B OLYAI SO CIETY

MATHEMATICAL STUDIES 18
BOLYA
Y I SOCIETY
T MATHEMATICA
C L STUDIES

Series Editor:
Gábor Fejes Tóth
Publication Board:
Gyula O. H. Katona
László Lovász
Péter Pál Pálfy
András Recski
András Stipsicz
Domokos Szász
Managing Editor:
Dezső̋˝ Miklós

1. Combinatorics, Paul Erdős is Eighty, Vol. 1

D. Miklós, V.T. Sós, T. Szőnyi (Eds.)
2. Combinatorics, Paul Erdős is Eighty, Vol. 2
D. Miklós, V.T. Sós, T. Szőnyi (Eds.)
3. Extremal Problems for Finite Sets
P. Frankl, Z. Füredi, G. Katona, D. Miklós (Eds.)
4. Topology with Applications
A. Császár (Ed.)
5. Approximation Theory and Function Series
P. Vértesi, L. Leindler, Sz. Révész, J. Szabados, V. Totik (Eds.)
6. Intuitive Geometry
I. Bárány, K. Böröczky (Eds.)
7. Graph Theory and Combinatorial Biology
L. Lovász, A. Gyárfás, G. Katona, A. Recski (Eds.)
8. Low Dimensional Topology
K. Böröczky, Jr., W. Neumann, A. Stipsicz (Eds.)
9. Random Walks
P. Révész, B. Tóth (Eds.)
10. Contemporary Combinatorics
B. Bollobás (Ed.)
11. Paul Erdős and His Mathematics I+II
G. Halász, L. Lovász, M. Simonovits, V. T. Sós (Eds.)
12. Higher Dimensional Varieties and Rational Points
K. Böröczky, Jr., J. Kollár, T. Szamuely (Eds.)
13. Surgery on Contact 3-Manifolds and Stein Surfaces
B. Ozbagci, A. I. Stipsicz
14. A Panorama of Hungarian Mathematics in the Twentieth Century, Vol. 1
J. Horváth (Ed.)
15. More Sets, Graphs and Numbers
E. Győri, G. Katona, L. Lovász (Eds.)
16. Entropy, Search, Complexity
I. Csiszár, G. Katona, G. Tardos (Eds.)
17. Horizons of Combinatorics
E. Győri, G. Katona, L. Lovász (Eds.)
Béla Bollobás
Robert Kozma
Dezső Miklós (Eds.)

Handbook of
Large-Scale
Random Networks

  JJÁNOS BOLY
YAI MATHEMATICA
C L SOCIETY
T
Béla Bollobás
Department of Pure Mathematics & Mathematical Statistics
University of Cambridge, Centre for Mathematical Sciences
Cambridge CB3 0WB, UK
and
University of Memphis, Dept. Mathematical Sciences
Memphis, TN 38152, USA
e-mail: [email protected]

Robert Kozma
University of Memphis, Dept. Mathematical Sciences
Memphis, TN 38152, USA
e-mail: [email protected]

Dezső Miklós
Alfréd Rényi Inst. of Mathematics
Reáltanoda u. 13–15
Budapest 1053
Hungary
e-mail: [email protected]

Mathematics Subject Classification (2000):

05A16,05C, 05C80, 60C05, 60J80, 68R10, 92B20, 92D20, 92D25, 92E10, 94C15

Library of Congress Control Number: 2008929577

ISSN 1217-4696
ISBN 978-3-540-69394-9 Springer Berlin Heidelberg New York
ISBN 978-963-9453-10-4 János Bolyai Mathematical Society, Budapest
This work is subject
b to copyright. All rights are reserved, whether the whole or part of the mate a rial
is concerned, specifif cally the rights of translation, reprinting, reuse of illustrations, recitation, broad-
casting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this
publication or parts thereof is permitted only under the provisions of the German Copyright Law of
September 9, 1965, in its current version, and permission for use must always be obtained from Springer.
Violations are liable for prosecution under the German Copyright Law.
Springer is a part of Springer Science+Business Media
springer.com
© 2008 János Bolyai Mathematical Society and Springer-Verlag
Printed in Hungary
The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific
f statement, that such names are exempt from the relevant protec-
tive laws and regulations and therefore free for general use.
Cover design: WMXDesign GmbH, Heidelberg; Gabriella Bollobás, Memphis, TN
Printed on acid-free paper 44/3142/db – 5 4 3 2 1 0
 Contents

Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

Chapter 1.
B. Bollobás and O. Riordan: Random Graphs and Branching
Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2. Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3. The Phase Transition in G(n, p) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4. The Phase Transition in Inhomogeneous Random Graphs . . . . . . . . . . 68
5. Branching Processes and Other Global Properties . . . . . . . . . . . . . . . . . 84
6. Appropriateness of Models: Metrics on Graphs . . . . . . . . . . . . . . . . . . . . 90

Chapter 2.
P. Balister, B. Bollobás and A. Sarkar: Percolation, Connec-
tivity, Coverage and Colouring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
2. The Gilbert Disc Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
3. The k-nearest Neighbour Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4. Random Tessellations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5. Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

Chapter 3.
R. Cohen and S. Havlin: Scaling Properties of Complex Networks
and Spanning Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
1. Random Graphs and Complex Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
2. The Bollobás Configuration Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6 Contents

3. Percolation on Random Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

4. Weighted Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5. Fractal Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6. Fractal Properties of Network Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . 163
7. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

Chapter 4.
A. Rudas and B. Tóth: Random Tree Growth with Branching
Processes – a Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
1. Random Tree Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
2. Local Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
3. Branching Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
4. Global Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

Chapter 5.
M. Catanzaro, M. Boguñá and R. Pastor-Satorras:
Reaction-diffusion Processes in Scale-free Networks . . . . . . . . . . . . . . . . 203
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
2. Analytical Description of Diffusion-annihilation Processes in
Complex Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
3. Numerical Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
4. Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233

Chapter 6.
J. Thakar, C. Christensen and R. Albert: Toward Under-
standing the Structure and Function of Cellular Interaction Net-
works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
1. Definition of Cellular Interaction Networks . . . . . . . . . . . . . . . . . . . . . . . . 239
2. Detecting or Inferring Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
3. Network Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
4. Properties of Cellular Interaction Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . 249
5. Graph Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
6. Modeling the Dynamics of Cellular Networks . . . . . . . . . . . . . . . . . . . . . . 258
7. Examples of Network-Based Dynamics Models . . . . . . . . . . . . . . . . . . . . . 259
8. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
Contents 7

Chapter 7.
W. J. Freeman and R. Kozma, with an appendix by B. Bol-
lobás and O. Riordan: Scale-Free Cortical Planar Networks . . . . . 277
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
2. Allocortex: The Simplest Model for the Phase Transition in
Criticality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
3. Application of RGT to Neocortex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
4. Structure and Dynamics in Cortex Models . . . . . . . . . . . . . . . . . . . . . . . . . 302
5. Requirements the Mathematical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
6. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
7. Appendix: Mathematical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307

Chapter 8.
T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó: Recon-
structing Cortical Networks: Case of Directed Graphs with High
Level of Reciprocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
2. Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
3. Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361

Chapter 9.
´
G. Palla, D. Abel, I. J. Farkas, P. Pollner, I. Derényi and
T. Vicsek: k-clique Percolation and Clustering . . . . . . . . . . . . . . . . . . . . 369
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
2. k-clique Percolation in the E-R-graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
3. Percolation of Weighted k-cliques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
4. Directed k-clique Percolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
5. Applications: Community Finding and Clustering . . . . . . . . . . . . . . . . . . 390

Chapter 10.
G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Erdi: ´
The Inverse Problem of Evolving Networks – with Application to
Social Nets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
2. The Model Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
8 Contents

3. Solving the Inverse Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415

4. Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426

Chapter 11.
A. Lőrincz: Learning and Representation: From Compressive Sam-
pling to the ‘Symbol Learning Problem’ . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
2. Sparse Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
3. Factored Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
4. Robust Control, Event Learning and Factored RL . . . . . . . . . . . . . . . . . 460
5. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479

