab´asi. Maslov. Raghavan. T. Translated and revised from the English.-L. P. Berlin.-P. Albert. monograph in preparation. Grossman. Computer Communications Rev. B. Newman. . American Sociological Review 66 (2001) 317.S. Hyde and S. Princeton University Press. pp.-L.E. F´ıs. E 64 (2001) 026118.V. 29 (1999) 251–262. Onnela. Chakraborti. Tomkins. Physica A 333 (2004) 529–540.nl/~rhofstad/ [23] H. Physica A 355 (2005) 21–35. in: Proceedings of the International Conference on Combinatorics and Computing. [27] R. Di Matteo. Mendes. Animal Ecology. 1–18. Eur. S. Walker. D. Vitanyi. tue.E. Zaliznyak. Exact. [20] C.N. [19] C. Oltvai and A. On power-law relationships of the internet topology. [17] T. Faloutsos.N. Acad. K.-L. Nature 411 (2001) 41. Advances in Physics 51 (2002) 1079. [30] M. [26] B. Faloutsos. 1927. Kaski and J. De Castro and J.W. 23 (1999) 563–582.R. Watts. Computer Networks 31 (1999) 1481–1493. London. Raghavan. Natur. R.N. File can be downloaded from http://www. Kert´esz. Rev. Strogatz and D. Mason. Colombiana Cienc. Jeong. Rajagopalan and A. Random Graphs and Complex Networks. [29] S. Kumar. Trawling the web for emerging cyber communities. Princeton Studies in Complexity. S. Nature 407 (2000) 651.

He. Y. Evolution and Structure of the Internet. E. Inc. Brenner. The Science of a Connected Age.B. West.L. E67 (2003) 026112. W. [37] E. Brown and B. Cao & D. Cao. 2993 (2013). Cambridge. Science 284 (1999) 1677. Enquist. [40] G.H.-L.A. V´azquez and A. [39] D.J. Phil.N. . Pimm. 2004. Chapman & Hall. Rev. Jin.J. E 68 (2003) 015101(R). Thomson and S. R. Enquist. Pastor-Satorras. Soc. [43] X. [42] J. Scientific Reports 3. Ravasz and A. White. [36] S. London B 314 (1986) 1.. Six Degrees. A. ´ Serrano and M. Vespignani. J. Wang.J. [41] G. X. Brown and B.W. Watts.G. Trans.30 BIBLIOGRAPHY [34] R. West. J.B.H. New York. A Statistical Physics Approach. 87 (2001) 258701. London. Southgate. Lett. n´a. 1982. J. Phys. Pastor-Satorras and A. [35] R. Phys. Phys. Science 276 (1997) 122. Rev. Vespignani. Bogu˜ [38] M. D. Cambridge University Press. Food Webs. Norton & Co. Rev. Barab´asi. 2003.

where edges occur randomly and independently. See Fig. s s s s s s @ @s s s s s s s s s Figure 2. Examples 1 and 5 contain isolated vertices. They play an important role in analysing and explaining the empirical properties observed in real-world networks.1: Examples 1–3 are complete graphs. examples 4 and 6 are multi-graphs.Chapter 2 Random Graphs In this chapter we describe some key concepts in graph theory. and are used as null-models. Edges are undirected. They can also be used to make predictions. A graph that is not simple is called a multi-graph.2 we analyse the simplest random graph model.) In Section 2.1 Graphs. A graph that is simple and has all pairs of vertices connected by an edge is called a complete graph.1 we introduce graphs and random graphs. The size of a 31 . 2. Graphs with directed edges are called directed graphs. Example 5 has two clusters.1. Random graphs are models for complex networks (randomness is often synonymous to complexity). random graphs. Not all pairs of vertices need to be connected by an edge. Such vertices are called isolated. 2. A graph is called simple when there are no self-edges (= no edges between a vertex and itself) and no multiple edges (= at most one edge between a pair of vertices). Examples 1–3 and 5 are simple graphs. and look at four particular scaling features as these graphs become large. E) consists of a set of vertices V (also called nodes or sites) and a set of edges E (also called links or bonds) connecting pairs of vertices. four scaling features A graph G = (V. Some vertices may have no edge at all. (More detailed scaling features will be discussed in Chapter 4. A cluster or connected component is any maximal subset of vertices whose vertices are connected by edges (= maximally connected component). due to Erd˝ os and R´enyi. They are inspired by real-world networks. In Section 2.

3) with N0 the set of non-negative integers. i6=j 1 where i ↔ j means that i and j are connected. The degree sequence of a graph G is the vector ~k = (ki )i∈V (2. i6=j d(i.2) i∈V where |V | is the cardinality of V and δki is the point distribution concentrated at ki .e. HG is the distance between two vertices drawn uniformly from all pairs of connected vertices. (2. a tree graph has clustering coefficient 0. i. (2. The degree of a vertex is the number of edges attached to it.i2 i3 are present} . i3 forms a wedge when the edges i1 i2 and i2 i3 are present. a linear graph has typical distance roughly one third of its length. An isolated vertex has degree 0.i2 i3 . j) denotes the graph distance between i and j (= the minimal number of edges in a path between i and j).4) represent the fraction of vertices with degree k.. k ∈ N0 . The clustering coefficient of G is the ratio CG = ∆G ∈ [0.32 CHAPTER 2.i3 ∈V (2. and a triangle when the edges i1 i2 . WG (2. In words.i2 .i2 .j∈V : i↔j. i1 . and various different choices have been made with the aim to model real-world networks of . i2 i3 and i3 i1 are present. i2 .7) i. and d(i.5) where ∆G = X i1 . Note that fG is a probability distribution on N0 . whose weights fG (k) = |V |−1 |{i ∈ V : ki = k}|. δki (k) = 1{k=ki } . RANDOM GRAPHS cluster is the number of vertices it contains. j) HG = P ∈ [1. (2. The degree distribution is the probability distribution X fG = |V |−1 δk i .i3 ∈V 1{i1 i2 . The typical distance in G is the ratio P i.1) with ki the degree of vertex i. There are many possible ways in which this can be done. An isolated vertex is a cluster of size 1.2. k ∈ N0 . WG = X 1{i1 i2 . A triple of distinct vertices i1 . A vertex with a loop has degree 2..i3 i1 are present} .6) i.) A complete graph has clustering coefficient 1. ∆G is 3! = 6 times the number of triangles in G and WG is 2! = 2 times the number of wedges in G. but with the same flavour. (2. A random graph is a graph where the vertices and/or edges are chosen randomly. 1].3 a different definition will be used. (In Section 4. ∞).j∈V : i↔j. This definition is sometimes referred to as the wedge-triangle clustering coefficient. The complete graph has typical distance 1.e.

Since networks tend to grow. following van der Hofstad [2. i. If the latter holds with K log n replaced by an upper bound that is o(log n). Sparse means that most vertices have a degree that stays bounded as n → ∞.12) for some C ∈ (0.1. where E denotes expectation and kfGn − f k∞ = sup |fGn (k) − f (k)| (2. f (k) = k −τ +o(1) as k → ∞. This is referred to as a random graph process. In such cases the upper bound is often K log log n.13) for some K ∈ (0. ∞). (4) G is called a small world when lim P(HGn ≤ K log n) = 1 n→∞ (2. Chapter 1]: (1) G is called sparse when lim E(kfGn − f k∞ ) = 0 n→∞ (2.11) for some τ ∈ (1. Scale free means that the graph “looks similar on all scales”. In what follows we give a precise mathematical definition of four scaling features of random graph processes.9) for some non-random probability distribution f on N0 . Small world means that typical distances grow only very slowly with the size of the graph (and are almost independent of the size). it is natural to consider sequences of random graphs G = (Gn )n∈N . GRAPHS. (2.e. We use the symbol P to denote the probability distribution of G.2. ∞). [Begin intermezzo] . then G is called an ultra-small world. Not highly clustered means that locally the graph “looks like a tree”. FOUR SCALING FEATURES 33 different types.8) where n denotes the number of vertices in Gn . Chapter 4 contains several exercises where the reader is requested to compute empirical properties of examples of graphs. (2) G is called scale free with exponent τ when it is sparse and lim k→∞ log f (k) =τ log(1/k) (2.. (3) G is called highly clustered when lim n→∞ E(∆Gn ) =C E(WGn ) (2. 1].10) k∈N0 is a distance between fGn and f (called the supremum-norm on the space of probability distributions on N0 ). RANDOM GRAPHS.

Chapter 6 describes algorithms to simulate ERn (p). the removed edges are called “closed”. The symbol O stands for “is at most of the same order as”: an = O(bn ) when lim supn→∞ an /bn < ∞. 1) and remove it with probability 1 − p. The symbol Θ stands for “is of the same order as”: an = Θ(bn ) when both an = O(bn ) and bn = O(an ). light blue). This is also written as an  bn . respectively. See Fig. RANDOM GRAPHS In asymptotic analysis. 3/200. The symbol o stands for “is of smaller order than”: an = o(bn ) when limn→∞ an /bn = 0. Here. and marked the beginning of random graph theory.2 Erd˝ os-R´ enyi random graph The simplest example of a random graph is the Erd˝ os-R´enyi random graph. [End intermezzo] 2. for each n ∈ N. O and Θ.d. (The retained edges are called “open”. independently for different edges.  . Courtesy Remco van der Hofstad.) The resulting graph is a random subgraph of Kn . (= independent and identically distributed) random variables taking the values 1 with probability p and 0 with probability 1 − p. Note that E(Ye ) = P(e is retained) = p. The three largest clusters are ordered by the darkness of their edge colors (dark blue. Figure 2. three symbols are used frequently: o. blue.  where the sum runs over the n2 edges of the complete graph Kn . The remaining edges all have the lightest shade (grey). and is denoted by ERn (p). we consider the complete graph Kn on n vertices.2 for two realizations of ER200 (1/200) and ER200 (3/200).i. and show that it satisfies the law of large numbers and the central limit theorem in the limit as n → ∞ (look P up on Wikipedia what this means).1 Find the distribution of the number of edges in ERn (p)? Compute its mean and its variance. and Ye = 1{e is retained} are i.34 CHAPTER 2. It was introduced in 1959 by the Hungarian mathematicians Paul Erd˝os and Alfred R´enyi [1]. and for each of the n2 edges we decide to retain it with probability p ∈ (0. Hint: Use that the number of edges is e Ye . Homework 2.2: Two realizations of Erd˝os-R´enyi random graphs with 100 vertices and edge probabilities 1/200. 2.

[Begin intermezzo] A branching process is a simple model for a population evolving over time.˝ ´ 2.d. each individual in the population independently gives birth to a random number of children. Let X m= kf (k). Chapter 4]. i. while if m > 1. Thus.n∈N0 is an array of i. 2. Xi. chosen according to a prescribed probability distribution f = (f (k))k∈N0 called the offspring distribution.2.n is the number of children of individual i in generation n). Yet. but has a large connected component containing a positive fraction of all the vertices when λ > λc (“supercritical regime”). f (k) is the probability that an individual has k children. ERDOS-R ENYI RANDOM GRAPH 35 The Erd˝ os-R´enyi random graph is not really suitable as a model of a realworld network. (2. where for convenience we pick Z0 = 1. in each generation. it captures a basic feature of a real-world network: complexity. • Θ(n) when λ > 1. It can also be shown that for λ = λc there are multiple clusters of size Θ(n2/3 ).n .15) k∈N0 One of the key results for branching processes is that if m ≤ 1. At λ = λc there is a percolation transition: the small clusters coagulate into a large cluster. (2. • Θ(n2/3 ) when λ = 1. ∞) and let n → ∞.e.2.e. n ∈ N0 . while all the other clusters are of size Θ(log n). it turns out that the extinction probability η = P(∃ n ∈ N : Zn = 0) (2.. Then Zn satisfies the recursion relation Zn+1 = Zn X Xi..14) i=1 where (Xi. we make a brief digression into the mathematics of branching processes. Namely.1 Percolation transition We follow the exposition in van der Hofstad [2. random variables with common distribution f (i. In fact. there is a critical value λc = 1 such that ERn (λ/n) consists of a large number of small disconnected components when λ < λc (“subcritical regime”). Let Zn denote the number of individuals in the n-th generation. for which typically neither the number of vertices is fixed nor the edges are retained or removed independently. then the population dies out with probability 1 (unless f = δ1 ). then the population has a strictly positive probability to survive forever. the largest cluster has size • Θ(log n) when λ < 1.i.16) .n )i∈N. The Erd˝os-R´enyi random graph exhibits an interesting phenomenon: ERn (p) has a percolation transition when we pick p = λ/n with λ ∈ (0. Before we explain the intuition behind the above percolation transition. It can be shown that for λ > λc there is only one cluster of size Θ(n). Suppose that.

and show that η = 1 when 0 < p ≤ 21 and η = (1 − p)/p when 21 < p < 1. and supercritical when m > 1.1 Show that η = 0 if and only if f (0) = 0. Pick any vertex. and are painted green as well. [End intermezzo] We are now ready to explain the intuition behind the percolation transition in the Erd˝ os-R´enyi random graph. because the exploration process rarely creates loops. Gf (x) = X xk f (k). (2. The result is a random sequence (Nd )d∈N (2. The idea is that this exploration process is close to a branching process when n is large. 2. we speak of binary branching. Write N1 to denote their number. Repeat this procedure. . These vertices all lie at distance 2 from ?.18) that we can think of as an exploration process. RANDOM GRAPHS is the smallest solution of the equation (see Fig. Exercise 2.) Gf (x) s f (0) s s x η Figure 2.3: Plot of the generating function x 7→ Gf (x) for the case where m = G0f (1) > 1. Next. and are painted green as well. call this vertex ?. 1]. consider the vertices connected to the N1 vertices just painted green.17) k∈N0 A branching process is subcritical when m < 1. x ∈ [0.3) η = Gf (η).2 When the offspring distribution is given by f (k) = (1 − p)1{k=0} + p1{k=2} for some p ∈ (0. a vertex is counted each time it has a connection to a green vertex. Write N2 to denote their number where. counting the vertices that are connected to ? at successive distances (with multiplicities) and painting them green.36 CHAPTER 2. Exercise 2. Compute Gf (x). (The case f = δ1 is uninteresting and is excluded. Next. consider the vertices that are connected to ?. 1). and paint it green. but exclude ?. in order to avoid complications. These vertices all lie at distance 1 from ?. critical when m = 1.

2 The distribution of N1 is BINOMIAL(n − 1. The exploration process eventually covers the entire set of vertices.˝ ´ 2. the approximation Nd ≈ N Homework 2.20) Show that the mean of fλ equals λ.19) k Show that as n → ∞: (a) E(N1 ) = p(n − 1) ∼ λ and Var(N1 ) = p(1 − p)(n − 1) ∼ λ. ERDOS-R ENYI RANDOM GRAPH 37 Homework 2. E(Nd ) reaches values in the range Θ(1/n). Hence we see that ? lies in a small cluster when λ < 1. E(Nd ) reaches values in the range Θ(n) (after which ˜d begins to break down). Hence. for large n. • λ < 1: When d reaches values in the range Θ(log n). ˜d is good (recall • λ > 1: As long as d = o(log n).λ/n (k) = fλ (k).2(b) shows that the set of vertices not yet covered gradually depletes. n − 1. but that for large n this effect is hardly noticeable.p (k) = P(N1 = k) = p (1 − p)n−1−k . Show that lim fn−1. p). (b) E(N2 ) = E(N1 p(n − 1 − N1 )) = p(n − 1)E(N1 ) − pE(N12 ) ∼ λ2 .3 The probability that ? has degree k equals fn−1. (2. When d reaches values in the range Θ(log n). the binomial distribution with parameters (n − 1. . for large n the Erd˝ os-R´enyi random graph locally looks like a random tree. . . but has a positive probability of lying in a large cluster when λ > 1. n→∞ k ∈ N. (2. Hence. Thus. Homework 2. However. with fλ = POISSON(λ) the Poisson distribution with parameter λ given by fλ (k) = e−λ λk . k! k ∈ N. Hint: Use that limn→∞ (1 − n−1 )n = e−1 . the branching process (N The above argumentation is heuristic: for finite n both the exploration processes and the degrees associated with different vertices are dependent. among the n vertices of the graph there are Θ(1) vertices that lie in a cluster of size Θ(log n). p) given by   n−1 k fn−1. Homework 2. as more and more vertices are covered. . the random sequence (Nd )d∈N is close to a branching process ˜d )d∈N whose offspring distribution has mean λ.2) and we have E(Nd ) ≈ E(N˜d ) ≈ λd .  Thus.2. limd→∞ N with positive probability when λ > 1.p (k) = P (N1 = k). k = 0. (c) E(Nd ) ∼ λd for every d ∈ N. ? has a strictly the approximation Nd ≈ N positive probability to lie in a cluster of size Θ(n). for large n this dependence is weak and we may approximate the exploration processes from different vertices as independent branching processes in which each vertex at distance d has a number of vertices at distance d + 1 attached to it whose distribution is close to POISSON(λ). According to the above in(N ˜d = 0 with probability 1 when λ < 1 and limd→∞ N ˜d = ∞ termezzo. In other words.  ˜d )d∈N has offspring distribution fλ . .

X2 ) with independent components is a coupling. Show that lim n−1 E(WERn (λ/n) ) = 12 λ2 . Indeed. it follows that Nd = Θ(n) when d = Θ(K log n).  Consequently. The Erd˝ os-R´enyi random graph is a small world when λ 6= 1.. in the subcritical regime λ < 1. i. X1 ) and (X2 .4 Compute the average number of wedges E(WERn (λ/n) ) and the average number of triangles E(∆ERn (λ/n) ) in the Erd˝ os-R´enyi random graph. For instance. Given X1 and X2 . respectively.3 Show that (Nd )d∈N is stochastically smaller than (N ˜d for all there exists a coupling of the two random sequences such that Nd ≤ N d ∈ N with probability 1.e. . Exercise 2. We say that X1 is stochastically smaller than X2 when there exists a ¯1 ≤ X ¯ 2 with probability 1. In this way the random graph can be made to be scale free and highly clustered as well. the Erd˝ os-R´enyi random graph is not highly clustered. the Erd˝os-R´enyi random graph is not scale free. fλ (k) decays faster than polynomially in k as k → ∞. then the pair (X1 .  [Begin intermezzo] A coupling of two random variables X1 and X2 is any pair of random variables ¯1. X ¯ 2 ) coincide (X with the probability distributions of X1 and X2 . As long as Nd ˜d in distribution.. Since fλ has a thin tail.21) which implies that C = 0. Since is small compared to n. Exercise 2. the exploration process from vertex ? must stop after at most Θ(K log n) iterations. the Erd˝os-R´enyi random graph is sparse. there are many ways to construct a coupling. n→∞ (2. X ¯ 2 ) such that the marginal probability distributions of (X ¯1. typical distances are at most K log n with K = 1/ log λ by the following heuristic argument. but also the pairs (X1 . this fact is obvious because the largest cluster has size Θ(log n). Since there are not more than n vertices. In Chapter 3 we look at more realistic models to construct random graphs with these properties. 2.2. i. Finally. RANDOM GRAPHS The following comparison is valid for any n without approximation. coupling such that X [End intermezzo] We will encounter coupling again in Chapters 8–9.e. if X1 and X2 have the same distribution.38 CHAPTER 2. we know that Nd is close to N d ˜ E(Nd ) = λ . In the supercritical regime λ > 1. X2 ) with identical components are. it is possible to consider a generalised Erd˝ os-R´enyi random graph in which the parameter λ is chosen randomly according to a distribution with a power law tail.2 Scaling features Since the degree distribution of ERn (λ/n) converges to fλ . ˜d )d∈N . n→∞ lim E(∆ERn (λ/n) ) = λ3 .

R´enyi. tue. I. Volume I. monograph in preparation. [2] R.win. Publ. On random graphs. van der Hofstad. Erd˝ os and A. File can be downloaded from http://www.nl/~rhofstad/ 39 . Random Graphs and Complex Networks. Math. Debrecen 6 (1959) 290–297.Bibliography [1] P.

Chapter 3 Network Models In this chapter we describe two examples of random graphs that are more realistic models of real-world networks than the Erd˝os-R´enyi random graph encountered in Chapter 2. i. . Suppose that we take the degree sequence as the starting point of our model. A practical situation may arise from a real-world network of which we know the degrees but not the topology. 3.1 The configuration model Motivation In this section we investigate random graphs with a prescribed degree sequence. but these can be removed afterwards. i.e. . (3. To that end.1 we look at the configuration model. . .1. . n} a pre-specified degree ki ∈ N0 .2 Construction We follow van der Hofstad [9. Chapter 7]. the degrees are given to us beforehand. . The former is a static realisation of a random graph (like the Erd˝os-R´enyi random graph). the latter is a dynamic realisation.1 3. and we are interested in generating a uniformly random graph with precisely the same degrees (where uniformly random means that all realisations have the same probability)..2 at the preferential attachment model.. or does it inherently have more structure? The configuration model described below was introduced by Bollob´as [5]. 3.. .e. inspired by earlier work of Bender and Canfield [4]. kn ). It generates the desired random graph by matching half-edges in a uniformly random manner. . it is the result of a growth process.e. for n ∈ N we associate with each vertex i ∈ V = {1.1) and we connect the vertices with edges in some way so as to realise these degrees. This comes at the expense of possibly creating self-loops and multiple edges. forming a prespecified degree sequence ~k = (k1 . i. In Section 3.1. we think of placing ki half-edges (“stubs”) incident to 40 . An interesting question we may want to settle is: Does the real-world network resemble a uniformly random graph with the same degree sequence. in Section 3.

By conditioning on the graph being simple. The total degree is i=1 ki . Sometimes this is referred to as the repeated configuration model. Chapter 7 describes algorithms to simulate CMn (~k). (3.9)) lim E(kfCMn (~k) − f k∞ ) = 0 n→∞ (3. kn ) with k1 ≥ k2 ≥ . One way to do this is to match the half-edges in a uniformly random manner.3 Graphical degree sequences A natural question is: Which sequences of numbers can occur as the degree sequence of a simple graph? A sequence ~k = (k1 .1). The resulting random multi-graph is denoted by CMn (~k) and is referred to as the configuration model. and matching the different half-edges in some way so as to form full edges.1. However. conditionally on the paired half-edges so far. . the total Pn number of edges is 21 i=1 ki . we end up with a random graph that has the pre-specified degree sequence. which is why the total degree must be even. we will see that if the degrees are not too large. ≥ . As long as. the final outcome is the same in distribution. 3. more precisely. since we may think of the conditioning as repeatedly forming the graph until it is simple. if (see (3. This leads to what is called the configuration model (see Fig. . Exercise 3. . we can use the configuration model as a particular way to realise a sequence of random graphs that is sparse and scale free with any desired exponent. THE CONFIGURATION MODEL 41 vertex i. . the next half-edge is paired to any of the remaining half-edges with equal Pnprobability. Hence. which is referred to as the erased configuration model. The pairing procedure may not lead to a simple graph: self-edges and multiple edges may occur. .1.11)–(3. .3) for some pre-specified probability distribution f (where E is the average with respect to the randomness of the graph).3. the degree distributions in these two models also converges to f .12) below) lim Var(fCMn (~k) ) = Var(f ) < ∞ n→∞ (3. It does not matter in which order the half-edges are paired in the pairing procedure.  The degree distribution associated with CMn (~k) is (recall (2. then the resulting graph is simple with a strictly positive probability.2)) fCMn (~k) = n−1 n X δki . 3. Show that not all pairings give rise to a different graph.1 Show that there are (2m − 1)!! = (2m − 1) × (2m − 3) × · · · × 3 × 1 different ways of pairing 2m half-edges.2) i=1 By choosing ~k such that (recall (2. Another approach is to remove the self-edges and multiple edges afterwards. It can be shown that when n → ∞.4) with Var(f ) the variance of f . and they are completely harmless in the limit as n → ∞. To keep the computations simple we stick to the original construction. for large n the conditioning and the erasing do not alter the degrees by much.

3.42 CHAPTER 3. NETWORK MODELS Figure 3. Courtesy Oliver Jovanovski. 5). The pictures show how 16 pairs of half-edges are randomly matched to become 16 edges. .1: Simulation of the configuration model with n = 7 vertices and degree sequence ~k = (5. 5. 5. 4. 5.

8) n→∞  1 .8) (which incidentally is slightly weaker than the condition k∈N0 k 2 f (k) < ∞ of finite second moment) guarantees that ! n . if lim nF (n) = 0. Exercise 3. . ~ is graphical) = lim P(D (3. if lim n→∞ nF (n) = ∞. This becomes relevant when the degree sequence ~k in the configuration model is itself drawn as an i.2 Give an example of a non-graphical sequence ~k = (k1 . sequence ~ = (D1 . . + k4 is even. . The sufficiency is harder to see.5) i=l+1 The necessity of this condition is easy to see. . . n→∞ 2 where F (n) = P k≥n f (k).  Arratia and Liggett [1] investigate the probability that an i. . k4 ) for which k1 + . . under the assumption that 0 < k even f (k) < 1 (i. (3. both even and odd degrees are possible). . n→∞ P k even f (k) < 1. . and we refer to Choudum [7] for a proof. sequence.   0.e.6) is graphical.d. Indeed. l = 1.9) i=1 Consequently. n − 1. the tail condition limn→∞ nF (n) = 0 in the P second line of (3. ki ).i. The first term on the right-hand side is the maximal total degree of the first l vertices coming from edges between them. THE CONFIGURATION MODEL 43 kn is called graphical when it is the degree sequence of some simple graph. In that case automatically lim E(kfCMn (D) ~ − f k∞ ) = 0. the left-hand side is the total degree of the first l vertices. . . Dn ) D (3. .7) n→∞ P It turns out that. while the second term is a bound on the total degree of the first l vertices coming from edges that connect to the other vertices. and explain in a picture why it is non-graphical.. say according to a pre-specified probability distribution f on N0 .1.i.d. . It is not hard to show that if ! n X lim P Di even = 21 . . then (3. Pn Erd˝os and Gallai [8] proved that a sequence ~k is graphical if and only if i=1 ki is even and l X n X ki ≤ l(l − 1) + i=1 min(l.3. (3.

X .

~ lim P D is graphical .

e. we make it possible for CMn (D) . i. the probability that even. (3.10) n→∞ i=1 ~ for which Pn Di is In other words. Di even = 1.. by retaining only those realisations of D i=1 ~ to be simple.

It can be shown that if f (0) = 0 and k∈N0 k f (k) < ∞.44 CHAPTER 3. then n . NETWORK MODELS ~ CM P n (D)2 is simple is strictly positive.

X   ~ is simple .

.

lim P CMn (D) Di even = exp[− 12 ν 2 − 14 ν 4 ] ∈ (0. ∞). However. • Θ(n2/3 ) when ν = 1. Indeed.i. the probability that a vertex at distance 1 from ? has k vertices not linked to ? equals f¯(k) (see Fig. 3. except that during the exploration process vertices and half-edges get gradually depleted (a phenomenon we already encountered in the . s s ? s s k s ] Figure 3. 3. the probability that a vertex with k + 1 half-edges is linked to ? is proportional to k + 1. We again consider the case where the degree sequence D with distribution f having a finite second moment.12) where N is the normalisation constant. The offspring distribution of a given vertex ? is equal to f .1. the fraction of vertices with k + 1 edges is f (k + 1). This shows that the repeated configuration model is a feasible way to generate simple random graphs with a prescribed degree distribution.2: The vertex ? linked to a neighbour ] that has k neighbours not linked to ?. namely.11) n→∞ i=1 with P ν= k∈N k(k − 1)f (k) ∈ [0. Then the largest cluster of ~ has size CMn (D) • Θ(log n) when ν < 1.d.) The intuition behind the above result is as follows. (3. By the uniform matching of half-edges.4 Percolation transition Like the Erd˝ os-R´enyi random graph. the configuration model has a percolation ~ is i. The same is true for vertices at distances ≥ 2 from ?.13) with N the normalisation constant. Hence. transition. k∈N kf (k) P (3.2). • Θ(n) when ν > 1. The probability that the other k half-edges end up being linked to half-edges of vertices at distance 2 from ? is 1 (in the limit as n → ∞). f¯(k) = N k ∈ N0 . 1] (3. the offspring distribution of the vertices at distance 1 from ? is different. this equals f¯ given by 1 (k + 1)f (k + 1). (The critical scaling actually requires that f has a finite third moment.

 Homework 3.d. In van der Hofstad [10.d.2. which is the average degree.1. 3. Since (compare (3. degrees is small-world.1 shows that CMn is locally tree-like.i. You are more likely to be friends with a person who has many friends than with a person who has few friends. In the language of social networks this inequality can be expressed as: On average your friends have more friends than you do! This sounds paradoxical. in a triangle in the limit as n → ∞.2 3. but it is not. respectively.7). this explains why the percolation transition occurs at ν = 1 (recall Homework 2. then ν > k∈N kf (k). for any ν > 1.  lim P HCMn (D) ∀ K > ν/(ν − 1). Chapter 5] it is shown that the configuration model with i. PREFERENTIAL ATTACHMENT MODEL 45 Erd˝ os-R´enyi random graph).5 Scaling features The configuration model can be made sparse and scale free by construction: since the degree distribution is pre-described it can be chosen so as to satisfy the conditions in (2.9) and (2.15) ~ ≤ K log n = 1 n→∞ where we recall (2.11)). But for large n this effect is minor and so we can think of f¯ as the forward degree of vertices in the exploration process.14) ν= k∈N0 is the average forward degree. (3. namely. This causes a bias. (3. 3) (recall (2.11). If the degree distribution f has exponent τ ∈ (2. the number of triangles grows much slower with n than the number of wedges.1 Is the configuration model with i.12) and (3. and compute the probability that ? lies in a wedge.2.. which is precisely what (3.13) captures.i. degrees highly clustered? Hint: Recall (2. so that X X kf (k) < ∞.12).3. The intuition behind this result is similar to that for the Erd˝ os-R´enyi random graph.2. in the sense that it gives rise to random graphs with degree distributions that can be matched to . with ν taking over the role of λ.13)) X k f¯(k) (3. P Note that if f 6= δ1 . Use Fig.1 Preferential attachment model Motivation The configuration model describes networks satisfactorily. i. k 2 f (k) = ∞.16) k∈N0 k∈N0 then the configuration model is even ultra small-world : distances are at most of order log log n. 3.2 and the intermezzo on branching processes in Chapter 2). Homework 3. 3.e.

the connections to the newcomer are spread uniformly over the population. In this case. suppose that new vertices are added (each carrying a fixed number of edges. it does not explain how these networks came to be the way they are.17) such that for every n the graph has n vertices. via a feature called preferential attachment. . We will also argue that preferential attachment leads to small-world random graphs that are “locally tree-like”. Therefore. For example. So. (3. say 1) that want to connect to older vertices. . In an Erd˝ os-R´enyi random graph. but it does give us the possibility to investigate and model how they grow. mn edges and total degree 2mn (see Exercise 3. Each edge is connected to a specific older vertex with a probability that is proportional to the current degree of that older vertex. A possible explanation for the occurrence of scale-free behaviour was given by Barab´asi and Albert [3]. 3. Below we will argue that preferential attachment naturally leads to sparse and scale-free random graphs. where people who are well known are likely to become even more known. Most real-world networks grow. But is this realistic? Is the newcomer not more likely to get to know individuals who are socially active and therefore already have a large degree? Probably so! We do not live in an egalitarian world. and this is responsible for new connections to the newcomer. which are static models of random graphs (even though it would not be hard to reformulate these graphs as a growth process where vertices and edges are added successively). Thus.6 below). δ) n∈N . rather than taking equal probabilities for our newcomer to become acquainted with other individuals in the population. each time enlarging the network by one vertex. Growth is an aspect that is not taken into account in the Erd˝os-R´enyi random graph or the configuration model. 3.18) .2. Rather. m ∈ N and δ ∈ [−m. PA1 (1. Chapter 8].46 CHAPTER 3. δ) consists of a single vertex v1 with a single self-loop (which has degree 2). vn } (3. Phrased in a more mathematical way. the WWW has increased from a few web pages in 1990 to several billion web pages at present. However. how do real-world networks grow? Think of a social network describing a population in which new individuals arrive one by one. We begin by defining the model for m = 1 (see Fig. like the Erd˝ os-R´enyi random graph and the configuration model. we should allow for a bias towards individuals who already know many other individuals. For example. and produces a random multi-graph process. NETWORK MODELS the power-law degree distributions found in real-world networks. A newcomer will start to socialise with other individuals in the population. The preferential attachment we consider depends on two parameters. ∞). denoted by  PAn (m. viewing networks as evolving in time is not in itself enough to explain the occurrence of power laws.3). The power-law exponent of the degree distribution depends on the parameters of the model. Let {v1 . we live in a self-reinforcing world. . . a preferential attachment model is such that new vertices are more likely to attach to old vertices with a high degree than to old vertices with a low degree.2 Construction We follow van der Hofstad [9.

i = 1. PREFERENTIAL ATTACHMENT MODEL 47 1+δ 3+2δ s v1 s v1 2+δ 3+2δ s v1 s v2 s v2 Figure 3. yet tends to infinity with probability 1 as n → ∞. . Also verify that i=1 Di (n) = 2n for all n. δ). . The first iteration is a single vertex v1 with a single self-loop. drawn from {v1 . δ) =    (1+δ) n(2+δ)+(1+δ) .3 Verify that Di (n) ≥ 1 for P all n ≥ i.3: The first two iterations in the construction of PAn (1. then j=i Ij is stochastically smaller than Di (n).5 Show that PAn (1.4 Verify that the attachment probabilities in (3.  Exercise 3. vn . n→∞ (3. Exercise 3. Subsequent iterations involve adding vertices one by one and linking them via a single edge to the already existing vertices with probabilities that depend on the current degrees of these vertices.2 Fix i ∈ N. .  Exercise 3. .20) with probabilities  P vn+1 → vi | PAn (1.   Di (n)+δ n(2+δ)+(1+δ) .19) are random and typically change as more vertices are added: Di (n) depends on the vertex label i and the stage of the iteration n. We add a single vertex vn+1 carrying a single edge. which amounts to a shift of the proportionality in the preferential attachment. Dn (n)} (3. Note that the parameter δ is added to the degrees. δ) is as follows.21) Note that the degrees in (3. (3. −1) consists of a self-loop at vertex v1 while each other vertex is connected to v1 by precisely one edge. with probabilities that depend on the degree of v1 (which is 2 after the first iteration). vn+1 } (3.3. . . denote the vertices of PAn (1.21) sum up to 1. . . . i = n + 1. and let {D1 (n).22) Hint: Show that if (Ij )∞ {0. so that Di (n) + δ ≥ 0 for n all n ≥ i because δ ≥ −1. . δ). .  .  Homework 3. δ).2.19) denote their degrees (a self-loop raises the degree by 2). the growth rule to obtain PAn+1 (1. . 1}-valued random j=i is a sequence of independent Pn variables with P(Ij = 1) = (1+δ)/[j(2+δ)+(1+δ)]. n. Conditionally on PAn (1. Show that   P lim Di (n) = ∞ = 1. The second iteration adds a vertex v2 and links this via a single edge either to itself or to v1 . This edge is connected to a second endpoint.

δ/m) are collapsed into a single vertex vj [m] in PAn (m. δ) is attached to vertex vj [m] with a probability proportional to the total weight of the vertices {v(j−1)m+1 . n i = n + 1. Since for each j ∈ {1. . . .25) We collapse {v1 . vmn }. etc.2 imply for m ∈ N\{1}? Simulations are shown in Figs. What do Exercises 3. vm } into a single vertex v1 [m]. an edge in PAn (m. (3. . Figure 3. . We continue by defining the model for m ∈ N\{1}. δ) = D (n)  i . NETWORK MODELS For δ = 0 the probabilities in (3. . such that the total degree is equal to 2mn. because the attachment probabilities depend linearly on the degrees of the random graph PAn (1. which uses the model for m = 1 as follows. . vjm }. we obtain PAn (m.23) i = 1. . which according to the second line of (3. . . 3.5 and Homework 3.21) is called affine. 30. n. Courtesy Remco van der Hofstad. . . collapse {vm+1 . Since the sum of the degrees of these vertices is equal to the degree of vertex vj [m]. P vn+1 → vi | PAn (1. . n. . . and for δ = −1 to P vn+1    0. . . this probability in turn is proportional to the degree . (3. . . δ/m). δ) =  Di (n)−1 .5. . δ/m) is attached to vertex vi with a probability proportional to the weight of vertex vi . After all vertices are collapsed.4: Preferential attachment random graph with m = 2. v2m } into a single vertex v2 [m].4 and 3. . . Note that an edge in PAmn (1.6 Show that PAn (m. 100. (3. . vn [m]). . → vi | PAn (1. . 2n+1 i = n + 1. . δ = 0 and n = 10.24) i = 1. . δ). (3.26) Exercise 3. and denote its vertices by {v1 . The preferential attachment mechanism in (3. vjm } in PAmn (1. δ) is a multi-graph with n vertices and mn edges. . . n} the vertices {v(j−1)m+1 . We start with PAmn (1. .21) is equal to the degree of vertex vi plus δ/m.48 CHAPTER 3. .21) simplify to  1   2n+1 . δ). . . .3–3. . . δ) with vertices (v1 [m].

It is possible to define the model with m ∈ N\{1} directly. 30. t > 0. Hence (PAn (m. Thus.δ = (τ − 1) Γ(m + δ + (τ − 1)) .2.27) xt−1 e−x dx.3 Scaling features The following two results are taken from van der Hofstad [9.29) is scale free with exponent τ = 3 + (δ/m) ∈ (2. (2) The tail behavior of fPA is given by fPA (k) = cm. In the above construction the degrees are updated each time an edge is attached. . (3. but the construction is a bit more involved. m − 1.2. . δ) n∈N is sparse with limiting degree distribution fPA given by   0. 3.3. δ) plus δ.δ k −τ [1 + O(1/k)]. Chapter 8] and makes the model more flexible. fPA (k) = Γ(k+δ) Γ(m+2+δ+(δ/m))  2+ δ  . .5: Preferential attachment random graph with m = 2. k = 0. .  (1) The random graph process (PAn (m. 100. This is referred to as intermediate updating of the degrees. which has received a lot of attention in the literature and was formally defined in Bollob´as and Riordan [6]. k ≥ m. ∞). δ = −1 and 10. . m Γ(k+3+δ+(δ/m)) Γ(m+δ) where Γ is the Gamma-function defined by Γ(t) = R∞ 0 (3. The extra parameter δ was introduced by van der Hofstad [9. PREFERENTIAL ATTACHMENT MODEL 49 Figure 3. Chapter 8] and are valid for any m ∈ N and δ > −m. The model with δ = 0 is the Barab´ asi-Albert model. Γ(m + δ) (3. Exercise 3. without the help of the model with m = 1. δ)  n∈N cm.28) with τ = 3 + (δ/m). k → ∞. δ) grows in an affine manner. With the help of partial integration and induction it can be shown that Γ(j) = (j − 1)! for j ∈ N.7 Look up the properties of t 7→ Γ(t) on Wikipedia. Courtesy Remco van der Hofstad. also PAn (m. of vertex vj [m] in PAn (m.

. as long as preferential attachment and growth are present it will maintain its hub-dominated scale-free topology.2 (= every vertex eventually sees its degree tend to infinity) not in contradiction with the fact that the degree distribution converges to fPA (= sparseness)? For m = 1 the above formulas simplify to   0. δ) n∈N is not highly clustered because the random graph is locally tree-like. preferential attachment offers a possible explanation as to why power-law degree distributions occur in realworld networks. It can also be shown that (PAn (m. the number of triangles grows much slower with n than the number of wedges.δ k −τ [1 + O(1/k)]. the proof is not easy and there is no good control on the constant K. Chapter 7] it is shown that (PAn (m. (3. NETWORK MODELS Exercise 3.. (3. n = 1. δ) n∈N is smallworld for any m ∈ N and δ > −m. Γ(3+2δ) Γ(k+3+2δ) Γ(1+δ) .δ = (2 + δ) 100000. i. 10000 10000 1000 1000 100 100 10 10 1 1 1 5 10 50 100 5001000 1 10 100 1000 Figure 3. corresponding to nfPA (k). The horizontal axis is the degree k. Courtesy Remco van der Hofstad. τ = 3 + δ. the vertical axis is the number of vertices with degree k. 000.32) Γ(1 + δ) Figure 3. Unfortunately. 000. δ = 0 and n = 300. respectively. 000..30) k ≥ 1. the scale-free topology is evidence of organising principles acting at each stage of the network formation.50 CHAPTER 3. For m ∈ N\{1} and δ ∈ (−m. 000 on a log-log scale.. the power law degree distribution is explained via a mechanism for the growth of the graph. fPA (k) =  (2 + δ) Γ(k+δ) k = 0. Therefore. As Barab´asi [2] puts it: “. unlike the configuration model..6 shows a realisation of the degree sequence of PAn (2. k → ∞.2.  In van der Hofstad [10.31) with Γ(3 + 2δ) . (. 0) it is even ultra small-world.4 Dynamic robustness The important feature of the preferential attachment model is that.e.6: The degree sequence of a preferential attachment random graph with m = 2.8 Why is the result in Homework 3. 100000. 0) for n = 300. c1. and fPA (k) = c1.) No matter how large and complex a network becomes. 3. (3. 000.” . 000 and n = 1.

For Internet it may be profitable for new routers to be connected to highly connected routers. Many more possible explanations have been given for why power laws occur in real-world networks. the idea of preferential attachment in the context of the evolution of species dates back to Yule [14] in 1925. if γ ∈ (1. also in other examples of real-world networks some form of preferential attachment is likely to be present. . Moreover. ∞). in the WWW when a new webpage is created it is more likely to link to an already popular site. all giving rise to powerlaw degree distributions. On the other hand. Indeed. since power laws are intimately related to the affineness of the attachment probabilities. While preferential attachment is natural in social networks. then there are finitely many vertices with degree > 1/(γ −1) and infinitely many vertices with degree < 1/(γ −1) (Oliviera and Spencer [13]). it turns out that if the probability for a new vertex to attach itself to an old vertex with degree k is chosen proportional to k γ with γ ∈ (0. 1). if 1/(γ − 1) is non-integer. but it is overstating the point a bit.2. then fPA falls off like a stretched exponential and scale freeness is lost (Krapivsky. and many adaptations of the above simple preferential attachment model have been studied in the literature. PREFERENTIAL ATTACHMENT MODEL 51 This is correct. In fact. Redner and Leyvraz [12]). such as Google. since these give rise to short distances. Even in biological networks some form of preferential attachment exists.3. then there is a single vertex that is connected to nearly all the other vertices (Krapivsky and Redner [11]). For example. than to the personal web page of a single individual.

monograph in preparation. Bender and E. Redner and F. Graphs with points of prescribed degrees (in Hungarian). Phys.R.R. Liggett. F. File can be downloaded from http://www. Leyvraz. Arratia and T.nl/~rhofstad/ [10] R. Bull. Gallai. London B 213 (1925) 21–87.L. Internet Math. Connectivity transitions in networks with superlinear preferential attachment.i. Math. Roy. Soc. Perseus Publishing. Yule. European J. Albert. Emergence of scaling in random networks. 2002. Rev. [9] R. Cambridge. Oliveira and J.L. tue. Canfield. [4] E.A.d.S. [8] P. monograph in preparation. Trans. Organization of growing random networks. Volume I. 85 (2000) 4629. Bollob´ as and O. [13] R.win. [3] A. A probabilistic proof of an asymptotic formula for the number of labelled regular graphs.-L.C.Bibliography [1] R. Redner. Massachusetts. Soc. File can be downloaded from http://www. The diameter of a scale-free random graph. A simple proof of the Erd˝os-Gallai theorem on graph sequences. Rev. Appl. Mat. Random Graphs and Complex Networks. 52 . Journal of Combinatorial Theory (A) 24 (1978) 296–307. Connectivity of growing random networks. Krapivsky and S. Spencer. 2 (2005)121–163. Probab. Combinatorica 24 (2004) 5–34. Choudum. [7] S.win. Barab´ asi and R. Lett. Combin. 15 (2005) 652–670. 1 (1980) 311–316. degree sequence to be graphical? Ann. Erd˝ os and T.-L. J. The asymptotic number of labelled graphs with given degree sequences. van der Hofstad. Volume II. 33 (1986) 67–70. Phil. Bollob´ as. How likely is an i. E 63 (2001) 066123. based on the conclusions of Dr. S. [14] G. [5] B. Riordan.M. Barab´ asi. Linked: The New Science of Networks.A. Austral. Willis. Lapok 11 (1960) 264–274.U. [12] P. Science 286 (1999) 509–512.nl/~rhofstad/ [11] P. [2] A. tue. Krapivsky. van der Hofstad. [6] B. A mathematical theory of evolution. Phys. Random Graphs and Complex Networks.

In some cases. In this chapter we will therefore prefer the latter choice. Other reviews presenting an empirical introduction to networks from a rather general point of view can be found in review articles [1. or undirected.1 and then discuss various empirical properties in some detail throughout Section 4. The aim is to complement the mathematical definitions of the previous chapters with a phenomenological basis and to provide a solid empirical reference for the following chapters.e. the terms network and links are more used when referring to real-world objects. 8. and at the same time it introduces a series of new definitions aimed at capturing more structural details. 3. This should not alarm readers: being aware of the existence of different quantitative expressions for the same abstract notion is actually an instructive exercise. In general. if an orientation (i. These different expressions reflect the existing variety of definitions in the scientific literature about complex networks (and in most other fields as well). Correspondingly. we first introduce some basic notions in Section 4. 2.2. 4. 9]. the links of a network can be either directed. here we need to distinguish between directed and undirected networks. these definitions (or the notation) are slightly different than the corresponding definitions we gave in Chapter 2. because. 4] and books [5. While the use of the terms graph and edge are preferred in the definition of abstract mathematical models. an arrow) is specified along them. and also as a compact summary of the most commonly observed empirical properties. 4. even if we will occasionally employ the former one as well. if no orientation is specified.Chapter 4 Network Topology In this chapter we discuss some of the most important empirical properties observed in real-world networks. 6. To this end. for instance. the whole network is denoted as directed (see Fig. 7. Most of these empirical properties are computed on real-world examples of the type discussed in Chapter 1. The chapter puts some of the definitions already introduced in Chapters 2 and 3 ‘at work’ on empirical networks.1 Basic notions First we introduce some basic definitions. and between so-called ‘local’ and ‘global’ properties. This chapter is meant as a general introduction to the structural characterization of real-world networks.1a) or 53 .

since they allow transit in both directions. for instance. but with opposite direction. a) A directed network. while in general aij 6= aji for directed networks (aij = aji = 1 if and only if the links between i and j are reciprocated). the identity of each vertex matters. Exercise 4. if n is the number of vertices. an undirected network can always be thought of as a directed one where each undirected link is replaced by two directed ones pointing in opposite directions (see Fig.1b). which is also the undirected version of network a). For this reason. b) An undirected network. NETWORK TOPOLOGY Figure 4. . we denote the adjacency matrix elements by the different symbol bij and use the definition   1 if a link between i and j is there bij ≡ (4. is there.1.54 CHAPTER 4.1c).2)  0 else.1 Write the adjacency matrices corresponding to the three networks in Fig. More precisely. For directed networks. Note that for undirected networks bij = bji . . Here all links are reciprocated. 4. . A link in a directed network is said to be reciprocated if another link between the same pair of vertices. For undirected networks. Therefore. For this reason.1: Simple examples of networks. each vertex is explicitly labelled with an integer number i = 1. Here the links between vertices 1 and 2 and between 1 and 4 are reciprocated. In a real-world network. 4. undirected links are bidirectional ones. All the topological information can then be compactly expressed by defining the n × n adjacency matrix of the network. in this case. . the adjacency matrix aij of the resulting directed network is . whose entries tell us whether a link is present between two vertices (this is what is ordinarily done. c) The directed version of network b). n. to store network data in a computable form). we denote the adjacency matrix elements by aij and define them as follows:   1 if a link from i to j is there aij ≡ (4. an undirected network can be regarded as a special case of a directed network where all links are reciprocated. undirected (see Fig.1)  0 else.  As mentioned above. each with n = 6 vertices. 4. an undirected network can be regarded as a directed one.

All these networks are regular. c) Fully connected network (mean-field approximation). different from the . but the connections are randomly established between vertices. and then consider its undirected projection. 4. A different class of regular networks is that of random regular graphs. For instance. in a highly ordered fashion. 4.1a we can obtain Fig.is also possible. 4. By contrast. 4. namely the class of (deterministic) regular networks . These are networks where all vertices are connected to the same number z of neighbours. b) Twodimensional lattice with only first-neighbour interactions. 4. It should be noted that such graphs are another example of random graphs.1a and 4.1a. the network shown in Fig. In this particular case.  Before introducing some specific real-world networks and presenting their empirical properties.1c.1 Imagine a generic directed network where not all links are reciprocated. 4.1c (by assuming that the latter is the original directed network).1a unless we are given more information. simply given by aij ≡ bij .1b and 4.and second-neighbour interactions. (4.1b (by assuming that the former is the original directed network). we briefly mention the simplest and most ‘familiar’ kind of networks that scientists from different fields have been traditionally experienced with. From Fig. aij is a symmetric matrix. and no ‘disorder’ is introduced. Test your relation on the networks shown in Figs.1b is the undirected version of that shown in Fig. and then for the networks shown in Figs.1b we can always obtain Fig. 4.4.1b. ‘familiar’ undirected networks. and vice versa. the mapping of a directed network onto an undirected one . Homework 4. a) Periodic onedimensional chain (ring) with first. BASIC NOTIONS 55 Figure 4.1.where an undirected link is placed between vertices connected by at least one directed link .2: Examples of simple. 4. Find the mathematical relation between the entries {bij } of the adjacency matrix of the projected undirected network and the entries {aij } of the adjacency matrix of the original directed network. even if in general it cannot be reversed due to a loss of information. Note that this mapping can be reversed in order to recover the original undirected network: from Fig. where all vertices still have the same number of neighbours. but from the latter we cannot go back to Fig. 4.3) where bij is that of the original undirected network.

while exhibiting a behaviour that is much richer than that of their deterministic counterparts.2 Empirical topological properties We now come to the description of various empirical topological properties of real networks. 4. we first consider the properties specified by the first neighbours of a vertex (‘first-order’ properties). second. resulting in what is sometimes referred to as the mean-field scenario. they are a particular case of the Configuration Model introduced in Chapter 3. . . Several reviews exist in the literature [1. The failure of regular networks motivates the introduction of more complex models. Here we follow an approach in which we characterize the topology of real networks progressively.and second-neighbour interactions (z = 4). along with their predictions for the dynamical behaviour of any process defined on them. Complete networks are instead used when infinite-range connections are assumed. NETWORK TOPOLOGY Erd˝ os-R´enyi model discussed in Chapter 2. find out what is the most convenient labelling of vertices in each network and describe it.2 we show three examples of regular (undirected) networks: a periodic chain with first. They are among the simplest specifications of networks and represent only a small subset of the full range of possible topological configurations. In fact. traditional assumptions such as nearest-neighbour or mean-field connections cannot be considered as good choices for most real-world networks. In this case. used whenever a set of elements is assumed to be connected to its first. The highly ordered structure of these networks translates into certain regularities of their adjacency matrices. and so on until we come to those relative to the whole network (‘global’ properties). and lth neighbours (nearest-neighbour connections). 4. . 6. nor regular networks. z = n − 1. 3. . 7] presenting this subject from various viewpoints. This problem persists even after introducing randomness in regular networks: random regular graphs. . More precisely. Chains and square lattices are particular examples of the more general class of D-dimensional discrete lattices. 5. from local to global properties. 4. z). where the degree sequence is now chosen to be the constant vector ~k = (z. Therefore.56 CHAPTER 4. Before doing that. a twodimensional lattice with only first-neighbour interactions (z = 4) and a complete network (where each vertex is connected to all the others: z = n − 1). 4. i. In Fig.e.2. Exercise 4. . then those specified by the first and second neighbours (‘second-order’ properties). . some of which have been presented in Chapter 3 and some of which will be introduced in Chapter 5. each vertex is connected to z = 2Dl other vertices. 2. are still not good models of real-world networks.2 Write the adjacency matrices for the three networks shown in Fig. or in other words without introducing randomness. .  The examples of regular networks considered above can be built deterministically. The rest of this Chapter aims at showing that real networks are not consistent with neither ER random graphs.

4.2. EMPIRICAL TOPOLOGICAL PROPERTIES

4.2.1

57

First-order properties

By ‘first-order’ properties we mean the set of topological quantities that can be
specified by starting from a vertex and considering its first neighbours. This
information is captured by simply considering the elements of the adjacency
matrix, or functions of them.
(In- and Out-)Degree
In an undirected network, the simplest first-order property is the number ki of
neighbours of a vertex i, or equivalently the number of links attached to i. The
quantity ki is called the degree of vertex i. In terms of the adjacency matrix
bij , the degree can be defined as
ki ≡

X

bij .

(4.4)

j6=i

In a directed network, it is possible to distinguish between the in-degree kiin
and the out-degree kiout of each vertex i, defined as the number of in-coming
and out-going links of i respectively. In this case, if aij denotes an entry of the
adjacency matrix, then the degrees read
X
kiin ≡
aji ,
(4.5)
j6=i

kiout

X

aij .

j6=i

In an undirected network, the n-dimensional vector ~k of vertex degrees is
called the degree sequence (recall the definition (2.1) in Chapter 2). In a
directed network, the n-dimensional vectors ~k in and ~k out are called the indegree sequence and out-degree sequence respectively.
Exercise 4.3 Calculate the (in- and out- where applicable) degree sequences of
the three networks shown in Fig. 4.1. Then calculate the degree sequences of the
three networks shown in Fig. 4.2. Describe the effects of a possible relabelling
of vertices in the two sets of networks. 
Empirical degree distribution
A very important quantity for the characterization of the first-order topological
properties of a real-world network is the (normalized) histogram of the values
{ki }ni=1 , i.e. the empirical degree distribution P (k) expressing the fraction
of vertices with degree k, or equivalently the probability that a randomly chosen
vertex has degree k (recall the definition (2.2)). In a directed network, the
corresponding objects are the empirical in-degree distribution P in (k in ) and the
empirical out-degree distribution P out (k out ).
It turns out that, for a large number of real-world networks, the empirical
degree distribution displays the power-law form
P (k) ∝ k −γ

(4.6)

58

CHAPTER 4. NETWORK TOPOLOGY

with 2 ≤ γ ≤ 3. In directed networks, the in- and out-degree distributions often
display the same form:
P in (k in ) ∝ (k in )−γin ,

P out (k out ) ∝ (k out )−γout ,

(4.7)

where γin and γout have in general different values, both typically between 2
and 3.
For the practical purpose of plotting empirical degree distributions and estimating their exponents, the (empirical) cumulative distributions are commonly
used:
X
X
X
P> (k) ≡
P (k 0 ), P>in (k in ) ≡
P in (k 0 ), P>out (k out ) ≡
P out (k 0 ).
k0 ≥k

k0 ≥kout

k0 ≥kin

(4.8)
In this way, the statistical noise is reduced by summing over k 0 .
Exercise 4.4 Plot the empirical cumulative (in- and out- where applicable)
degree distributions for the three networks shown in Fig. 4.1. 
If the empirical degree distribution has the power-law behaviour of Eq. (4.6)
or (4.7), then the empirical cumulative distributions are again power-laws, but
with an exponent reduced by one:
P> (k) ∝ k −γ+1 ,

P>in ∝ (k in )−γin +1 ,

P>out ∝ (k out )−γout +1 .

(4.9)

Homework 4.2 Prove the above statement, approximating k, k in , k out with
continuous variables. In contrast, consider an exponential empirical degree distribution of the form P (k) ∝ e−ak (with a > 0) and find the large-degree expression for the empirical cumulative distribution P> (k) (approximating k again
with a continuous variable). 
In Fig. 4.3 we show the empirical cumulative degree distribution for three
networks: a snapshot of the Internet, a protein network and a portion of the
WWW. The power-law behaviour is witnessed by their approximate straight-line
trend in log-log scale. As we mentioned, the exponent of the empirical degree
distribution of real-world networks is systematically found to be in the range
2 ≤ γ ≤ 3, a fact that is verified by all plots in the figure (where 1 ≤ γ − 1 ≤ 2).
A note on scale-free distributions
Power-law distributions are very important from a general point of view since
they lack a typical scale [13, 14]. More precisely, they are the only distributions
satisfying the scaling condition
P (ak) = f (a)P (k),

k ∈ N,

a > 0.

(4.10)

Homework 4.3 Prove the above statement. 
The functional form of a power-law distribution is therefore unchanged, apart
for a multiplicative factor, under a rescaling of the variable k. Due to this
absence of a characteristic scale, power-law distributions are also called scalefree distributions. An important consequence of scale-free behaviour is the

4.2. EMPIRICAL TOPOLOGICAL PROPERTIES

59

Figure 4.3: Empirical cumulative degree distribution for three different networks. a)
P> (k) for the Internet at the autonomous system level in 1999 [10]. b) P> (k) for
the protein interaction network of the yeast Saccaromyces cerevisiae [11]. c) P>in (kin )
for a 300 million vertex subset of the WWW in 1999 [12]. All curves are approximately straight lines in log-log scale, indicating that they are power-law distributions
(modified from ref. [4]).

presence of so-called ‘fat tails’: compared to distributions that decay (at least)
exponentially (e.g. Gaussian or Poisson distributions), power-law distribution
decay much slower and assign a much bigger probability to ‘rare events’ (the
outcomes in the tail of the distribution).
The empirical scale-free behaviour means that in real-world networks there
are predominantly many low-degree vertices but also a fair number of highdegree ones, which are connected to a significant fraction of the other vertices.
The fraction of vertices with very large degrees (i.e. ‘hubs’) is not negligible
and gives rise to a whole hierarchy of connectivities, from small to large. If we
imagine that a dynamical process takes place on the network (see for instance
Chapters 8, 9, 11, 12), the scale-free property as a remarkable effect: once the
process reaches a high-degree vertex, it then propagates to a large portion of
the entire network, resulting in extremely fast dynamics.
By contrast, note that regular networks introduced in Section 4.1 have a
delta-like ‘empirical’ degree distribution of the form P (k) = δk,z where z is
the degree of every vertex (see Fig. 4.2). For D-dimensional lattices with lthneighbours interactions, z = 2Dl and no vertex is connected to a significant
fraction of the other vertices. For fully connected networks, z = n − 1 and every
vertex is connected to all the others. In all these cases, no hierarchy is present
and the network is perfectly homogeneous.
Average degree and number of links
It is possible to consider the average degree k¯ as a single quantity characterizing the overall first-order properties of a network, and then compare different
networks with respect to it. In an undirected network the average degree can
be expressed as
P
ki
2Lu
k¯ ≡ i =
,
(4.11)
n
n
where
XX
1 XX
Lu ≡
bij =
bij
(4.12)
2 i
i j<i
j6=i

k¯out and L for the two directed networks in Fig.4 Prove that.15) k0 k¯ = X k0 k 0 P in (k 0 ) = X k 0 P out (k 0 ) (4. the mean value of the degree is finite while the variance of the degree is infinite. expressed in terms of the adjacency matrix entries.6 Using the result of Exercise 4. the average degree k¯ reads k¯ = X k 0 P (k 0 ). (4. 4.11). Exercise 4. Homework 4. in terms of the degree distribution. +∞).14) consistently from Eq.12). NETWORK TOPOLOGY is the total number of undirected links in the network. (4. for an empirical undirected degree distribution of the form P (k) ∝ k −γ where 2 < γ ≤ 3 (as in most real-world networks).  . Check that k¯ = 2Lu /n as stated in Eq. n n ≡ k¯in = ≡ k¯out P i (4. (4. and this is reflected in the requirement i 6= j in Eq. (4.12).13) where L≡ XX i aij (4. Note that in principle the total number of links also includes self-loops.60 CHAPTER 4. However.1.  For directed networks. and recall that the variance is defined as k 2 − k¯2 where k and k 2 are the first and second moment respectively]. with two directed links replacing each undirected one.1. [Hint: rewrite Eq. and our notation yields L = 2Lu as expected.3. it is easy to see that the average in-degree k¯in equals the average out-degree k¯out . calculate k¯ and Lu for the undirected network in Fig.16) k0 for undirected and directed networks respectively.3) allows us to recover Eq. (4. (4.15) by approximating P (k) with a continuous probability density defined for k ∈ (1. expressed in terms of the entries of the adjacency matrix.3.  It should be noted that we chose a different notation for Lu and L to avoid confusion when an undirected network is regarded as directed.14) j6=i is the total number of directed links. For both networks. here and in the following we assume that there are no self-loops in the network. Note that.13). check that k¯in = k¯out = L/n as stated in Eq. 4.5 Using the result of exercise 4. Exercise 4. (4. which are links starting and ending at the same vertex (corresponding to nonzero diagonal entries of the adjacency matrix). In that case the mapping described by Eq. (4. calculate k¯in . and both quantities can be expressed as k¯ kiin Pn out k L = i i = .

most real-world networks are found to be sparse. the maximum possible number of links (with  self-loops excluded) is given by the total number of vertex pairs. the case cu (n) → 0 is often referred to as the ‘sparse network’ limit.2. Ddimensional lattices display cu (n) ∝ n−1 → 0 (note that these networks are undirected. 4. Bearing this warning in mind. we conclude that all the real-world networks in Fig. 4. n(n − 1) (4. We see that networks of the same type tend to be clustered together. Therefore the link density is defined as cu ≡ 2Lu . (4.2 Second-order properties By ‘second-order’ topological properties we denote those properties which depend not only on the direct connections between a vertex and its nearest neighbours.4.19).4). In an undirected network. except the WTW which is a dense network. but also on the indirect connections from a vertex to the ‘neighbours of its neighbours’. n(n − 1) (4.4 we encountered a second-order property when we looked at the distribution of vertices at distance 2 from a given vertex ?.4 are sparse.  In the limit n → ∞. EMPIRICAL TOPOLOGICAL PROPERTIES 61 Link density The number of links is an interesting property in itself. independent of n. therefore in principle we should speak of ‘large size’ regime rather than ‘infinite size’ limit. in the limit n → ∞ of large network size.18) It is instructive to plot the values of the link density c for different realworld networks in a single figure. unlike those in Fig. 4. being a measure of the ‘density’ of connections in the network. Real networks are however of finite size.1. In graph theory.2. the latter being only a formal extrapolation of Eq. An example of such plot is shown in Fig. in a directed network the maximum number of links is given by twice the number of possible vertex pairs (since each pair can be occupied by two links with opposite directions) and the link density is therefore defined as c≡ L .5 Show that. In order to compare networks with different numbers of vertices. which is n2 = n(n − 1)/2 if the number of vertices is n. Homework 4. these limits can be defined rigorously for mathematical models (see for instance Chapter 2). 4. as a function of the number n of vertices.19) There is however one important exception: the temporal snapshots of the World Trade Web lie out of the curve and tend to have a constant link density. (4. the number of links is usually divided by its maximum possible value in order to obtain the link density. Therefore the computation of these properties involves products of two adjacency matrix elements bij bjk .4.17) By contrast. while cu (n) → cu (∞) > 0 (where cu (∞) is a finite constant) is the ‘dense network’ limit. In Section 3. . Indeed. and that all points approximately follow the trend c(n) ∝ n−1 .

real networks display a more complex behaviour and are characterized by nontrivial degree correlations which make the form of P (k 0 |k) deviate from Eq. In the trivial case with no correlation between the degrees of connected vertices. Except for the WTW. the second-order properties can be obtained in terms of the first-order ones.62 CHAPTER 4. (4. the most complete way to describe second-order topological properties is to consider the two-vertices conditional empirical degree distribution P (k 0 |k) specifying the probability that a vertex with degree k is connected to a vertex with degree k 0 .4: Link density c versus number of vertices n for several real-world directed networks. Degree-degree correlations An important example of second-order structure is given by the degree correlations: is the degree of a vertex correlated with that of its first neighbours? Statistically speaking. or in other words the conditional probability is equal to the unconditional (marginal) probability that a vertex is connected to a vertex of degree k 0 : k0 P (k 0 |k) = ¯ P (k 0 ). since P (k 0 |k) is a two-parameter curve and is affected by statistical fluctuations (however two-parameter plots of this type have been studied [15]). NETWORK TOPOLOGY Figure 4. which also partly averages out the statistical noise.20) However. all points roughly follow the dashed line n−1 . is given by defining the average nearest-neighbour degree (ANND . k (4. Average nearest-neighbour degree Estimating the empirical form of the conditional probability directly from real data is difficult.20). as we will show in the following. A more compact description.

5: Assortative and disassortative mixing in a generic network. when considered as a function of k. (4. (4.7 Compute the value of kinn for each vertex of the network shown in Fig.2. the ANND is independent of k. even though interesting by themselves. Exercise 4. These behaviours are schematically depicted in Fig.5. k¯nn (k) is an increasing function. 4.22) k0 . The slope of this curve gives information about the nature of the degree correlations: if.1b.21) ki b j6=i ij It is then possible to average kinn over all vertices with the same degree k and plot the result against k to obtain the one-parameter curve k¯nn (k). Real networks are systematically found to be either assortative or disassortative. the ANND is denoted by kinn and defined in terms of the adjacency matrix as P P P bij bjk j6=i bij kj j6=i nn P k6=j ki ≡ = .4. as measured by the increasing or decreasing trend of the average nearest-neighbour degree k¯nn (k) as a function of the degree k. 4. z). this means that the degrees are positively correlated (high-degree vertices are on average linked to high-degree ones) and the network is said to display assortative mixing. For an undirected network.  Note that in regular networks (see Section 4. EMPIRICAL TOPOLOGICAL PROPERTIES 63 <knn> assortative mixing disassortative mixing k Figure 4. In the uncorrelated or ‘neutral’ case. This means that the first-order topological properties such as the degree distribution. while if k¯nn (k) decreases the degrees are negatively correlated and the network is said to display disassortative mixing. however the k¯nn (k) plot reduces to the single point (z. still do not capture the relevant complexity of real networks. in the following) of a vertex i as the average degree of the neighbours of i.1) kinn = z ∀i and the degrees are perfectly correlated. We note that the quantity k¯nn (k) can be expressed in terms of the conditional probability P (k 0 |k) as X k¯nn (k) = k 0 P (k 0 |k).

a) ¯nn (k) plot for the 1998 snapshot of the Internet (circles). k¯nn. just like the Internet. The k¯nn (k)-curve is particularly interesting when it displays the empirical form k¯nn (k) ∝ k β . For instance.5 ). the undirected ANND obeys Eq. However.in and the average nearest-neighbour out-degree kinn. More specifically.6b [17].out (k out ) and k¯nn (k) calculated on a snapshot of the World Trade Web in the year 2000 are reported in Fig.out as P P P P aij ajk j6=i k6=j aji akj j6=i nn. 4. (4. k ¯out (kout ) portional to k−0.in (k in )-curve and the k¯nn.21) and the corresponding k¯nn (k)-curve.21) hold for directed networks as well. b) The three plots k nn ¯ and k (k) for a snapshot of the World Trade Web in 2000 (the solid line is again pro¯nn (k) curve for the subset of the undirected portional to k−0. [16]).6a) [16] and is therefore a disassortative network. From the above expression we recover the expected constant trend for the uncorrelated networks described by Eq. it is possible to define the average nearest-neighbour in-degree kinn. The power-law scaling holds for all three of them. the Internet topology displays the above trend with β = −0.in nn.out (k out )curve.in (k in ).22) yields k¯nn (k) = k 2 /k¯ independently of k.64 CHAPTER 4. the quantities k¯nn. (4. (4. (4.6: Plots of the average nearest-neighbour degree for two real networks.5 (modified from Ref. the solid line is proThe k ¯in (kin ). which inserted into Eq. (4. The inset of the same figure shows the k¯nn (k) curve computed on a subnetwork of the undirected WTW where pairs of vertices are connected only if in the original .1) and then consider the undirected ANND defined in Eq. [17]).out P P k6=j ki ≡ . NETWORK TOPOLOGY Figure 4. ki ≡ (4.20). In particular.24) j6=i aji j6=i aij respectively. the inset reports the k network defined only by the reciprocated links (after Ref. it is also possible to regard the directed network as undirected (using the mapping you have found in exercise 4. meaning that highdegree autonomous systems are on average connected to low-degree ones and viceversa.23) with β = −0. and correspondingly the k¯nn. 4. Relations similar to (4.5 (see Fig.5.23) For instance.

The number L↔ of reciprocated links can be defined in terms of the adjacency matrix as L↔ ≡ n X X aij aji . L i j<i (4. k 2 − k¯2 (4. while biological networks. which is a characteristic of directed networks. 20]. As anticipated in Section 4.27) i=1 j6=i It is interesting to compare the above expression with Eq.28) . real-world networks are generally found to be either assortative (ra > 0) or disassortative (ra < 0). a link from a vertex i to a vertex j is said to be reciprocated if the link from j to i is also present. (4. To this end. Assortativity coefficient As for the first-order properties. over all links of an undirected network. 20] as the correlation coefficient between the degrees at either ends of a link. For instance. Consistently with the analysis of the ANND curve.23). L (4. (4. The trend is similar to the other trends. of the product of the degrees of the nodes at two ends of a link: kk ≡ 1 XX bij ki kj . social networks turn out to be assortative. while it is rare to find an uncorrelated network (ra ≈ 0). one can introduce the assortativity coefficient [19. and the WTW is therefore a disassortative network in all the above representations. [17].1. it is possible to define single quantities characterizing the overall second-order properties of the network as a whole. only the pairs of vertices for which two reciprocated links exist (aij aji = 1) contribute to L↔ .26) A similar expression can be derived for directed networks. let us define kk as the average. it is possible to define the reciprocity r of the network as r≡ L↔ . As expected.2. confirms the disassortative behaviour but questions the actual occurrence of a scaling form as described by Eq.25) We can then defined the assortativity coefficient as ra ≡ kk − k¯2 . EMPIRICAL TOPOLOGICAL PROPERTIES 65 directed network they are joined by two reciprocated directed links pointing in opposite directions (see Section 4. Since 0 ≤ L↔ ≤ L.4. while each pair of connected vertices (aij = 1) gives a contribution to the number of directed links. Reciprocity We conclude our discussion of the second-order properties with the notion of reciprocity. (4. based on a more detailed data set than that used in ref.14). Another extensive analysis of the WTW [18].1). Interestingly. the WWW and the Internet turn out to be disassortative [19.

defined (for an undirected network) as the number of links connecting the neighbours of vertices i to each other. It can also be regarded as the probability of finding a link between two randomly chosen neighbours of i. which is the link density c. Ci is the link density (see Section 4. the analysis of real networks reveals interesting properties that we report below.6 For the two directed networks in Fig. Homework 4. k 00 |k) that a vertex with degree k is simultaneously connected to a vertex with degree k 0 and to a vertex with degree k 00 .2. If reciprocated links occurred by chance. the relevant quantity for an undirected network is now the three-vertices probability P (k 0 . real-world networks systematically display a nontrivial degree of reciprocity. Local clustering coefficient The most studied third-order property of a vertex i is the (local) clustering coefficient Ci . second and third neighbours.3 Third-order properties The third-order topological properties of a network are those which go ‘the next step beyond’ the second-order ones. then the number of interconnections between the . In this case too. For instance. The computation of third-order properties involves products of three adjacency matrix elements bij bjk bkl .1. and is therefore related to the occurrence of (closed) paths of three links. 4. In other words. while social networks [22]. Therefore if r = c the reciprocity structure is trivial. The measured value of r allows us to assess if the presence of reciprocated links in a network occurs completely by chance or not. Foodwebs and shareholding networks display 0 < r < c [21]. the WWW [12]. note that r represents the average probability of finding a link between two vertices already connected by the reciprocal one. since they regard the structure of the connections between a vertex and its first. while if r > c (or r < c) reciprocated links occur more (or less) often than predicted by chance. 4. where information always travels both ways along computer cables). divided by the total number of pairs of neighbours of i (therefore 0 ≤ Ci ≤ 1). Compare these two numbers and conclude whether there is a tendency towards or against reciprocation in the two networks. citation networks always display c > 0 and r = 0. calculate the reciprocity r and the link density c. To see this.  Real-world networks generally exhibit a nontrivial degree of reciprocity [21]. the World Trade Web [17. then this probability would be simply equal to the average probability of finding a link between any two vertices. Indeed.2. since recent papers can cite less recent ones while the opposite cannot occur. where all links are reciprocated (such as the Internet. and can therefore be thought of as a ‘local link density’. the extreme case c < r = 1 corresponds to (not fully connected) undirected networks. In the general language of conditional degree distributions. In conclusion.1) of the subnetwork defined by the neighbours of i. The clustering coefficient is a third-order property since it measures the number of ‘triangles’ a vertex belongs to. email networks [23].66 CHAPTER 4. Finally. NETWORK TOPOLOGY so that 0 ≤ r ≤ 1. if bij denotes an entry of the adjacency matrix of the network. 21] and cellular networks [21] generally display c < r < 1.