Chapter 12.
M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csa-
logány and A. Lukács: Telephone Call Network Data Mining:
A Survey with Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
2. Link Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
3. Network Topology: Navigation in the Small World . . . . . . . . . . . . . . . . . 502
4. Clustering Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
5. Stacked Graphical Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
 Preface

With the advent of digital computers more than half a century ago, re-
searchers working in a wide range of scientific disciplines have obtained an
extremely powerful tool to pursue deep understanding of natural processes
in physical, chemical, and biological systems. Computers pose a great chal-
lenge to mathematical sciences, as the range of phenomena available for
rigorous mathematical analysis has been enormously expanded, demanding
the development of a new generation of mathematical tools. There is an
explosive growth of new mathematical disciplines to satisfy this demand, in
particular related to discrete mathematics. However, it can be argued that
at large mathematics is yet to provide the essential breakthrough to meet
the challenge. The required paradigm shift in our view should be compara-
ble to the shift in scientific thinking provided by the Newtonian revolution
over 300 years ago. Studies of large-scale random graphs and networks are
critical for the progress, using methods of discrete mathematics, probabilis-
tic combinatorics, graph theory, and statistical physics.
Recent advances in large scale random network studies are described
in this handbook, which provides a significant update and extension be-
yond the materials presented in the “Handbook of Graphs and Networks”
published in 2003 by Wiley. The present volume puts special emphasis on
large-scale networks and random processes, which deemed as crucial for fu-
ture progress in the field. The issues related to random graphs and networks
pose very difficult mathematical questions. Authors of this volume present
the state-of-art in this field, including scale-free random graph models and
large-scale real-world networks in various application areas. This handbook
covers all aspects of large-scale random networks, including mathemati-
cal foundations and rigorous results of random graph theory, modeling and
computational aspects of large-scale networks, as well as areas in physics, bi-
ology, neuroscience, sociology, and technical areas. Applications span from
microscopic, mesoscopic, and macroscopic models.
This handbook exposes the reader to the newest methods available for
the description and analysis of large-scale, complex structures. The book is
divided into 3 major parts. The first part provides a comprehensive overview
of the mathematical foundations of large-scale graphs and networks. The
10 Preface

second part introduces applications of the theory in biological sciences.

The third part describes further applications in a broad range of fields
including physics, computer networks and other technical systems, and
social networks.
Chapter 1 by Béla Bollobás and Oliver Riordan summarizes basic con-
cepts of random graphs in the context of the evolution of large-scale net-
works. The ‘classical’ random graph models, in particular G(n, p), are ‘ho-
mogeneous’, in the sense that the degrees tend to be concentrated around
a typical value. Many graphs arising in the real world do not have this
property, having, for example, power-law degree distributions. Thus there
has been a lot of recent interest in defining and studying ‘inhomogeneous’
random graph models. Svante Janson, Oliver Riordan and Béla Bollobás
introduced a very general model of an inhomogeneous random graph with
(conditional) independence between the edges, which scales so that the num-
ber of edges is linear in the number of vertices. This scaling corresponds to
the p = c/n scaling for G(n, p) used to study the phase transition; also, it
seems to be a property of many large real-world graphs. Our model includes
as special cases many models previously studied. Thsi chapter describes re-
sults concerning the general model.
Chapter 2 by Paul Balister, Béla Bollobás, and Amitesh Sarkar discusses
general concepts related to coverage problems with particular applications
in wireless networks.
Chapter 3 by Reuven Cohen and Shlomo Havlin provides a thorough
mathematical description of scaling properties of complex networks and
spanning trees. They present a relation between three properties of net-
works: the fractal properties of percolation clusters at criticality, the optimal
path between nodes in the network under strong disorder and the minimum
spanning tree. A mapping is established between percolation and minimum
spanning tree. They discuss fractal properties of models and of real world
scale free networks. These networks show self similarity and a finite fractal
dimension when measured using the box counting method. It is shown that
the degree distribution of the boxes is conserved under renormalization. We
also discuss the implications of fractalityon transport properties using the
renormalization approach.
Chapter 4 by Anna Rudas and Bálint Tóth provides a survey of random
tree growth with branching processes. One of the most important mech-
anisms underlying the inhomogeneity of many networks in the real world
seems to be the combination of growth with ‘preferential attachment’, i.e.,
the phenomenon that as new nodes are added to the network, they are more
Preface 11

likely to link to existing nodes that already have many links. The original
mathematical models for this phenomenon, including the Barabási–Albert
network model, assume exact proportionality. However, there is no reason
to expect this, and Krapivsky and Redner proposed and studied network
models based on preferential attachment using probabilities given by an ar-
bitrary ‘weight function’ of the degrees. Even in the special case of trees,
such models are very interesting: indeed, two much earlier models of random
trees fall in this general class, namely random recursive trees and random
plane-oriented recursive trees. Chapter 4 describes a beautiful mathemat-
ical theory of random trees grown in this way, using an arbitrary weight
function. Using a transformation to continuous time, they show that these
trees can be very well understood in terms of branching processes. They
also present results describing the limiting local behaviour of these trees,
and give simple direct proofs.
Part II starts with Chapter 5 by Michele Catanzaro, Marian Boguna, and
Romualdo Pastor-Satorras, in which they review reaction-diffusion equa-
tions in scale-free networks. They review the main results recently obtained
in modeling of binary fermionic reaction-diffusion processes on scale-free
networks. We show how to derive rate equations within the heterogeneous
mean-field formalism, and how information can be obtained from them both
for finite networks in the diffusion-limited regime and in the infinite network
size lime. By means of extensive numerical simulations, we check the mean
field predictions and explore other aspects of the reaction-diffusion dynam-
ics, such as density correlations and the effects of the minimum degree or a
tree-like topology.
Chapter 6 by Claire Christensen, Juilee Thakar and Réka Albert de-
scribes models of cellular regulatory networks. Genetic, biochemical and
molecular biology techniques have been used for decades to identify bio-
logical interactions; newly developed high-throughput methods now allow
for the construction of genome-level interaction maps. In parallel, high-
throughput expression data paired with computational algorithms can be
used to infer networks of interactions and causal relationships capable of
producing the observed experimental data. Graph-theoretical measures and
network models are more and more frequently used to discern functional and
evolutionary constraints in the organization of biological networks. Perhaps
most importantly, the combination of interaction and expression informa-
tion allows the formulation of quantitative and predictive dynamic models.
Some of the dominant experimental and computational methods used for
the reconstruction or inference of cellular networks are reviewed, also the
12 Preface

biological insights that have been obtained from graph-theoretical analysis

of these networks, and the extension of static networks into various dynamic
models capable of providing a new layer of insight into the functioning of
cellular systems is discussed.
Chapter 7 by Walter J. Freeman, Robert Kozma, and Béla Bollobás
describes large-scale brain networks and analyzes scale-free properties in
planar geometry characteristic of the cortical tissue. The cortical tissue
has a complex anatomical structure of interconnected neural population
producing well-studied behaviors, namely oscillations in space and time.
Two major aspects of the structure-function relationship are described.
First, for a given structure, they study the dynamics generated by the given
network. Next, they introduce structural changes produced by learning and
adaptation in the cortex and they describe how the behavior of the network
varies under such conditions. The concept is called pioneer neurons evolving
along cortical subplates. The mathematical formulation of this process is
given as well.
Chapter 8 by Tamás Nepusz, Fülöp Bazsó, Peter Hussami, László
Négyessy and Gábor Tusnády deals with the reconstruction of cortical net-
works using the intuition provided by Szemerédi regularity lemma. Areas
of the cerebral cortex, which are specialized for different functions, form a
highly clustered, densely connected small world network via neuronal con-
nections. Mapping the connections between the areas is the primary way to
understand their functional interactions. However, the cortical network is
not fully charted because of methodological and practical difficulties. Sze-
merédi’s Regularity Lemma states that every graph can be approximated
by the union of random-like bipartite graphs. By comparing the approx-
imation with the original connectivity graph, one can to point out which
groups of areas tend to connect with each other densely and which groups
are not likely to connect at all. Based on this approach, the connectivity
of brain networks is modeled in cases when experimental investigations are
not feasible at present.
Chapter 9 is the first contribution in Part III written by Gergely Palla
and Tamás Vicsek on k-clique percolation and clustering. One of the most
interesting features of real-world networks, especially social networks, is the
presence of communities, i.e., small subsets of the nodes within which the
density of connections is much higher than in the network as a whole. In
studying the communities in a network, a key question is whether they tend
to be isolated, with rather few communities overlapping, or whether they
percolate, i.e., whether there are large chains of communities one overlapping
Preface 13