5).j bij bki (4.30) where it becomes manifest that the numerator counts the number of triangles in which vertex i participates. The clustering coefficient Ci is then obtained by dividing this number by the number of possible pairs of neighbours of i.1. (4. . a) Network of synonymy between English words (circles).j bij bjk bki Ci ≡ . b) The undirected versions of the World Trade Web described in Section 4. (4. 4. It follows that P P j6=i k6=i. and the denominator counts the number of wedges in which vertex i participates. it is ¯ possible to obtain a C(k)-curve whose trend gives information on the scaling of the clustering coefficient with the degree [24].  For directed networks. (4. [17]).29) can be rewritten as P P j6=i k6=i.2. [24]). Clustering coefficient versus degree A statistical way to consider the clustering properties of real networks is similar to that introduced for the degree correlations.7: Plot of the C(k)-curve for two real networks.29) ki (ki − 1) The above expression is a local (vertex-specific) version of Eq.29) holds for directed networks as well.2 (the inset shows the subnetwork with only reciprocated links). the computation of the clustering coefficient can be carried out on the undirected version of the network.7 Show that Eq. Homework 4.j bij bjk bki /2. the dashed line is proportional to k−1 (after Ref.2. Compute the value of Ci for each vertex of the network shown in Fig. which equals ki (ki − 1)/2 if ki is the degree of i.5). with bij given by the expression you found in Homework 4. (2. (2. Therefore Eq. the solid line is proportional to k−0. By computing the average value of Ci over all vertices with a given degree k and plotting it versus k. EMPIRICAL TOPOLOGICAL PROPERTIES 67 ¯ Figure 4. P P neighbours of i is given by j6=i k6=i.4.1b and compare the calculated values with the (single) value obtained using Eq. j6=i k6=i.j bij bjk bki Ci = P P . [Hint: use the fact that b2ij = bij ].7 (after Ref.

as for the k¯nn (k)-curve. C¯ ≡ n i=1 (4. finite ¯ An analysis of some real networks also reveals that for large n) value of C.8 Show that.31) For instance. Dense subnetworks can be thought of as ‘modules’ into which the whole network is subdivided.31). scaling of C/c . 4. This is shown in Fig. neighbour connections. the word network of English synonyms [24] and the aforementioned (incomplete) representation of the World Trade Web [17] display the above power-law trend with τ = 1 and τ = 0. even if its decreasing trend is confirmed.33) ¯ u ∝ n0. reduces to a single point with coordinates (z. the analysis of a more ¯ comprehensive version of the network reveals that the C(k)-plot deviates from the functional form of Eq. since it signals that the network is hierarchically organized in such a way that low-degree vertices are surrounded by highly interconnected neighbours forming dense subnetworks. most real networks are on average highly clustered. as the k¯nn (k)curve. while highdegree vertices are surrounded by loosely connected neighbours forming sparse subnetworks. Low-degree vertices are more likely to belong to such modules. for regular D-dimensional lattices with up to l-th ¯ u = 0 if l = 1 and C/c ¯ u = 3(n−1)(z−2D) ∝ n if l > 1. reporting the data with best power-law fit C/c Exercise 4. (4. Note that it is different from the (also network-wide) definition of clustering coefficient in Eq. The empirical analysis of most real networks reveals a ‘large’ (i. Average clustering coefficient It is of course possible to compute the average (network-wide) clustering coefficient C¯ over all vertices: n 1X Ci . C/c (4.e.96 .5) of Chapter 2. (2. For the WTW we note however that. C/c 4Dz(z−1)  In conclusion. ¯ By contrast. ¯ u displays an approximate linear dependence on the the rescaled quantity C/c number of vertices n: ¯ u ∝ n. Ci ).7 respectively (see Fig. and that this trend is sometimes consistent with a power-law behaviour of the form ¯ C(k) ∝ k −τ .68 CHAPTER 4. the analysis of real networks reveals that in many cases the average clustering of k-degree vertices decreases as k decreases. just like regular lattices with l > 1.32) This quantity represents the average probability to find a link between two randomly chosen neighbours of a vertex (clearly 0 ≤ C¯ ≤ 1). Both classes of networks display a qualitative linear ¯ u with n. The decrease of Ci with the degree ki is a topological property often referred to as clustering hierarchy [24]. (4. NETWORK TOPOLOGY Remarkably. while high-degree vertices are more likely to connect different modules together. 4. note that for regular networks the C(k)-curve.7).8.

the power grid. Full circles represent data from the 18 networks summarized in ref. the networks of word cooccurrence and word synonymy. the neural network of the nematode C.8: Log-log plot of the ratio between the average clustering coefficient C¯ and the link density cu as a function of the size n of the network. The solid line represents the best power-law fit to the data.2. coli. [25]). 6 scientific coauthorship data sets. the collaboration network between movie actors. the WWW.4. EMPIRICAL TOPOLOGICAL PROPERTIES 69 Figure 4.96 (modified from ref. [2]: 2 food webs. . elegans. having slope 0. 2 maps of the Internet. Empty circles represent data from 16 additional food webs [25]. the substrate network and the reaction network of the bacterium E.

two vertices i and j are said to belong to the same strongly connected component (SCC) if it is possible to go both from i to j and from j to i. while no path from j to i is there. it is possible to define the in-component of vertex i as the set of vertices from which a path exists to i. i. {z } | n factors Global properties often have the most important effect on processes taking place on networks. and the out-component of i as the set of vertices to which a path exists from i. which better characterizes the communication properties in a network. In other words. is the shortest distance between vertices. the study of higher-order properties of real networks generally goes directly to global properties.2.2 all vertices belong to the same connected component. plus a number of much smaller components containing the few remaining vertices. Since in a network with n vertices the longest path required to go from a vertex to any other vertex contains at most n − 1 links. Finally.4 CHAPTER 4. This means that the spread of information on real networks is efficient. The presence of the largest component is interesting also for theoretical reasons. .70 4. bin−1 in bin i1 . Here we consider two (out of the many) examples of global network properties: connected components and average distance.2. which are intimately related to each other.1 we have already mentioned that two vertices in an undirected network are said to belong to the same connected component (or cluster) if a path exists connecting them through a finite number of steps.34) bi1 i2 bi2 i3 . those that (at least in principle) require the exploration of the entire network to be computed.3. since they are responsible for the way information ‘spreads’ over the network and for the possible emergence of collective behaviour of vertices (some of these aspects will be covered in Part II). . The size of a connected component is the number of vertices in it. However. Shortest distance Another important property. For each . 4. since it is related to the occurrence of a phase transition in models where links are drawn with a specified probability (see Chapter 8). For directed networks. an empirical property of most real networks is the presence of one very large component containing most of the vertices.e. We have already encountered the SCC in our discussion of the WWW in Subsection 1. Connected components In Section 2. since starting from a vertex in the largest component it is possible to reach a large number of other vertices in the same component. it follows that global properties involve products of at most n adjacency matrix elements: (4. Note that for each of the regular networks shown in Fig. it is possible that a path going from a vertex i to a vertex j exists. or n if one includes loops of length n. There is in principle no limit to the number and size of connected components in a network. NETWORK TOPOLOGY Global properties Although it is in principle possible to proceed with the analysis of fourth-order properties and so on.

4. this would yield d¯ = ∞ for all networks where the (strongly) connected component does not span the entire network.13)). To prevent this outcome. This property. In such a way. Its importance lies in the remarkable deviation from the behaviour of regular networks in any Euclidean dimension D. A rough logarithmic trend is observed.36) where now i and j run over the entire set of n vertices. known as the small-world property. ln k (4. both properties are typically observed in real-world networks. It turns out that.  by considering all 2 m = m(m − 1) ‘ordered’ pairs of vertices in a (strongly) 2 connected component C of size m. dij becomes a quasimetric). As we mentioned above. the shortest distance dij is defined as the minimum number of links that must be crossed to go from i to j. Eq. the average distance is generally very small. the average distance of the component is defined as the arithmetic average XX 1 d¯ ≡ dij . Most of them are based on different versions of the . (2.2. However. via the expression d¯−1 ≡ XX 1 d−1 n(n − 1) i j ij (4.35) can be extended to the entire network by performing an average over all the n(n − 1)/2 pairs of the n vertices.9 where a plot of d¯ln k¯ versus n is reported for a set of real networks. a result that is not very informative about the differences in other aspects of the topology of such networks. we have seen various ways to define a measure of importance for the vertices in a network. is shown in Fig. the average distance can be alternatively defined as the harmonic mean over all n(n − 1) ordered pairs of vertices. meaning that d¯ scales with n according to the approximate law ln n d¯ ≈ ¯ . which instead display d¯ ∝ n1/D and are therefore characterized by a much larger average distance. The small-world effect is sometimes defined (in a ‘stronger’ sense) as the simultaneous presence of a small average distance and a large average clustering coefficient. actually. in principle the definition (4.37) The above equation is usually taken as a quantitative statement of the smallworld effect (see also Chapter 2. EMPIRICAL TOPOLOGICAL PROPERTIES 71 pair of vertices i and j in a (strongly) connected component of a (directed) network. even in a network with an extremely large number of vertices. Note that for directed networks in general dij 6= dji (which means that. Then. Then. (4.35) m(m − 1) i∈C j∈C The shortest distance between two vertices belonging to different (strongly) connected components can be formally defined as infinite. 4. d¯ will be finite even for networks where the (strongly) connected component does not coincide with the whole network and its value will discriminate among different topologies. Betweenness centrality So far. The empirical behaviour of d¯ is very important.

the contribution to the betweenness centrality of a third vertex i will be the number of shortest paths (between the two given vertices) that pass through i. and more sparsely connected among each other.10. [2] (see the cited reference for the symbol legend). Another widely used. the vertex that is crossed the most times is also the most central in terms of its betweenness. Whenever two or more shortest paths of equal length exist between the same two vertices. vertices with high betweenness play the role of ‘bridges’ across regions of the network that are highly connected internally. The dashed line represents the curve ln n. The betweenness of a vertex i in a network measures the number of shortest paths between all possible pairs of vertices that pass through i [8]. Formally. divided by the total number of shortest paths between the two vertices. law d¯ = ln n/ ln k. along with several ‘internal’ vertices with . NETWORK TOPOLOGY Figure 4. a few ‘bridging’ vertices with very high betweenness are typically detected.i where the sums run from 1 to the total number n of vertices and one must take care that j is different from k and that both j and k are different from i. we can write the betweenness Bi of a vertex i as X X Njk (i) Bi = . showing that real data approximately follow the ¯ even if with some exceptions (modified from Ref. In real-world networks. while Njk (i) counts how many such paths pass through i.72 CHAPTER 4. but non-local. (4. As a result.38) Njk k6=i j6=k.9: Log-linear plot of the product between the average distance d¯ and the ¯ as a function of number n of vertices for a set of logarithm of the average degree ln k real networks studied in Ref. The quantity Njk counts the total number of shortest paths between j and k. concept of centrality: a vertex is more important if it is ‘close’ to many other vertices. choice is the notion of betweenness centrality (or betweenness for short). [2]). the presence of such regions is typically observed (see next subsection). We have already considered the degree as a direct and completely local measure of centrality in terms of the number of first neighbours of a vertex. Correspondingly. As shown in the example in Fig. 4.

10: The betweenness of the central black vertex is computed by considering all shortest paths (distances) between all the possible pairs of vertices. while others do not. i. a mathematical model where some topological property is taken as input from the data. one of which contributes 1/2 to the betweenness of the black vertex and one of which does not contribute to it.38)) and iteratively removing the nodes (or links) with the largest betweenness. while others do not. as it typically requires the calculation of quantities that require repeated iterations across the whole network. For instance. and even when a single definition is adopted. or higher than an expected value (obtained under certain null hypotheses).4. This method starts by computing the betweenness of all nodes (or of all links. Between the two grey vertices (A and B) in the figure there are two different shortest paths. there are various methods to identify the communities of a particular network [26]. Communities are subsets of vertices whose internal link density is higher than the average density across the entire network. Community structure Very important structures that can be identified in a network are communities of densely connected vertices. A community can consist of any number of vertices (from a few vertices up to a large fraction of the network).e.2. but where communities are absent by construction [26]. some definitions allow for overlapping communities that share one or more vertices. A simple approach employs the concept of betweenness centrality (see previous subsection) to define and detect communities in large networks [26]. The ‘best partition’ of the network into communities is sought for by maximizing a so-called modularity function defined as a sort of ‘difference’ between the real network and its null model. recalculating all the values of the betweenness after each removal. and a network can therefore be partitioned into heterogeneously sized communities. In such a way. some definitions allow for hierarchical communities that can be further resolved into smaller sub-communities. Other methods are based on the comparison of the real network with a null model. Similarly. There is no unique definition of a community. Detecting communities is a non-local task. via an appropriate modification of the definition (4. Some null models of networks will be introduced . lower betweenness. EMPIRICAL TOPOLOGICAL PROPERTIES A 73 B Figure 4. ‘the bridges between communities are cut’ and the network gets partitioned hierarchically into smaller and smaller communities.

We see that the community detection method identifies three communities that largely reflect the political viewpoints of the books: “liberal” (circles).14 in Chapter 1 for a network of books (about US politics) that are frequently co-purchased online. The particular method used to resolve the communities in the figure is based on nonnegative matrix factorization [27].74 CHAPTER 4. . The communities are allowed to overlap (some vertices in the figure are of mixed colour). The method also detects vertices (coloured in pink) that are “outliers”.e. i. Real-world networks typically display strong community structure. is a convenient and widely used null model to detect communities. that do not belong to any community. Vertices in the same community are assigned the same colour. NETWORK TOPOLOGY in Chapters 5 and 10. Other methods are based on the spectral properties of empirical adjacency matrices or on matrix algebra [26]. if implemented in such a way that the empirical degree sequence of the real-world network can be taken as input. 1. An example is provided in Fig. “neutral” (triangles) and “conservative” (squares). The Configuration Model introduced in Chapter 3.

Vespignani. 46(5). and A.I. 87. Advances in Physics 51. Wiener. 323-351. [2] R. Freeman.F. Garlaschelli and M. [15] S. Nature 410. R. (2007). S. [12] A. Power laws. Oxford University Press. Rajagopalan.Bibliography [1] S. A. Newman. Mandelbrot. Physica A 333. 75 .-L.H. Mod. A. Norton. Buchanan. [5] A. Albert and A. 258701 (2001). Barab´asi and Z. Tomkins and J. Nexus: Small Worlds and the Ground. S. Rev. Shenker and W.B. R. Kumar. G. Rev. New York (2003). Barab´ asi. S. Maslov.-L. 309 (2000). Bogu˜ [17] M. Mendes. Maghoul. Lett. Chang. E 68. [9] Newman. 41 (2001). Norton. [13] Newman. Phys. MA (2002). Dorogovtsev and J. Govindan.J. Sneppen. E. Lett. [3] S. Willinger.breaking Science of Networks. [14] B. Jeong. [6] M. Chen. Scale-free networks: complex webs in nature and technology. Watts. Broder. Networks: an introduction. [18] D. Pastor-Satorras. (2010). Pareto distributions and Zipf’s law. [4] M. Loffredo. A.J. Jamin. 47 (2002). R. F. [16] R.-L. S.J.F. 1079 (2002). 188701 (2004). SIAM Review 45. 93. 015101(R) (2003). Contemporary physics. San Francisco (1983). Zaliznyak. 529-540 (2004). P.N. Computer Networks 33. M. n´ a. Six Degrees: The Science of a Connected Age. IEEE Computer Society (2002). Rev. Rev. Phys. ´ Serrano and M. Proceedings of the 21st Annual Joint Conference of the IEEE Computer and Communications Societies.N. 74. [7] D. Barab´ asi. K. Phys. Cambridge. Oltvai. 167 (2003).E. Raghavan.A. Linked: The New Science of Networks. [11] H. 268 (2001). V´ azquez and A. The Fractal Geometry of Nature. Oxford University Press. M. New York (2002). Strogatz. Phys. H. Stata. Nature 411. Perseus. [8] Caldarelli. Mason. (2005). [10] Q.

[26] Fortunato. D. Rev. Natl.J.76 BIBLIOGRAPHY [19] M. [27] X. Dunne. Rev. Cao. Proc. Newman. Martinez. 026112 (2003). Lett. Phys.J. Newman. Sci. Acad.D. 208701 (2002). E 67. Phys.A. R. E 66. [23] M. S. . Forrest and J.E.E. 268701 (2004). 89. Cambridge. Wasserman and K. 93. Newman. (2010). 2993 (2013). [25] J. Phys. 1994). Williams and N. Y.E. Social Network Analysis (Cambridge University Press. Jin. He. USA 99. Faust. [20] M. [21] D. S. 026126 (2003). X.I. Ravasz and A. E 67. Phys. 75-174.-L. Community detection in graphs. 12917 (2002). 486(3). Barab´asi. Lett. Cao & D. Rev. Phys. Physics Reports. [22] S.J. [24] E. Garlaschelli and M. Rev. Balthrop. Loffredo. Wang. Scientific Reports 3. 035101(R) (2002).J. Rev.

e. Since. adjacency matrices). they sample the space of graphs uniformly (under some constraint). Starting with some preliminary observations about the Erd˝os-R´enyi random graph and the Configuration Model introduced in Chapters 2 and 3. for these two classes of networks. All these models have one feature in common: they are stochastic. (4. In this Chapter we study graph ensembles in more detail.e. respectively. GM } of M graphs (i. for our purposes here.1) and 4. respectively.e. This leads to unbiased expectations for the higher-order properties of a network. . . . Such ensemble is a collection1 G ≡ {G1 . any graph G is uniquely specified by its adjacency matrix. we gradually arrive at the definition of so-called maximum-entropy ensembles of networks.3). . the symbol G denotes a different family of graphs than that denoted in Chapter 2. If we fix the parameters of a stochastic network model. (5. graphs) of the model itself define a so-called ensemble of random graphs. 1 Note that. starting from local information. where each graph Ga is assigned a probability P(Ga ) such that X G∈G P(G) = M X P(Ga ) = 1. The importance of maximum-entropy ensembles is the fact that.Chapter 5 Network Ensembles In Chapters 2 and 3 we have already encountered various network models. and generalizes the notation aij and bij introduced in Eqs.1) a=1 The number M is known as the cardinality of the ensemble G. This is property if of crucial importance not only for the theoretical reason of redefining the Configuration Model in order to allow for strongly heterogeneous degree sequences (as we show in Section 5. but also for the practical problem of pattern detection in real-world networks (as we will discuss in Chapter 10). i. we may think of G as an adjacency matrix with entries gij . 77 . as we discussed in Chapter 4. they are based on some degree of randomness.2. all the possible realizations (i. This notation applies to both directed and undirected graphs.

2)  0 i=j P where. The proof of this result is the goal of the following series of homeworks. they are equiprobable. from the results of Chapter 4. and we cannot use the results obtained for n → ∞ in Chapter 2. (5. Using the adjacency matrix notation introduced in Chapter 4. once n is fixed. each given value of p produces a corresponding expected number of links. Given a set of n vertices.4). where the sum runs over all the possible values of Lu in Lu the ensemble of graphs with n vertices. Check that P M (L n u ) = Mn . It therefore follows from Eq. each pair of vertices is connected by an undirected link with probability p (independently of all other pairs). given p. In this perspective. we want to compare the predictions of the ER model with the observed properties of a real-world network with a given number n of vertices and a given number Lu of undirected links.4) This strategy is useful if. Note that n will be necessarily finite. the ER model has an important and desirable property: all the graphs with the same value of n and Lu are generated with the same probability. if we would like the network to have a given expected number of links E(Lu ). NETWORK ENSEMBLES Equiprobability in the Erd˝ os-R´ enyi model To realize the importance of unbiasedness. and the model is fitted to these constraints by choosing the same n as in the real network and p as in Eq. We already know. for instance. Homework 5. However. This implies that. we restate the model as follows. We might reverse the point of view and say that. i. when n is the (fixed) number of vertices.78 5. 2 (5. the expected value of the entry gij of the n × n adjacency matrix of a graph G generated by the model is   p i 6= j E(gij ) ≡ hgij i = (5.1 Write the cardinality Mn of the ER ensemble. Calculate the number Mn (Lu ) of simple and undirected graphs (without self-loops) with n vertices and Lu edges. here and in what follows. with E(Lu ) ≡ Lu .1 CHAPTER 5.e. the expectation value x xP(X = x) of a (discrete) random variable X is denoted by E(X) or hXi.12) that the expected number of undirected links is E(Lu ) = p n(n − 1) .  . No self-loops are created.3) So. in particular their broad degree distribution and their large clustering. we first highlight some important properties of the Erd˝ os-R´enyi (ER for short) model introduced in Chapter 2. large real-world networks imply that many asymptotic results will hold at least approximately. that the comparison between the real network and the ER model will be unsuccessful: the ER model is not able to reproduce many properties of most real-world networks. (4. n(n − 1) (5. the empirical values of n and Lu are treated as ‘constraints’. then we should set the probability p to the corresponding value p= 2E(Lu ) . However.

5. In Section 5.12) adapted to the notation used here. 5. e. Let us consider undirected networks first. Write the resulting expression for P(G) as a function of p. in the ER model. The degree sequence can be picked from any desired degree distribution P (k).2 Calculate the probability P(G) to generate a particular graph with binary adjacency matrix G as a function of p in the ER model (write your answer explicitly in terms of the entries P{gij } of the matrix G). Now the idea is to generate an ensemble of random directed graphs in such a way that the desired in. the CM does not make explicit hypotheses on how networks organize themselves in a given structure.  Homework 5. generating a suitably randomized ensemble of graphs once some low-level information is assumed as an input. However. We then focus on the specific implementations that have been proposed to realize this idea.3 in the context of maximum-entropy ensembles of graphs. the degree sequence must be graphical. (4. Note that.and out-degree sequences are simultaneously realized. IMPLEMENTATIONS OF THE CONFIGURATION MODEL 79 Homework 5. calls for a more sophisticated implementation of this model. 2] (CM for short) introduced in Chapter 3.  Another important property of the ER model is that its natural implementation (connecting each pair of initially disconnected vertices with probability p.g. n and Lu (G) explicitly. which will be further illustrated in Chapter 7. Show that P(G) depends on G only through Lu (G) ≡ i<j gij . the CM is easily generalized by assigning each vertex i a given in-degree kiin and a given out-degree kiout .3 The above homeworks allow you to conclude that. This limitation. i.2 Implementations of the Configuration Model With this section we return to the so-called Configuration Model [15. Use this result to write the probability of generating any graph with n vertices and Lu undirected links.1. It rather is a null model.2 we extend the discussion of Chapter 3 by showing that both the equiprobability and the feasibility property are violated in the simplest implementations of the Configuration Model. it is possible to define computational algorithms that do not ‘get stuck’ and always lead to a realization of the model (see Chapter 6). by drawing links at random between vertices in such a way that the desired degree sequence is realized. all graphs with the same value of n and Lu are generated with equal probability. as discussed in Chapter 3 (Subsection 3. i. For directed networks. among the possible Mn (Lu ) ones. The idea of the CM model is to assign each vertex i a desired degree ki and then generate the ensemble of all graphs compatible with the resulting degree sequence {ki }ni=1 . or equivalently disconnecting each pair of initially connected vertices with probability 1 − p) is always feasible.e. Note that the latter expression coincides with Eq. We first recall the abstract idea behind the model and make general considerations.2. realizable by at least one graph.e. a scale-free one with desired exponent. 1. unlike the preferential attachment model (see Chapter 3).3). This is provided in Section 5. As . All graphs compatible with the desired degree sequence should be realized with the same probability.

then the observed deviations from the model indicate interesting structural patterns that cannot be traced back to the null hypothesis (i.2. In what follows. with the aim of realizing a random graph with the desired degree sequence.e. As already anticipated in Chapter 3. (the simplest) canonical ensembles are analytically tractable and unbiased. not simply their expected values as in the ER model (see Chapter 2). NETWORK ENSEMBLES we will see in detail in Chapter 10. It turns out that microcanonical ensembles are much more difficult to deal with mathematically and are prone to bias. the number of links Lu = i ki /2 is automatically fixed as well.e. they are not explained by the degree sequence alone). Therefore the actual values of the number of links and of the vertex degrees are fixed. then one can conclude that the observed higher-order properties are a mere outcome of the specified form of the degree distribution.e. In a similar manner. If these extra links are not discarded. By contrast. we are going to describe various implementations of the CM model. and samples the graph ensemble defined by the given degree sequence. This implementation of the CM is therefore microcanonical. it makes a big difference whether the constraint is soft. If this is the case. as many link stubs (half links) as the prescribed degree ki are initially attached to each vertex i. enforced on each individual realization separately (as in the implementation of the CM model discussed in Chapter 3). while in the second case one speaks of microcanonical ensembles. Here.80 CHAPTER 5. If this is not the case. In other words. a problem with the above implementation of the CM is that it typically gives rise to undesired self-loops and multiple links between two vertices. in the CM model the constraint is the entire degree sequence {ki }ni=1 . enforced only as an average over all realizations (as in the ER) or sharp. i. see Chapter 4) as the real network. 5. the resulting . iterating this process generates as many realizations of the network as desired. the CM can also be used as a benchmark for empirical data: a comparison between the CM and a real-world network allows us to check whether some of the higher-order properties observed in the real-world network are consistent with those generated by the CM using the same degree sequence (which is a first-order property. which is specified by fixing the connection probability p. It is important to notice that. the ensemble of networks generated by the CM model is completely random except for the degree sequence. In principle. the CM model can be regarded as a generalization of the ER model in the following sense. while in the ER model the only constraint (besides the number n of vertices) is given by the number of links Lu . once the desired degree sequence is specified. being consistent with a random assignment of links compatible with the degree sequence.1 Link stub reconnection The implementation of the CM that we already discussed in Chapter 3 is the so-called ‘link stub’ reconnection process. i. The ensemble of networks generated by the ER model is completely random except for the number of links. Note P that. In the first case one speaks of canonical ensembles. However. These variants will gradually lead us from microcanonical to canonical implementations of the model. All these stubs are then randomly matched. in the above abstract formulation. which is specified from the beginning. as we will show below.

Computationally. the link stub reconnection method is unsatisfactory when broad empirical degree sequences are taken as an input. thereby generating an ensemble of random graphs with the same degree sequence as the original network. 5. IMPLEMENTATIONS OF THE CONFIGURATION MODEL 81 Figure 5. In this way an ensemble of randomized networks is generated. one can identify degree sequences for which these problems can be avoided (e. like in the link stub reconnection . having the same degree sequence as the original network and no multiple links or self-loops. Two edges.1b for undirected and directed networks. Sneppen and Zaliznyak [15] proposed to start from the whole real-world network (not just its degree sequence) and iteratively randomize it in a degree-preserving manner. Note that the degree of each vertex is unchanged (in the directed case. if one rejects the attempted matching of two link stubs that would result in double links or selfloops.2 The local rewiring algorithm In an attempt to overcome the above limitation.1a and Fig.1: Elementary step of the local rewiring algorithm for a) undirected and b) directed networks.g. the in. here (A. C). for sparse graphs in the asymptotic limit n → ∞. Note that this ensemble is microcanonical.2. C) and (D.5. respectively: two links are randomly chosen from the initial graph G1 and the vertices at their ends are exchanged in such a way that the new graph G2 has the same degree sequence of the initial one. since this comparison might highlight patterns that are merely due to the differences between the two topological classes. then the above problems cannot be easily avoided. then the step is aborted and two different links are randomly chosen again. If double links and self-loops are deliberately removed. The local rewiring algorithm consists of the iteration of the elementary step shown in Fig. At a mathematical level. are randomly chosen from graph G1 and the vertices at their ends are exchanged to obtain the edges (A. Their randomization process consists in what they call the local rewiring algorithm. graph ensemble cannot be consistently compared with real networks that do not admit such kinds of links.2. some important variants of the CM have been introduced. the algorithm will typically ‘get stuck’ in configurations where largedegree vertices have no more eligible partners to connect to (see Chapter 7).and out-degrees are separately conserved). see Chapter 3). in practice. if one wants to define the CM using the empirical degree sequence of a real-world network. If the ‘new’ links already exist in the network. However. then the original degree sequence can no longer be realized. B) and (D. B) in graph G2 . So. Maslov. 5. which deliberately avoids the occurrence of multiple links and selfloops in the randomized networks. operatively. 5.

This important point highlighted by Maslov. given the degree sequence. Indeed.82 CHAPTER 5. Roughly speaking.2) can be accounted for in this way. it does not explore the space of graphs compatible with the degree constraints uniformly [3.3 The Chung-Lu model Microcanonical ensembles are difficult to deal with analytically. The main aim of their approach is that of obtaining an explicit mathematical expression for the probability that two vertices are connected in the ensemble. (5. More precisely. because the ensemble averages of the degrees converge to their observed values: P X j kj E(ki ) = pij = ki = ki . However. NETWORK ENSEMBLES method. In order to overcome this problem. while other patterns such as the clustering properties are instead genuine [15]. one should introduce a suitable ‘acceptance probability’ for each attempted configuration. but the whole expected degree sequence is specified.2. as a natural extension of the ER model. Chung and Lu reinterpreted the configuration model in a canonical form. since two vertices cannot be connected more than once. or equivalently the probability that two vertices are connected. 4]. which require the comparison between a real-world network and some implementation of the CM. 2Lu (5. kj are the observed degrees of vertices i. Sneppen and Zaliznyak led them to show that much of the disassortativity observed in the Internet (see Section 4.5) where Lu is the observed number of links and ki . in the case of an undirected network.2.6) 2Lu j . a link between two vertices i and j is drawn with probability pij = ki kj . Howeverm. the calculation of this probability is computationally demanding because it depends on the current configuration. Recently. it has been proved mathematically that the local rewiring algorithm is biased. In particular. while they might be present in the link stub reconnection method. in the local rewiring algorithm the presence of links between high-degree vertices is suppressed. the root of the problem is the fact that the algorithm explores with higher probability the graph configurations that are ‘closer’ to the orginal network. j. determining a certain degree of ‘spurious’ disassortativity which is not due to a ‘basic’ anticorrelation between vertex degrees. it is not obvious how to write down an exact for the probability of occurrence of a given graph G in the ensemble. will be made extensively in Chapter 10. and should therefore be repeated at each step of the algorithm. not only the expected number of links. Chung and Lu [14] proposed a completely different variant of the CM where the graph ensemble is turned from microcanonical into canonical. since double links and selfloops are absent here. the two ensembles are different. Considerations of this type.e. in the implementations of the CM discussed so far. This choice is (at least apparently) reasonable. Since imposing a fixed number of links seriously complicates the analytical calculations in the microcanonical approach. i. 5. In this extension.

This choice ensures that hkiin i = kiin and hkiout i = kiout for all vertices. however. L (5. by establishing a directed link from vertex i to vertex j with probability pij = kiout kjin . and use the result to approximate the expected reciprocity hri in the Chung-Lu model. (5. Compare this value with the reciprocity of a random graph with the same number of vertices and directed links. Discuss the effects of the heterogeneity (i.2).8) to write the expected number E(L↔ ) ≡ hL↔ i of reciprocated links (see Chapter 4).7) is not entirely correct. Find out why and try to write a more refined expression.e. Homework 5. generalizing Eq. respectively. (5.6) is not entirely correct. Section 4. (5. that this is not entirely correct and subject to the same limitation discussed in Homework 5. since each pair of (distinct) vertices is considered only once.2. IMPLEMENTATIONS OF THE CONFIGURATION MODEL 83 Homework 5.  Note that the model can be formulated also for directed graphs.  Note that the factorized form of pij in Eq. Discuss in which limit the correct expression reduces to Eq.2.5 Using the correct expression.4.  The Chung-Lu model avoids by construction the occurrence of multiple links and self-loops. Discuss these effects for a scale-free degree distribution of the form P (k) ∝ k −γ with 2 < γ ≤ 3. Try to write a similar expression for the expected value E(Ci ) of the clustering coefficient of vertex i (see definition in Chapter 4. (5. kjin are the observed out-degree of vertex i and the observed in-degree of vertex j. Rearrange this expression to write the difference Lu − hLu i as a function of only the first and second moments of the degree distribution P (k) of the real-world network. j in Eqs. we are forced to consider only those degree sequences satisfying the constraint v uX p u n max{ki } ≤ 2Lu = t kj . (5.2.3). (5.8). and kiout .7) E(ki ) ≈ ki j kj which is independent of i and has the expected form k 2 /k¯ valid for uncorrelated networks (see Section 4. write the relationship between the observed number of links Lu and the expected number of links hLu i generated by the Chung-Lu model.4 Equation (5.6 Use Eq. Note.  Homework 5. the breadth) of the degree distribution.6).5.5) and (5.9) i j=1 . Exercise 5. Find out why and write the corresponding correct expression.6).8) P P where L = i kiin = i kiout is the observed number of directed links.5) also implies that no degree correlations are introduced: the expected average nearest-neighbour degree (see Chapter 4) reads P 2 P j kj j pij kj nn = P . (5. However. to ensure (as we should) that 0 ≤ pij ≤ 1 for all i.1 Equation (5.

Exercise 5.6) is a good approximation to the correct expression you found in exercise 5.  5. even if they turn out to be in general very different from the expected degrees {hki i}ni=1 and are therefore denoted by a different symbol. The quantities {xi }ni=1 play a role similar to that of the ‘desired’ degrees {ki }ni=1 in the Chung-Lu model. (5. Park and Newman started from the general problem of finding the form of the connection probability pij that generates a canonical ensemble of graphs with no multiple links and such that two graphs with the same degree sequence are equiprobable. i. and check whether the condition (5. (5. A connection ‘probability’ pij > 1 can be regarded as the establishment of multiple links between i and j.4 The Park-Newman model The limitation of the Chung-Lu model led Park and Newman [68] to modify the canonical approach in such a way that no restriction on the desired degree sequence is imposed. This limitation prevents us from using the Chung-Lu model for most empirical degree sequences. Discuss what you obtain if you use Eq.2 Consider a ‘marginal’ degree distribution where maxi {ki } = √ 2Lu . and this possibility is avoided only by imposing the above constraint. or . the violation of Eq. we want the connection probability to be a function pij = p(xi . (5.  Homework 5. (5.9) holds. where a central vertex is connected to all the other vertices (and these vertices are not directly connected to each other). As for the Chung-Lu model.4. in the general spirit of the Configuration Model.9) is very strong and is violated by most empirical degree distributions where a few hubs with very large degree are present. of either pij (if the link is realized. Use your result in Homework 5.84 CHAPTER 5. the problem of the Chung-Lu method is in some sense the canonical counterpart of the problem encountered in the link stub reconnection method.2. over all pairs of vertices.e.9) can be thought of as leading to multiple links.7 Consider a regular network with n vertices where ki = z ∀i and check whether the condition (5. the constraint expressed by Eq.5) to generate the graph ensemble in this case. NETWORK ENSEMBLES √ √ and similarly maxi {kiin } ≤ L and maxi {kiout } ≤ L for directed graphs. (5.9) holds. Unfortunately. pij = 1). in the Chung-Lu model the problem of the occurrence of multiple links is circumvented by restricting the possible degree sequences to those satisfying Eq. Discuss whether this is enough to ensure that Eq. The starting point is to write the probability P(G) of occurrence of a given graph G (with adjacency matrix entries {gij }) in the ensemble as a product. and at the same time no multiple links are generated. xj ) of some quantities xi . Discuss what you obtain if you use Eq. as we mentioned.5 to write the expression for the difference Lu − hLu i in this case. xj controlling the expected degrees of vertices i and j. Therefore. (5.  Homework 5.8 Now consider a star graph with n vertices.9). Since.5) to generate the graph ensemble in this case.

1). For the above expression to hold for all quadruples A. . D. but in a trivial way.e.3) to the CM.5. Rearranging for pij . 1 + f (xi )f (xj ) (5. they are sampled uniformly. C.e. This yields p(xi . Looking again at Fig. implying that graphs with the same degree sequence are equiprobable. i.1-5. Moreover. This nicely extends the properties we proved for the ER (see exercises 5. thus overcoming the limitation of the Chung-Lu approach.2. we have pij = p(xi . xj ). IMPLEMENTATIONS OF THE CONFIGURATION MODEL (1 − pij ) (if the link is not realized. The above expression can be used to find the form of pij warranting that two graphs G1 and G2 with the same degree sequence are equiprobable. Since different choices can be mapped to each other via a redefinition of x. it ensures that the desired property of unbiasedness. 1 + xi xj (5. (5. 5.12) Any form of f (x) is compatible with the requirement in Eq. the requirement that the graphs G1 and G2 occur with the same probability P(G1 ) = P(G2 ) translates into the requirement pDC pAC pDB pAB = . 1 − pij Q where P0 ≡ i<j (1 − pij ) is a product over all vertex pairs and is therefore independent of the particular graph G. where we lose control over the expected degree sequence. In other words.11). xj ) = f (xi )f (xj ) . B. It ensures that 0 ≤ pij ≤ 1 with no restriction on the degree sequence. Recalling that pij = p(xi .13) Equation (5. we see that fi = f (xi ). So the simplest nontrival choice for the CM is f (x) = x. xj ) = xi xj . 1 − pAB 1 − pDC 1 − pAC 1 − pDB (5. the form of pij must be such that pij /(1 − pij ) = fi fj .9 Show P that the graph probability P(G) depends on G only through {ki (G)} ≡ { j6=i gij }. B. This implies that graphs with the same degree sequence are equiprobable. C. is extended to the CM. gij = 0): Y Y P(G) = pij i<j|gij =1 = Y (5.13) is of fundamental importance.10) i<j|gij =0 (1 − pij ) i<j = P0 (1 − pij ) 85 Y i<j|gij =1 Y i<j|gij =1 pij 1 − pij pij .11) since the two graphs are identical except for the subgraphs defined by the four vertices A. D. Show that this is not true if the probability pij of the Chung-Lu model is used. Homework 5.1. since in the ER model actually any two graphs with the same number of links are equiprobable (see Section 5. This still ensures that any two graphs with the same degree sequence are equiprobable. we can choose the simplest nontrivial2 function f (x) = x for later convenience. i.  2 Note that the simplest choice f (x) = const would lead to the result p ij = const as in the ER random graph model. where fi is a quantity depending on i alone. that we proved for the ER model. the probability to generate a graph G depends only on the degree sequence of G.

Secondly.15) 1 + xy 0 The behaviour of hk(x)i is proportional to x for small values of x and then ‘saturates’ to the maximum value n − 1 for large x. The trend is initially linear and then saturates to the asymptotic value k → n − 1 (atfer ref. They found that this assumption has two important consequences on the topology: firstly. the above expression is intended as a system of n nonlinear coupled equations where {hki i} = {ki } are known quantities and {xi } are the unknowns (we will consider this case explicitly in Chapter 10). the absence . or to find the particular values of {xi } that produce an expected degree sequence equal to the observed one.2: Average degree hk(x)i of vertices versus x corresponding to the choice σ(x) ∝ x−τ for three values of τ . consistently with the requirement of no multiple or self loops. In the latter case. This can be best appreciated by considering a continuous approximation where the distribution of x over all vertices is assumed to be the continuous density σ(x): Z ∞ xy Ex (k) ≡ hk(x)i = (n − 1) dy σ(y).3a). after integrating out the variables for the other vertices.2). Equation (5. Park and Newman [68] studied the model assuming a power-law distribution σ(x) ∝ x−τ with various values of the exponent τ (see Fig.86 CHAPTER 5.3b). [68]). As expected. NETWORK ENSEMBLES Figure 5. (5. 5. The expected degrees read E(ki ) = X j pij = X j xi xj 1 + xi xj ∀i (5. 5.14) and the above expression can be used to either generate an ensemble of networks with degree distribution determined by the (free) parameters {xi }. but then diplays a cut-off ensuring k ≤ n − 1 (see Fig. We now come to an important result. 5. the degree distribution P (k) behaves as a power-law with the same exponent τ of σ(x) for small values of x.14) shows that the expected degree of a vertex with a given value of x can be written as a function of x. the average nearest neighbour degree turns out to be a decreasing function of the degree (see Fig.

and gij = 0 if not.11) can be approximated as p(xi . The Park-Newman model leads to a very interesting analogy with the statistical physics of quantum particles. the ‘occupation number’ of a microstate in a system of fermions can be only 0 (no particles in that state) or 1 (one particle in that state). the system of equations (5. resulting in some ‘spurious’ disassortativity. In quantum physics.2. The requirement of no multiple links (gij ≤ 1) is equivalent to the exclusion principle that there is at most one particle per state. 5.14) decou√ ples and the solution is xi = ki / 2Lu ∀i.3b). it is found that each microstate of a system can only be occupied by one fermion at a time (this is known as the exclusion principle).10 Let us consider the ‘sparse graph limit’ where xi xj  1 ∀i.2. Eq. IMPLEMENTATIONS OF THE CONFIGURATION MODEL 87 Figure 5. and leads to Eq. Here isolated symbols correspond to numerical simulations. In this regime.3: a) Cumulative degree distribution P> (k) corresponding to the choice ¯nn (k) for the σ(x) ∝ x−τ for three values of τ . Homework 5.11) which is the analogous of the so-called Fermi distribution. The analogy relates links of binary graphs to so-called Fermi particles or ‘fermions’. [68]).5. Moreover. as suggested by Maslov et al. xj ) ≈ xi xj . [15] by making use of the local rewiring algorithm (see Section 5. even if it approaches a power-law behaviour asymptotically (see Fig.  . These results allowed Park and Newman to confirm that. (5. microscopic particles that obey quantum mechanics and are therefore subject to some ‘discreteness’ constraint. b) Average nearest neighbour degree k same three choices of the exponent τ . while solid lines are the analytical predictions (after ref. In jargon. (5.16) Show that. each pair of vertices i and j can be regarded as a ‘microstate’ whose occupation number is gij = 1 if a ‘particle’ (link) is there. In this analogy.e. i. j. the studied mechanism generates a k nn (k) which is not strictly a power-law.2). of multiple links generates an ‘effective repulsion’ between high-degree vertices. part of the disassortativity displayed by the Internet can be accounted for by this mechanism. (5. with the above approximation.

18) which is equivalent to the√directed version √ of the Chung-Lu model defined in Eq. (5. we are led to a condition analogous to Eq. we might ask a natural question: if we consider a different constraint (other than the number of links or the degree sequence). and no degree correlations are introduced. with xi = kiout / L and yj = kjin / L.and in-degrees kiout .8). . how can we be sure that we end up with an appropriate method to generate the ensemble? Is there some constructive method to generate graph ensembles with given constraints? In this section. kjin in the directed version of the Chung-Lu model defined in Eq. Again. We finally briefly describe the directed case. in this limit the expected degrees converge to the specified values of x. It is also easy to show that the degree distribution P (k) becomes a rescaled form of σ(x). just like for the ER model. NETWORK ENSEMBLES The above exercise implies that the model defined by Eq. (5.16). Curiously. This implies that in this case pij /(1 − pij ) = fi gj where fi = f (xi ) and gj = g(yj ) are functions of xi and yj alone respectively. yj ) = xi yj 1 + xi yj (5.16) is equivalent to the Chung-Lu model defined by Eq.11) with q replaced by p. the assortativity) is unbiased. Therefore the spurious disassortativity disappears in this limit.g. all nontrivial choices can be mapped onto the linear case through a suitable redefinition of x and y. As already discussed in Section 5. By looking at Fig. The Park-Newman approach solves the practical and conceptual problems of other implementations of the CM by ensuring that the ensemble is properly sampled. in the above ‘quantum’ analogy this regime corresponds to the ‘classical limit’ where the discreteness of the quantum world can be neglected. (5. 5.5). Now the probability that a directed link from i to j is there is a function pij = p(xi . (5.3 Maximum-entropy ensembles All the examples in the last section show that even a conceptually simple idea (generating a random ensemble of networks with a specified degree sequence) can encounter big difficulties when naively implemented. 5.3.and out-degree sequence are equiprobable.1b and requiring that graphs with the same in. so that any inference about the higher-order properties (e.17) and in this case the sparse graph limit yields pij ≈ xi yj . (5. At this point. Therefore we have pij = p(xi . and the Fermi distribution can be replaced by the Boltzmann distribution which (in a suitable representation) has the expression (5.2. (5.88 CHAPTER 5. we show that such a method exists and is based on the Maximum Entropy Principle. yj ) of two quantities xi and yj playing a role analogous to that of the desired out.8).

then S = S1 + S2 where Si = − Gi Pi (Gi ) ln Pi (Gi ) denotes the entropy of the individual event Gi (i = 1. gives the maximum value S = ln M .20) G where xα is the observed (known) value of the α-th property. A deterministic (certain) process.3 Prove the last statement. Exercise 5. especially as developed by Jaynes [5]. gives S = 0. expressed by Eq.e. one for which one outcome has probability one while all other outcomes have probability zero. i. An additional constraint is given by the normalization of the probability. i. If the constraints represent structural properties taken from observations (like the number of links or the degree sequence in the ER and CM respectively).e. one where all the outcomes have exactly the same probability P(G) = M −1 .e.e. MAXIMUM-ENTROPY ENSEMBLES 5. Note that the constraints are enforced canonically. (5. we are maximally ignorant about the system. this will provide us with a measure of the degree of randomness that we are left with. According to this principle. then our least biased inference or ‘best guess’ about the (unknown) rest of the system shoul be obtained by finding the probability P(G) that maximizes S. Introducing one Lagrange multiplier θα for each constraint xα (plus an additional multiplier for the normalization constraint) and taking the functional . Shannon’s entropy will quantify the residual uncertainty that we are left with about the network. whenever we have only partial information (summarized in the knowledge of a set of m observables {xα }m α=1 ) about a system. or ‘unpredictability’ of a random process is provided by Shannon’s entropy.19) G where the sum runs over all the M possible outcomes of the process. 2) and the sum runs over the possible outcomes of such event. i. an important measure of the ‘uncertainty’. By contrast.3. which is the minimum possible entropy.5. (5. Use the result to calculate the entropy of the ER model with n vertices as a function of the probability p. the joint probability P(G) = P(G1 ∩ G2 ) can be factorized as P(G) = P1 (G1 )P2 (G2 ) where P1 and P2 are the marginal probabilities P for the individual events G1 and G2 .  If we measure Shannon’s entropy on a graph ensemble.1 89 The Maximum Entropy Principle In information theory. Another important property of the entropy is additivity: if the event G requires the simultaneous occurrence (intersection) of two events G1 and G2 (i. An important application of Shannon’s entropy is the Maximum Likelihood Principle. i. If P(G) is the probability of the outcome G. except for the known constraints.(5. Shannon’s entropy is defined (up to a proportionality constant which is irrelevant for our later purposes) as X S≡− P(G) ln P(G). once we enforce the constraints that define the model itself.3.e. This reflects the fact that. G = G1 ∩ G2 ) and if these events are independent. a completely unpredictable (uniform) process.1). The constraints are expressed in the form X E(xα ) ≡ xα (G)P(G) = xα ∀α. as ensemble averages. subject to the known constraints. after we measure those properties.

which were first introduced in social network analysis [22. 7] where it was showed that traditional tools borrowed from statistical physics could successfully contribute to investigate and sometimes even solve them explicitly [19]. This establishes another beautiful connection. (5.25)  In the following. 8] to generate ensembles of graphs matching a given set of observed topological properties. Applying the Maximum Entropy principle to graph ensembles leads to the so-called Exponential Random Graph (ERG) models. in terms of the probability P(G) of the microscopic configurations.90 CHAPTER 5. . From the knowledge of a few aggregate properties {xα }m α=1 (such as the degree sequence). by analogy with (statistical) physics. which is enforcing the normalization constraint for the probability. statistical physics looks for a least biased estimate of the microscopic properties of the system.23) G is the so-called partition function. In the context of networks. Z ∂θα ∂θα (5. Moreover.11 Prove the following relation (useful in the following): E(xα ) = − 1 ∂Z ∂Ω = . Indeed. the beauty of the Maximum Entropy approach is that of showing that the entirety of statistical physics can be reformulated exactly as an inference problem from limited information. This consideration establishes a fascinating connection between information theory and statistical physics.22) α=1 is a linear combination of the constraints that. we will call energy or Hamiltonian.24) where we have introduced the free energy Ω ≡ − ln Z. NETWORK ENSEMBLES derivative of S with respect to P. one can show that the result of the constrained maximization of S is e−H(G) P(G) = . 6. while X Z≡ e−H(G) (5. ERGs were then ‘rediscovered’ within an explicit statistical-mechanics framework [19. The above expressions coincide with those of traditional statistical physics. we are interested in doing the same operation. we will show that the general method allows to extend the approach to different constraints. Homework 5.21) Z where m X H(G) = θα xα (G) (5. (5. from the knowledge of only a few macroscopic quantities (like the total energy) of a system. we want to construct completely random ensembles of graphs. we consider specific examples of ERGs and we show that the ER model and the Park-Newman implementation of the CM can be recovered as particular cases of maximum-entropy ensembles. this time to graph theory. Indeed.

33) (5. i<j i<j where we have introduced the vertex-pair partition function Zij = 1 + e−θij .1 Y (1 + e −θij )= Y Zij .3. With this choice the partition function reads XY X X P e−θij gij (5.28) Finally. With such a choice.27) e−H(G) = e i<j −θij gij = Z = {gij } = {gij } i<j {gij } Y X e −θij gij = i<j gij =0. (5. (5. is pij = E(gij ) = ∂Ωij 1 = ∂θij 1 + eθij and the expected total number of links in the network is X X E(Lu ) = gij = pij .29) i<j where Ωij = − ln Zij .26). representing the probability that such vertices are connected. the ‘free energy’ is Ω = − ln Z = − X ln Zij = i<j X Ωij (5.26) i<j Note that in this model the total energy H(G) of the graph is the sum of the energies θij corresponding to its individual links.3. j. i<j (5.32) i<j ER Random Graph We now show that the ER random graph model can be recovered as particular cases of the exponential model defined by the Hamiltonian (5. the expected ‘occupation number’ of the pair of vertices i. the Hamiltonian reads X H(G) = θ gij = θLu (G).2 91 Simple undirected graphs We start by considering the rather simple.30) Equations (5. From the free energy it is possible to compute all the relevant quantities.34) . case when the Hamiltonian can be expressed in the form X H(G) = θij gij . (5. This is obtained when all energies are equal: θij = θ.30) completely define the model.5.31) (5. For instance.26-5. still quite general. i<j (5. MAXIMUM-ENTROPY ENSEMBLES 5. Each energy θij can be regarded as the ‘cost’ of placing a link between i and j.

38) and by introducing xi ≡ e−θi . Configuration Model We now consider the additive case θij = θi + θj . Other variants (such as the Chung-Lu one) cannot be derived from a Maximum Entropy approach.41) .4.31) now reads pij = 1 1 + eθi +θj (5.2.22). we are fixing the desired degree sequence and we expect this case to be equivalent to the version of the configuration model described in Section 5. (5. corresponding to the Park-Newman version of the configuration model.35) and. it is easy to see that the above equation is the same as Eq.26) becomes X H(G) = θij gij (5. (5. 5. x1 = Lu and θ1 = θ in Eq.40) i6=j i6=j and the free energy as Ω = − ln Z = − X i6=j ln Zij = X i6=j Ωij . as expected. (5. (5. we have pij = p = 1 1 + eθ (5. (5. NETWORK ENSEMBLES This corresponds to m = 1. i<j (5.31).3 Directed graphs We now briefly consider the directed case. we recover the constant form for the connection probability characterizing the ER model. Eq. Looking at Eq.36) which results in H(G) = X X X (θi + θj )gij = θi gij = θi ki (G).39) i6=j and calculations analogous to those presented above allow to write the partition function as Y Y Z= Zij = (1 + e−θij ) (5. in accordance with the Maximum Entropy Principle. and we are therefore only requiring that the expected number of links hLu i can be set to any desired value Lu by tuning the parameter θ. thus showing their biasedness. In other words. (5.12).37) i i6=j Note that in this case we are requiring to set the expected value of each degree hki i to any desired value ki by tuning the corresponding parameter θi . Indeed. The Hamiltonian (5. This requirement corresponds to the ER random graph model.3.92 CHAPTER 5. This confirms that that version is unbiased.

1 + eθ (5. ranging from 0 to +∞. we consider an exercise involving an ensemble of weighted graphs.45) This choice is equivalent to the directed version of the configuration model defined in Eq. MAXIMUM-ENTROPY ENSEMBLES 93 The probability that a directed link from i to j is there is pij = E(gij ) = ∂Ωij 1 . A weighted graph G is still described by the entries {gij } of an adjacency matrix. we recover a directed version of the ER random graph model: H(G) = θ X gij = θL(G) ⇒ pij = p = i6=j 1 .3. (5. = ∂θij 1 + eθij (5.31).43) i6=j In the constant case θij = θ. 1 + eαi +βj (5.5. yj ≡ e−βj . As a last example.17). Following the derivation leading to Eq.3.4 Consider the Hamiltonian H(G) = X θij gij .44) In the more general additive case. since for directed graphs θij can be asymmetric.  . (5. (5.and the out-degree of each vertex separately: H(G) = X X  (αi + βj )gij = αi kiout (G) + βi kiin (G) i6=j i ⇒ pij = 1 . These two parameters control the in. write the expected weight E(gij ) of the connection between vertices i and j.4 Weighted graphs Using the same general recipe. where xi ≡ e−αi . (5. we have θij = αi + βj . these entries are now assumed to be nonnegative integers. Exercise 5. However. Assume that the possible values of gij are non-negative integers.42) and the expected number of directed links is X E(L) ≡ hLi = pij . 5. one might also consider other ensembles of graphs.46) i<j where G is a generic weighted graph.

E 64. Jaynes.S. 620 (1957).Bibliography [1] M.J.E. L¨ assig.C Coolen. Rev. Random Structures and Algorithms 6. Roberts. Stat.J. [6] J. Krzywicki. Watts. Lett. 103567 (2009). J. Rev. Burda. Phys. [4] E. 161 (1995).C. Phys. T. Phys. [3] A. Molloy and B. 89. A. Phys. Physical review 106(4). E 85.C. A. Annibale. Jurkiewicz and A. De Martino. 026106 (2004). 94 . E 69. J. Rev. Newman. Berg and M. Rev. Coolen. A. S. 228701 (2002). Phys. B 136. [7] Z. 046103 (2012). 026118 (2001). [2] M.C. Reed.H. [5] E. Strogatz and D.

Therefore. These analyses can help us to: • visualize static graphs to gain a visual impression of the shape of the graph • compute static network properties of graphs. such as how graphs grow or shrink in time In this chapter we will look deeper at how we can implement graphs in the computer.Chapter 6 Random Graph Implementation In the previous chapters we have discussed at some length examples of graphs and of their properties. the homework exercises will concentrate on implementing graphs using a low level language. and the output (if any). We will use Gephi for graph visualization. The main idea is to come closer to an understanding of the nature of real world graphs. and Python are lower level tools. The level of abstraction that you work at is more detailed. and we will introduce Netlogo to study dynamic behavior of graphs. as the basics are enough to give an idea of how more elaborate systems work. hand in the source code. In the remainder of these notes we will use three computer tools. to understand the limitations of the high level tools. Therefore. the degree of a vertex. so that we can perform computer analyses of the graphs. In the history of our understanding of these graphs computer models and computer analysis tools have played an important role. In all programming homework exercises of this course. C++. the shortest path between two vertices. or Python to compute network properties. Gephi and Netlogo are high level tools. however. This will be. The high level tools are built using the low level languages. we use a programming language such as C. Java. Java. the input to the program (if any). C++. C. In the computer science chapters we will delve a bit deeper into these tools. and it takes more effort to achieve things. Provide in your 95 . such as the number of edges. Given the limited time that there is in a part of one-semester course. we will only cover some basic notions. Your program should contain documentation in the source code (comments). it pays to learn a bit about the low level languages. quite useful. or the clustering coefficient of a graph • understand dynamic properties.

The boolean values represent the presence of an edge. where m is the variable holding the matrix. VI. A logical choice for the datatype of m could be an (two-dimensional) array of BOOLEAN or an int.1 Adjacency Matrix The programs in this chapter are used to study network properties such as counting edges or computing shortest paths. you will create a functioning program that will be able to do increasingly useful things. Our first program will generate a graph. Wordpad.1 Random graph We will start with implementing a small regular network so that we can compute a few important properties of the network. The program will store it in memory. TRUE means an edge is present. An edge between vertex i and vertex j can be represented as m[i][j] = TRUE in a C-like language. Exercise 6. 1 Just as with your introduction to programming class. with real world graphs. Next. Java. Emacs.96 CHAPTER 6. The aim of these exercises is to gain proficiency in manipulating random graphs. We will see how the graphs can be constructed. First. or one in the computer labs. . RANDOM GRAPH IMPLEMENTATION report a short descripition of what you did. 6. the machine should have a text editor and a compiler or a development environment such as Eclipse. and will compute network properties. graphs can be represented in a computer program by an adjacency matrix. write a program in C. The data structure of an adjaceny matrix of a graph of N vertices can be represented in a programming language is a two-dimensional array of boolean values. and how essential aspects of your approach work (except for very small exercises).1 We will only be using simple text-based command-line programs in this course. with the Configuration Model. C++. FALSE means there is no edge. arrange to get access to a machine on which you can program.1. TextEdit. and. or Xcode. it is time to start writing the program. and with real world graphs.1 Question: How much memory does a boolean data type take in your favorite language? How much does an integer take? How much memory does it take to store a graph of 1024 vertices? How can this be reduced? At what cost? Now that we have discussed the data structure. are texteditors that save their output as plain text that can be read by a compiler). later on. 6. Visual Studio. You can use your own computer. As has been noted before. Your basic skills from first year programming classes should be all you need. By following the exercises. In later chapters we see how to work with larger graphs. or Python using an adjacency matrix. and how certain network properties can be checked to see if the graph exhibits regular or real world properties. (Wordprocessors such as Word and Pages save their text in ways that a compiler cannot read.

0. Print the resulting graphs. RANDOM GRAPH 97 Using the principles of modular or object oriented programming will result in cleaner code. Then. We will start with computing the mean and variance of the edge distribution of the networks that we have. The initialization routine takes as input a floating point value. Print a 0 if no edge is present. Homework 6.1. Print all vertex/vertex combinations. we will compute a few properties of the network. repair your program. or randomgraph.2 What do you expect is the mean of the edge distribution of the regular graphs for the values of p that you have used in the previous Homework Exercise? What is the expected variance? . Write a small function that reads the value of p as input. 1’s represent edges.) Homework 6. For the theory of this section.2 and p = 1. 6.0. p = 0.3 Mean and Variance Next. either interactively. or from the command line. The printout should be formatted in readable way. (3) Write a print routine that prints out the graph. (Refer to your introductory programming courses for help on this. assign with probability p the presence of an edge to the adjacency matrix. (2) In your program. You now have a working graph-manipulation program. globally declare the matrix m and initialize this matrix to the graph that you drew in the previous step. you would have to edit and recompile the program. Congratulations. which is not very flexible.1. and will make it easier to change your data structure implementation in the final chapter. We will work towards properties that allow us to discern a difference between regular networks and complex networks. and check if the output graph matches the graph on paper in step (1) in a logical sense. To print another graph. Exercise 6. for all pairs i and j.1. which is the probability p that an edge is present between vertex i and vertex j. draw an unweighted undirected graph with 10 vertices and 20 edges. refer to section 2. say 5-10.2 Random Graph This program uses a hardcoded initialization of the adjacency matrix. If the program does not correctly display the graph. your program does not have to provide a graphical representation of the graph.c. and a 1 if an edge is present.6. p = 0. First declare the graph size and the connection probability as a global constant. (4) Compile and run the program. 6.1 (1) On a piece of paper. we will modify the program so that it generates a random graph and stores it in the adjacency matrix. Next. Hint: You may want to use a doubly nested loop to achieve this.c).05. Start with a small number of vertices. Use the standard random number generator that draws numbers from a uniform distribution.1.2 Write an initialization routine that creates a random graph. Compile and run the program for p = 0. Make a copy of your program under a new name (such as: exercise2.

stackoverflow. Do not print the edges. and variance each increase or decrease compared to n = 10? Why do you think is that the case (provide an argument for each case)? Time and Space Complexity For an algorithm to be useful it has to be correct (it computes what it is supposed to compute) and it has to be efficient (it should compute its result in as little time and space as possible). This notion of efficiency is formally called “computational complexity. not constant). The program should compute and print the following numbers: the min. mean. just “complexity. Generate 100 random graphs.com. O(n) should be read as: “the complexity of the algorithms is of the order of n. A single for loop.3 Write two routines to compute (1) the mean and (2) the variance of the number of edges in a series of graphs. when the meaning is clear. and variance of the degree distribution of each graph. 2 Note that the word “complex” is used in many contexts.”2 For most algorithms the computational complexity depends on the size of the input problem. such as your statistics books. A doubly nested for loop from 1 to n has quadratic time complexity (since the operations inside the outer-most loop now also take linear time. Homework 6. only the mean and variance if the number of edges for the graphs. (Defintions and algorithms of mean and variance can be found in the usual places. To put it differently. and complexity is expressed in “big O” notation. 0.0 edge probabilities. median. Is the mean of 100 graphs different from 10 graphs? What about the variance? Why? Exercise 6. then the run time of the algorithm also increases by that factor.4 Computing Graph Properties Next we will consider computing the distribution of the degree of the vertices of a graph. . we are here dealing with the computational complexity of complex network computations. 6.1.” and often. exponential. An algorithm with a quadratic time complexity takes a 100 times longer to run. where the O is followed by some function. median. max.5. The time complexity of an algorithm indicates how the running time scales with the size n of the input.” which means that there is a constant C such that the complexity is bounded by C × n.05. Generate 10 random graphs. Repeat this exercise with n = 100. 0. running from 1 to n has a run time that is linear in n (assuming the operations inside the for loop take constant time).4 Write a program to compute the degree distribution of a random graph with n = 10 vertices and p = 0.3 Is your answer correct? Why do you believe that your answer is correct? Name two strategies to use so you can have more faith in the correctness of your program.98 CHAPTER 6. max. and compute mean and variance of the number of edges. mean. RANDOM GRAPH IMPLEMENTATION Homework 6. or Wikipedia). A linear time complexity means that when the size of the matrix increases by a factor of 10. or logarithmic function. and compute mean and variance of the number of edges. 1. usually a linear. Do the values of min.

. Google for Gephi. In order to reduce the homework workload. You should now see a nice graph of the relations between the characters in this famous play. Gephi draws a colored diagram of the network. 6. that can be quite useful to get a feel for the graphs that you are working with. just click OK on the import report. or even faster. Les Miserables. that run in linear time. go back to the Overview.1. for Windows you may need to install the latest Java version (available at https://java. there are no Compulsory exercises for Gephi. or go to http://gephi. switch to Data Laboratory. such as the ones that we have our own algorihtms for. [Russel and Norvig(1995)] for examples). In the remainder of this course you will from time to time be asked to use Gephi to create a picture of the network that you are working with. VISUALIZATION 99 The space complexity of an algorithm indicates how much storage an algorithm needs. Click on the Run buttons. A popular visualization tool is Gephi. or that you have generated with your program. 6. and thus has quadratic space complexity in n. 6. and launch a sample network. and see what happens.6. Windows). For these properties approximation algorithms may have been devised. in sub-linear time (see. etc. for example. However. An adjacency matrix is a two dimensional array. Gephi is a nice tool to play around with.2.github.2 Running Gephi Startup the program.2 Visualization The structure of complex networks is facilitated by visualization tools. Network Diameter.io and download the package that is suited for your operating system (Linux. To the right of the pane are buttons to compute network properties such as Average Degree. the number of vertices. Your sample file is imported. right click on the large red circle. you must download it from the ususal download site. such as computing the shortest path between all pairs of vertices.com/en/download/). some network properties are harder to compute exactly. Mac OS X. such as counting the number of vertices or edges.2.2. We say that the run-time of the algorithm is linearly (or polynomially) related to the size of the network. that draw a graphical depiction of a graph. Your screen should look like the screenshot in Figure 6. Visualization is an important method for undertanding the structure of your network.1 Downloading Gephi In order to use Gephi. and require a run-time that has a quadratic or cubic relation to the size of the network.4 Click on Data Laboratory. and can be used to compute network properties. and you will see that the best connected character in Les Miserables is Valjean. Some network properties can be computed quickly. Exercise 6. Everything should work right away under Linux and Mac OS X.

and add nodes and edges.1: Gephi Exercise 6. Go to Overview.100 CHAPTER 6. go to Data Laboratory. to recreate the graph that you used in the previous exercises. RANDOM GRAPH IMPLEMENTATION Figure 6.5 Create a new Project (under the File menu). . and look at your graph.

Prentice Hall. Norvig (1995): Artificial intelligence: a modern approach. S. Series in Artificial Intelligence.Bibliography [Russel and Norvig(1995)] Russel.J. and P. 101 .

In this chapter we will revisit the configuration model. The next exercise is to modify our program to generate edges randomly according to the pre-specified degree sequence. The configuration model can be used to generate graphs that exhibit more real world properties. using a pre-specified sequence that defines the degree of each vertex. For the regular graphs this was not the case. 7. In a histogram there should be a bar for each value of the degrees. and does not display real-world properties.1 Degree Sequence Histogram Print a histogram of the degree of vertices in the graph. The array is indexed 102 . We also experimented with computing properties of the graphs such as the clustering coefficient. the configuration model assigns the edges differently. In this chapter we will also take further steps in the road towards real-world networks by writing a program to generate networks using the Configuration Model. to see if the graphs exhibited real world properties. Exercise 7. You can print a histogram by declaring an array to store the counts of the vertex degrees.1. See also section 3.1 for more background on the Configuration Model.1 Configuration Model Implementation In the previous chapter we created a program for generating regular random graphs. provided that the degree sequence is satisfied. These will hopefully show us the obvious. The edges are then assigned randomly. 7. We will now write some more code to compute network properties for the random networks. As you will recall.1 Pre-specified degree sequence In the previous chapter the edges were assigned randomly to the pair i and j.Chapter 7 Configuration Model Implementation In the previous chapter we have discussed how to write a program to generate and manipulate Erd˝ os-R´enyi networks using an adjacency matrix. that the network is regular.

You can choose how you draw the edges. The program will now. we are ready to implement the configuration model. Write the output to a file. you can write a small routine that prints (source. or print it as a series of asterisks. (3) In your program. if that destination vertex still has “room” (i.2 To do so. Exercise 7.e. (6) Run this program a few times and give its output (the graph) in your report. We will do so shortly. Loop through all vertices. attempt to create the correct number of edges. connecting to a randomly chosen destination vertex. . for each vertex. program > file. (2) Write down the degree sequence of this graph. Then print the array value as a number.2 Visualization How can you know that your program is working correctly? One way to see if your program behaves as expected is to have your program output the edges.. • randomly choose a destination vertex whose degree is too low according to its pre-specified degree. to have it visualize the graph. There are other ways to see if your program behaves correctly. or print it to standard output and redirect the output to a file which you then import into Gephi. print the graph’s edges. stop) • repeat this process until the actual degree of the source vertex matches that of the pre-specified degree In the end. draw a graph of 10 vertices. declare an array of integers (of size 10) for the degree sequence. and create an edge between the two vertices • (update the degree-count of the vertices) • (if there is no destination vertex left whose degree still allows incoming edges. destination) pairs of the vertex id’s. Now that we have worked with a degree sequence. and increment the array element at the right index if there is a vertex with the right degree value. CONFIGURATION MODEL IMPLEMENTATION 103 by the degree k of each vertex.1. (5) So: For each (source) vertex. Homework 7. the prespecified degree has not yet been reached). (4) Initialize the array to the degree sequence. 7.1 (1) On a piece of paper.1. and import the that into Gephi. such as by writing a routine to check the consistency of the output.7.

We can write a checking routine to find mismatches between the actual and the pre-specified degree sequence. One way to address this issue is the Repeated Configuration Model. Due to the way that the edge destinations were chosen. Do the reported results match the results of the degree sequence checker? How are they related? Now draw two more graphs on a piece of paper.2 Check Routines Does the output of your program match the pre-specified degree sequence for the graphs that you tried? Homework 7. however. Have the routine report the number of vertices with mismatching degrees.104 7. removing them would subtly change the randomness of the graph. 7.2 CHAPTER 7. Run the routine a few times on the graph.1 Self-Loops and Multi-Edges Due to the way that the edge destinations were chosen. We need another solution. What do you see? . The checking routine may have found mismatches between the actual and the pre-specified degree sequence. In this way. The second network is to have 20 vertices and resemble a real network.3 Clustering Coefficient Compute the clustering coefficient of the graphs. Compare the clustering coefficient of the regular Erd˝ os-R´enyi graph and of the Configuration Model graph. The first one is to have 10 vertices and resemble a real world network (hubs. We will start with a small warmup to check the degree and to count the number self-loops and multi-edges. and the number of missing edges. which allowed random graphs to be generated according to a pre-specified degree sequence. CONFIGURATION MODEL IMPLEMENTATION Repeated Configuration Model Previously. Multi-edges and self-loops are undesirable. why? If not. we could make sure that a graph is generated that exhibited some properties of real world graphs (such as. power law degrees). which we will implement in this chapter.2. multi-edges and selfloops may have occurred. clusters.2 Check routine Write a routine that checks the degrees of the vertices against the degree sequence. why not? Homework 7. as chapter 5 discussed in depth. we have made an implementation of the Configuration Model. multi-edges and selfloops may have occurred.2. Does the graph have the required degree sequence? If so. and your degree sequence follows a power law.3 Rewiring to remove self loops and multi edges Write a program that reports the number of loops and the number of multi edges. a certain degree distribution). Exercise 7. obviously. 7. small world.