the next. There are many possible mathematical abstractions of these rather
vague notions; one of the simplest is to choose a parameter k, define a
community to be a k-clique, i.e., a set of nodes all of which are connected
to each other, and to consider two k-cliques as connected if they overlap in
the strongest possible sense, i.e., share k − 1 nodes. In Chapter 9, Palla,
Ábel, Farkas, Pollner, Derényi and Vicsek present a survey of work on this
notion of clique percolation. It turns out that this notion leads to new
and very interesting mathematical questions about classical random graphs;
they start by surveying results on these questions. In the second part of the
chapter they turn to applications of clique percolation to practical problems
of finding communities in real-world networks.
Chapter 10 by Gábor Csárdi, Katherine J. Strandburg, Jan Tobochnik,
and Péter Érdi describes the inverse problem of evolving social networks.
The evolution of a network involves the addition and/or deletion of edges
and vertices. These processes are modeled using a discrete time, stochastic
framework. In the case of the inverse problem one asks what is the best
description of the evolution of a given network, or a set of networks, in
a well-defined mathematical framework? The input of this approach is
the data collected about the evolution of a network, the output is a set
of parameters for a stochastic model framework. The present model uses
parameters of some kernel-functions and the inverse problem is formalized
in this framework. various approaches are introduced for optimal estimation
of the parameters, and demonstrate the utility of the methodology in several
applications.
Chapter 11 by András Lőrincz addresses general issues of learning and
representation and overviews a wide range of approaches relevant to large-
scale networks. Individual learning and communication are amongst enig-
matic problems of computer science, artificial intelligence as well as neuro-
science. Intuitively, learning is worthless without verifying what has been
learned. For example, in evolutionary learning, the learned information is
encoded in genes and it is replicated during reproduction. The replicated
information is subject to verification during the life of the newborn individ-
ual. This chapter reviews reconstruction networks and the potential role of
Szemerédi’s Regularity Lemma in learning in such networks. Szemerédi Reg-
ularity Lemma deals with graph partitioning and recent findings indicate
that partitions can be found constructively and finding a partition takes
polynomial time. The chapter describes the use of Szemerédi Lemma by
revisiting the Lőrincz–Buzsáki model of the hippocampal-entorhinal loop,
which known to be responsible for the formation of declarative memory.
14 Preface

It is shown that learning in this structure can happen efficiently at non-

combinatorial cost.
Chapter 12 is written by András A. Benczúr, Károly Csalogány, Miklós
Kurucz, András Lukács, and László Lukács on the complex structure of so-
cial communities as evidenced by telecommunication data resurces. Their
results provide an intuition on how contacts within real social networks arise
and suggest properties unexplained by current network evolution models.
They build the corresponding social interaction graph and extract informa-
tion on the strength of social contacts. They investigate several clustering of
users. Communities algorithmically extracted from the networks are com-
pared with basic clustering by external socio-demographic parameters such
as geographic location. Alternative clustering is provided using a different
telephone service provider over the social network.
This Handbook is the result of the “U.S. – Hungarian Workshop on
Large-Scale Random Graph Methods for Modeling Mesoscopic Behavior
in Biological and Physical Systems,” held at the Alfred Rényi Institute of
Mathematics of the Hungarian Academy of Sciences, August 28 – Septem-
ber 2, 2006, Budapest, Hungary, co-sponsored by the National Science
Foundation, USA and the Hungarian Academy of Sciences. Their support
is appreciated. In particular thanks are due to Ms. Bonnie Thompson,
NSF for her help thorough this project and to Dr. Mark Suskin, NSF
European Office for his active participation at the Workshop. The aim of the
workshop was to expose the participants to the newest methods available
for the description and analysis of large-scale, complex structures using
methods of combinatorics and graph theory, in combination with nonlinear,
adaptive systems theory, and statistical physics. This volume contains
selected topics from the workshop and additional contributions from leading
authors working in the subject area, in order to provide a comprehensive
view of the field. We hope the readers will find the materials useful in their
research and also as part of advanced academic studies. Our intention with
publishing this volume is to stimulate further intensive research in this field.

July 9, 2008 Béla Bollobás

Memphis – Cambridge – Budapest Robert Kozma
Dezső Miklós (Eds.)
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 15–115.

Chapter 1
Random Graphs and Branching Processes

BÉLA BOLLOBÁS∗ and OLIVER RIORDAN

During the past decade or so, there has been much interest in generating and
analyzing graphs resembling large-scale real-world networks such as the world
wide web, neural networks, and social networks. As these large-scale networks
seem to be ‘random’, in the sense that they do not have a transparent, well-defined
structure, it does not seem too unreasonable to hope to find classical models of
random graphs that share their basic properties. Such hopes are quickly dashed,
however, since the classical random graphs are all homogeneous, in the sense that
all vertices (or indeed all k-sets of vertices) are a priori equivalent in the model.
Most real-world networks are not at all like this, as seen most easily from their
often unbalanced (power-law) degree sequences. Thus, in order to model such
graphs, a host of inhomogeneous random graph models have been constructed
and studied.
In this paper we shall survey a number of these models and the basic results
proved about the inhomogeneous sparse (bounded average degree) random graphs
they give rise to. We shall focus on mathematically tractable models, which often
means models with independence between edges, and in particular on the very
general sparse inhomogeneous models of Bollobás, Janson and Riordan. The first
of these encompasses a great range of earlier models of this type; the second, the
inhomogeneous clustering model, goes much further, allowing for the presence of
clustering while retaining tractability.
We are not only interested in our inhomogeneous random graphs themselves,
but also in the random subgraphs obtained by keeping their edges with a certain
probability p. Our main interest is in the phase transition that takes place around
a certain critical value p0 of p, when the component structure of the random
subgraph undergoes a sudden change. The quintessential phase transition occurs
in the classical binomial random graph G(n, c/n) as c grows from less than 1 to
greater than 1 and, as shown by Erdős and Rényi, a unique largest component,
the giant component, is born.
A ubiquitous theme of our paper is the use of branching processes in the study
of random graphs. This ‘modern’ approach to random graphs is crucial in the
∗
Research supported in part by NSF grants DMS-0906634, CNS-0721983 and CCF-
0728928, and ARO grant W911NF-06-1-0076.
16 B. Bollobás and O. Riordan

study of the very general models of inhomogeneous random graphs mentioned

above. To illustrate the power of branching processes, we show how they can
be used to reprove sharp results about the classical random graph G(n, c/n),
first proved by Bollobás and L
 uczak over twenty years ago. When it comes
to inhomogeneous models, we shall have time only to sketch the connection to
branching processes.
Finally, we close by discussing the question of how to tell whether a given
model is appropriate in a given situation. This leads to many fascinating ques-
tions about metrics for sparse graphs, and their relationship to existing models
and potential new models.