Homework 7. Section 5. 20 vertices). The time complexity can be assessed in at least two ways: by inspection.2 contains a more elaborate description of alternative implementations.2. you may want to output the graph to visualize it with Gephi.4 Degree Sequences Print a histogram of the degree sequences.2. We will now implement the repeated configuration model to remedy this problem. Thus. Run the repeat configuration model for the two graphs (10 vertices. and to compare it with the un-repeated configuration model. . Run the repeated configuration model.1. how often the original configuration model is called. it is discarded. What do you see? 7. How often does your program have to restart (for each graph)? Is termination guaranteed? At this point. Compare the histogram with the degree sequence historgram of the Erd˝ os-R´enyi graph of Exercise 7. Change your program so that when a graph with loops and multi edges is discovered. and by experiment. Do this experiment 20 times. and average the result. and the program starts over to try with a new graph. and count how many times it restarts.5 (a) Assess the time complexity of the repeated configuration model by code inspection/logical reasoning.3 Repeated Configuration Model You have most likely found that the resulting degree sequences were not always realized by the program. The repeated configuration model addresses a shortcoming of the original configuarion model by calling the model repeatedly. the number of vertices of the graph? Does it depend on other factors? (b) Assess the time complexity of the repeated configuration model experimentally. the time complexity of the repeated configuration model is larger than of the original configuration model.7. REPEATED CONFIGURATION MODEL 105 Exercise 7.4 Repeated Configuration model Create a new copy of your program. Does the time complexity depend on n. It also contains a remark about the computational complexity of the repeated configuration model (“Edge stub reconnection”). Exercise 7.

Part II Applications of Networks 106 .

1) where (Zd )∗ is the set of edges (see Fig. 1). either deterministically or randomly. d ≥ 2.1: Z2 (vertices) and (Z2 )∗ (edges) with random weights. Associate with each edge e a random variable w(e). 8. (8. This gives a random field of weights w = (w(e))e∈(Zd )∗ . Pick p ∈ [0. and partition Zd into p-clusters by connecting all vertices that 107 . drawn independently from the uniform distribution on (0. which is a model for the spread of a virus through a large regular network in which edges have random transmission capacities. This is a model for a “malicious attack” on a network by a “hacker”. In Section 8.3 we look at the configuration model and investigate how vulnerable its connectedness is when vertices are being removed.2 we look at invasion percolation on an infinite lattice. Draw edges between neighbouring vertices. Figure 8. A standard reference for percolation theory is Grimmett [2]. 1]. 8. which is the study of connectivity in large networks.1). In Section 8.1 Ordinary percolation Consider the d-dimensional integer lattice Zd . which is a model for connectedness of a large regular network in which edges are randomly removed. In Section 8.Chapter 8 Percolation This chapter is devoted to percolation theory.1 we look at ordinary percolation on an infinite lattice.

p x←→y (8. This is called the percolation function.4).2: Qualitative plot of the percolation function. Continuity is expected to hold also at p = pc .3). 1). C1 (0) = Zd . θ(1) = 1. a phase transition occurs: p < pc : all clusters are finite. It is further known that pc = 21 for d = 2 (see Fig.2) if and only if there is a path π connecting x and y such that w(e) ≤ p for all e ∈ π. At p = pc . i. the probability that the origin is connected to infinity via edges with weight ≤ p.6) It is known that pc ∈ (0. p > pc : there are infinite clusters.e.) Let Cp (0) denote the p-cluster containing the origin. i.. p 7→ Cp (0) is non-decreasing. 8. 1] : θ(p) = 0}.e. and that p 7→ θ(p) is continuous for all p 6= pc and strictly increasing on (pc .7) It is known that in the supercritical phase there is a unique infinite cluster with P-probability 1. However.108 CHAPTER 8. but this has only been proved for d = 2 and d ≥ 11. (8. Define pc = sup{p ∈ [0.1 Prove (8..3) with P denoting the law of w.4) so that (see Fig. good numerical approximations are available for pc .2) θ(0) = 0. p 7→ θ(p) is non-decreasing. Figure 8. 8. as well as 1 expansions in powers of 2d for d large. . 1). (8. (8. (8. called the critical percolation threshold. We have C0 (0) = {0}. while no explicit expression for pc is known for d ≥ 3. (A path is a collection of neighbouring vertices connected by edges. and define θ(p) = P(|Cp (0)| = ∞) (8.5) Exercise 8. Ordinary percolation may serve as a model for the connectedness of a network in which edges are randomly removed. PERCOLATION are connected by edges whose weight is ≤ p.

s. which is random because w is random.11) n→∞ The intuition behind this fact is the following.2 Invasion percolation Again consider Zd and (Zd )∗ with the random field of weights w. The set of invaded edges at time n has cardinality n. 2. (The reason for the inequality is that the vertex at the other end may have been invaded before.2. Then lim sup Wn = pc a. An interesting fact about invasion percolation is the following..10) Thus. Look at all the edges touching I(0). 8. because it cannot be invaded a second time.e. in going from I(n − 1) to I(n). Invade the origin: I(0) = {0}. |BN | with BN = [−N. the virus affects only a small part of the network. etc. The first question we may ask is whether CIPC = Zd with probability 1. N ]d ∩ Zd . in the sense that lim N →∞ 1 |BN ∩ CIPC | = 0 a. (8. Pick p ∈ (pc . 3. which means that the statement is true with P-probability 1).) The invasion percolation cluster is defined as CIPC = lim I(n). (see Fig. An invaded edge no longer counts for the probing of the weights.4). Repeat 2 with I(1) replacing I(0). (8. x}.8.9) (a. Then the union of all the p-clusters contains a unique infinite component (recall Exercise 8. . First we argue why the limsup is ≤ pc . with x = argminy∈Zd : kyk=1 w({0. Note that the sequence I is uniquely determined by w (because no two edges have the same weight). CIPC turns out to be an asymptotically thin set. In fact.2). The answer is no: CIPC ( Zd a.s. In this way we obtain a sequence of growing sets I = (I(n))n∈N0 with I(n) ⊂ Zd the set of invaded vertices at time n and |I(n)| ≤ n + 1. and invade both that edge and the vertex at the other end: I(1) = {0. 1).s. Grow a cluster from 0 as follows (k · k denotes the Euclidean distance on Zd ): 1. i. Invasion percolation may serve as a model for the spread of a virus through a computer network : the virus is “greedy” and invades the network along the weakest links. choose the edge with the smallest weight.8) This is an infinite subset of Zd . INVASION PERCOLATION 109 Exercise 8. n→∞ (8. y}).s. (8. Let Wn denote the weight of the edge that is traversed in the n-th step of the growth of CIPC . is an abbreviation for almost surely.2 Why is this uniqueness not obvious? 8.

55. the third largest cluster green.110 CHAPTER 8.45.4: The first three steps of the invasion: I(n) for n = 0.45 and p = 0.3: Simulation of ordinary percolation on a 25 × 25 block in Z2 for p = 0. Figure 8. the second largest cluster blue. PERCOLATION Figure 8. 2. The largest cluster is colored red. Note that the red cluster spans across the block for p = 0.55 but not for p = 0. 1. .

suppose that Wn ≤ p˜ for all n large enough for some p˜ ∈ (0. all edge weights incident to the boundary of Cp are > p.1 Work out the details of the above argument.15) with . and so after time τp the invasion will be “stuck inside Cp forever”.8. Figure 8. pc ). Since the invasion took place along this edge. INVASION PERCOLATION 111 which we denote by Cp . all edges incident to I(τp ) ∪ Cp have weight > p. (Note that we tacitly use that the critical percolation threshold of the full space is the same as that of the half-space. All edge weights in Cp are ≤ p. Thus.) Now. the edge invaded at time τp . it sees a “never-before-explored” region containing a half-space. The lower black circle is the origin (the starting location of the invasion).5: Picture of invasion percolation becoming trapped inside Cp (shaded region). it follows that lim sup Wn ≤ pc a. it also shows that Wn ≤ p for all n large enough a. 8.13) n→∞ Next we argue why the reverse inequality holds as well. All edges incident to Cp have weight > p. Let τp denote the first time a vertex in Cp is invaded: τp = inf{n ∈ N0 : I(n) ∩ Cp 6= ∅}. all edges connecting the vertices of Cp have weight ≤ p. (8. 1) that it hits Cp at such a break out time. all edges incident to I(τp − 1) (which includes this edge) have weight > p too. Then CIPC ⊆ Cp˜ ∪ I(σp˜) (8. Indeed. 1) (see Fig. Since p > pc is arbitrary. (8. Therefore it will eventually hit Cp with probability 1. which occurs at time τp . being incident to Cp . Indeed. 8. Not only does this show that CIPC = I(τp ) ∪ Cp ( Zd . which is not obvious but is true. Note that the asymptotic density of Cp is θ(p) ∈ (0. The upper black circle is the vertex where the invasion enters Cp . Homework 8.2) and that Cp does not necessarily contain the origin (see Fig.12) We claim that P (τp < ∞) = 1. However.14)  σp˜ = inf m ∈ N0 : Wn ≤ p˜ ∀ n ≥ m (8.s.s. There is an independent probability θ(p) ∈ (0. each time I “breaks out” of the box around 0 it is currently contained in.5).2. has weight > p.

. see Angel. the largest cluster has size Θ(n) as n → ∞.16) which proves (8. . Namely. For further background.112 the first onwards and this Note CHAPTER 8. respectively.s. respectively. Very little is known about the probability distribution of CIPC .s. even though invasion percolation has no parameter. Indeed. But |Cp˜| < ∞ and |I(σp˜)| < ∞ a.3 Vulnerability of the configuration model Let us leave the world of infinite lattices and return to the world of finite random graphs. 8. that lim sup N →∞ 1 |BN ∩ CIPC | ≤ θ(p) a. from time σp˜ the invasion must stay inside Cp˜. In Chapters 2 and 3 we saw that percolation may occur in the Erd˝ os-Ren´ yi model and in the configuration model.10) because limp↓pc θ(p) = 0 by the (supposed) continuity of p 7→ θ(p). In Chapter 11 we will encounter two further examples of models that exhibit self-organized criticality: the sandpile model and the Bak-Sneppen model. contradicts |CIPC | = ∞. Informally this can be expressed by writing CIPC = lim Cp . 8. |BN | ∀ p > pc . Figure 8. Goodman.6: Simulation of CIPC on Z2 . respectively. while for λ < 1 and ν < 1. (8. for λ > 1 and ν > 1. p↓pc (8. it behaves as if it were critical. PERCOLATION time from which onwards Wn stays below p˜.17) Thus.6). The above argument shows that invasion percolation exhibits self-organized criticality: CIPC is in some sense close to Cpc for ordinary percolation (see Fig. We found that the critical percolation thresholds in these models are λ = 1 and ν = 1. den Hollander and Slade [1]. with n the number of vertices. it has size Θ(log n).

3. (8.20) k∈N kf (k) and so the attack is successful if and only if P 1 1≤k≤k∗ k(k − 1)f (k) P ≤ . each taking values in [0.4.2 Give a heuristic explanation why the denominator in (8. The key quantity determining the occurrence of percolation is (recall (3. Hint: Recall the argument in Section 3. Dn that are i. Two choices are of interest: (1) π(k) = π ∈ (0.d. where the terrorist removes all vertices with degree larger than k∗ . 1) for all k ∈ N. . where the terrorist removes a fraction 1 − π of the vertices without looking at the degrees.22) ζ(τ ) . ∞). In case (2) we have P 1≤k≤k∗ k(k − 1)f (k) P ν¯ = ..21) For the special case whereP f (k) = k −τ /ζ(τ ). This corresponds to a deterministic attack. then the answer is no and we say that the attack was successful. 1]. does the the largest cluster still have size Θ(n) or not? This question was answered by Janson [3]: The answer is yes if and only if P k(k − 1)π(k)f (k) > 1.19) ν¯ = k∈N0P l∈N0 lf (l) Thus.19) P is not k∈N0 kπ(k)f (k). We ask ourselves the following question: • Given is a sequence π = (π(k))k∈N0 of probabilities. This corresponds to a random attack.1. with k∗ ∈ N some threshold value. Suppose that a “hacker attacks the network” by randomly removing vertices: vertex i is retained with probability π(Di ) and is removed with probability 1 − π(Di ) (together with all its incident edges). k → ∞.18) ν = k∈N k∈N kf (k) Suppose that ν > 1. (2) π(k) = 0 for k > k∗ and π(k) = 1 for k ≤ k∗ . k ∈ N. (8. with exponent τ ∈ (3. ∞). . random variables drawn from a prescribed probability distribution f . a fraction 1 − (1/ν) of the vertices is removed. k(k − 1)f (k) ν k∈N (8.e. with ζ the Riemann zetafunction given by ζ(τ ) = k∈N k −τ . After the attack. We will be particularly interested in choices of f having a polynomial tail f (k) ∼ Ck −τ . (8. In case (1) we have ν¯ = πν.i. . i.8. and so the attack is successful if and only if π ≤ 1/ν. independently for different vertices.12)) P k(k − 1)f (k) P ∈ (0. . the denominator in the left-hand side of (8.21) equals 1 [ζ(τ − 2) − ζ(τ − 1)]. if ν¯ ≤ 1. Homework 8. (8. VULNERABILITY OF THE CONFIGURATION MODEL 113 We focus on the configuration model with n vertices and with vertex degrees D1 .

114 CHAPTER 8.7: Plot of τ 7→ k(τ ). τ →∞ (8. k(τ ) 2 τ 3 Figure 8. Hint: Show that limu↓1 (u − 1)ζ(u) = 1 and limu→∞ 2u [ζ(u − 1) − ζ(u)] = 1.24) We have (see Fig. 8. k(τ ) = (τ − 3) ζ(τ − 2) − ζ(τ − 1) 1 − ν (8. for τ ↓ 3 the network becomes extremely vulnerable because only few vertices with a high degree need to be removed in order to take down the network.7) lim k(τ ) = ∞.25) Thus.23) This gives us the approximate criterion that the attack is successful if and only if k∗ / k(τ ) with −1/(τ −3)    1 . (τ − 3)ζ(τ ) k∗ → ∞. while for τ → ∞ the network becomes extremely robust because all vertices with degree > 2 need to be removed in order to take down the network. τ ↓3 lim k(τ ) = 2. . PERCOLATION while the denominator minus the numerator scales like [1 + o(1)] 1 −(τ −3) k∗ .3 Prove (8. Exercise 8.25). (8.

On percolation in random graphs with given vertex degrees. Goodman. Ann. 115 . 1989. Springer.R. Invasion percolation on regular trees. Probab. Angel. [3] S. Electronic J. den Hollander and G. Percolation.Bibliography [1] O. J. F. Probab. Slade. [2] G. Janson. Berlin. 14 (2009) 86–118. 36 (2008) 420–466. Grimmett.

where ηt = {ηt (x) : x ∈ Zd } (9. i ∈ N.4 we investigate a closely related problem. In Section 9. P(T1 > t) = e−λt . called the contact process. 1}Z ..2) = lim (1 − λ∆t)t/∆t = e−λt . d ≥ 1.e. t] into pieces of length ∆t each and letting ∆t ↓ 0.3) 116 . In Section 9. ∆t↓0 which matches the expression in (9. are i.2 at finite lattices.. t] (9. This process is an example of a larger class of random processes referred to as interacting particle systems. t ≥ 0. t ≥ 0. by dividing up the time interval [0. i. in Section 9. we see that  P(T1 > t) = P no ring occurs in [0.1 9.Chapter 9 Epidemiology In this chapter we look at a model for the spread of an infection over a network. ∞) is defined as the increasing sequence of random times (Ti )i∈N0 such that T0 = 0 and Ti −Ti−1 . [begin intermezzo] The Poisson process with rate λ ∈ (0. We imagine that the clock “rings at rate λ”.1). the probability that a ring occurs in an infinitesimally small interval dt is λdt. Indeed.e. [end intermezzo] 9. namely.1 we look at infinite lattices.d. (9. how a rumour spreads through a network. in Section 9.3 at finite random graphs.1.i.1) We may think of Ti as the time at which a “random clock” rings for the i-th time. i. random variables with common distribution EXP(λ).1 The contact process on infinite lattices Construction d The contact process is a Markov process (ηt )t≥0 with state space Ω = {0. A standard reference for interacting particle systems is Liggett [8].

e. . η) = c(x + y. x ∈ Zd . η → ηx (9. η 0 ) if η(x) = η 0 (x) = 1. and ηt (x) = 0 means that vertex x is healthy at time t. η 0 ) if η(x) = η 0 (x) = 0. (9. τy η is the configuration η viewed relative to vertex y: (τy η)(x) = η(x − y). ηt (x) = 1 means that vertex x is infected at time t.10) η  η0 →  c(x.8) with τy the shift of space over y.2 Shift-invariance and attractiveness Note that c(x.. the state at x flips up faster in η 0 than in η. 9.4) playing the role of the rate at which the state of vertex x “flips” in the configuration η.e. (Because Zd is infinite. ∞). y∼x c(x. if η(x) = 1..1. THE CONTACT PROCESS ON INFINITE LATTICES 117 denotes the configuration at time t. i.1. The configuration ηt changes with time. if η(x) = 0. η) ≥ c(x. which is natural when the interaction between individuals is shift-invariant. Systems with this property are called attractive.e.10) says that when η is everywhere smaller than η 0 .9. written η  η 0 . (9.) The configuration space Ω comes with a natural partial order : we say that η is everywhere smaller than η 0 . Details can be found in Liggett [8]. Property (9. In the contact process these rates are chosen as  P   λ η(y).. x ∈ Zd . The parameter λ measures how contagious the infection is. η ∈ Ω. and this models how an infection spreads through a population of individuals: the individuals form the vertices of the network. a little work is needed to show that the contact process is well-defined. The evolution is modelled by specifying a set of local transition rates c(x. the contacts between the individuals form the edges of the networks.9) Property (9. (9.6)   1. η). i.8) says that the flip rate at x only depends on the configuration η as seen relative to x. τy η) ∀ y ∈ Zd (9. η) ≤ c(x. η) = λ ∈ (0.7) This order is called partial because some pairs of configurations are ordered while others are not. (9. (9. Also note that   c(x. infected vertices become healthy at rate 1 and healthy vertices become infected at rate λ times the number of infected neighbours.5) with η x the configuration obtained from η by changing the state at vertex x (either 0 → 1 or 1 → 0). when η(x) ≤ η 0 (x) ∀ x ∈ Zd . i. but flips down slower.

118

CHAPTER 9. EPIDEMIOLOGY

Homework 9.1 Show that contact process “preserves the partial order ”,
i.e., two realisations of the contact process starting from η, η 0 with η  η 0 can
be coupled in such a way that an infinitesimally small time later the two configurations are still everywhere ordered (with probability 1). Hint: Recall the
intermezzo about coupling in Chapter 2.
In what follows we will see that properties (9.8) and (9.10) allow for a number
of interesting conclusions about the equilibrium behaviour of the contact process,
as well as its convergence to equilibrium.

9.1.3

Convergence to equilibrium

Write [0] and [1] to denote the configurations η ≡ 0 and η ≡ 1, respectively.
These are the smallest, respectively, the largest configurations in the partial
order , and hence
[0]  η  [1],
∀ η ∈ Ω.
(9.11)
Since the contact process preserves the partial order , we can obtain information about what happens when the system starts from any η ∈ Ω by comparing
with what happens when it starts from [0] or [1]. Indeed, writing
Pηt (·) = P(ηt ∈ · | η0 = η)

(9.12)

for the probability distribution of ηt given that η0 = η, we have
[0]

is stochastically non-decreasing,

[1]
Pt

is stochastically non-increasing,

t 7→ Pt
t 7→

[0]

(9.13)

[0]

i.e., if s < t and ηs , ηt have probability distributions Ps , Pt , respectively, then
there exists a coupling of ηs and ηt such that ηs  ηt with probability 1.
Homework 9.2 Give a heuristic argument for this monotonicity based on the
result of Homework 9.1.
An immediate consequence of (9.13) is the existence of the limits
[0]

ν = lim Pt ,
t→∞

[1]

ν = lim Pt ,
t→∞

(9.14)

which are referred to as the lower equilibrium, respectively, the upper equilibrium.

9.1.4

Critical infection threshold

Note that [0] is a trap for the dynamics (if all sites are healthy, then no infection
will ever occur), and so we have
ν = δ[0] .

(9.15)

An immediate consequence of (9.13) is that there is a critical infection threshold
λd ∈ [0, ∞] such that
λ ≤ λd :

ν = δ[0]

(“extinction of infection”),

λ > λd :

ν 6= δ[0]

(“survival of infection”).

(9.16)

9.2. THE CONTACT PROCESS ON LARGE FINITE LATTICES

119

Thus, for large λ there is an epidemic while for small λ there is not.
Let p(λ) denote the density of the infections in ν. The critical infection
threshold
λd = inf{λ ∈ (0, ∞) : p(λ) > 0} = sup{λ ∈ (0, ∞) : p(λ) = 0}

(9.17)

separates the phase of extinction of the infection from the phase of survival of
the infection. The function λ 7→ p(λ) is non-decreasing and continuous (see
Fig. 9.1). The continuity at λ = λd is hard to prove.

Figure 9.1: Qualitative plot of the density function.
Here are three facts about the critical infection threshold, the proof of which
requires delicate coupling arguments:
dλd ≤ λ1 ,

2dλd ≥ 1,

λ1 < ∞.

(9.18)

These inequalities combine to yield that λd ∈ (0, ∞) for all d ≥ 1, so that the
phase transition occurs at a non-trivial value of the infection rate parameter.
Sharp estimates are available for λ1 , but these require heavy machinery. For
instance, it can be shown that the one-dimensional contact process survives
when  

2  
λ2
λ
1
80
>
.
(9.19)
λ+1
λ+1
81
This yields the bound λ1 ≤ 1318 (see Durrett [4] for details). The true value
is λ1 ≈ 1.6494, which can be shown with the help of simulations and with the
help of of approximation techniques.

9.2

The contact process on large finite lattices

Suppose that we consider the contact process on a large finite lattice, say
ΛN = [0, N )d ∩ Zd ,

N ∈ N.

(9.20)

For convenience we may endow ΛN with period boundary conditions, so that
every vertex has the same environment. The contact process (ηt )t≥0 on ΛN is
again well-defined, and is again shift-invariant and attractive. However, since

120

CHAPTER 9. EPIDEMIOLOGY

ΛN is finite, we have ν = ν = δ[0] , i.e., the critical infection threshold is infinite
for all N ∈ N: on a finite lattice every infection eventually becomes extinct.
An interesting question is the following. Starting from the configuration [1]N
where every individual is infected, how long does it take the dynamics to reach
the configuration [0]N where every individual is healthy? In particular, we can
ask how large is the average extinction time
τ[0]N = inf{ηt = [0]N }.

E[1]N (τ[0]N ),

(9.21)

We expect this time to be growing slowly with N when λ < λd and rapidly
with N when λ > λd , where λd is the critical infection threshold for Zd . The
following results are shown in Durrett and Liu [5], respectively, Durrett and
Schonmann [6]: There exist C− (λ), C+ (λ) ∈ (0, ∞) such that
λ < λd :
λ > λd :

lim

N →∞

E[1]N (τ[0]N )
log |ΛN |

= C− (λ),

log E[1]N (τ[0]N )
|ΛN |
N →∞

lim

(9.22)
= C+ (λ).

Thus, in the subcritical phase the time to extinction is logarithmic in the volume
of the lattice (i.e., very slowly increasing with the volume), in the supercritical
phase it is exponential (i.e., very rapidly increasing with the volume). This is a
rather dramatic difference.
Homework 9.3 Give a heuristic explanation for the scaling of the average
extinction time in the two phases.
It can be shown that in the supercritical phase  

τ[0]N
lim P[1]N
> t = e−t
N →∞
E[1]N (τ[0]N )

∀ t > 0,

(9.23)

i.e., the extinction time is asymptotically exponentially distributed.
Homework 9.4 Why is (9.23) reasonable? Hint: Because in a unit time interval extinction occurs with only a very small probability, the time until extinction is exponentially distributed on the scale of its average (recall the explanation
of (9.1)).

9.3

The contact process on random graphs

Chatterjee and Durrett [3], Mountford, Mourrat, Valesin and Yao [9] look at the
contact process on the configuration model and show that, for every λ ∈ (0, ∞)
and every f with τ ∈ (2, ∞), the average time to extinction grows exponentially
fast with n (the number of vertices). This says that the contact process on the
configuration model with a power law degree distribution is “always supercritical”: regardless of the value of λ the average extinction time grows very rapidly
with the size. Apparently, the presence of vertices with large degrees makes it
easy for the infection to survive: hubs easily transmit the infection. In Hao and
Schapira [7] it is shown that the same behaviour occurs for τ ∈ (1, 2].
Similar results have been obtained for a selected class of other random
graphs, such as regular trees and the supercritical Erd˝os-Ren´
yi random graph.
However, it turns out to be hard to obtain sharp estimates. It would be interesting to understand what happens for the preferential attachment model.
Partial results have been obtained by Berger, Borgs, Chayes and Saberi [1].

Consequently both the typical distance and the typical travel time decrease. ∞)-valued random variables. Assign to each edge e of the configuration model a random time Ye . Indeed. τ −1 Since α > α ¯ when ν > 1 and τ ∈ (3. which is plausible because of the small-world property of the configuration model. For the special case where f (k) = k −τ /ζ(τ ). For the configuration model it is known that the typical distance Hn (defined in Chapter 2) scales like α log n with   ν . non-degenerate (0. van der Hofstad and Hooghiemstra [2] that in the supercritical regime ν > 1. and spreads through the graph employing time Ye to travel accross edge e.. The rumour is inserted at a randomly drawn vertex V1 . we see that the rumour has a tendency to spread along edges with an atypically small crossing time.d. This can also be viewed as a model for the spread of a rumour through a social network : once an individual picks up a rumour. the tail of the degree distribution gets thicker and thicker and the network acquires more and more hubs. ν−1 (9.i. (9. In fact. i. it turns out that the rumour passes through a small set of hubs that connect up with almost every other vertex in the network. the rumour needs a time of order log n to spread through the network.d. ∞). both α and α ¯ decrease with ν and increase with τ .26) n→∞ with Z1 . and so on. which represents the time that is needed by the rumour to travel along the edge (in either direction). lim [Tn − α ¯ log n] = Z n→∞ in distribution. (9. Assume that (Ye )e∈E are i.4 121 Spread of a rumour on random graphs Suppose that we consider a modified contact process in which we do not allow infected sites to become healthy..24) π : V1 →V2 e∈π where the infimum runs over all paths π from V1 to V2 . with probability distribution EXP(1). (9. it transmits this rumour to his/her friends. 3). i.e. we have ζ(τ − 2) ν = ν(τ ) = − 1. exponential with mean 1. 3).4.25) where α ¯ = 1/(ν −1) and Z is a non-degenerate R-valued random variable. the typical travel time between two vertices X Tn = inf Ye . for every τ ∈ (2. which in turn transmit the rumour to their friends. τ ∈ (3. once a site is infected it stays infected forever and passes its infection on to neighbouring sites.9. i. Thus. when the edge degrees have an infinite second moment the rumour spreads in a bounded time. as τ decreases. Pick a randomly drawn vertex V2 and ask how long it takes for the rumour to reach V2 starting from V1 .. SPREAD OF A RUMOUR ON RANDOM GRAPHS 9. τ ∈ (2. ∞). It is shown in Bhamidi. ∞).28) ζ(τ − 1) . Thus. k ∈ N.i. for every τ ∈ (3. lim Tn = Z1 + Z2 in distribution.e.27) α=  2(τ −2) . It is further shown in [2] that in the supercritical regime ν > 1. We model this situation as follows.e. (9. Interestingly. Z2 i.

.122 CHAPTER 9.1 Show that τ 7→ ν(τ ) is non-increasing. EPIDEMIOLOGY Exercise 9.

16 (1988) 1570–1583. Valesin and Q. Belmont CA. The contact process on a finite set. Probab. 37 (2009) 2332–2356. van der Hofstad and G. D. Chatterjee and R. Liggett. J.T. New York.-C. Ann. Liu. [8] T. Lecture Notes on Particle Systems and Percolation. Metastability for the contact process on the configuration model with infinite mean degree. Bhamidi. Wadsworth Pub. 20 (2010) 1907–1965. 1985. Co. Springer. 1988.Bibliography [1] N. [3] S. Hooghiemstra. Berger. 16 (1988) 1158–1173.V. 42 (2014) 1–40. Schonmann. Ann. The contact process on a finite set II. Ann. First passage percolation on random graphs with finite mean degrees. R. C. to appear. Durrett and R. [9] T. Appl.. Probab. USA. Probab. Grundlehren der mathematische Wissenschaften 276.M. [5] R. [2] S. Durrett. Ann. Durrett. Mountford. Saberi. [6] R. Ann. Schapiro. Chayes and A. Asymptotic behavior and distributional limits of preferential attachment graphs. Durrett and X. 123 . Mourrat. Probab. Interacting Particle Systems. [7] C. Contact processes on random graphs with power law degree distributions have critical threshold 0. Exponential extinction time of the contact process on finite graphs. Yao. J. Probab. Borgs. Hao and B. [4] R.

and apply it to network models. or whether a statistically correct criterion exists which selects a unique parameter value. 124 . Since this procedure requires maximum-entropy models to be fitted to empirical data. and surely many more than the number of model parameters. one of the main goals in the study of complex networks is that of reproducing the empirical topological properties of real-world networks by means of relatively simple theoretical models. allows one to address this problem successfully. we will first introduce an important and powerful statistical criterion.1 The maximum-likelihood principle As we have already discussed a number of times. which has a rigorous statistical basis. In this chapter. 10.Chapter 10 Pattern detection in networks In Chapter 5 we introduced various network ensembles built according to the Maximum Entropy principle. given a real-world network and a mathematical model of a graph. We show that the ML criterion also yields an unbiased way to correctly randomize a network. overcoming the structural bias introduced by other methods. it is important to realize whether this choice is really arbitrary. this is done by selecting one or more ‘target’ topological properties and looking for the parameter values that make the expected value of these properties match the corresponding observed value. we need to tune the free parameters of the model to those values that optimally reproduce the empirical properties of the network. But since we can target virtually as many topological properties as we want. Usually. we are going to use those ensembles as null models that allow us to detect empirical patterns in real-world networks. In this section we show that the Maximum Likelihood (ML) method. namely the Maximum Likelihood principle. We will then describe a pattern detection method based on this principle. and reveal the presence of higherorder mechanisms that cannot be explained by the null models themselves. In general. Such patterns are defined as statistically significant deviations from the prediction of maximum-entropy ensembles.

3) pM = n(n − 1) as in Eq. Thus the average clustering coefficient C as the probability that any two vertices sharing a common neighbour are mutually connected.2) In general.32). .10.17)) as target properties yields exactly the same value for pM . then one would get pM = C¯ ∗ . one might in principle derive (analytically or via numerical simulations) the dependence of Eθ~ (πi ) on θ~ and then look for the ‘matching value’ θ~M of the parameters θ~ that realizes the equality Eθ~M (πi ) = πi (G∗ ) ∀i. In the literature. choosing the average empirical degree k¯∗ = 2L∗u /n (see Eq. in the set of graphs spanned ~ For a given ‘target’ by the model. (10. which yields 2L∗u (10. (10. 5.2) statistically correct? Finally. (4.1. THE MAXIMUM-LIKELIHOOD PRINCIPLE 10.1. network models provide us with the expected values hπi iθ~ obtained as ensemble averages: Eθ~ (πi ) ≡ hπi iθ~ ≡ X ~ πi (G)P(G|θ). (4.30) i can be viewed as the probability that two randomly chosen neighbours of vertex i are connected ¯ defined in Eq. (4. (10.1 1 We recall from Chapter 4 that the clustering coefficient C defined in Eqs. the above system of equations might not admit a (unique) solution.1). topological property π(G) displayed by a graph G (in general a function of the matrix G).1 125 Motivation In general.29) and (4. (4. (10. And even if it does. However.1. We assume that a real network G∗ with n vertices and L∗u ≡ Lu (G∗ ) undirected links (see Eq. or a set of target properties {πi (G)}i .32) can be viewed to each other. we can use again the simple example we considered in Sec. different choices of target properties would in general result in a different value for pM . if the target property was taken to be the average clustering coefficient C¯ defined in Eq. is the criterion leading to Eq. (5. a common choice for the matching value pM of the parameter p is the one ensuring that the expected number of links hLu ip = n(n − 1)p/2 equals the empirical value L∗u . any network model depends on a set of parameters that we col~ Let P(G|θ) ~ be the conditional probability of lectively denote by the vector θ. once the parameters are set to the value θ. Clearly. it follows that the expected ˝ ¯ under the Erddos-R´ value of C enyi model is simply p.4) since the expected value of the clustering coefficient in the Erd˝os-R´enyi model coincides with the connection probability p. (4. occurrence of a graph with adjacency matrix G.12)) is compared with an Erd˝ os-R´enyi random graph model where the only (unknown) parameter is the uniform connection probability θ = p. (4. ˝ Since in the Erddos-R´ enyi model the probability that any two vertices are connected is p. For instance.11)) or the link density cu = 2L∗u /n(n − 1) (see Eq. independently on whether the vertices share a common neighbour. which target properties have to be chosen anyway? To concretely illustrate some of the above limitations.1) G When comparing the model with a particular real–world network G∗ .

126 CHAPTER 10. evaluated at the particular parameter value θ~∗ . which gives the same value θ∗ for the maximum. {θ~M } for the parameters θ. and the value of the parameter that maximizes the likelihood. consider a (discrete for simplicity) random variable V whose probability distribution f (v|θ) (defined as the probability that V = v) depends on a parameter θ. the Maximum Likelihood (ML in the following) criterion. is also denoted with an asterisk. the expected value of a quantity X(G). Throughout the rest of this Chapter.5) The ML condition for the optimal choice θ~∗ is found by requiring " ~ ~ θ~∗ ) = ∂λ(θ) ∇λ( ∂ θ~ # = ~0 (10. such as the subjective choice of fitting curves and of the region where the fit is performed. For a physically realized outcome V = v ∗ . PATTERN DETECTION IN NETWORKS As we next show. ~ λ(θ) (10.e. which is a statistically rigorous and widely used concept. Finally. Among all the possible ‘matching values’ ~ the one preferred by the ML principle is θ~∗ . hXi∗ . ∗ G is generated by the model with parameter choice θ~ is ~ ≡ log P(G∗ |θ). It is often simpler to define the ‘log-likelihood function’ λ(θ) ≡ log f (v ∗ |θ) and maximize it. the optimal choice for θ is the value θ∗ maximizing f (v ∗ |θ).e. This reminds us that the parameters are fixed by the data.e. Therefore. In the context of network modelling. i. but which may be more closely reproduced by a different value of the exponent or even by different curves as the fitting region is changed.1. indicates a unique choice for the optimal parameter value in the above example as well as in general. and is also recognized as more reliable than other fitting methods. i. X ∗ ≡ X(G∗ ). and simplifies the full notation ~ which illustrates that θ~∗ is ultimately a function of θ~∗ = arg maxθ~ ln P(G∗ |θ). By contrast. will be also denoted by an asterisk. i. thus achieving optimal inference from the empirical knowledge available. and makes the latter subject to the former. f (v ∗ |θ) represents the likelihood that such outcome is generated by the model with parameter choice θ. ∗ G . The ML approach reverses the role of data and parameters. This method avoids the drawbacks of other fitting methods. .2 Generalities In general.6) ~ θ ~∗ θ= and checking the second derivatives to be negative in order to ensure that this indeed corresponds to a maximum. the empirical value of a network property X(G) measured on a real network G∗ is denoted with an asterisk. This is particularly important in the case of networks and other systems exhibiting broad empirical distributions which may look like power laws with a certain exponent (which is also subject to statistical error) in some region. the ML approach always yields a unique and statistically rigorous parameter choice. 10. given the data. for fixed v ∗ . θ~∗ . where we typically have a model gener~ the log-likelihood that a real network ating a graph G with probability P(G|θ).

just like in the previous example for the Erd˝ os-R´enyi model. 10.5.10) In principle. the above correct choice is also the simplest and most frequently used one. and Eq. the optimal one is precisely z = (2L∗u )−1 . the ML condition rules out the possibility to construct a random graph with the same value of the average clustering coefficient as the real network. However. THE MAXIMUM-LIKELIHOOD PRINCIPLE 10.8) i<j where the becomes a∗ij ’s are the entries of the adjacency matrix of graph G∗ . (10. the ML value for p is the one we obtain by requiring hLu i = L∗u . which would be obtained by choosing p = C ∗ as an alternative matching value for the parameter p.9) i<j A biased example For instance. For instance. ~ 1 − pij (θ) (10. This coincides with the criterion we used in Eq.2 if useful) and show that the ML value p∗ that maximizes λ(p) is given by 2L∗u p∗ = .1. as in Eq. For the Erd˝os-R´enyi model. we will often consider a more general class of models obtained when the links between all pairs of vertices i. In general.4) in Chapter 5. among the possible values for the parameter z. where now (2L∗u )−1 is replaced by a free parameter z: pij (z) = zki∗ kj∗ .5) ~ = λ(θ) X i<j a∗ij log X ~ pij (θ) ~ + log[1 − pij (θ)]. in the Erd˝os-R´enyi model. (10. we would like to find that. (5. let us consider a modified version of the Chung-Lu model we introduced in Eq. (5.1. Write the log-likelihood function λ(p) = log P(G∗ |p) to generate a real-world network with adjacency matrix G∗ (refer to exercise 5.7) n(n − 1)  The above exercise shows that.1 Consider the Erd˝ os-R´enyi random graph model with connection probability p.5). more complicated models may be intrinsically ill– defined. ∗ 0 < z < (kmax )−2 . 10. as we now show. where 0 < p < 1. because we already know from sec.3 127 Erd˝ os-R´ enyi random graph Homework 10. and also that the expected degree sequence will coincide with the observed one.1. Let us see whether this can be actually achieved. In this case Y ~ = ~ a∗ij [1 − pij (θ)] ~ 1−a∗ij . . includes some examples discussed in chapter 5.10. Note that this class and independent probabilities pij (θ).4 More complicated models In the rest of the chapter.2.3 that this choice would ensure that the expected number of links hLu i coincides with the observed value L∗u .4. (10. j are drawn with different ~ where 0 < pij (θ) ~ < 1. choosing different reference quantities would not yield a statistically correct value consistent with the ML principle. as there may be no possibility to match expected and observed values of the desired reference properties without violating the ML condition. P(G∗ |θ) pij (θ) (10.

(10. PATTERN DETECTION IN NETWORKS Homework 10. (10. (10.6) and the requirement hLu i = L∗u are now equivalent. (10. 15.13) [5. consider the following variant of Eq. (1 − gij ) 1 − z ∗ ki∗ kj∗ i<j (10. (10. just like in the Erd˝ os-R´enyi model. show that the ML criterion leads to the parameter choice z ∗ defined by the equation L∗u = X i<j z ∗ xi xj . if we want the expected number of links to match the empirical one. (5. (10.2) to match expected and observed properties violates the ML condition. 1 + zxi xj z > 0. Homework 10. we have to force z away from the value z ∗ indicated by the ML principle.13).  P i<j (10. 7] pij (z) = zxi xj . which adds to the problematic aspects already discussed in sec.13) where the positive values {xi } are assumed to be fixed for the moment. (10. .2. Eq. the same equation automatically shows that the observed number of links L∗u is in general different from the expected value generated by the model: hLu i∗ = X X pij (z ∗ ) = i<j z ∗ ki∗ kj∗ .14) pij (z ∗ ) now coincides with the The above exercise shows that the model defined by Eq.13) is unbiased: the ML condition (10. write the log-likelihood λ(z) ≡ log P(G∗ |z) as in Eq.  The above exercise shows that the ML value of z is set by Eq.2 For the model defined by Eq. the expected value hLu i∗ = observed value L∗u . we cannot match the expected number of links with the observed number! Viceversa.11).11) and not by z = (2L∗u )−1 .11) which is the counterpart of Eq. (10. Similar considerations apply to the degree sequence ~k.3.3 For the model defined by Eq.12) i<j This means that if we want the ML condition to be fulfilled. In other words. 1 + z ∗ xi xj As a result. xi > 0 ∀i (10.128 CHAPTER 10.10) is an example of a biased model where the use of Eq. while z is a free parameter. Given the value z ∗ set by Eq.9) and show that the ML criterion leads to the parameter value z ∗ defined by the equation L∗u = X z ∗ ki∗ kj∗ ∗ . An unbiased example As another example. whose empirical value ~k ∗ is no longer replicated by the expected value h~ki∗ since z ∗ 6= (2L∗u )−1 . (10.7) for this model.5. This highlights another limitation of the multiplicative form of the Chung-Lu model.10). (10.

e. Empirical properties that are significantly (i. and so on. the CM is implemented in a way that does . For this reason. the maximum-entropy ensembles of graphs introduced in Chapter 5 play an important role in network analysis. the average and standard deviation are instead calculated analytically (with no need to sample the ensemble). there is no a-priori preferred level of organization separating ‘higher-order’ from ‘lower-order’ properties.2 129 Detecting structural patterns in networks Detecting patterns in real-world networks means identifying nontrivial structural properties. not explained by the local properties defining the maximum-entropy null model. properties that cannot be explained by simple random graph models and therefore indicate the presence of nontrivial mechanisms of network formation.2. popularity. wealth. One way of detecting such patterns is by isolating the higher-order empirical topological properties that are not simply explained by lower-order ones (for a discussion of the notion of ‘first-order’ properties. because these properties (such as node degree.2. ‘second-order’ properties. and then using the fitted ensemble to provide the expectation values and standard deviations of various higher-order properties. DETECTING STRUCTURAL PATTERNS IN NETWORKS 10. i. An important feature of this method is that. As we discussed in detail in Chapter 5.10. We will then compare these model predictions with the empirical data. see chapter 4) are likely to be directly affected by basic properties of nodes.1 Maximum likelihood in the configuration model The simplest ensemble where the local properties of all nodes are controlled for is the configuration model (CM). where we use the ML principle to make rigorous inference starting from the empirical knowledge available. for real-world networks. In principle. However. resulting in a method that is both unbiased and fast. including nontopological properties such as (depending on the nature of the network) size. an accepted criterion is that of considering ‘local’ (i. this requires that we make an important step. first-order) properties as the fundamental building blocks of the network organization. and are systematically used as a benchmark. However. that we have introduced in Chapters 3 and 5. Since such properties are usually very heterogeneously distributed over nodes. such an approach does not allow us to study ensembles of graphs whose realized properties coincide precisely with the observed properties of a real-world network. While a purely mathematical approach to the models introduced in Chapter 5 allows us to easily generate and characterize such graphs ensembles analytically (for instance by assuming some probability distribution for the model parameters and studying the resulting network properties). by many standard deviations) different from the expectations are interpreted as nontrivial patterns.e. although one can in principle measure the averages and standard deviations of higher-order properties by sampling many graphs from the maximum-entropy ensemble using the graph probability P(G|θ~∗ ). it is important to control for their effects by comparing the real network with a maximum-entropy ensemble having the same local properties.e. In this section we are therefore forced to reverse the perspective and take a statistical approach. please refer to chapter 4). or null model. it is not easy to ensure that. This will be achieved by first fitting the model ensemble to the data using the ML principle. etc. 10. starting from an empirical degree sequence. importance.

In Chapter 5 we mainly used Eq. allow us to define a method to compare the observed properties of a real-world networks with those of a CM that is fitted precisely on the empirical node degrees.  Homework 10.17) is formally identical to Eq. Importantly. The presence of n parameters allows us (at least in principle) to tune the expected value of each degree of the network to the observed value. Homework 10.14) as a way to infer the expected degrees.16) .14). By contrast.16) Show that the ML principle leads to the parameter value ~x∗ defined as the solution of the following set of n nonlinear coupled equations: ki∗ = X j6=i x∗i x∗j 1 + x∗i x∗j ∀i. PATTERN DETECTION IN NETWORKS not lead to a biased construction. as it only requires summing – or integrating.2 – the values of the connection probability (over all nodes except i). In the rest of this chapter. together with the maximum likelihood principle discussed in sec. Show that ~x∗ as a function of the entries {gij is such that the expected degree hki i∗ = j6=i pij (~x∗ ) of each node i is in general different from the observed degree ki∗ = ki (G∗ ).17) thus proving that in this model the expected degree sequence h~ki∗ coincides precisely with the empirical one ~k ∗ . (10. there can be opposite interpretations to such formulae. we show how the results discussed in Chapter 5. (5.13) respectively. Note that this operation is very simple. of eqs.17) the degrees are fixed by observation.4 Consider the model defined by pij (~x) = xi xj . but (as already noted when discussing the latter). (10. xi > 0 ∀i. (10. (10.16) can do so in a manner that is compatible with the ML principle.(10. 2 Note that this model coincides with the Chung-Lu implementation of the CM. (10.16) in the above two exercises are in some sense a generalization.10) and (10. as already noted in our discussion of Eq. given some theoretical probability distribution of the values of ~x. because now ~x∗ is the solution of n nonlinear coupled equations.15) and (10. However. (5. Write the log-likelihood λ(~x) ≡ ln P(G∗ |~x) of the model (where G∗ is the empirical network) and write the equation fixing the ML values ~x∗ = {x∗i } of the parameters ∗ } of thePadjacency matrix of G∗ . (5. we consider two exercises that help us appreciate once more the subtleties of the implementations of the CM. (10.5 Repeat the previous exercise for the model defined by pij (~x) = xi xj .130 CHAPTER 10. the exercises above instruct us that only the model defined by Eq. see sec. 1 + xi xj xi > 0 ∀i. and the parameters ~x∗ have to be found accordingly. To this end. to the case of n free parameters. Eq.5.  Equations (10.1.10. in Eq. This complicates the problem significantly.15) where ~x = {xi } is an n-dimensional vector of positive free parameters 2 .

DETECTING STRUCTURAL PATTERNS IN NETWORKS 131 However. which is generally much less than n: k=n X k0 P (k 0 ) x∗k x∗k0 (x∗k )2 − ∗ ∗ 1 + xk xk0 1 + (x∗k )2 (10.j hgij i hgjk i hgki i ∗ ∗ P k6P E (Ci ) ≡ hCi i ≈ . then we can approximate the value of such constraints as constant and equal to the empirical value (i. (4. the expectation value of the ANND defined in Eq. This is because the maximum-ensemble has been constructed precisely in order to keep the value of those constraints as close as possible to the empirical value. k 0 take only the empirical values of the degrees. implying x∗i = x∗j . Calculating averages and standard deviations Once the parameters ~x∗ are found. In particular. we can put them back into pij (~x) and use the resulting p∗ij ≡ pij (~x∗ ) to calculate the expected value hXi∗ of any higher-order property X(G) of interest. This provides us with some ‘error bar’ using which we can distinguish between properties that are statistically consistent with the expectations of the null model and properties which are not consistent. (4. (10. (10. all vertices with the same degree k have the same value x∗k . the degrees {ki (G)} in the case of the CM). While this is in general rigorously true only in microcanonical ensembles.30) is P P ∗ ∗ ∗ j6=i =i. thus representing higher-order patterns.21) will be approximated as P P hgij i∗ hgjk i∗ j6=i Pk6=j E∗ (kinn ) ≡ hkinn i∗ ≈ (10. any two vertices i.2. The last term in the above equation removes the self–contribution of a vertex to its own degree.e. if X(G) has some nonlinear dependence on the constraints (i. ki (G) ≈ ki∗ ).17). so that nP (k) is the number of vertices with degree k. j with the same degree ki∗ = kj∗ give rise to the same equations for x∗i and x∗j . In general. .10. are still much smaller than the fluctuations of any other (unconstrained) quantity. So eqs. in Eq. it is also possible to estimate the standard deviation p σ ∗ [X] ≡ hX 2 i∗ − (hXi∗ )2 of the properties of interest. this can be done analytically only if X(G) is a linear function of the entries {gij } of the adjacency matrix of G. However.17) reduce to as many nonequivalent equations as the number of distinct degrees actually displayed in the network.19) ∗ j6=i hgij i where hgij i∗ = p∗ij .(10. we will treat the value of the degrees as fixed and equal to the empirical value. and k. In general. Similarly. although nonzero. one expects that the fluctuations of the constraints in canonical ensembles. to calculate the expectation values of the higher-order properties of interest. the expectation value of the clustering coefficient defined in Eq.18) where P (k) is the empirical degree distribution.j hgij i hgki i Using the same sort of approximation.20) ∗ ∗ j6=i k6=i. we can reduce the number of such equations by noting that. Therefore.e.

1: Application the ML method to binary undirected networks.15 0. The panels report kinn versus ki (left) and Ci versus ki (right) for: a) and b) the network of the largest US airports (n = 500) [33]. 12].0 0 20 40 60 80 100 120 140 k 0 20 40 60 80 100 120 140 k c d 0.02 0. c.4  <c>. c.14 0. k nn 0. k nn.8  <c>.08 0.8 0.6 0.2 0.0 <k nn>.6 0.1 0. c.2 0.0 0 0 50 100 k 150 0 200 50 150 200 j i 600 500 400 300 200 100 0 100 k <c>.4 0.05 0. c 0. and highlight the average level of correlation in the case when there is no dependence on the degree.04 0.3 0.0 200 <k nn>. The red points are the empirical data. c. the black solid curves are averages over the configuration model obtained using the local rewiring algorithm [11.132 CHAPTER 10. i) the Internet at the AS level (n = 11. k nn 120 100 80 60 40 20 0  <c>. k nn 30 k k 15 10 5 0 0 10 20 30 k 40 50 0 10 g 20 30 k 40 50 h 1. k nn. c <k nn>. c <k nn>. k nn 30 25 20 15 10 5 0 0.142) [35]. and the blue dashed curves are the analytical expectations (± one standard deviation).00 0 500 1000 1500 2000 k 0 50 100 150 200 250 k Figure 10. c <k nn>.174) [37] and j) the protein-protein interaction network of Saccharomices cerevisiae (n = 4. PATTERN DETECTION IN NETWORKS b a 1.10 0. k nn. The green curves are the flat expectations under the random graph model.12 0.4 0.10 0. c 150 100 50 0.2 0.0 0 10 20 30 40 0 50 10 20 40 50 f e 25 0. k nn. c) and d) the synaptic network of Caenorhabditis elegans (n = 264) [34].00 20  <c>. . e) and f) the protein-protein interaction network of Helicobacter pylori (n = 732) [35]. k nn 0. g) and h) the network of liquidity reserves exchanges between Italian banks in 1999 [36] (n = 215).06 0.

meaning that structural constraints can fully explain the observed behaviour of higher-order network properties. suggesting the presence of mechanisms producing negative correlations. the two approaches yield very similar results most of the time. As can be seen. This means that the real network is more correlated than the baseline randomized expectation. The results shown in fig.1 we show an application of the ML method to the analysis of various networks. Secondly. For the two largest networks (the protein interactions in S. the microcanonical implementation of the CM is biased.10. When they differ. Firstly. By contrast. the trends displayed by the CM are not flat as those expected in the Erd˝os-R´enyi case. Instead. a synaptic network [34]. This confirms that residual structural correlations. we also report the average values obtained sampling the microcanonical ensemble with the ‘local rewiring algorithm’ [11. and the expected values over the ensemble of random graphs with the same number of links (corresponding to the Erd˝os-R´enyi random graph model). naively suggesting that correlations can never be traced back to the degree sequence alone. an interbank network [36] and the Internet at the Autonomous Systems level [37]. By contrast. simply due to the enforced constraint. The presence of these correlations does not require any additional explanation besides the existence of the constraints themselves. These are among the most studied networks of this type. namely the network of the 500 largest US airports [33]. (10. For the sake of comparison.10. DETECTING STRUCTURAL PATTERNS IN NETWORKS 133 Empirical results In fig. cerevisiae and the Internet). in the airport network the randomized curves lie below the real data (except for an opposite trend of hkinn i for low degrees). the measurement of kinn and ci on each network separately. they unveil different correlation patterns when compared to the randomized trends.1 allow to interpret the effect of the degree sequence on higher-order properties. plus a final averaging.15) which would yield flat trends as well. This is very different from the picture one would get by using the (wrong) expectation of Eq.2. the deviations are presumably due to the fact that. as the microcanonical approach would require too much computing time. the ML method only requires the preliminary estimation of the {x∗i }. with the expected values hkinn i∗ and hCi i∗ obtained analytically using the ML method. are still present after the rewiring has taken place. while the trends observed in all the networks considered are always decreasing. The same is true for the correlation structure . pylori ’s protein network the expected curves lie above the real data. 12] (see Chapter 5). in the H. Then the calculation of hkinn i∗ and hci i∗ is analytical and takes exactly the same time as that of the empirical values. We also highlight the region within one standard deviation around the average by plotting the curves hkinn i∗ ± σ ∗ [kinn ] and hCi i∗ ± σ ∗ [Ci ]. we only report the expectations obtained using the ML method.10. It should be noted that the microcanonical method requires the generation of many randomized networks (each obtained after many rewiring steps). We compare the correlation structure of the original networks. as ordinarily measured by the dependence of kinn (G∗ ) and Ci (G∗ ) on ki (G∗ ) (see chapter 4). two protein interaction networks [35]. and indicates that additional mechanisms producing positive correlations must be present on top of structural effects. The real interbank data are almost indistinguishable from the randomized curves. as we mentioned in Chapter 5.

~y ) = xi yj . (4. The local constraints are now represented by the joint sequence of out-degrees and in-degrees {kiout . (10.out ∗ ∗ nn. ∗ Empirical results We can now apply the method to various directed networks. ~y ∗ allow us to obtain hXi∗ and σ [X] analytically and quickly. defined in Eq.and out-degree sequences coincide precisely with the empirical ones. kiin }. Given a particular realworld network G∗ and a measured topological property X(G∗ ). and makes the comparison between real data and randomized ensembles essential. This means that the mere measurement of the topological properties is uninformative.25) ∗ j6=i haji i . ~y of parameters: pij (~x. without sampling the ensemble explicitly.21) Write the log-likelihood of the model (taking care of the fact that the network is directed) and show that the ML principle requires that these parameters are set to the particular values ~x∗ . which are specified by an asymmetric (in general) adjacency matrix.24) ∗ j6=i haij i P P haji i∗ hakj i∗ j6=i Pk6=j E∗ (kinn. while S.2 Directed graphs We now consider binary directed networks. confirming previous results [12]. the ML method allows to analytically obtain the expectation value hXi∗ and standard deviation σ ∗ [X] across the ensemble of binary directed graphs with. the same directed degree sequences ~k out (G∗ ) and ~k in (G∗ ) as G∗ .24). P P haij i∗ hajk i∗ j6=i nn. the quantities ~x∗ .in ) ≡ hkinn. 1 + xi yj xi > 0. ~y ∗ that solve the following set of 2n coupled nonlinear equations: X j6=i X j6=i x∗i yj∗ 1 + x∗i yj∗ = kiout (G∗ ) x∗j yi∗ 1 + x∗j yi∗ = kiin (G∗ ) ∀i ∀i (10.134 CHAPTER 10. We will denote this model as the directed configuration model (DCM). by studying the second-order topological properties measured by the outward ANND and the inward ANND.  As in the undirected case.2.1 The DCM is defined in terms of two n-dimensional vectors ~x. Exercise 10. PATTERN DETECTION IN NETWORKS of the Internet. 10.23) thus proving that the expected in. on average.22) (10.out Pk6=j E (ki ) ≡ hki i ≈ (10. yi > 0 ∀i.in i∗ ≈ (10. All these considerations show that seemingly similar trends can actually reveal very different types of structural organization. cerevisiae’s protein network is completely different from its randomized variants.

k nn <kout out out <kinnn>. kout . e) and f) the Little Rock Lake food web (n = 183) [39]. k nn <kout out out <kinnn>. kinnn. k nn . The panels report kinn.out versus kiout (right) for: a) and b) the directed neural network of Caenorhabditis elegans (n = 264) [34].in versus kiin (left) and kinn. k nn . DETECTING STRUCTURAL PATTERNS IN NETWORKS 135 b a 20 nn nn nn <kout >. kinnn.10. kout <kinnn>.2: Application of the ML method to directed networks. kinnn 80 60 40 20 80 60 40 20 0 0 0 50 100 0 150 50 100 kin 200 250 f e 40 20 30 15 nn >. the black solid curves are expectations under the directed configuration model using the local rewiring algorithm. c) and d) the metabolic network of Escherichia coli (n = 1078) [38]. kinnn. kinnn 15 15 10 5 10 5 0 0 0 10 20 30 40 0 50 5 10 15 kin kout c d 20 25 100 100 nn >. Red points are the empirical data.2. using the directed configuration model. and the blue dashed curves are the exact expectations obtained using the ML method (± one standard deviation). The green curves are the flat expectations under the directed version of the Erd˝os-R´enyi model. . kinnn 150 kout 20 10 0 10 5 0 0 20 40 60 kin 80 100 0 10 20 30 kout 40 50 60 Figure 10.

23) using the empirical degrees is particularly evident.out ] obtained using the ML method. To further highlight this important point. Small networks cannot be described by approximating the mass probability function of their topological properties (such as the degree) with a continuous probability density like in the Park-Newman approach described in Chapter 5.out i∗ are approximately constant. but at a different level). in our analysis of directed networks we have considered second-order topological properties. This unexpected result highlights once again that the measured values of any topological property are per se entirely uninformative.136 CHAPTER 10.3 (where for simplicity we omit the comparison with the LRA). the metabolic network of E.in ] and hkinn. This means that while one would be tempted to interpret the three observed trends as signatures of different patterns (zero. 15].3).2 we plot the observed values kinn. Reciprocity and motifs So far.out (G∗ ) versus kiout (G∗ ). This means that the baseline behavior for both quantities is flat and uncorrelated (as in the directed random graph. negative and positive correlation). while it is expected that random networks with specified degrees display a disassortative behavior [12. We also confirm that while some networks (C. Finally. As before. This is because the ML method extracts the hidden variables directly from the specific real world network. we confirm that the (directed) CM can display not only flat or decreasing trends. and the Little Rock Lake food web [39]. but also increasing ones. rather than drawing them from ad hoc distributions. coli [38]. we also show the microcanonical average obtained using the LRA and the expectation under the directed Erd˝ os-R´enyi random graph model (DRG) with the same number of links. in this case we find that the three randomized trends behave in totally different ways. and thus the ensemble of randomized networks is disassortative as for the undirected graphs considered above. coli ) are almost consistent with the null model. Interestingly. we can select three more food webs characterized by a particularly small size (see fig.in i∗ ±σ ∗ [kinn. we find a very good agreement between the two approaches. Therefore in this case the difference between the expectations obtained by drawing the ~x and ~y values from analytically tractable continuous distributions and those obtained by solving eqs.in (G∗ ) versus kiin (G∗ ) and kinn. and the randomized ensemble is even assortative. PATTERN DETECTION IN NETWORKS In fig. in the food web the constraints enforce unusual positive correlations. but the ML method yields the correct prediction in incredibly shorter time. the assortative trend is totally surprising.(10. By contrast. both hkinn. However. the most interesting point for the present analysis is that. As we show in fig.10. in the metabolic network the expected curves are decreasing. Importantly. while for the undirected networks considered above all randomized trends were decreasing. for three real directed networks: the neural network of C.10. in this case all three webs do not deviate dramatically from the null model. Again. In the neural network. others (Little Rock ) deviate significantly.10. as well as the expectations hkinn. invalidating the results obtained under other hypotheses. and can only be interpreted in relation to a null model.out i∗ ± σ ∗ [kinn. The resulting values can be distributed in a very complicated fashion. elegans and E. third-order properties can be studied by in- . In principle. elegans [34] (now in its directed version).in i∗ and hkinn. actually in all three cases the observed behavior can be roughly replicated by the same mechanism and almost entirely traced back to the degree sequence only.

. c) and d) the Mondego Estuary web (n = 46) [40]. k nn <kout out <kinnn>.out out ki versus ki (right) for: a) and b) the Narragansett Bay web (n = 35) [40]. kinnn 15 10 5 4 2 0 0 0 5 10 15 kin 20 25 30 0 2 4 6 c 8 kout 10 12 14 d 30 15 nn >. For the latter. e) and f) the St.10. k nn <kout out <kinnn>.2. k nn <kout out <kinnn>. where also the number of reciprocated links of each vertex is specified. directed food webs. kinnn 25 20 15 10 10 5 5 0 0 0 10 20 kin 30 0 40 5 10 8 8 6 6 nn >.3: Application of the ML method to small-sized. The panels report kinn. kinnn 20 25 f e 4 2 4 2 0 0 0 10 20 30 0 40 5 10 kin g 15 kout 20 25 20 25 h 8 8 6 6 nn >. kinnn 15 kout 4 4 2 2 0 0 0 10 20 30 kin 40 0 5 10 15 kout Figure 10. The green curves are the flat expectations under the directed Erd˝ os-R´enyi model.in versus kiin (left) and nn. DETECTING STRUCTURAL PATTERNS IN NETWORKS a 137 b 8 6 nn >. Red points are the empirical data and the blue dashed curves are the exact expectations (± one standard deviation) under the directed configuration model obtained using the ML method. Marks River web (n = 54) [40]. in g) and h) we also compare the empirical data with the expectations under the reciprocal configuration model. k nn <kout out <kinnn>.

17]) to preserve not only the total numbers kiin and kP of ini ↔ coming and outgoing links of each vertex. defined as the possible non-isomorphic connected subgraphs of 3 vertices in a directed graph.138 CHAPTER 10. ki← (G∗ ). we note that directionality also implies that. 44]. PATTERN DETECTION IN NETWORKS Figure 10. In other words. k ≡ a ij i j6=i ij P (number of non-reciprocated incoming links) and ki↔ ≡ j6=i a↔ ij (number of ← ↔ reciprocated links). One can show that in this case one needs to solve the following 3n coupled . This is an example of model with nonlocal (secondorder) constraints which can still be treated analytically using the ML method. 9]. there is a proliferation of possible third-order patterns. and comparing the empirical abundances with the expected ones under the null model. for each vertex i.4: The 13 triadic motifs. a more refined way to randomize directed networks includes the possibility to enforce additional constraints on the reciprocity structure [10. aij ≡ aji (1 − aij ) and aij ≡ aij aji . 42]. but also the number ki ≡ j aij aji of reciprocated links (pairs of links in both directions) [43. besides the DCM considered above. due to the directionality of links. ki↔ (G∗ )} as the reciprocal configuration model (RCM). This specification P is equivalent to enforce. it is possible (and important in many out applications [10.10. However. a more complete analysis consists in counting (across the entire network) the occurrence of all the 13 possible directed motifs [10] involving three vertices (see fig. Given a real directed network G∗ . As we show below. the three quantities 9] P [43. Before presenting the results. troducing directed generalizations of the clustering coefficient [41.4). the ML method is very effective in such a case. For this reason. ← ← ki→ ≡ j6=i a→ (number of non-reciprocated outgoing links). we denote the null model with specified joint reciprocal degree sequences {ki→ (G∗ ). where a→ ij ≡ aij (1 − aji ).

~z∗ . 6. estuaries. while the motifs m = 1.27) zi∗ zj∗ 1 + x∗i yj∗ + x∗j yi∗ + zi∗ zj∗ = ki↔ (G∗ ) ∀i (10. Large values of z[Nm ] indicate motifs that are either over. In this case. 8 for Little Rock Lake). In particular. The food webs considered here are from different ecosystems (lagoons. marshes. The value of z[Nm ] indicates by how many standard deviations the observed and expected numbers of occurrences of motif m differ. lakes. For instance. 13 are significantly reduced from the DCM to the RCM. Also note that in a network with no reciprocated . and thus to obtain the z-score z[Nm ] ≡ Nm (G∗ ) − hNm i∗ σ ∗ [Nm ] (10. and the most striking features of real webs become the over-representation of motifs m = 2.29) analytically. the analysis of motifs reveals a dramatic difference between the predictions of the two null models. the under-representation of motif m = 9 (the 3-loop) is the most common pattern across all webs. 11 display large values of z with opposite signs across different webs under the DCM.28) X j6=i X j6=i X j6=i 1+ x∗i yj∗ The expectation value of any topological property.10. Also. Indeed. we find no significant difference with respect to the DCM considered above (fig. 10 (plus m = 6. This can be done for both the DCM and the RCM. DETECTING STRUCTURAL PATTERNS IN NETWORKS 139 equations: x∗i yj∗ + x∗j yi∗ + zi∗ zj∗ = ki→ (G∗ ) ∀i (10. grasses). 9. under the RCM all networks display a very similar pattern. the signs of all statistically surprising motifs (i. Marks River food web. for both null models. We also show the two lines z = ±2 to highlight the region within 2 standard deviations from the model’s expectations. as well as its standard deviation. and that are therefore not explained by the lower-order constraints enforced.10.5 we show the z-scores for all the possible 13 non-isomorphic connected motifs with three vertices in 8 real food webs.or under-represented under the particular null model considered. as we now show.2.e. and becomes stronger as the reciprocity of the web increases. with a prevalence of aquatic habitats. The z-scores for the motifs m = 2. the ML method allows to calculate the expected number hNm i∗ and standard deviation σ ∗ [Nm ] exactly. bays. The presence of (intrinsically directed) predator-prey relationships implies that reciprocity is a very important quantity in food webs [17]. 13 (plus m = 3.10. 10. can now be calculated analytically in terms of the three n-dimensional vectors ~x∗ . in fig.3g-h we repeat the analysis of the directed ANND of the St. when |z| & 2) become consistent with each other under the RCM (except for m = 13). ~y ∗ . 11. Thus the RCM should fluctuate less than the DCM.26) x∗j yi∗ 1 + x∗i yj∗ + x∗j yi∗ + zi∗ zj∗ = ki← (G∗ ) ∀i (10.3e-f). 7.10. 3. this is confirmed by our analysis. As a consequence. If Nm denotes the number of occurrences of a particular motif m. now comparing the observed trend against the RCM. 13 for the Little Rock Lake web) and the under-representation of motifs m = 5. However. In fig.

5: Application of the ML method to the analysis of directed motifs involving three vertices in 8 real food webs. PATTERN DETECTION IN NETWORKS Chesapeake Bay 15 z=HNm -<Nm >LΣNm 10 Little Rock Lake 5 0 Maspalomas Lagoon -5 -10 Florida Bay -15 0 2 4 6 8 10 12 14 m St Marks Seagrass 15 Everglades Marshes z=HNm -<Nm >LΣNm 10 5 Grassland 0 -5 -10 Ythan Estuary -15 0 2 4 6 8 10 12 14 m Figure 10.140 CHAPTER 10. . Bottom panel: z-scores obtained enforcing also the reciprocal degree sequence (reciprocal configuration model). Top panel: z-scores obtained enforcing only the in-degree and out-degree sequences (directed configuration model).

In this way we are sure to obtain an unbiased model.6 Show that the ML principle implies that the optimal choice ~ θ∗ for the parameter θ~ in the model defined by Eq. Eq. Let {πα }α be our desired set of reference properties (constraints). as its stricter constraints allow to analyze 3-vertices motifs once 2-vertices motifs (i.e.2). in this class of models. (5. the number of motifs with at least a pair of reciprocated links is zero.30) ~ ≡ P πα (G)θα is the graph Hamiltonian and Z(θ) ~ ≡ P exp[−H(G|θ)] ~ where H(G|θ) α G ∗ ~ ≡ log P(G |θ) ~ = is the partition function [7].13) is another ~ = P θi ki (G) with xi ≡ e−θi [7]. (10.10. (10. Thus we confirm that the upgrade to the RCM is necessary. DETECTING STRUCTURAL PATTERNS IN NETWORKS 141 links. the hidden–variable model defined by Eq. if Lu is chosen as reference.30). the expected number of these motifs remains zero. and the model is biased. The log-likelihood function is λ(θ) ∗ ~ ~ −H(G |θ) − log Z(θ). By contrast. The random graph is a trivial special case where π(G) = Lu (G) and H(G|θ) = θLu (G) [7]. By contrast. their expected number under the DCM is always positive. Under the RCM. i. we should decide ~ from the beginning what these reference properties are. special case where ~π (G) = ~k(G) and H(G|θ) i and so it is unbiased too. the ML condition is equivalent to Eq. all possible dyadic patterns) are correctly accounted for.2. (10. This means that the whole class of maximum-entropy ensembles is unbiased. The possibility to treat the RCM analytically using the ML method is an important ingredient of this analysis. Homework 10. (10. Similarly. 10.3 General case We conclude with an exercise aimed at generalizing the results that we have enumerated so far for specific ensembles. include them in H(G|θ) ~ and define P(G|θ) as in Eq.2. This gives us the following recipe: if we wish to define a model whose predictions will then be matched to a set of properties {πα (G∗ )}α observed in a real–world network G∗ .5) cannot be traced back to Eq.31) G  The above exercise shows that. (10. and this is the reason why it is unbiased. .30). θ~∗ = θ~M .e. and let us define in terms of them the maximum-entropy model (see Chapter 5) ~ ~ = e−H(G|θ) ~ P(G|θ) /Z(θ) (10.30) is given by the solution to the following set of coupled equations: X ~∗ πα (G∗ ) = πα (G)e−H(G|θ ) /Z(θ~∗ ) = hπα iθ~∗ ∀α (10.

Newman. E 73. W. Rev. Nat. Rev. 125 (2002). [15] Park. Pastor–Satorras. 63 (2007).E. Physica A 350.H. P. Rev. profile of the Internet.D. Phys. [14] Chung. Phys. Leinhardt. Chung and L. N. 93. Phys. of Combin. 258702 (2002). 142 . Caldarelli. 026112 (2003). R.E. 066117 (2004). E 66. Rev. J. Phys. [18] R. [10] R. E 70. [4] F. Battiston. Bogu˜ n´ a and R. Watts Phys. J. Stat. [16] Catanzaro. 066121 (2002). Amaral. 910 (2002). N. Soc. 6.J. of Combin. & Newman. Capocci. [9] D. Phys. [8] P. Proc. [19] Park. R. Phys. Kashtan. M. & Lu. 3. U. 529–540 (2004). F. 027103 (2005).N. J. De Los Rios and M. M. Holland and S. S. [6] D. J.A. Alon Science 298. Newman.J. Garlaschelli and M. Assoc. 036112 (2003). 824-827 (2002). Science 296. Sneppen. 026118 (2001).. Caldarelli. [17] D. [3] M. Loffredo. Maslov and K. Rev. M. Camacho. B 274. M. Mu˜ noz. Lett.A. & Zaliznyak. S. Garlaschelli. Rev. Phys. Jiang & L.P. Rev. D.E. 33 (1981). S. L. [7] J. S. Amer. Rev. Ann.E. Castri.. Sales-Pardo L. K. Physica A 333. Amaral. Strogatz & D. [5] D. Phys. A. [2] B. I. Garlaschelli and M.W. Chklovskii. A. & Newman. 188701 (2004). [13] M. [12] Maslov. Park and M.J. M. Itzkovitz. E 71. E 68. Milo. Boguna. Phys. Ann. Lu.Bibliography [1] G. 015101(R) (2006). 491 (2005).J. A. Loffredo. 89. Shen-Orr. 6.I. Guimer´ a. 76. & Pastor–Satorras. V. M.B. 125 (2002). S.J. E 70. Sneppen. Rev. Stouffer. E 64. E 68. 066117 (2004) and references therein. Servedio and G. [11] S. Phys. Lett. 1931-1940 (2007). S¨ oderberg. Rev.