1. Introduction

In the last 10 years or so, the graph-theoretic study of real-world networks

of all kinds has become a major field. A very broad range of networks
have been studied, from abstract ones, such as social networks, to physical
networks such as power grids. These all have in common the property
that one cannot hope to explain their detailed structure exactly from a
mathematical viewpoint. For this reason, the appropriate comparison is
with random graphs of some kind. Since, by the 1990s, there was a rich
theory of random graphs, it was not unreasonable to see whether the large-
scale real-world networks appearing in various areas resemble any of the
random graphs that had been studied by mathematicians. In fact, it would
have been quite astonishing if there had been much resemblance: after all,
the real-world networks are far from homogeneous (there is no reason why
‘vertices’ widely separated geographically and in time should have very
similar characteristics), while, as we shall describe in the next section,
almost all random graphs studied up to the early 1990s were homogeneous
in the sense that, a priori, all vertices are equivalent. (We shall expand on
this in a moment.)
Watts and Strogatz [195] were perhaps the first to note a different kind
of discrepancy: in the standard random graphs of mathematicians, the con-
nection between two basic graph parameters (clustering and diameter) was
very different from that observed in various real-world networks. Return-
ing to inhomogeneity, a little later, Faloutsos, Faloutsos and Faloutsos [101],
among others, drew attention to power-law type degree distributions in net-
works such as the internet, and Barabási and Albert [13] proposed a random
Random Graphs and Branching Processes 17

construction for graphs with such ‘scale-free’ degree distributions, and jus-
tified their construction by heuristic analysis and computer experiments.
The initial surprise that the networks seem to exhibit the ‘small-world phe-
nomenon,’ which is mathematically anything but surprising, was rapidly
superseded by attempts to give random constructions of ‘scale-free’ graphs
expected to be closer to real networks than any of the standard random
graphs.
As a starting point in the modelling process, one wants a simple mathe-
matical model. The best known and most studied is G (n, p); for a discussion
of the history of this model, see the next section. One of the features of
a random G(n, p) ∈ G (n, p) is its homogeneity: for example, the degree
sequence is in a certain sense fairly flat (it is well approximated by a Pois-
son distribution). This contrasts with many real-world networks, which, as
noted above, often have power-law degree distributions. Indeed, it was pri-
marily this observation, and the oft-cited papers of Watts and Strogatz [195]
and Barabási and Albert [13], that formed the starting point for much of the
recent work on new models. (Note, however, that observations of power laws
in the real world go back much further: in 1926, Lotka [140] claimed that
citations in academic literature follow a power law; in 1997, Gilbert [116]
suggested a probabilistic model supporting ‘Lotka’s law.’ Power-law dis-
tributions were also investigated by Zipf [201] in 1949, and Simon [177] in
1955.)
Although the term is not used consistently, networks with power-law
distributions of degrees and other parameters are now often known as ‘scale-
free’. Since many real-world networks are scale-free, one of course wants
as starting points mathematical models with this property, in particular,
inhomogeneous models. Over the last 10 years, very many such models
have been introduced; we shall describe some of these in the next section.
Throughout we focus on sparse graphs, where the number of edge grows
(roughly) linearly with the number of vertices. These are the natural models
of large-scale real-world networks such as social networks, where the vertex
set may number in the billions, but typical degrees are not too large.
While much of the work on scale-free random graphs is concerned with
detailed modelling of real networks, which almost inevitably leads to models
that are too complicated for rigorous analysis and must be studied heuris-
tically or by simulation, quite a few simple and appealing mathematical
models have also been (in some cases re-) introduced. In addition to the
obvious contributions of the theory of random graphs to the understand-
ing of real-world networks, these models motivated by real-world networks
18 B. Bollobás and O. Riordan

have greatly enriched the theory of random graphs. It is this latter phenom-
enon that we shall focus our attention on. Even this is too broad a topic to
cover in a short survey, so we shall focus on one particular (very important)
aspect, namely the role played by branching processes in the analysis of
sparse random graphs, both homogeneous and inhomogeneous. The selec-
tion of models we consider will necessarily be far from comprehensive, and
will naturally tend to favour those we find most interesting, or have worked
on ourselves.
The rest of the paper is organized as follows. In the next section we
describe the various models we shall consider, starting with the ‘classical’
models, and moving on to a selection of mathematically tractable inhomo-
geneous models. In Section 3 we show that branching processes may be
used to give a new, simple proof of a sharp result about the phase transi-
tion in G(n, p). In Section 4 we survey corresponding results in a variety of
inhomogeneous models; since the models are more complicated, the results
are (unsurprisingly) weaker. In Section 5 we briefly mention applications
of branching processes to the study of other aspects of random graphs; this
section is included to dispel the a priori reasonable impression that branch-
ing processes are suitable only for studying the phase transition. Finally,
in Section 6 we turn to a rather different but very important topic, namely,
how to tell whether a given model is appropriate in a given situation. This
turns out to be a question about metrics on (here sparse) graphs.
In this article we use standard graph theoretic notation as in [33], writ-
ing, for example, V (G) and E(G) for the vertex and edge sets of a graph
G, |G| for the number of vertices, and e(G) for the number of edges. For
asymptotic notation,
 we write f (n) ∼ g(n) if f (n)/g(n) → 1 as n → ∞, we
write f (n) = O g(n) if there is a constant C such that f (n) ≤ Cg(n), we 
write f (n) = o g(n) if f (n)/g(n) → 0 as n → ∞, and f (n) = Θ g(n)
if f = O(g) and g = O(f ). For sequences of random variables, we use
standard asymptotic notation as in Janson, L  uczak and Ruciński [127], for
p  
example: Xn → c if Xn converges in probability to c, Xn = Op f (n) if
  p
Xn /f (n) is bounded in probability, and Xn = op f (n) if Xn /f (n) → 0.
As usual, we say that an event A holds with high probability, or whp, if
its probability tends to 1 as n → ∞. Formally, one of course considers a
sequence of events An , with An depending on the n-vertex random graph
under consideration. Although we often spell this out, whenever we say that
a random graph has a certain property, we mean that it has this property
whp.
Random Graphs and Branching Processes 19

2. Models

In this section we describe the various random graph models we shall con-
sider in the rest of the paper, starting with some comments on the history
of the field.

2.1. Classical models

There are two truly ‘classical’ models of random graphs: G (n, p) and
G (n, m); it is extraordinary that they were introduced at about the same
time, but independently of each other. In 1959, Gilbert [115] introduced
the ‘binomial’ model G (n, p) with n labelled vertices (say, with vertex
set [n] = {1, . . . , n}), and edges present independently with probability
p = p(n). Just about simultaneously, Erdős and Rényi [93] started the sys-
tematic study of the space G (n, m) of all graphs with n labelled vertices
and m = m(n) edges, with any two such graphs G(n, m) equiprobable.
Often, one writes Gn,p for a random element of G (n, p), and Gn,m for
a random element of G(n, m). Here we shall use the alternative notation
G(n, p) or G(n, m); this is less compact, but easier to read when the para-
meter p, for example, is replaced by a more complicated expression.
Although G (n, p) is a quintessential probability space, Gilbert studied
the random graph G(n, p) ∈ G (n, p) with the aid of generating functions,
giving exact but unwieldy expressions for the probability of events such as
the event that G(n, p) is connected. On the other hand, although Erdős
and Rényi asked questions about enumeration, such as how many of the
n
 
 G (n, m) = 2
m

graphs in G (n, m) are connected, they strove for asymptotic answers by

using probabilistic techniques. Thus Erdős and Rényi treated graph pa-
rameters such as the connectivity of a graph, its chromatic number and
the number of triangles it contains as random variables on G (n, m), and
adapted the methods of classical probability theory to the study of these
random variables. As a matter of fact, Erdős and Rényi did not use much
high-powered probability theory in their work on random graphs: their main
probabilistic tools were the first and second moment methods, i.e., the use of
the expectation and variance of a random variable, which they enhanced by
20 B. Bollobás and O. Riordan

clever combinatorial arguments. The combined efforts of Erdős, the combi-

natorialist par excellence, and Rényi, the probabilist, were ideally suited to
such an attack on G (n, m). (Needless to say, Erdős was also an outstanding
probabilist, and Rényi was an excellent combinatorialist.)
In fact, random graphs had made their first appearance a few years ear-
lier. In 1950, Uhlenbeck pointed out in his Gibbs Lecture that a number
of problems in statistical mechanics lead in a natural way to various prob-
lems concerning graphs, and some years later he wrote several papers on
the combinatorial problems that arose in this way. Thus, in studying the
condensation of gasses, Riddell and Uhlenbeck [165] were led to problems
concerning the number of certain connected graphs, and they attacked these
problems with the aid of generating functions. This prompted Ford and Uh-
lenbeck, in a series of papers [106, 105, 107, 108] (the second with Norman),
and Austin, Fagen, Penney and Riordan [10] to prove further results about
various families of connected graphs, including the family of two-connected
graphs with n vertices and m edges. The random graph model introduced
and studied by Austin, Fagen, Penney and Riordan was very similar to
G (n, m): they made
n m selections of edges independently
n and at random,
with each of the 2 edges having the same chance 1/ 2 of selection at each
trial. Thus, after m selections, a random graph is obtained with m labelled
edges (signifying the ‘time’ when each edge was ‘born’), with multiple edges
allowed. When m = Θ(n), as in the cases we shall be most interested in
here, this model is very close indeed to the G (n, m) model studied by Erdős
and Rényi.
Although Gilbert [115] did prove that for a fixed probability p, 0 < p < 1,
the probability that G(n, p) is disconnected is asymptotic to nqn−1 , where
q = 1 − p is the probability that two vertices are not joined by an edge, the
formulae Uhlenbeck, Austin, Gilbert and others obtained with the aid of
generating functions for the exact number of graphs in certain classes tended
to be too complicated to yield asymptotic expressions. The revolutionary
idea of Erdős and Rényi was that there is no need to obtain exact formulae in
order to deduce asymptotic results: rather, such results are best obtained
by probabilistic means. The introduction of such a probabilistic point of
view is perhaps their greatest and most lasting achievement. In addition
to introducing a probabilistic point of view into combinatorics, Erdős and
Rényi wrote a series of papers [93, 94, 95, 96, 97, 98] in which they proved
a host of important results and so founded a rich theory of random graphs.
They realized, years before it became common knowledge, that for many
a function m = m(n) there is a ‘typical’ random graph G(n, m), in the
Random Graphs and Branching Processes 21

sense that almost all random graphs G(n, m) have certain properties (their
diameter is 2, the largest clique they contain has 6 vertices, etc.), so when
we talk about a random G(n, m) or G(n, p) we may as well have a certain
concrete graph in mind.
Perhaps the greatest single result of Erdős and Rényi is aboutthe phase

transition in the component structure of a random graph G(n, (c/2)n )
(or, equivalently, of a random graph G(n, c/n) from the binomial model)
as the constant c ‘increases’ from c < 1 to c > 1: if c < 1 then whp (with
high probability,
  i.e. with probability tending to 1) the largest component of
G(n, (c/2)n ) has only O(log n) vertices, but if c > 1 then whp our random
graph has ρ(c)n + o(n) vertices, where ρ(c) > 0. Although it took close to
25 years for the mathematical community to realize that this theorem was
not an isolated curiosity but a fundamental result, for the past 20 years or
so it has been the starting point for much important research. This Erdős–
Rényi result will be our starting point as well in this paper: when we return
to it in Section 3, we shall say more about the result itself and subsequent
developments.
Not surprisingly, in the past fifty years or so a good many probabilistic
tools have been found that can be brought to bear on problems concern-
ing random graphs: while Erdős and Rényi made do with expectation and
variance (and, occasionally, higher moments), today’s probabilistic combi-
natorialist has a rich variety of tools at his disposal, including martingales,
branching processes, correlation inequalities, limit theorems and Fourier
methods. The theme running through our article is exactly the use of one of
these methods: we shall emphasize how branching processes can be applied
to a variety of problems. In particular, we shall use branching processes to
give a new and simple proof of a sharp form of the phase transition above.
Note that G (n, m) is a ‘finer’ model than G (n, p): the latter is just the
‘weighted union’ of the spaces G (n, m), with the ‘weights’ or probabilities
 
coming from the binomial distribution Bi(N, p) with N = n2 , so the
  m N −m
weight of G (n, m) is N m p (1 − p) . In fact, we can do even better
than G (n, m): we can couple all these spaces into a single random graph
process. Informally, this coupling was already used by Erdős and Rényi; it
was properly introduced and studied formally by Bollobás only many years
later. A random graph process on V = [n] = {1, . . . , n} is a Markov chain
G = (Gt )N whose state space is the set of all 2N graphs on V . Each graph
0
Gt has t edges, so G0 is the empty graph on V , the graph without any
edges. Given Gt , the next graph Gt+1 is obtained by adding an edge, with
22 B. Bollobás and O. Riordan

all additions equiprobable. Clearly, the mth term of a random graph process
has exactly the distribution of G(n, m) ∈ G (n, m).
Random graph processes can be used to show that, as far as the random
graphs G(n, m) are concerned, certain properties are very close to each
other. For example, the local property of having no isolated vertices is
closely tied to the global property of being connected, in the following
sense. Given G = (Gt )N , write τ1 (G)
for the minimal t such that Gt has
0
for the minimal t such that Gt is connected.
no isolated vertex, and τ2 (G)

Clearly, τ1 (G) ≤ τ2 (G) for every G, and a slight extension of a classical
result of Erdős and Rényi says that we have equality whp: the probability
= τ2 (G)
that τ1 (G) tends to 0 as n → ∞. Thus the main obstruction to
connectedness is the existence of an isolated vertex.
At the 1990 Quo Vadis, Graph Theory conference in Fairbanks, Alaska,
Bollobás and Erdős proposed a class of variations of random graph processes
in the faint hope of improving bounds on Ramsey numbers. For example, let
G0 ⊂ G1 ⊂ · · · ⊂ G be a random sequence of graphs such that V (Gt ) = [n]
and e(Gt ) = t for every t, where Gt+1 is obtained from Gt by adding to it
a randomly chosen edge that does not create a triangle in the new graph
Gt+1 . The process ends in a graph G to which it is impossible to add
an edge without creating a triangle. What can one say about the maximal
triangle-free graph G ? For example, what is the distribution of the random
variable ? And of the independence number of G ?
Similarly, let HN ⊃ HN −1 ⊃ · · · ⊃ H be such that V (Ht ) = [n] and  
e(Ht ) = t for every t, so that HN is the complete graph on [n] with N = n2
edges, and Ht−1 is obtained from Ht by deleting at random one of the edges
in the triangles of Ht . Here the process stops with a graph H containing no
triangle. What can we say about H ? What is its independence number?
Needless to say, a ‘random’ choice may mean a uniform choice or a
choice according to some probability distribution. For example, in the
second process the probability of deleting a particular edge from Ht may be
proportional to the number of triangles containing it. Clearly, instead of a
triangle, we may take any fixed graph or any family of graphs. There are
numerous papers about processes such as these, including Erdős, Suen and
Winkler [99], Bollobás and Riordan [46], Osthus and Taraz [156], Ruciński
and Wormald [171, 172, 173] and Greenhill, Ruciński and Wormald [118].
Random Graphs and Branching Processes 23

2.2. Random graphs with a fixed degree sequence

After the graphs G(n, p) and G(n, m), the best known random graph is
perhaps the random r-regular graph G(n, r-reg). The definition of the
appropriate space G (n, r-reg) is just as that of G (n, m): it is the set of
all r-regular graphs
 on [n], with all graphs equiprobable, so every graph has
probability 1/ G (n, r-reg) . (Needless to say, when we consider r-regular

graphs, we always assume that rn is even and that n > r; otherwise, there
is no r-regular graph on [n].)

As this definition is so natural, and regular graphs have been studied

a great deal, one might wonder why this space was not studied by Erdős
and Rényi, or, for that matter, by anybody else for almost twenty years
after their pioneering work. The answer is very simple. First, except in
some simple cases, it was not even known roughly how many r-regular
graphs there are on n vertices, for a fixed value of r. The second reason
is much more important: in order to study the probability that G(n, r-reg)
has chromatic number k, say, we have to have some means of counting
subfamilies of regular graphs. For example, we should have a chance of
counting the number of r-regular graphs that are k-colourable and have
girth at least .

The first objection was overcome by Bender andCanfield [17]: in 1978

they proved an asymptotic formula for  G (n, r-reg) . The random graph
G(n, r-reg) was born a little later, when Bollobás [25] re-proved the Bender–
Canfield asymptotic formula by probabilistic means, and defined a very
simple space, the so-called configuration space, that could be used to study
even complicated properties of random regular graphs.