Ahnert. R. E 74. Barthelemy.. Serrano & A. E 74. Nat. 101. B.A. 110-115 (2006). E 78. G. ´ Serrano. Tech. & Pastor–Satorras. Tombor. [21] Serrano.I. S. [35] http://dip.. A. Rev. 276 . Iori & G. 2003).uni-lj. E 79.edu/dip/Main. 776. Rev.. PRL 103. Lett. Kawamura. Iwasaki. Rev. 066112 (2006). [23] Barrat. Flammini.J.BIBLIOGRAPHY 143 [20] Serrano. E 78. Caldarelli. [34] K. Melnik.. Phys. Nat. Pastor-Satorras & A. Phys. Garlaschelli. Bogun´ a. 11. networks. & Manna. Vespignani. Oshio. Porter. Rev. http://arxiv. [25] Bhattacharya. 026101 (2008). Pastor–Satorras.doe-mbi. New J. [38] H.I. Y. 367-392 (1991). B 38. Newman. Saramaki. Rev.282 (2007). Phys. A.htm [41] G. P. Vespignani. J. Colizza. A. (Keio University. [33] V. E 76. Mucha. M. [31] M. & Boguna. P02002 (2008). Garlaschelli & M. [27] D.si/pub/networks/data/bio/ foodweb/foodweb. Kaski. Phys.E. 3747-3752 (2004). R. 2. G. K. M. Panzarasa J.N. Osana. Phys.. 036114 (2009). Pastor-Satorras. Y. Eur. Phys. S. Rev. 055101(R) (2006).-A. M. Rep. [22] M. PNAS 101. . of Phys. 058701 (2009). T. Garlaschelli & M. 026107 (2007). Z. E. 101 (2005).1439.A. Phys. Phys. [40] http://vlado. Morita.cgi [36] G. E 78. [32] D. A. [26] Bianconi. De Masi. 102. 036112 (2008). [28] D.. R. K. Colizza. Mukherjee. E. S.ucla.J. Oltvai and A. Gomi. modelling. M. J. Loffredo.fmf. Vespignani. J. M.. 073005 (2009). Phys. Rev. Phys.org/abs/1001. 205-209 (2004). R.D. Fink. A. V. Opsahl. S. P. Martinez. & Vespignani. K. Omata. [30] M. Oka and K. [24] T. 038701 (2009). Barab´asi. Keio Future 3. Hackett. of CCeP. Proc. Ecological Monographs 61. [29] S. Loffredo Phys. Stat. Albert. M. 168702 (2008). Phys. 015101(R) (2008). Ramasco. [39] N. Rev. A. [42] S. P.A.-L. [37] V. 3. Boguna. M. Nature 407. Fagiolo. Colizza. J. Lett. Akiyama. Jeong. 651 (2000). Mech. A. J. Rev. K. R. Gleeson. G. M. AIP Conf.

M. 75-174 (2010). I. (JosseyBass. [44] V.E. E 73. 1-45. 015101(R) (2006). Loffredo. M. . S. [47] J. Heise. Kert´esz. 056131 (2004). 1975). T. networks.-P. D. 027105 (2007). Phys. [49] P. E 75. Rev. Phys. 268701 (2004). Phys. Phys. Physics Reports 486(3-5). E 80. Onnela. Fortunato. Caldarelli. Garlaschelli and M. Stefancic. Holland. Rev. Garlaschelli. Phys. Leinhardt.J. Newman. 016117 (2009). in Sociological Methodology. pp. 93. San Francisco. E 70. Rev. Saramaki.144 BIBLIOGRAPHY [43] D. K. Kaski and J. D. Fink & G. Ed. [46] S. Ahnert. [48] S. Lett. Kivela. Zlatic & H. Rev. [45] M. J. Rev. E.

More recently. These results imply that adaptive networks generated by this interplay may represent an entirely novel class of self-organized complex systems. 3.Chapter 11 Self-Organized Networks 11. 5] present reviews of these results. or on the reverse effects that vertex–specific dynamical variables have on network structure. iii) the study of the effects that the topology has on various dynamical processes (such as the spread of epidemics) taking place on static networks. In order to highlight the novel phenomena that originate from the interplay between the two mech145 . it has been argued that the complexity of real–world networks is in the most general case the result of the interplay between topology and dynamics. one finds that the feedback between topology and dynamics can drive the system to a steady state that differs from the one obtained when the two processes are considered separately [6]. Remarkably. 2. So. iv) networks as computing entities. ii) the mathematical modelling of (either static or growing) network formation. exploiting the idea that all these aspects should in the end be related to each other. These points of view reflect the main approaches in the literature [1. 4. it has been suggested that one should consider the mutual influence that these processes have on each other. while most studies have focused either on the effects that topological properties have on dynamical processes. whose properties cannot be straightforwardly understood in terms of what we have learnt so far. In particular.1 Introduction So far. we have approached networks from different points of view: i) the definition and computation of the static topological properties of real-world networks. because networks can process the information about their environment and respond to it adaptively. and that one is therefore allowed to consider the evolution of the fast variables while the slower ones are ‘quenched’. In this chapter we shall present a self–organized model [6] where an otherwise static model of network formation driven by a so-called vertex fitness [7] is explicitly coupled to a so-called extremal dynamics process [8] providing a dynamical rule for the evolution of the fitness itself. a few attempts to provide a unified approach to networks and their dynamics have been proposed. This amounts to relax the (often implicit) hypothesis that dynamical processes and network growth take place at well separated timescales.

the shape of coastlines [10].1 is iterated an infinite number of times. the branching of rivers [12].2 we recall some aspects of scale invariance and Self–Organized Criticality (SOC). named after the Polish mathematician Waclaw Sierpi´ nski who introduced it in 1915 [17]. by comparison with another area chosen as the unit of measure. In section 11. or fractality. scientists have tried to understand the possible origins of fractal behaviour. one of the simplest fractals defined by recursion is the Sierpinski triangle. one obtains an object whose empty regions extend at any scale (up to the maximum area delimited by the whole triangle). where the idea that network properties may depend on some fitness parameter associated to each vertex was proposed. etc. in section 11.1 Geometric fractals Due to this ubiquity.2 Scale invariance and self–organization Self–similarity. where self–similar geometric objects are constructed iteratively. is the property of an object whose subparts have the same shape of the whole. in an incredible number of situations the objects of interest can be represented by self–similar structures over a large. Finally. For instance. 10]. and a variety of biological structures [16]. At first. Based on these examples. it turned out that examples of fractal structures (even if approximate due to natural cutoffs) are actually ubiquitous in nature. 11. A way to solve this problem is to .2. the growth of cities [14]. even if finite. anisms.3 we briefly review the so–called fitness model of network formation [7].146 CHAPTER 11.4 we present the self–organized model obtained by coupling these mechanisms. mathematicians introduced quantities in order to distinguish rigorously between fractals and ordinary compact objects.11. Only later.e. It is therefore difficult to measure its area in the usual way. When the procedure shown in Fig. range of scales. 11. the discharge of electric fields [11].1: First steps in the iteration procedure defining the Sierpinski triangle. i. deposition processes [13]. SELF-ORGANIZED NETWORKS Figure 11. fractures [15]. due to the activity of Benoit Mandelbrot [9. In section 11. The first preliminary studies have focussed on mathematical functions built by recursion (Koch’s snowflake. Indeed. self–similarity appeared as a peculiar property of a limited class of objects.). Examples include commodity price fluctuations [9]. we first review the main properties of such mechanisms when they are considered separately. and in particular the biologically–inspired Bak–Sneppen model [8] where the extremal dynamics for the fitness was originally defined on static graphs. Sierpi´ nski triangle and carpet.

2 Self–Organized Criticality Despite their importance in characterizing the geometry of fractals. Firstly. Another situation where self–similarity is observed is at the critical point of phase transitions.  ln 3 ln 2 ' 1.58496. (11. these features are obtained when a driving parameter of the nonlinear dynamical system is set to a crossover value at which chaotic behaviour sets on. Note that at the first iteration we only need three triangles of side length 1/2 to cover the object (while for the whole triangle we would need four of them). In general. D = DE = 2. magnetic systems display a sharp transition from a . D is lower than 2 because the Sierpinski triangle is less dense than a compact bidimensional triangle.. This has shifted the interest towards dynamical models..  Homework 11. the nonlinear system is said to be at the “edge of chaos”.2.2. Formally. For a compact (non-fractal) object. but also for the measurement of its area. Therefore the fractal dimension measures the difference between the compactness of a fractal and that of a regular object embedded in a space of equal dimensionality. for a compact 2-dimensional triangle. if N () is the number of DE -dimensional volumes of linear size  required to cover an object embedded in a metric space of Euclidean dimension DE . attractors in the phase space of a nonlinear dynamical system can have a fractal geometry.2 Prove that. D is also larger than 1 because it is denser than a one-dimensional object (a line). real–world situations. purely iterative models are not helpful in order to understand whether a few common mechanisms might be responsible for the fractal behaviour observed in so many different. Note that the above formula can be rewritten in the familiar form of a power law by writing. N () ∝ −D (11. for the Sierpi´ nski triangle.. D = Note that now D < DE = 2. SCALE INVARIANCE AND SELF–ORGANIZATION 147 consider a limit process not only for the generation of the fractal.11. open dissipative systems are in many cases associated with fractals for more than one reason. then the fractal dimension is defined as ln N () . the fractal dimension D has the same value as the Euclidean dimension DE . their evolution can proceed by means of scale–invariant bursts of intermittent activity [18] extending over both time and space. Homework 11. and seemingly unrelated.2) This highlights the correspondence between the geometry of a fractal and scale– invariant laws. For instance.1) D = lim →0 ln 1/ which approaches an asymptotic value giving a measure of the region occupied by the fractal. 11. When this occurs.1 Prove that. In the present example. Indeed. At the second iteration we need nine covering triangles of side 1/4 (while for the whole triangle we would need sixteen of them). secondly. The (scale–dependent) number of objects required to cover a fractal is at the basis of the definition of the fractal dimension D. for small .

At a phenomenological level. functions of the height zi can be defined: for example the height itself. one considers a simple square lattice. the discrete Laplacian operator of height (second discrete derivative). where parameters have to be tuned to their critical value. in order to explain the ubiquity of self–similar systems one should understand why they appear to behave as if their control parameter(s) were systematically fine–tuned to the critical value(s). As long as zi remains below a fixed threshold. spins are spatially arranged in aligned domains whose size is power–law distributed. There are several examples of simplified models showing SOC. and shown to lead to SOC only if fine–tuned to zero. Any site i of the lattice is assumed to store an integer amount zi of sand grains. [19] and references therein). that drive the control parameter to the critical value as a spontaneous outcome of the dynamics. 22]1 . determining a nonzero overall magnetization. two classes of SOC models attracted many studies: the class of sandpile models [26] and the class of models based on extremal dynamics such as the Bak–Sneppen [8] and Invasion Percolation [27] models. a class of open dissipative systems defined over a finite box Λ in a d–dimensional hypercubic lattice. Exactly at the critical transition temperature. This supports the hypothesis that SOC models are closely related to ordinary critical systems. Sandpiles One prototype is represented by sandpile models [26]. SELF-ORGANIZED NETWORKS high–temperature disordered phase. This means that domains of all sizes are present. and most of them have a common structure. it is the system itself that evolves autonomously towards the critical state. In both cases. so that the height zi is increased by one. the fundamental difference being the feasibility of this tuning. 21. with no need for an external fine–tuning. the difference of height between nearest neighbours (first discrete derivative of the height). Therefore the heights evolve according to zi → zi − ∆ki 1 Interestingly (11. But as soon as zi exceeds the threshold. This point led to the idea that feedback effects might exist. At every time step one grain of sand is added on a randomly chosen site i.148 CHAPTER 11. 2 Different . the column of sand becomes unstable and “topples” on its nearest neighbours. 25]. In what follows we briefly review these examples. and so on.3) a similar claim has been made for networks as well [23]. In d = 2 dimensions. Why this happens is still a matter of debate. This paradigm is termed Self–Organized Criticality (SOC) (for a review see Ref. with a scale– invariant pattern. In practice. it has been proposed that a temperature–like parameter can actually be introduced for these systems [24. even if some authors claimed that this behaviour may be based on the minimization of some energy potential [20. where microscopic spins point in random directions and generate no macroscopic magnetization. corresponding to the height reached by the sandpile at that site. to a low–temperature ordered phase where almost all spins point in the same direction. SOC aims at explaining the tendency of open dissipative system to rearrange themselves in such a way to develop long–range temporal and spatial correlations. In this scenario. Also. nothing happens 2 .

The latter is the duration of the fundamental event defined by these two processes: • a set of sites becomes critical due to the previous toppling event. is the Bak–Sneppen (BS) model [8. the model is defined by the following steps: • N species are arranged on the sites of a 1-dimensional lattice (a chain.2. one must first define the unit timestep. they may topple in their turn. Similarly. this loss balances the continuous random addition of sand. Q(t) ∼ t−ξ [26]. The size of an avalanche can be defined. Then the lifetime of an avalanche can be defined as the number of unit timesteps between two sand additions. In its simplest formulation. In order to define the lifetime of an avalanche. One can define both a size and a characteristic time for an avalanche. a behaviour reminiscent of the diverging susceptibility characterizing critical phenomena. This process is called toppling. • all such critical sites undergo a toppling process. and the heights of their neighbours are updated. Therefore the model succeeds in reproducing the critical behaviour. in which case the procedures starts again with the addition of a new grain. often associated to phase transitions.11. as the total number of toppling sites (one site can topple more than once) or the total number of topplings (it is clear that these two definitions give more and more similar results as the space dimension increases). SCALE INVARIANCE AND SELF–ORGANIZATION 149 where ∆ki    2d   = −1     0 k=i k nearest neighbor of i (11. for instance. 28]. or a ring if periodic boundary conditions are enforced). In the steady state of the process. but with a self–organized mechanism requiring no external fine tuning of the control parameter. All the toppling events occurring between two consecutive sand additions are said to form an avalanche. Note that the grain addition can be viewed as the action of an external field over the system. and this effect can propagate throughout the system until no updated site is active. at the steady state. even if with strong simplifications. . The spatial correlations that develop spontaneously at all scales indicate that the system reacts macroscopically even to a microscopic external perturbation. for topplings on the boundary sites (i ∈ ∂Λ) some amount of sand falls outside and disappears from the system. While the amount of sand remains constant when toppling occurs in the bulk. The Bak–Sneppen model A model that attempts to explain some key properties of biological evolution. the avalanche processes can be viewed as the response (relaxation) of the system to this field. The fundamental result of the sandpile model is that. As the neighbouring sites acquire new grains.4) otherwise. both the size s and the lifetime t of avalanches are characterized by power law distributions P (s) ∼ s−χ .

2. drawn from the same uniform distribution on the unit interval.66702 ± 0. field theory [36]. The Bak–Sneppen model is a prototype mechanism generating fractal phenomena as an effect of extremal dynamics [30]. Being so well studied. the update is interpreted as a mutation of the least fit species towards an evolved species representing its descendant or offspring. 32]. both predator and prey along the food chain). the longer the time between two modifications of the genetic code. mean-field approximations [28. 38].11. In this model an (evolutionary) x-avalanche is defined as a causally connected sequence of mutations of barriers. the model shows that a power–law distribution of extinction events may be interpreted as the outcome of a single internal macroevolutionary process acting at all scales.003 [29] (see Fig. are assigned them.00008 [29] (see Fig. 1]. In any case.11. 39. [44] and references therein. right panel). left panel). 38. For a recent review on this model see ref. the system self–organizes to a critical stationary state in which almost all the barriers are uniformly distributed over a certain threshold value τ = 0. The basic idea behind the model is that the species with the lowest fitness is the one that is most likely to go extinct and replaced by a new one. some analyses of which have indicated that extinction sizes are power–law distributed.2. and is modelled as a random update of their fitness as well. 40. It has also been defined on higher–dimensional lattices and more general graphs. 30] and probabilistic approaches (run time statistics) [37. 35]. renormalization group techniques [34. The effect of this change on the fitness of the nearest neighbours is not known a priori (it may be beneficial or not). all below a fixed value x. Rather than considering large–scale extinctions as triggered by external catastrophic events (such as meteorites or major environmental changes) and small–scale extinctions as caused by evolutionary factors. 43]. For x ≈ τ the avalanche distribution is a power law P (s) ∝ s−χ with an exponent χ = 1. If the procedure described above is iterated. In this way the size of an x-avalanche is uniquely defined as the number of mutations between two consecutive configurations where all barriers are above x. one can interpret xi as the barrier against mutation for the genotype of species i: the higher the barrier. SELF-ORGANIZED NETWORKS • a fitness value xi (sometimes interpreted as a fitness barrier ) is assigned to each species i. The species with lowest barrier is therefore the first to evolve. Alternatively. including complex networks [8. the Bak– Sneppen model is ideal for studying the effects introduced by an additional . The Bak–Sneppen model has been studied within a variety of different frameworks ranging from numerical simulation [29.150 CHAPTER 11. It also provides a possible explanation for the phenomena of mass extinctions observed in the fossil records [31]. • the site with the lowest barrier and its nearest neighbours are updated: new random fitness values. theoretical analysis [33]. the reason for updating the nearest neighbours is the same: the mutation of one species changes the state of all the interacting species (for instance. the fitness distribution evolves from a uniform one in the interval [0. drawn randomly from a uniform distribution in the interval [0. 41. 42. 1] to a uniform one in the interval [τ. Finally. 28. 1].073 ± 0. In other words.

THE FITNESS MODEL 151 Figure 11. By contrast. Another successful model.3 The fitness model In chapter 3 we have already discussed various models aimed at generating realistic network topologies. the discrete sum can be approximated by an integral. Right: the probability distribution P (s) for the size of a critical τ -avalanche.4. y) [7.2: Left: plot of the probability distribution of fitness values at the steady state in the Bak–Sneppen model with 500 species. Caldarelli et al. is based on a suitable extension of the Erdos-Renyi random graph defined in chapter 2. . feedback mechanism between fitness dynamics and topological restructuring. so (unsurprisingly) no heterogeneity emerges. In particular. the so-called Fitness Model. xj ). N ) is assigned a fitness xi drawn from a specified distribution ρ(x).11. each pair of vertices i and j is sampled. 11. 66. For instance. [7] have proposed a model where each vertex i (i = 1. the expected degree of vertex i is hki i = X j6=i pij = X f (xi . The expected topological properties of the network can be easily computed in terms of ρ(x) and f (x. Then.5) j6=i For N large. one can define a static model where heterogeneity is explicitly introduced at the level of vertices. all vertices are assumed to be statistically equivalent. . . y)ρ(y)dy (11. Thus the expected degree of a vertex with fitness x is Z k(x) = N f (x. .3.6) . and a link is drawn between them with a fitness–dependent probability pij = f (xi . it is at the basis of the adaptive model [6] that we shall present in detail in section 11. xj ) (11. In the latter. For this reason. 67].

the above connection probability reduces to the bilinear choice f (x. Similarly.11).5) one finds that hki i ∝ xi .(4. Also.10) hki i(hki i − 1) jk pij pki As for eq.7) k then the latter can be easily obtained in terms of the former as P> (k) = ρ> [x(k)] (11. the above expressions can be easily rephrased in terms of integrals involving only the functions f (x. This leads to [68. We also know that.21) is P P j pij hkj i jk pij pjk nn hki i ≈ = P (11. with a cut–off at large degree values that correctly takes . If one chooses a power–law fitness distribution ρ(x) ∝ x−γ .29-4.(11. it is therefore clear that the degree distribution will have exactly the same shape: P (k) ∝ k −γ . y) and ρ(x).12) In this case. when z  1.3.11) where z is a positive parameter controlling the number of links.(11. a sparse graph is obtained where structural correlations disappear. We know from chapter 5 that the resulting network is completely random. which is obtained by requiring that all the realizations of the fitness– dependent network having the same degree sequence occur with the same probability.152 CHAPTER 11. 69] f (x. In the more general case corresponding to eq.1 Particular cases The constant choice f (x.6). the expected value of the average nearest neighbours degree defined in eq.(11. the same choice for ρ(x) yields again a power–law degree distribution. upon which all the results depend. SELF-ORGANIZED NETWORKS where the integration extends over the support of ρ(x). y) = zxy (11. irrespectively of the form of ρ(x).30) is P P jk pij pjk pki jk pij pjk pki = P hCi i ≈ (11. apart for its degree sequence having a definite (fitness-dependent in this case) expected value.(11. If one consider the cumulative fitness distribution and the cumulative degree distribution defined as Z Z +∞ +∞ ρ(x0 )dx0 ρ> (x) ≡ P (k 0 )dk 0 P> (k) ≡ x (11. y) = zxy 1 + zxy (11.8) where x(k) is the inverse of the function k(x) defined in eq.5). from eq. 11. y) = p is the trivial case corresponding to an ErdosRenyi random graph (see chapter 2).9) hki i j pij and the expected value of the clustering coefficient defined in eqs.(4. The simplest nontrivial choice coincides with the configuration model (see Chapter 5).

By contrast. 74]. 73. now plays the role of a positive threshold.11) and eq. 67]. In all these cases. recent approaches to the modelling of complex networks have considered the idea that the topology evolves under a feedback with some dynamical process taking place on the network itself (see for instance refs. y) = Θ(x + y − z) ρ(x) = e−x (11.12). 74. corresponding to eq. is combined to fast relaxation. [6.11. Bianconi and Marsili [73] have defined a model where slow network growth. A SELF–ORGANIZED NETWORK MODEL 153 into account the requirement k ≤ N for dense networks. dissipation is also introduced. For sparse networks. defined as the random rewiring of links connected to congested (toppling) vertices. Among the various contributions. corresponds to the finite– temperature regime. • a small–world effect. it has been shown that both eq. Equation (11. In summary.(11. these correlations disappear. • hierarchically organized clustering coefficients. 47. 75.4 A self–organized network model As we have anticipated. To avoid the collapse to a complete graph.13) where z. This choice yields again a power–law degree distribution P (k) ∝ k −γ (where now γ = 2). . eq. 73.(11. A particular choice of these quantities drives the system to a stationary state characterized by a scale–free topology and a power–law distribution for toppling avalanches. • correlations between neighbouring degrees.11). 76. three groups have considered a possible connection with Self–Organized Criticality [6. the average distance is small. which again controls the number of links. for a series of reasonable choices the networks generated by the fitness model display • a scale–invariant degree distribution. 72. allowing toppling nodes to lose all their links at a given rate. Remarkably. both arising as structural correlations imposed by the local constraints. Another interesting choice is given by f (x. 11. 66.4.13) corresponds to the zero–temperature regime where the structural correlations disappear and the graph reaches a sort of “optimized” topology [71]. Equation (11.13) are particular cases of a more general expression obtained by introducing a temperature–like parameter [71]. The outcomes of the model depend on the dissipation rate and on the probability density function for the toppling probabilities to be assigned at each vertex. with ρ(x) ∝ x−γ . 77]).11) also generates disassortativity and hierarchically distributed clustering. defined as the gradual addition of links between randomly chosen vertices.(11.(11. and hierarchically distributed clustering c(k) ∝ k −2 (times logarithmic corrections) [7. where the temperature can be reabsorbed in a redefinition of x and z. anticorrelated degrees with k nn (k) ∝ k −1 .

and at the stationary state become very similar and scale–free. either in the topology or in the dynamical variables. the overwhelming majority of studies have instead considered the two processes separately. and ad hoc assumptions must therefore be made. In cases when there is indeed such a sharp separation of timescales. as in other previous studies [61]. Unfortunately. 4. Rather. when considering the spreading of epidemics on a network one should assume an arbitrary fixed topology. In addition. On the other hand. The rest of the chapter is devoted to a detailed description of this model. Garlaschelli. In this way. the A ends of links in the network that have not been rewired for the longest time are rewired to the initiator of the avalanche.1 Motivation We have already mentioned that the topology of a network affects dramatically the outcomes of dynamical processes taking place on it [1.2. the idea behind the fitness model presented in section 11. 2. Moreover.2. 11. Capocci and Caldarelli [6] have introduced another fully self– organized model where the Bak–Sneppen dynamics defined in section 11.154 CHAPTER 11. one should assume an arbitrary distribution for the fitness variables. Similarly. are unnecessary. but where each vertex has a different critical height equal to its degree.3 captures the empirically observed result [52. these results imply that in general one should consider the mutual effects that dynamics and topology have on each other. or the opposite situation where the dynamical variables evolve much more slowly than the topology (and are therefore assumed fixed as in the fitness model itself). Clearly. They find that the mutual interplay between topology and dynamics drives the system to a state characterized by scale–free distributions for both the degrees and the fitness values. SELF-ORGANIZED NETWORKS Fronczak. They considered the sandpile dynamics defined in section 11. even when the timescales are indeed well separated. when a network is formed according to the fitness model. For instance.4. these approaches are helpful. the avalanche area distribution and the degree distribution evolve in time. in which case the decoupled approach gives no insight into the real process. Fronczak and Holyst [74] have proposed a model where no parameter choice is required in order to drive the system to the critical region. 53. 78] that the topology of many real networks is strongly dependent on some vertex–specific quantity. by postulating either a scenario where the topology evolves over a much longer timescale than the dynamics.2. they assumed that after an avalanche of size A.2 takes place on a network whose topology is in turn continuously shaped by the fitness model presented in section 11. The choice of both the dynamical rule and the graph formation process was driven by the interest to highlight the novel effects arising uniquely by the feedback . But in many cases the topological evolution and the dynamics may occur at comparable rates. These motivations lead to the definition of a self–organized model [6] where ad hoc specifications of any fixed structure. it is the interplay between dynamics and topology that autonomously drives the system to a stationary state. 5].3. it is clear that the variables involved in the slower of the two processes must be specified as external parameters. These unexpected properties differ from what is obtained when the two models are considered separately.

3 Analytical solution Despite its complexity. while little changes lead to almost equiprobable interactions. all the set of connections between i and the other vertices j 6= i are drawn anew with updated probability f (x0i . 28.e. y) [6].14) where ηj is uniformly distributed between 0 and 1. 41. 43]. 40. 11. and between all pairs of vertices i and j a link is drawn with probability f (xi . 42. In particular. each of the N vertices is assigned a fitness value xi . at the stationary state the fitness values are uniformly distributed above a critical threshold τ . 40. whenever a species i is assigned a new fitness x0i . macroevolution is believed to be at the same time the cause and the effect of food web dynamics [47]. An example of this evolution rule is depicted in figure 11. By contrast. two extremely well understood models where chosen. 39. Then. the model is exactly solvable for any choice of the connection probability f (x. i. 43]. At the initial state the network is generated as in the fitness model. a species is expected to develop a new set of interactions with the other species. at time t + 1 the fitness of i as well as that of its neighbours is updated.4. 38. 28. 42. On regular lattices [8. the assumption of a static graph leads to the ecological paradox that. This means xj (t + 1) = ηj if j = i or aij (t) = 1 (11. A SELF–ORGANIZED NETWORK MODEL 155 introduced between them. xj ). The only dependence on the particular topology is the value of τ [8. τ vanishes for scale–free degree distributions with diverging second moment [41.3. drawn anew from the uniform distribution on the unit interval. If at time t the vertex i has the minimum fitness. 42. random graphs [28]. This automatically implies that major mutations (a large change in xi ) are associated with very different connection probabilities. 41.2). 43]. 11. 42. one can write down a so-called master . Indeed. the extremal fitness dynamics of the Bak–Sneppen model (see section 11.2 Definition In order to define an improved.11.4. Similarly. the topology generated by the fitness model can be completely calculated for any distribution of the fitness values. and on the other hand the fitness network model (see section 11. In particular. On a generic graph. and ki + 1 new fitness values (drawn from the same uniform distribution) are assigned them. the outcomes of the Bak–Sneppen model on several static networks are well studied [8. as for the one–dimensional model.4. On one hand. 39. 43] networks it has been shown that. after a mutation.2.3. As we have shown in section 11. initially drawn from a uniform distribution between 0 and 1.3). While these more complicated networks are closer to realistic food webs (i. evolving model. 39]. the evolved species inherits the same connections of the previous species. small–world [40] and scale–free [41. one can assume that the Bak– Sneppen dynamics is combined with a fitness–driven link updating.e. xj ) (where the xi ’s are the initial fitness values). Therefore. At each timestep the species i with lowest fitness and all its ki neighbours undergo a mutation. the real-world predator-prey networks [48]). as in the one–dimensional case. after a mutation.

Note that we are not assuming from the beginning that τ > 0 as is observed for the Bak–Sneppen model on other networks (with finite second . y). The range where ρ(x) ≈ q(x)/N holds can be defined more formally by introducing the fitness value τ such that  x≤τ N ρ(x)  = 1 lim (11.17) N →∞ q(x)  >1 x>τ This means that in the large size limit the fitness distribution for x < τ is determined by the distribution of the minimum.3: Example of graph evolution in the self–organized model. the value of τ can be determined by the normalization condition Z 1 ρ(x)dx = 1 (11. If a stationary distribution (time–independent) distribution ρ(x) exists. equation for the evolution of the fitness distribution ρ(x. t) =0 ∂t ⇒ rin (x) = rout (x) (11.18) 0 as we show below. The minimum–fitness vertex (black) and its two neighbours (gray) undergo a mutation: three new fitness values are assigned them (light grey). t) ∂t (11. the probability densities) of vertices with fitness x entering and exiting the system at time t respectively. t) − rout (x. After an expression for ρ(x) is derived. t) at time t: ∂ρ(x. ρ(x) must be very close to q(x)/N (the distribution of all fitness values must be approximated by the correctly renormalized distribution of the minimum). it is found by requiring ∂ρ(x. it is useful to introduce the probability density q(m) that the minimum fitness takes the value xmin = m. To this end.16) where at the stationary state the quantities no longer depend on time. t) and rout (x. then the above condition will give the stationary form of ρ(x) for any choice of f (x. SELF-ORGANIZED NETWORKS Figure 11. t) = rin (x. and new links are drawn between them and all the other vertices.156 CHAPTER 11. y) and ρ(x). For x small enough. If one manages to write down rin (x) and rout (x) in terms of f (x. t) are the fractions (strictly speaking.15) where rin (x.

where rin (x|m). This means rout (x|m) = Θ(τ − x) δ(x − m) + Θ(x − τ )ρ(x)f (x. for which the independence on x does not hold. A SELF–ORGANIZED NETWORK MODEL 157 moment of the degree distribution). then on average 1 + k(m) new fitness values are updated. m) N (11.(11. If the minimum fitness is m. Now consider rout (x). Also. note that lim q(x) = 0 for x > τ. For x > τ .21) where Θ(x) = 1 if x > 0 and Θ(x) = 0 otherwise.19) N independently of x. This directly implies rin (x) = Z τ 0 q(m)rin (x|m)dm = 1 + hkmin i N (11. that can be calculated in a way similar to eq. For x ≤ τ . Let us consider rin (x) first. since the minimum is surely replaced and the probability of having other vertices with fitness in this range is zero.6). Thus the normalization condition for q(x) reads Z τ q(x)dx = 1 as N → ∞. one has 1 + k(m) rin (x|m) = (11. N →∞ since eq. m) that a vertex with fitness x is connected to the vertex with minimum fitness m [6].(11.18) yields τ = 0.11. where k(m) is the expected degree of the minimum–fitness vertex. rout (x|m) = δ(x − m)/N (where δ(x) is the Dirac delta function). An integration over q(m)dm .20) Rτ where hkmin i ≡ 0 q(m)k(m)dm is the expected degree of the vertex with minimum fitness (irrespective of the value of the minimum fitness itself).4. the fraction of vertices with fitness x that are removed equals ρ(x) times the probability f (x.17) implies that the minimum is surely below τ . y) eq. Since each of these 1 + k(m) values is uniformly drawn between 0 and 1. signalling the absence of a nonzero threshold. It may well be that for a particular choice of f (x. rout (x|m) are conditional probabilities corresponding to the densities of vertices with fitness x which are added and removed when the value of the minimum fitness is m.(11. 0 The knowledge of q(m) allows one to rewrite rin (x) and rout (x) as Z rin (x) = q(m)rin (x|m)dm and r out Z (x) = q(m)rout (x|m)dm respectively. a quantity that can be derived independently of k(m) as we show below.

158

CHAPTER 11. SELF-ORGANIZED NETWORKS

yields
r

out

τ

Z
(x)

q(m)rout (x|m)dm

=
0


 q(x)/N
=
 ρ(x) R τ q(m)f (x, m)dm
0

x≤τ

(11.22)

x>τ

Finally, one can impose eq.(11.16) at the stationary state. For x ≤ τ , this
yields q(x) = 1+hkmin i independently of x. Combining this result with q(x) = 0
for x > τ as N → ∞, one finds that the distribution of the minimum fitness m
is uniform between 0 and τ :
q(m) = (1 + hkmin i)Θ(τ − m)

(11.23)

Requiring that q(m) is normalized yields
hkmin i =
and
q(m) =

1−τ
τ

(11.24)

Θ(τ − m)
,
τ

which implies
ρ(x) =

q(x)
1
=
N
τN

for x ≤ τ

and, from eq.(11.20),
rin (x) =

1
τN

∀x.

(11.25)

For x > τ , eq.(11.16) implies
ρ(x)

=


0

=
=
=

rout (x)
q(m)f (x, m)dm

rin (x)
q(m)f (x, m)dm
0
1

τ N 0 q(m)f (x, m)dm
1

.
N 0 f (x, m)dm

(11.26)

Therefore the exact solution for ρ(x) at the stationary state is found [6]:


 (τ N )−1
x≤τ
ρ(x) =
(11.27)
1


x>τ

N 0 f (x, m)dm
where τ is determined using eq.(11.18), that reads
Z 1
dx

=N −1
f
(x,
m)dm
τ
0

(11.28)

11.4. A SELF–ORGANIZED NETWORK MODEL

159

The above analytical solution holds for any form of f (x, y). As a novel result,
one finds that ρ(x) is in general no longer uniform for x > τ . This unexpected
result, which contrasts with the outcomes of the Bak–Sneppen model on any
static network, is solely due to the feedback between topology and dynamics.
At the stationary state the fitness values and the network topology continue to
evolve, but the knowledge of ρ(x) allows to compute the expected topological
properties as shown in section 11.3 for the static fitness model.