In fact, there is no difference between studying random r-regular graphs

and random graphs with a given degree sequence d = (di )n1 , provided r
and the members of the degree sequence are fairly slowly growing functions
of n, so let us describe the configuration model for the space G (n, d) of
random graphs with degree sequence d = (di )n1 , with max di bounded, say.
Naturally, the space G (n, d) is just the set of all graphs on [n] in which the
degree of vertex i is di for every i, with the graphs taken equiprobable.
We
shall assume that the following trivial conditions are satisfied: 2m = i di
is even (as we do not want G (n, d) to be empty), di ≥ 1 for every i (why
carry isolated vertices?) and 2m − n → ∞ (so that there are a fair number
of ‘excess’ edges to make the space ‘interesting’).
24 B. Bollobás and O. Riordan

Let W1 , W2 , . . . , Wn  |Wi | = di . A configuration F

be disjoint sets with

is a partition of W = ni=1 Wi into m = 12 di pairs; the configuration

space is the set Φ of all N (m) = (2m)!/(m!2m ) = (2m − 1)!! configurations,
each with probability 1/N (m). Given F ∈ Φ, let ϕ(F ) be the graph with
vertex set [n] in which i is joined to j whenever there is a pair {x, y} in
F with x ∈ Wi and y ∈ Wj ; sometimes, one considers the corresponding
multi-graph (with loops allowed), defined in the natural way, with one edge
for each pair in F . To generate a random element G(n, d) of G (n, d), pick
a random element F of Φ and take ϕ(F ): if this graph has degree sequence
d, i.e., if ϕ(F ) is a simple graph, then all is well: we keep ϕ(F ) as a proper
choice of a random G(n, d), otherwise we discard it and repeat the process.
It is crucial that the map ϕ does not ‘distort’ the probability:
 every element
G of G (n, d) occurs with the same multiplicity  ϕ−1 (G) = i di ! .
Furthermore, if d is a ‘moderate’ degree sequence (for example, if max di
is bounded by a constant as n → ∞), then ϕ(F ) has a fair chance of being
simple, i.e., of belonging to G (n, d). It is even more important that it is
very easy to tell when this happens. Indeed, let X = X(F ) be the number
of pairs in F chosen from the same Wi , and Y the number of pairs of
pairs, (i1 , i2 ) and (j1 , j2 ), say, with i1 , i2 ∈ Wi and j1 , j2 ∈ Wj for some
i = j. Then the probability that a configuration F gives us a graph ϕ(F ) in
G (n, d) is precisely the probability of the event E = {X = 0 and Y = 0} in
Φ. As Z = X + Y is easily seen to have asymptotically Poisson distribution,
conditioning on E rarely increases the complications we have to overcome
when studying the probability of a certain property of G(n, d).
1
di
 
i 2 we have E(Z) ∼ λ + λ , and
More precisely, with λ = 2m 2

 
 G (n, d) ∼ e−λ−λ2 (2m − 1)!! di !−1 .
i

When studying G(n, d), we do not even need the formula above. In order
to prove that a random graph G(n, d) has a certain property Q ⊂ G (n, d)
whp, it suffices to show that a configuration F ∈ Φ has the property
Q∗ = ϕ−1 (Q) ∪ {Z ≥ 1} ⊂ Φ whp.
The configuration space above has a perhaps even catchier description.
From each vertex i ∈ [n] grow di ‘half-edges’ so that altogether we have 2m
half-edges. Pair off the half-edges randomly, so that m edges are produced.
The graph obtained need not be simple since it may have loops (counted
by X), and multiple edges (counted by Y in a strange way), but when it is
Random Graphs and Branching Processes 25

simple, all graphs in G (d) arise with the same probability, so we generate
exactly G(n, d).
With the appearance of the configuration space the floodgates opened:
a host of results were obtained about random regular graphs and graphs
with given moderate degree sequences (see, e.g., [26, 27, 39, 28, 54, 45, 32,
102, 113]). One should emphasize that the use of the configuration model of
Bollobás is not the only way to attack problems about random graphs with
given degree sequences: far from it; in fact, in deeper questions it is only
the very simple and natural starting point. In particular, for faster growing
degree sequences the configuration model is less effective, but combined with
other tools such as the switching technique of McKay [147], deep results can
be proved in this case as well. In particular, McKay and Wormald [149, 148]
showed how to generate uniformly graphs with given fast-growing degree
sequences. One of the most beautiful results about random regular graphs
was proved by Robinson and Wormald [170]: for r ≥ 3 fixed, a random
r-regular graph is Hamiltonian whp; earlier, Bollobás [28] and Fenner and
Frieze [102] had proved this for r at least some constant r0 . Recently, Frieze,
Krivelevich and Smyth [112] gave a condition on the degree sequence d which
implies that whp the chromatic number of G(n, d) is Θ(d/ log d), where
d is the average degree. Results about G(n, r-reg) for r = r(n) growing
rapidly with n were proved by Frieze [114], Krivelevich, Sudakov, Vu and
Wormald [138], Cooper, Frieze and Reed [79], and Cooper, Frieze, Reed and
Riordan [80], for example.
To close this section, let us introduce a model, G (n, k-out), which does
not quite belong in this section, but is fairly close to G (n, r-reg), only
considerably simpler. Let 1 ≤ k ≤ n − 1; we tend to think of k as
constant and n large, n → ∞. To define a random element G(n, k-out) of
→
−
G (n, k-out), we first define a random directed graph G (n, k-out) on [n] as
follows: each vertex sends
n−1edges to a set of k other vertices chosen uniformly
at random from all k possibilities; the choices associated to different
starting vertices are made independently. Note that this random directed
graph has precisely kn directed edges, with precisely k edges leaving every
vertex. However, the in-degrees can be very different, and for some pairs of
vertices {x, y} we may have an edge from x to y and one from y to x. Now,
to get G(n, k-out), simply forget the orientations of the edges, replacing any
multiple (back and forth) edges by single edges.
This random graph is not far from regular in the sense that although
it has at most kn edges (so, for a fixed k, only linearly many edges), every
vertex has degree at least k. (On the other hand, a random graph G(n, kn)
26 B. Bollobás and O. Riordan

has many isolated vertices whp.) This model G (n, k-out) is much simpler
than G (n, r-reg) since it has built-in independence: writing Di for the set
of k vertices dominated by i ∈ [n] (so that i sends directed edges to the
vertices in Di ), the n random variables D1 , . . . , Dn are independent.

2.3. Inhomogeneous models

Turning to the new inhomogeneous models, one of the first and most stud-
ied is the ‘growth with preferential attachment model’ of Barabási and Al-
bert [13], often known as the BA model. In this model, the graph is ‘grown’
by adding one vertex at a time, and each new vertex sends m edges to
existing vertices, chosen with probabilities proportional to their degrees at
the time. This description is not quite complete; the description in [13] is
also incomplete, and in addition does not make sense; for details see [47],
for example. In order to study the Barabási–Albert model mathematically
one thus first needs to make it precise. This was done in [49], where the
LCD model was introduced, based on linearized chord diagrams. This is an
exactly defined model fitting the vague description above, which has the ad-
vantage of having a static description as well as an evolving one. The degree
sequence of the LCD model was analyzed by Bollobás, Riordan, Spencer and
Tusnády [53], who showed that it does indeed follow a power law.

The Barabási–Albert model was introduced as a (potential) explanation

of the occurrence of power laws. Other such models soon followed, including
the copying model of Kumar, Raghavan, Rajagopalan, Sivakumar, Tomkins
and Upfal [139], the Buckley–Osthus model [61] (a generalization of the
LCD model), and a very general model combining uniform and preferential
attachment introduced by Cooper and Frieze [78]. Although all these mod-
els were motivated in particular by a graph that is naturally seen as directed
(the web graph), they are themselves undirected. A directed version was
introduced by Bollobás, Borgs, Chayes and Riordan [35], with preferential
attachment based on in- and out- degrees. Since 1999, many (perhaps hun-
dreds of) such ‘growing network’ models have been introduced, incorporat-
ing many different features motivated by particular examples of real-world
networks; we have given only a very partial list. In the initial study of these
models the main focus was on the degree distribution, showing that it does
indeed follow a power law in each case. For surveys of some of the earlier
work in this field see, for example, Dorogovtsev and Mendes [85, 87], Albert
Random Graphs and Branching Processes 27

and Barabási [3], or [47]. These are somewhat out of date: there is now so
much material that it would be difficult to survey.
In a slightly different direction, instead of trying to understand the
origin of power-law degree sequences, one can attempt to understand their
consequences. For this reason Aiello, Chung and Lu [2, 1] proposed a
model for ‘massive graphs’ obtained simply by fixing a power-law degree
distribution in advance, and then picking a random graph with this degree
sequence. This is a particular case of the configuration model G(n, d)
discussed in Subsection 2.2.
Returning to explaining power laws, in their original paper [13], Barabási
and Albert asked whether it is the growth or the preferential attachment
that gives rise to the power law, and also considered the growing m-out
graph, in which vertices are added one at a time, and each sends m edges to a
set of m existing vertices chosen uniformly at random. One does not obtain
a power law in this case; the model is nonetheless very natural, and has
received considerable attention. It may be thought of as a growing version
of the usual m-out model defined at the end of Section 2.2. It is natural
to consider variants of the growing m-out graph where the number of edges
added at each stage is random, rather than constant. Such graphs have
been considered by Devroye, McDiarmid and Reed [84] and Riordan [166],
for example.

2.4. Models with independence between edges

Many of the models described above turn out to be difficult to analyze

precisely, especially when it comes to studying more complicated properties
of these graphs. The basic problem is that the events that different possible
edges are present are not independent. For this reason, one often studies
‘average’ or ‘mean-field’ versions of the models, with roughly the same
individual edge probabilities, but independence between edges. The term
mean-field is rather unfortunate, as the normal use in physics is to denote a
model where all particles interact equally, which in the graph context would
correspond simply to G(n, p).
It turns out that the probability pij that an edge ij is present in the
√
Barabási–Albert model is roughly m/ 2 ij ; in the LCD model one can
in principle give an exact formula for pij , although this is rather complicated
for m ≥ 2; see [168], for example. In any case, it turns out that for m fixed
28 B. Bollobás and O. Riordan

and 1 ≤ i < j ≤ n one has

m  
pij = √ 1 + O(1/i) .
2 ij
√
The natural mean-field version of the LCD model is thus the c/ ij-graph
 √ c> 0, for 1 ≤ i < j ≤ n the edge ij is
defined as follows: given a parameter
present with probability min c/ ij, 1 , and these events are independent.
This graph, with c = m/2, is often studied as a proxy for the Barabási–
Albert model.
There is a similar mean-field version of the growing m-out graph, corre-
sponding roughly to adding a Poisson number of edges for each new vertex
rather than a fixed number m. More precisely, the uniformly grown random
graph or c/j-graph is simply the random graph on [n] in which each edge
ij, 1 ≤ i < j ≤ n, is present with probability min {c/j, 1}, independently of
the other edges.
That the uniformly grown random graph is particularly natural is shown
by the number of times it has been introduced. In 2001, Callaway, Hopcroft,
Kleinberg, Newman and Strogatz [63] introduced an essentially equivalent
model, the CHKNS model, and studied the point at which the giant com-
ponent emerges. It turns out that this model had been introduced long
before, by Dubins in 1984 (see [129, 176]), who posed a question essentially
equivalent to theirs, answered by Kalikow and Weiss [129] and Shepp [176],
and generalized by Durrett and Kesten [90]. The connection between these
results (concerning the connectedness of the infinite version of the model)
and the question of CHKNS was pointed out independently by Bollobás,
Janson and Riordan [40] and by Durrett [89]. Dorogovtsev, Mendes and
Samukhin [88] also studied a family of models including as a special case
the CHKNS model; we return to these results in Subsection 4.3.
A generalization of the growing m-out and uniformly grown random
graphs was considered in [166], namely the uniformly grown Z-out random
graph, where Z is a random variable. Here the distribution of the number
of edges added at each stage is essentially arbitrary.
Yet another ‘mean-field’ version of an existing model is the ‘random
graph with given expected degree sequence’ studied by Chung and Lu [67,
68, 71, 69, 70]. To define this model, one starts with a sequence w =
w1 , . . . , wn of non-negative weights satisfying some mild conditions. The
random graph G(w) is obtained by selecting each possible edge ij
inde-
pendently, with the probability that ij is selected given by wi wj / k wk .
Random Graphs and Branching Processes 29

The reason for the name is that the expected degree of vertex i is then wi .
Closely related models were studied by Khokhlov and Kolchin [137, 135],
Norros and Reittu [155] and Britton, Deijfen and Martin-Löf [60], among
others.
Turova [187, 188, 189, 190] introduced a ‘dynamical random graph’ G(t),
t ≥ 0, defined as follows. Fix three parameters γ > 0, λ > 0 and μ ≥ 0.
Starting with a single vertex at time t = 0, each existing vertex produces
new vertices according to a Poisson process with intensity γ. The newborn
vertices start their lives as isolated vertices, but as soon as there are at least
two vertices, each vertex sends out edges according to another independent
Poisson process with intensity λ; the other endpoint is chosen uniformly
among all other existing vertices. Vertices live for ever, but edges die with
intensity μ, i.e., the lifetime of an edge has an exponential distribution with
mean 1/μ. It turns out that conditional on the vertex set, and on the
time at which each vertex was born, the different possible edges are present
independently of each other.

2.5. A general sparse inhomogeneous model

The various models described in the previous subsection, together with

many others, were finally brought under a single umbrella by Bollobás,
Janson and Riordan [41], who introduced a very general model for sparse
inhomogeneous random graphs with (conditional) independence between
edges that includes the examples above as special cases. The technical
details are somewhat involved, so we first describe a simple special case.
The BJR model of [41] is based around a kernel, i.e., a symmetric
measurable function κ from (usually) [0, 1]2 to the non-negative reals. To
construct the graph G(n, κ), one first chooses the ‘types’ x1 , . . . , xn of the
vertices from the ‘type-space’, here [0, 1]. In the simplest case, x1 , . . . , xn
are simply chosen independently and uniformly from [0, 1]. Given the vertex
types, each possible edge ij is present with probability
 
(1) pij = min 1, κ(xi , xj )/n ,

independently of the other edges. In [41], some restrictions are imposed on

κ; as noted in [44], most of these restrictions are not necessary in the case
where the xi are iid. In this case it is enough that κ is integrable.
There are many variations on the details. The example above uses [0, 1]
with Lebesgue measure as the type space. More generally, in [41] one consid-
30 B. Bollobás and O. Riordan

ers a general ground space, i.e., a pair (S , μ) where S is a separable metric

space and μ is a Borel probability measure on S . The types x1 , . . . , xn need
not be independent, or indeed random. To  achieve maximum
 generality one
considers a vertex space V , i.e., a triple S , μ, (xn )n≥1 , where (S , μ) is a
ground space and, for each n ≥ 1, xn is a random sequence (x1 , x2 , . . . , xn )
of n points of S , such that
 
νn (A) =  {i : xi ∈ A} /n → μ(A)
p
(2)

whenever A ⊂ S is measurable and the measure of the boundary of A is

0. [There is no other restriction on the relationships of the sequences xn for
(n)
different n, so formally we should write xi for the ith element of xn .]
Given a kernel κ and a vertex space V , the corresponding graph is
defined as follows: given (x1 , x2 , . . . , xn ), let GV (n, κ) be the graph on [n]
obtained by selecting each possible edge independently, with the individual
edge probabilities given by (1).
Allowing greater generality in the choice of vertex types, one must
restrict the kernel correspondingly. Otherwise, for example, we could take xi
to be always rational, and κ(x, y) = 0 if x and y are rational and 1 otherwise,
in which case GV (n, κ) would be always empty even though κ = 1 almost
everywhere. The exact conditions needed are very mild; see Definition 2.7
in [41].
Unsurprisingly, allowing general ground spaces adds nothing to the
model; for a precise statement see Janson [123]. However, allowing gen-
eral vertex spaces allows the inclusion of many previous models as exact
special cases. For example, with S = (0, 1] and μ Lebesgue measure, one
√
can take the types to be fixed with xi = i/n. Then, taking κ(x, y) = c/ xy,
one recovers exactly the mean-field version of the Barabási–Albert model.
Similarly, taking κ = c/ max {x, y}, one obtains exactly the uniformly grown
random graph.
The full set-up of the BJR model in [41] is even more general: one can
consider a sequence κn of kernels that converge in a suitable sense, and take
 