11.4.4

Particular cases

In what follows we consider specific choices of the connection probability f (x, y).
In particular, we consider two forms already presented in section 11.3. Once
a choice for f (x, y) is made, one can also confirm the theoretical results with
numerical simulations. As we show below, the agreement is excellent.
The random neighbour model
As we have noted, the trivial choice for the fitness model is f (x, y) = p, which is
equivalent to the Erdos-Renyi model. When the Bak–Sneppen dynamics takes
place on the network, this choice removes the feedback with the topology, since
the evolution of the fitness does not influences the connection probability. Indeed, this choice is asymptotically equivalent to the so–called random neighbour
variant [28] of the Bak–Sneppen model. In this variant each vertex has exactly d
neighbours, which are uniformly chosen anew at each timestep. Here, we know
that for an Erdos-Renyi graph the degree is peaked about the average value
p(N − 1), thus we expect to recover the same results found for d = p(N − 1) in
the random neighbour model.
Homework 11.3 Show that, when f (x, y) = p,

 (τ N )−1
x≤τ
ρ(x) =
 (pτ N )−1
x>τ

(11.29)

where, for N → ∞,



1


1

τ=
(1 + d)−1

1 + pN


 0

if

pN → 0

if

pN → d > 0

if

pN → ∞

(11.30) 

The reason for the onset of these three dynamical regimes must be searched
for in the topological phases of the underlying network. For p large, there is
one large connected component that spans almost all vertices. As p decreases,
this giant cluster becomes smaller, and several separate clusters form. Below
the critical percolation threshold pc ≈ 1/N [4, 5], the graph is split into many
small clusters. Exactly at the percolation threshold pc , the sizes of clusters are
power–law distributed according to P (s) ∝ s−α with α = 2.5 [4]. Here we find
that the dense regime pN → ∞ is qualitatively similar to a complete graph,

160

CHAPTER 11. SELF-ORGANIZED NETWORKS

where many fitness values are continuously updated and therefore τ → 0 as in
the initial state (thus ρ(x) is not step–like). In the sparse case where pN = d
with finite d > 1 as N → ∞, then each vertex has a finite number of neighbours
exactly as in the random neighbour model, and one correctly recovers the finite
value τ = (1 + d)−1 found in ref. [28]. The subcritical case when p falls faster
than 1/N yields a fragmented graph below the percolation threshold. This is
qualitatively similar to a set of N isolated vertices, for which τ → 1. It is
instructive to notice from eq.(11.27) that the choice f (x, y) = p is the only one
for which ρ(x) is still uniform. This confirms that, as soon as the feedback is
removed, the novel effects disappear.
Homework 11.4 Show that, in the stationary state of this specification of the
R1
model, the average fitness over all nodes (defined as hxi ≡ 0 xρ(x)dx) equals
hxi =

1+τ
2 

The above exercise shows that, while at the initial state (t = 0) the average
fitness is hx(0)i = 1/2 (since the distribution is uniform in the unit interval),
at the stationary state (t = ∞) the average fitness increases to the asymptotic
value hx(∞)i = (1 + τ )/2 ≤ 1/2.
The self–organized configuration model
Following the considerations in section 11.3, the simplest nontrivial choice for
f (x, y) is given by eq.(11.11). For a fixed ρ(x), this choice generates a fitness–
dependent version of the configuration model [4, 70], where all graphs with the
same degree sequence are equiprobable. All higher–order properties besides the
structural correlations induced by the degree sequence are completely random
[68, 69]. In this self–organized case, the degree sequence is not specified a priori
and is determined by the fitness distribution at the stationary state. Inserting
eq.(11.11) into eq.(11.27) one finds a solution that for N → ∞ is equivalent to
[6]

 (τ N )−1
x≤τ
ρ(x) =
(11.31)
 (τ N )−1 + 2/(zN τ 2 x)
x>τ
where τ , again obtained using eq.(11.28), is


 1
r

 p
φ(zN )
τ=

φ(d)/d

zN


 0

zN → 0
zN = d

(11.32)

zN → ∞

Here φ(x) denotes the ProductLog function, defined as the solution of φeφ = x.
Again, the above dynamical regimes are related to three (subcritical, sparse and
dense) underlying topological phases. This can be ascertained by monitoring
the cluster size distribution P (s). It is found that P (s) develops a power–law
shape P (s) ∝ s−α (with α = 2.45 ± 0.05) when d ≡ zN is set to the critical

1 10 0 10 1 2 10 Cluster Size s 10 3 10 4 Figure 11. A SELF–ORGANIZED NETWORK MODEL Cluster Size Distribution P(s) 10 161 4 10 2 10 0 10 −2 10 −4 d = 1.7) and the behaviour of τ (zN ) are plotted. Note that. which therefore represents the percolation threshold.11. Here N = 3200. while for d > dc it takes increasing finite values.4.32 d=4 d = 0. the simulations are in very good accordance with the analytical solution. as we have discussed in section 11. The theoretical predictions in eqs. 11.4). This quantity is negligible for d < dc . where the cumulative fitness distribution ρ> (x) defined in eq.6. It originates from the fact that at any time the fittest species is also the most likely to be selected for mutation. Here.3.11. This behaviour can also be explored by measuring the fraction of vertices spanned by the giant cluster as a function of d (see fig. Far from the critical threshold (d = 0.(11.31) mean that ρ(x) is the superposition of a uniform distribution and a power–law with exponent −1.5).5. 11. P (s) is well peaked. 11. this implies a purely power–law behaviour ρ(x) ∝ x−1 for x > τ . in the sparse regime z  1 one has f (x. Also.(11.31) and (11. In the dense regime obtained for large z. one can plot the average size fraction of non–giant components. At dc = 1.32 ± 0. The analytical results in eq. value dc = 1. This shows one effect of structural correlations. this quantity diverges at the critical point where P (s) is a power law.4: Cluster size distribution. As shown in the inset of fig. (After ref.45 ± 0. . Therefore ρ> (x) is a logarithmic curve that looks like a straight line in log–linear axes. y) ≈ zxy.(11. since it has the largest probability to be connected to the least fit species.1 and d = 4). The decay of ρ(x) for x > τ is entirely due to the coupling between extremal dynamics and topological restructuring.32) can be confirmed by large numerical simulations.05. P (s) ∝ s−α with α = 2. This is opposite to what happens on fixed networks. Indeed. [6]).32. the uniform part gives instead a significant deviation from the power–law trend.05 [6] (see fig. This is shown in fig.

For large z the presence of structural correlations introduces a sharp cut–off for P (k). while for large z a saturation to the maximum value kmax = k(1) takes place.4 0. and again the agreement between theory and simulations is excellent. [6]).162 CHAPTER 11.6) one can obtain the analytic expression of the expected degree k(x) of a vertex with fitness x: k(x) = 2 1 + zx zx − ln(1 + zx) ln + zτ 2 1 + zτ x zτ x (11.(11. SELF-ORGANIZED NETWORKS 1 N = 100 N = 200 N = 400 N = 800 N = 1600 N = 6400 0. this implies that in the sparse regime P (k) has the same shape as ρ(x). 42.4 1.8) allows to obtain the cumulative degree distribution P> (k). Other effects are evident when considering the degree distribution P (k).2 1.5: Main panel: the fraction of nodes in the giant component for different network sizes as a function of d.6 2 Non−Giant Component Size Giant Component Fraction 0.8 0. For small z. the average fitness over all nodes equals hxi =  1 2 + (1 − τ ) 2τ N zN τ 2 .7. k(x) is linear. (After ref. Using eq.33) Computing the inverse function x(k) and plugging it into eq.(11. in the stationary state of this specification of the model.8 1. Inset: the non-giant component average size as a function of d for N = 6400.5 Show that.6 1.2 1 0 1 2 3 4 d = Nz 0 0 2 4 6 8 10 d = Nz Figure 11.3. As discussed in section 11. Homework 11. 43].11. Both quantities are shown in fig. Another difference from static networks is that here τ remains finite even if P (k) ∝ k −γ with γ < 3 [41.

1.005 0.001 1000 10000 100000. z = 1000 (N = 5000).05 0.02 0. novel and unpredictable results are found. This indicates that adaptive networks pro- . we have argued that in many cases of interest it is not justified to decouple the formation of a network from the dynamics taking place on it. This effect is due to the non-uniform character of ρ(x) at the stationary state.11. (After ref. Solid lines: theoretical curves. and by no means complete. From right to left. We have emphasized the importance of introducing similar mechanisms in the study of networks.4 0. open dissipative systems appear to evolve spontaneously to a state where the response to an infinitesimal perturbation is characterized by avalanches of all sizes.2 0. z = 1. z = 0.01 0. by presenting an extensive study of a self–organized network model. summary of the ideas that inspired much of the research on scale–invariance and self–similarity.05 0. [6]). z = 10. In particular. · 106 0. z = 0. from the early discovery of fractal behaviour to the more recent study of scale–free networks.1 0. the framework of Self–Organized Criticality succeeds in explaining the onset of fractal behaviour without external fine–tuning. We have highlighted the importance of understanding the emergence of the ubiquitously observed patterns in terms of dynamical models. In both cases. one is forced to introduce ad hoc specifications for the process assumed to be slower.01 Nz 0. 11.8 Tau 0.5. Inset: log–log plot of τ (zN ). Indeed. with a higher density of values just above the threshold τ . z = 100.5 Conclusions We have presented a brief. CONCLUSIONS 163 CDF 1 0. In particular.1. According to the SOC paradigm.2 0. The above exercise shows that at the stationary state the average fitness can now be smaller than the initial value 1/2.5 1 x Figure 11. we have shown that if the feedback between topology and dynamics is restored.1 0.6: Main panel: cumulative density function ρ> (x) in log–linear axes.0050.002 100 0.6 0.01. points: simulation results.

01. z = 100.7: Left: k(x) (N = 5000. vide a more complete explanation for the spontaneous emergence of complex topological properties in real networks. from right to left.8 1000 0. points: simulation results. z = 10.1. [6]). Right: P> (k) (same parameter values. (After ref. Solid lines: theoretical curves.164 CHAPTER 11. z = 0. SELF-ORGANIZED NETWORKS 10000 CDF 1 k 0. . z = 1000).01 0.2 0. inverse order from left to right).6 100 0. z = 1.4 10 0.001 0.1 1 x 10 100 1000 k Figure 11. z = 0.

How Long Is the Coast of Britain? Statistical SelfSimilarity and Fractional Dimension. Lett. (2003) 167. 3 813-817 (2007).. Nature 308 721 (1984). 4083-4086 (1993)..N.). [10] Mandelbrot B. Rev. 2000). Barab´ asi A. (2002) 258702.J. [16] Brown J. 52. Mu˜ noz M. [5] Albert R.F. [7] Caldarelli G. Mendes J. Lett. Capocci A. West G.J. 165 .E. (2001) 47–97. Nature Physics.. Fractal Cities: a Geometry of Form and Functions Academic Press. Fractal growth of Copper electrodeposits Nature 309. (1963). 71. Rinaldo A. Phys. J.B..B. Oxford University Press. San Diego (1994) [15] Mandelbrot B. Oxford (2003). Sneppen K. and Wiesmann H.. of Mod. Ball..E.. Cambridge University Press. [6] Garlaschelli D. (1997).. Passoja D. Oxford (2007).. R.. Rev. Business 36 394-419. Longley P. Large Scale Structure and Dynamics of Complex Networks (World Scientific Press. Phys. New York. The variation of certain speculative prices. [3] Dorogovtsev S.. (eds.. Rev. Vespignani A... Scaling in biology (Oxford University Press.C. Capocci A. Singapore 2007).M. Phys. SIAM Rev. 89. [13] Brady R. Fractal River Networks: Chance and Self-Organization. Newman.. [4] M. Science 156..Bibliography [1] Caldarelli G.F.. Rev. [8] Bak P. Paullay A...A.B. [2] Caldarelli G. I.H. 45..-L. Fractal character of fracture surface in metals. Caldarelli G.J. De Los Rios P.. Phys. Evolution of Networks: From Biological Nets to the Internet and WWW. (eds). 1033 (1984) [12] Rodriguez-Iturbe. Fractal Dimension of Dielectric Breakdown. [14] Batty M.. [11] Niemeyer L. 74.B. Scale-Free Networks Oxford University Press. Pietronero L. [9] Mandelbrot B. Lett. A. 636-638 (1967).. 225 (1984)...

Caldarelli G. Acad.. Rodr´ıguez-Iturbe I. [29] Grassberger P.. Phys. L13502 (2004).. Punctuated equilibria: an alternative to phyletic gradualism. Mean Field Theory for Ordinary and Hot sandpiles. Lett. [33] Dorogovtsev S.. Reprinted in N.. Braz. 337-343.. A 16. C. Lett.F... 34. [28] Flyvbjerg H. Phys. Phys. pp... [30] Dickman R. Sur une courbe dont tout point est un point de ramification.V. Phys. In T. Sneppen K.. Phys. Geoph. Lett. Maritan A. Mendes J. Not.. [25] Caldarelli G.K. (1998). Zapperi S. [22] Caylor K. Europhys. Phys. J.. Lett..166 BIBLIOGRAPHY [17] Sierpi´ nski W. [34] Marsili M... A 31. Lett. 4087 (1993).N. Lect.. Chapman and Hall... [20] Rigon R..J. [21] Marani M. Banavar J. 385 (1994). Europhys. [26] Bak P. Scanlon T. 252.. Phys. 625. E 62 295 (2000).A. Lett.. Vespignani A.. and Sol´e R. (1998). Feasible optimality of vegetation patterns in river basins. Lett. 1985 [19] Jensen H.A.. 381 (1987). The fossil record 2. J. 82-115 (1972)..M. Cambridge. 71. Models in Paleobiology. Invasion Percolation: a new form of Percolation Theory. 3181-3190 (1998). Gould S. Feasible optimality implies Hack’s law. [18] Eldredge N. (2003) [24] Caldarelli G. 295-307 (1998). Rev. San Francisco: Freeman Cooper. 59. Paris 160 302-305 (1915). Princeton: Princeton Univ. Rinaldo A. Rev. Rodr´ıguez-Iturbe I.. 114-126. London. Maritan A. E 55 3708 (1997). Phys. 80 5746 (1998). 3365-3376 (1983).M... Rev. R. Rev.. J.J. [27] Wilkinson D.F. Optimisation in Complex Networks.. 35 481-486 (1996). Bak P. Hot sandpiles. Stationary self-organized fractal structures in potential force fields.. Physica A. Rinaldo A. Press. [31] Benton M.G. [35] Mikeska B. A 200 277 (1995). Vendruscolo M.. Self-Organized Criticality Cambridge University Press.. Europhys. Weisenfeld K. 28.Mu˜ noz M. Willemsen J. Res. Lett.F....J. [23] Ferrer i Cancho R... in Phys. (1993). Rev... .. [36] Paczuski M. Sci. Bak P. Vendruscolo M.. Pogorelov Y.. Marsili M. Res.. Tang C.. [32] De Los Rios P. Phys. Water Res. J.31. ed. Maslov S. 30 27 (2000). Eldredge Time frames. 27 97 (1994). Schopf.

Rev. Goh. Proc.. Garlaschelli D. [51] Caretta Cartozo C. Comm. [42] Lee.. Faloutsos C. 165-168 (2003). Lett. . Rev. Physica A 342. 57(5).. (2001) 1896-1899. Science 287.. E 71. Felici M. & Stroud.A. Vendruscolo. Caldarelli. Biol. 066106 (2005). Rev. R. [46] Giot L et al. Castri. [53] D. Faloutsos P. & Kahng.. Asymmetric dynamics and critical behavior in the Bak-Sneppen model. [50] Burlando B. 93.H. Almaas. Universal scaling relations in food webs. & Vazquez. 86.. Pietronero L. A: Math. Lett. Servedio and G. Ricotta C. Journ.J. Y. [49] Burlando B. D.M.. Phys. [47] G.. 516-528 (2004). Loffredo. McKane. Journal Theoretical Biology 146 99-114 (1990). 224012 (2008). Extremal dynamics on complex networks: Analytic solutions. [44] Garcia G... A. (2005) 491-499. [45] Middendorf M.. 41. K. [38] M. [48] Garlaschelli D. E 72. Battiston. J. Theor. De Los Rios. 057102 (2005). B. Comp. Huberman B. [43] Masuda. Caldarelli G. [52] D. Lett. Marsili and M. Rev.I. Europhys. Dickman R.. 80.. A. Gabrielli A. [39] P. E 65 (2002) 046101. (2004) 188701. Sci. Phys. Pietronero. Phys. A protein interaction map of Drosophila melanogaster. G. V. M. Journal Theoretical Biology 163 161-172 (1993). On Power-law Relationships of the Internet Topology. 193. Gabrielli. Lett. [55] Adamic L. (1998) 345. 3192-3197 (2005). Caldarelli G. P. Rev. Wiggins C.. Nat.. Rev. Proc. L.. Caldarelli. Theor. M. [54] Faloutsos M. 29.. S.BIBLIOGRAPHY 167 [37] Caldarelli G. N. Barth´elemy M. Felici.G. E. A 350. Evolutionary dynamics in the Bak–Sneppen model on small–world networks.-I. Acad. (1998) 5746-5749. Caldarelli.A.. 2115 (2000). The Bak–Sneppen model on scale–free networks. [40] Kulkarni.251262 (1999). Phys. [41] Moreno. Higgs and A. S. Nature 423.J. Inferring network mechanisms: The Drosophila melanogaster protein interaction network. Phys. 765–771 (2002). Phys.. Y. Ziv E. Garlaschelli and M.D. Coevolutionary dynamics on scale-free networks. Pietronero L.P. ACM SIGCOMM. 102. ArXiv:cond-mat/9905066. Science 302 1727-36 (2003). V. Rev. Power-Law Distribution of the World Wide Web. Phys. Garlaschelli. & Kim.. Phys.

. 148701 (2003). Mendes. Phys. Phys.D. Rev. Fronczak P. Fronczak. Physica D 224(1-2). & Holyst. Rate equation approach for correlations in growing network models. J. The diameter of a scale-free random graph.-S. Europhys. Marchetti R. Rev.A.-L. [62] Barab´ asi A. Park and M. Kim. Loffredo. Ahnert. [67] V.J. Lett. E 73.E. 68. 224018 (2008)..168 BIBLIOGRAPHY [56] Caldarelli G. (2003) 036112. F. ArXiv:cond-mat/0609015. Lee. Rev. (2004) 529. S. Physica A 333.I. 41. [57] R. H. Maslov. E 68.. D. Theor. [72] Jain. Rev. Garlaschelli and M. Lett. 5684–5687 (1998). A. A. [59] D. Fink. E 70. & Krishna. and A. D. [70] S. Rev. Lett.. A. [61] K.. [74] Fronczak. [60] D. Riordan O.. Bogu˜ n´ a and R.-I. Critical phenomena in complex networks. 52. 156–165 (2006). Phys. Rev. J.P. & Gil. Opinion spreading and agent segregation on evolving networks. [63] Fronczak A.. V. S. 035105(R) (2004). R. [71] D. S. 509-512 (1999). Phys. Garlaschelli and M. Combinatorica.. Caldarelli. [64] Barrat A. 24. 81.0010v6. Sneppen. Goh. Pastor-Satorras. Mean-Field theory for clustering coefficient in Barab´asi-Albert networks. [66] M. A. Rev. Emergence of scaling in random networks. Phys. Self–organized criticality and coevolution of network structure and dynamics. G. K. E 70 (2004) 056126. Garlaschelli.. A: Math. B. Holyst J.. J. and Pietronero L. Newman. Rev. Phys. 86. Phys. 5-34 (2004). A. E. Rev.I. 113-118 (2004). Vespignani. Dorogovtsev. Albert R. N.. Physica A 338(1-2). [75] Zanette. Lett. E 68. ArXiv:condmat/0606805v1. Loffredo. F. S. 91. Rev. Phys. Kahng. Zaliznyak. Loffredo. T. Autocatalytic Sets and the Growth of Complexity in an Evolutionary Model. . P. arXiv:0705. Butt`a. [58] S. M. Science 286. 036127 (2005). M. Goltsev. Pastor-Satorras. 71.I. Garlaschelli and M. & Marsili. 046117 (2006). [73] Bianconi. Phys. Servedio. 386 (2000). 3200 (2001). E. [69] D. Pastor-Satorras R. E. Caldarelli and P. Phys. (2003) 026112.D. Clogging and self–organized criticality in complex networks.. [68] J. G. Phys. [65] Bollob´ as B. 046126 (2003). G.

D. Pacheco. Caldarelli. C. Garlaschelli. D. & Lenaerts.. Barrat. European Physical Journal B. PLoS Comput. Phys. Content-based networks: A pedagogical overview. . Cooperation Prevails When Individuals Adjust Their Social Ties. [77] B. Kozma. & Erzan. T. [78] Balcan. CHAOS 17.BIBLIOGRAPHY 169 [76] Santos. [79] G. A Self–organized model for network evolution. A. Biol. 026108 (2007). F. M. A. A. 2(10): e140 (2006). J. 016102 (2008). E 77. in press (2008). Capocci. Rev.

NetLogo is a free agent-based programming language 170 . 12. One of the tools that can be used to visualize the dynamics of unfolding complex processes is Netlogo. a sequence of random graphs. In a previous chapter we have seen the principle of preferential attachment.0.1 Netlogo Many computer scientists are fascinated by the dynamics of complex networks. how they grow. Do not print the graph. Generate G. . Higher order properties depend on neighbors of vertices. for n ∈ 0. the degree of vertices is one of the most useful properties. The Internet is a dynamic entity.1. 12. 1. . since it depends on a single vertex. 100 vertices. A second topic that will be covered in this chapter is that of the implementation of higher order properties.1 Question: Random graph process. Exercise 12. Recall from the beginning of this course the random graph process. Most real world networks evolve. using the same four edge probabilities p = 0. and how we can see how they change.0. . only print mean and variance of the number of edges.5.1 Network Dynamics We will start with network dynamics. 0. . In the next exercise. and can tell us more about properties of the network as a whole. friendship networks evolve. business networks evolve. Netlogo. In this chapter we will see examples of how networks change in time. you will implement the random graph process.1. 0.Chapter 12 Visualizing Dynamics and Higher Order Properties This chapter is about visualization. As has been noted in a previous chapter. Vertex degree is a first order property. We will use Gnuplot to visualize these lower and higher order properties. In this chapter we will introduce the second high level tool to study networks.

) . USA [1]. The NetLogo environment enables exploration of emergent phenomena. read the manual.1. or ask the course assistants for help. This makes it possible to explore the connection between the micro-level behavior of individuals and the macro-level patterns that emerge from their interaction. biology. These simulations address content areas in the natural and social sciences including biology and medicine. physics and chemistry. NetLogo lets students open simulations and ”play” with them. Evanston.12. or File/Models) to browse through some example models.0. Modelers can give instructions to hundreds or thousands of ”agents” all operating independently.northwestern. and have played a bit with it before the lecture starts. NetLogo is simple enough for students and teachers. and economics and social psychology. (Please refer to figure 12. When downloading. It is also an authoring environment which enables students. NetLogo is particularly well suited for modeling complex systems developing over time. mathematics and computer science. Illinois. we will not use 3D). It comes with an extensive models library including models in a variety of domains. You really want to have Netlogo running on your machine.1 for a related screenshot. It also comes with the Models Library. It is a wonderful system for exploration and creating an understanding of the dynamics of complex systems. Netlogo is created by Uri Wilensky of Nortwestern University. Ctrl-M. Installing NetLogo You should now download and install NetLogo onto your computer. It was authored by Uri Wilensky in 1999 and has been in continuous development ever since at the Center for Connected Learning and Computer-Based Modeling. teachers and curriculum developers to create their own models. physics. and go to the Models Library (Command-M. NetLogo has extensive documentation and tutorials. psychology. The homepage for Netlogo is http://ccl. choose version 5. or Windows. Mac OS X. look at the tutorial. and choose your operating system: Linux.5. NETWORK DYNAMICS 171 and integrated modeling environment. If you run into problems. and comes with an extensive library of example models. exploring their behavior under various conditions. a large collection of pre-written simulations that can be used and modified. chemistry. Alt-M. Now please startup NetLogo (the normal version. such as economics. To quote the documentation: NetLogo is a programmable modeling environment for simulating natural and social phenomena. that we gratefully acknowledge. which should get you to the homepage). fill in your name if you wish. Now go to the webpage (or search for Netlogo. system dynamics.edu/netlogo/ NetLogo is freely downloadable. yet advanced enough to serve as a powerful tool for researchers in many fields.

1: Open Fireworks . VISUALIZING DYNAMICS AND HIGHER ORDER PROPERTIES Figure 12.172CHAPTER 12.

while everybody else only has a few. This model generates these networks by a process of preferential attachment.2 Preferential Attachment Preferential attachment is a basic concept of our field. NETWORK DYNAMICS 173 Figure 12. It has been covered at some depth in section 3. We repeat the text from the Info tab of Netlogo.2). From chapter 2.12. This is the mechanism which is called . a few hubs have many connections.2: Preferential Attachment 12. Load the model (See figure 12. in which new network members prefer to make a connection to the more popular existing members. recall that in some networks.2. We will now load an instructive animation of preferential attachment. a nodes chance of being selected is directly proportional to the number of connections it already has. This model shows one way such networks can arise. a new node is added. More specifically. A new node picks an existing node to connect to randomly.1. ranging from the connections between websites to the collaborations between actors. Such networks can be found in a surprisingly large range of real world situations.1. but with some bias. or its degree. The model starts with two nodes connected by an edge. At each step.

but both axes are on a logarithmic scale. You can see the degree distribution of the network in this model by looking at the plots. that is. Pressing the GO ONCE button adds one new node. One simple way to think about power laws is that if there is one node with a degree distribution of 1000. This procedure attempts to move the nodes around to make the structure of the network easier to see. The RESIZE-NODES button will make all of the nodes take on a size representative of their degree distribution.1. and 100 nodes with a degree distribution of 10. Power law distributions are different from normal distributions in that they do not have a peak at the average. and they are more likely to contain extreme values (see Albert & Barab´asi 2002 for a further description of the frequency and significance of scale-free networks). and then want the network to have a more appealing layout. you can turn off the LAYOUT? and PLOT? switches and/or freeze the view (using the on/off button in the control strip over the view). then there will be ten nodes with a degree distribution of 100. If you press it again the nodes will return to equal size. What is the shape of the histogram in the top plot? What do you see in log-log plot? Notice that the log-log plot is only a straight line for a limited range of values. . When degree distribution follows a power law. The LAYOUT? switch controls whether or not the layout procedure is run. If you have LAYOUT? switched off. Barab´asi and Albert originally described this mechanism for creating networks.2 Let the model run a little while. Why is this? Does the degree to which the log-log plot resembles a straight line grow as you add more node to the network? 12. have many connections? How many have only a few? Does some low degree node ever become a hub? How often? Turn off the LAYOUT? switch and freeze the view to speed up the model. press GO. If you want the model to run faster. it appears as a straight line on the log-log plot. press the REDO-LAYOUT button which will run the layout-step procedure until you press the button again. Things to Notice The networks that result from running this model are often called scale-free or power law networks. How many nodes are hubs. These are networks in which the distribution of the number of connections of each node is not a normal distribution instead it follows what is a called a power law distribution.3 Percolation Transition We can use our graph program to study the percolation transition of the clustering of the network that was discussed in chapter 8. then allow a large network to form. The bottom plot shows the same data. To continuously add nodes. Exercise 12. The top plot is a histogram of the degree of each node.174CHAPTER 12. You can press REDO-LAYOUT at any time even if you had LAYOUT? switched on and it will try to make the network easier to see. The PLOT? switch turns off the plots which speeds up the model. The LAYOUT? switch has the greatest effect on the speed of the model. VISUALIZING DYNAMICS AND HIGHER ORDER PROPERTIES preferential attachment. but there are other mechanisms of creating scale-free networks and so the networks created by the mechanism implemented in this model are referred to as Barab´asi scalefree networks.

1 Gnuplot Gnuplot is a tool that allows easy plotting of statistical plots.2. Google Sheets. etc. Now we turn to Netlogo. p = 1/n. Explain what you see using the concept of percolation transition (refer to the chapter on random graphs). if you took soil samples. choose three values for p: pick p < 1/n. Read the Info tab. Compute the clustering coefficient for all three values of p. and p > 1/n. OpenOffice Calc. or a dedicated plotting tool such as Gnuplot.2.12.gpl that looks like the following: set term postscript enhanced color 20 set output ’example. Gnuplot expects an input file with some commands. and a setting where it just stops? If percolation stops at a certain porosity. load Earth Science/Percolation. Run the Simulation. Exercise 12. We do so in order to see if the network does indeed have the properties of real world networks. a sequence of pairs or numbers (x. y) is plotted on a 2 dimensional plane. OpenOffice.2 Network Properties Now that we have found a way to create a network that conforms to a prespecified degree sequence. As an example. Refer to your first year introduction to programming course for more (or use Excel. Does the value settle down roughly to a constant? How does this value depend on the porosity? Give the soil different porosity at different depths. Could you reliably predict how deep an oil spill would go or be likely to go? 12.eps’ set xlabel "degree" set ylabel "frequency" set xrange [0 : 10] set yrange [0 : 20] plot "example.4 From the Model Library. This subsection contains a small Gnuplot example to get you going. Apple Numbers.input" using 1:2 title "degree distribution" with lines axes x1y1 ls 1 .3 For n = 100. What do you notice about the pattern of affected soil? Can you find a setting where the oil just keeps sinking. let us create a command file example. Try different settings for the porosity. NETWORK PROPERTIES 175 Exercise 12. Note the plot of the size of the leading edge of oil. How does it affect the flow? In a real situation. it is time to study some network properties.). 12. its still possible that it would percolate further at that porosity given a wider view. The system that you can use to create the plots can be either a generic spreadsheet such as Microsoft Excel. The plots that we will make are quite simple.

VISUALIZING DYNAMICS AND HIGHER ORDER PROPERTIES 20 degree distribution frequency 15 10 5 0 0 2 4 6 8 10 degree Figure 12.gpl then you should get an output file named example.3.176CHAPTER 12.2. In the following exercise you will create plots for three graphs: for the 10-vertex Erd˝os-R´enyi .2 Empirical Network properties We can now start preparing graphs of network properties. As formula for the number of triangles and wedges we recall G X ∆G = i1 .i3 ∈V (12.i2 i3 are present} . X WG = 1{i1 i2 .i3 i1 are present} .1 In the chapter on random graphs the clustering coefficient of a graph has been defined as the ratio of triangles to wedges in the graph ∆G (CG = W ).1) Now add to your program the computation of this wedge-triangle clustering coefficient.i2 i3 .input that looks as follows: 1 2 3 4 5 4 8 9 3 2 When you run the command \$ gnuplot example.i2 .3: Gnuplot of the example file set term x11 set output Assume that you have an input file named example.eps that looks like Figure 12. Homework 12. 12.i3 ∈V 1{i1 i2 . i1 . Hint: the definitions of ∆G and WG suggest the use of nested for-loops for a straightforward algorithm.i2 .

Plot on the x-axis the value k of the degree from 0 to the highest occurring degreevalue. Answer any questions that were asked at plot number 2 and 4. some take a bit more work.2.5 It is instructive to see the behavior of these other properties on your networks (and you can of course define your own networks by defining new degree-sequences).axis the count of the number of nodes that have that degree. Create plots for each of the following network properties: empirical degree distribution. does the network display assortative mixing)? 3. and for the 10-vertex and for the 20-vertex Repeated Configuration Model graphs. What do your three networks do? Homework 12. Empirical Clustering Coefficient versus Degree As chapter 4 noted. Exercise 12. or kinn .12. 1. Note that in many cases the average clustering decreases as k decreases. NETWORK PROPERTIES 177 graph.2. Empirical Average Nearest-Neighbour Degree A second-order topological property is the average nearest-neighbour degree of a vertex i. Some are easy to implement. 2. Note that chapter 4 contains more higher order properties of networks. against the value of k on the x-axis. divided by the total number of pairs of neighbours of node i. empirical local clustering coefficient and clustering coefficient versus degree.2. Empirical Degree distribution The degree distribution is a first-order topological property. 12 plots in total. . empirical nearest-neighbour degree. 4. It can be computed by average kinn over all vertices with the same degree k and plot this value on the y-axis. a statistical way to consider the clustering properties of real networks. is to compute the average value of Ci over all vertices with a given degree k and plotting this value (on the y-axis) against k (on the x-axis). and on the y. Are degrees positively correlated (are highdegree vertices on average linked to high-degree vertices. See section 4. Empirical Local Clustering Coefficient The most studied third-order property of a vertex is the local clustering coefficient Ci . create the program to print the four plots of the network properties for the three networks.2 Plotting Network Properties From the above list. It is defined as the number of links connecting the neighbours of node i to each other.

northwestern. Center for Connected Learning and Computer-Based Modeling. U. 178 . IL. (1999). Evanston. http://ccl.Bibliography [1] Wilensky. NetLogo. Northwestern University.edu/netlogo/.

2 1 You may want to read up on adjaceny lists and linked list data structures on Wikipedia or on programming websites such as Stackoverflow. of the configuration model. Storing large networks that are sparse in an adjacency matrix would mean allocating storage for V 2 edges that would mostly be zero. An adjacency list is a linked list.1 Why? We will now change the primary data structure of the programs that we have used so far from an adjacency matrix to an adjacency list. this was our goal: to see if we could find a way to generate synthetic networks that nevertheless display the same properties as real-world networks. load them into our program. then it should now be easy to replace the implementation by another one. A more efficient method to use memory is to only store the non-zero edges. 13. In a sense. All networks so far have been artificially generated networks. 2 www. Real real-world networks consist of larger numbers on vertices and edges. The networks that we generated so far consist of modest numbers of vertices and edges that fitted easily in the memory of our computer.boost. For sparse networks with O(E) ≈ O(V ) edges. this would save O(V 2 ) space for large V .Chapter 13 Analysing Real Networks In the previous chapters we have studied the behavior of regular random networks. which you may recall from your introductory programming classes. and study their properties. In this final chapter we will have a look at the properties of real real-world networks. our programs must be able to handle real networks. or you may use a library implementation such as the Boost library. and of their network properties. Exercise 13.1 Adjacency Lists To do so. From a repository of real-world networks we will download a few networks.1 In C/C++ you may implement a linked list from scratch. If you have applied the principles of modularity and abstraction in the data structure implementation. Real real-world networks are also invariably sparse.org 179 .

Use a viewer or an editor (such as Wordpad. Use as a test-case one hundred thousand vertices. or SNAP. Go back to your programming course notes for tips on how to implement file management and reading and parsing input.stanford.001.180 CHAPTER 13. For which p value does the program stop working? For how many edges does it still work? Exercise 13. Clustering Coefficient versus degree. The web abounds with suitable solutions. . specifically the citation network cit-HepPh. Print the number of edges.2 SNAP real networks There are many repositories of real-world networks. ANALYSING REAL NETWORKS Python and Java provide similar solutions. Average Nearest Neighbour Degree. at http://snap. Create a short 5 line example graph to test your input routine.2 Implement an adjacency list data structure for your program.3 Properties of citation network Incorporate the input routine in your adjacency list program.2 Write a short input routine for reading in the graph from the input file. or Emacs) to view the files. Note that the files contain a few lines of commentary (preceded by a hash-mark #) and then lines in which the edges are specified as vertex-vertex pairs. you can unzip the textfile with gunzip. Download cit-HepPh and roadNet-CA. Check that it works correctly with the graphs that the program generates. there are so many solutions out there that some people may find implementing their own linked list implementation quicker. The *. and choose the Stanford Large Network Dataset Collection (the third big blue entry).1 Recompile and test your Erd˝ os-R´enyi program using adjacency lists. Local Clustering Coefficient.2 where you computed the mean an variance of 10 random graphs.3 Mean and variance of edges Incorporate the adjacency list implementation in your Erd˝ os-R´enyi program from Homework Exercise 6. Homework 13.) Exercise 13. or to edit the files and to manually remove the comments before feeding it to your program. 13. The structure of the files is simple. You can choose your own solution: to recognize comments and skip those lines. Plot the same four network properties as for Homework Exercise 12. Does your program work? Use progressively larger values of p. Homework 13. Read in real real world networks from Stanford Large Networks Project. We will now add an input routine to the program. or another solution that works.gz file type is gzip.001. Homework 13. One such repository is the Stanford Network Analysis Project. (In fact. Test the program. TextEdit. which reads the graph from an input file.2 for the Repeated Configuration Model: Degree Distribution. Compute and print the mean number of edges and the variance. VI. and a connection probability of p = 0. Go to SNAP.edu. Choose large values for V = 1000000 and small values for p = 0.

. can you create a pre-specified degree sequence for the repeated configuration model that comes closer to a real real world network than the previous network from exercise 7.5 Create the same plots for the roadNet-CA network.13.2.4 Are the plots different? How are they different? Is the Repeated Configuraton Model a good approximation of a real real-world network? Homework 13.6 Now that you have seen the properties of real real-world networks. SNAP REAL NETWORKS 181 Homework 13.3? Draw up this sequence. How is the roadmap network different from the citation network? Homework 13. and provide the four plots.

Bibliography 182 .