(3) pij = min 1, κn (xi , xj )/n

as the probability that ij is an edge in the n-vertex graph. As shown in [41,

Subsection 16.3], this allows the exact inclusion of the CHKNS model, for
example.
Random Graphs and Branching Processes 31

Returning to a single kernel κ, the ‘rank 1 case’, where κ(x, y) =

ψ(x)ψ(y) is simply a product of functions of one variable, corresponds (more
or less) to random graphs with given expected degree sequence; for the de-
tails we refer the reader once again to [41]. As noted earlier, many versions
of this model had been studied before by many people.
Turning to Turova’s model, as shown in [41], this is again a special case
of G(n, κ), with

2λ/γ  
κ(x, y) = max {x, y}μ/γ−1 − 1
1 − μ/γ

if μ = γ and
 
κ(x, y) = 2(λ/γ) log 1/ max {x, y}
if μ = γ. Here μ is Turova’s parameter, rather than the measure on the
type space S = (0, 1], which is simply Lebesgue measure.
Other special cases of GV (n, κ) arise in perhaps surprising contexts. For
example, Janson [121] discusses an example related to quantum theory.
One special case of the model GV (n, κ) is particularly simple, and has
a much longer history. Suppose that S is finite, say S = {1, 2, . . . , r}.
Then the assumption
  (2) simply says that the number of vertices of type
i is ni = nμ {i} + op (n). With r = 2, taking κ(1, 1) = κ(2, 2) = 0 and
κ(1, 2) = κ(2, 1) = c, choosing V suitably one obtains the random bipartite
graph with n/2 vertices in one part and n/2 in the other, in which each
possible edge is present with probability c/n. Finite-type random graphs
such as this have been used in many constructions in extremal graph theory.
The finite type case of GV (n, κ) was introduced (in slightly less gener-
ality) and studied by Söderberg [179, 180, 181, 182], who noted that the
branching process approach to studying the size of the giant component in
G(n, c/n) can be extended to this model. We shall return to this later.
Although we concentrate here on the sparse case, corresponding to gen-
eralizations of G(n, c/n), let us briefly mention the very important connec-
tion between kernels and dense graphs. Clearly, the normalizing factor 1/n
in (1) or (3) is somewhat arbitrary: given any normalizing density p = p(n),
one could consider the random graph  Gp (n, κ) defined in the same way as
the BJR model, but taking min 1, pκ(xi , xj ) as the probability of the
edge ij (see [51], for example). In this notation, we would write G1/n (n, κ)
for the BJR model; when the normalization is clear from context, we omit
the subscript.
32 B. Bollobás and O. Riordan

For kernels taking values in [0, 1], often called graphons, the p = 1 version
of this model (with the xi iid) was introduced by Lovász and Szegedy [141],
under the name ‘W -random graphs’. (They wrote W for a graphon.) Thus,
the natural idea of using a function on [0, 1]2 to generate inhomogeneous
random graphs arose independently in two different contexts. The rela-
tionship of W -random graphs to the BJR model is very similar to that of
G(n, c) to G(n, c/n), c constant: they are in some sense the ‘same’ model,
but the questions one considers in the two cases are very different. In fact,
the motivation for considering kernels (graphons) in the dense case was
not so much to define a new random graph model, as to understand lim-
its of sequences of (deterministic) graphs; this program was carried out by
Borgs, Chayes, Lovász, Sós, Szegedy and Vesztergombi in a series of pa-
pers [56, 55, 141, 142, 57, 58]; we return to this in Section 6.
Of course, one can pass from the dense graph G1 (n, κ) to (a special
case of) the sparse graph G1/n (n, κ) by deleting edges, keeping each with
probability c/n. Such random subgraphs of general dense graphs were
studied by Bollobás, Borgs, Chayes and Riordan [36].

2.6. Independence and clustering

In sparse random graphs, there is a tension between independence and

clustering: since the individual edge probabilities are of order 1/n, if the
events that different edges are present are independent, then the expected
number of triangles, for example, will be (typically) of order 1. In particular,
many of the models above contain almost no short cycles. (For a detailed
discussion of clustering in the LCD model see [47].) This is unsatisfactory,
as many of the real-world networks one would like to model display a high
degree of clustering, i.e., the ‘my friend’s friend is my friend’ phenomenon.
Indeed, this is one of the original observations of Watts and Strogatz [195].
On the other hand, models without independence tend to be very hard
to analyze. Fortunately, there is a way to introduce clustering while main-
taining independence: the key is to construct a random hypergraph, where
hyperedges are selected independently, and then replace each hyperedge by
a complete graph. More generally, one can form a random graph by putting
down copies of different ‘small’ graphs F , rather than just complete graphs.
This leads to the inhomogeneous clustering model, or inhomogeneous hyper-
graph model, introduced by Bollobás, Janson and Riordan [44].
Random Graphs and Branching Processes 33

The starting point is a kernel family (κ), consisting of one kernel κF for
each connected finite graph F (or, to be pedantic, for one representative
of each isomorphism class); the kernel κF is simply a non-negative measur-
able function on [0, 1]V (F ) that is symmetric under the action  of the auto-
morphism group of F . To construct the random graph G n, (κ) , choose
x1 , . . . , xn independently and uniformly from [0, 1], and then for each F and
each r-tuple v1 , . . . , vr of r = |F | vertices, insert a copy of F with vertices
v1 , . . . , vr (in this order) with probability κF (xv1 , . . . , xvr )/nr−1 . For full
details, see [44].
As shown in [44], while extremely general, this model is still simple
enough for analysis. Moreover, by choosing the kernel family suitably, one
can produce, for example, graphs with power-law degree sequences with a
wide range of parameters for the degree exponent, the clustering coefficient,
and the extent to which vertices tend to send edges to vertices of similar
degree.

2.7. Further models

Let us briefly mention a small selection of further recent random graph

models. As noted earlier, there are hundreds of these, so we cannot hope to
survey them; our selection is not meant to be representative, but is biased
towards those we know particularly well (having studied them ourselves),
and those with particularly interesting connections to branching processes.
A wide variety of mechanisms for generating power laws have been con-
sidered. One example is some form of ‘self-organized criticality’, another re-
lated one is the paradigm of ‘heuristically optimized trade-offs’ proposed by
Fabrikant, Koutsoupias and Papadimitriou [100]. Unfortunately, as shown
by Berger, Bollobás, Borgs, Chayes and Riordan [20, 21], the particular
model they proposed does not give power laws in the usual sense. However,
Berger, Borgs, Chayes, D’Souza and Kleinberg [22, 23] showed that a differ-
ent kind of optimization of trade-offs can lead to preferential attachment,
and hence to power-law degree distributions.
It is not absolutely clear what one means by an ‘inhomogeneous’ random
graph. There are many models which are homogeneous, in the sense that all
vertices are equivalent in the model, but where the relationships between
different pairs of vertices are different, in contrast to G(n, p). One par-
ticularly interesting example is the line-of-sight model proposed by Frieze,
Kleinberg, Ravi and Debany [111]: the vertex set is a random subset of the
34 B. Bollobás and O. Riordan

grid or discrete torus, and one joins two vertices if they lie on a common
horizontal or vertical line, and are within a given range. This corresponds to
communication between sensors in an ad hoc network in a structured envi-
ronment with obstacles. Branching processes, or rather branching random
walks, turn out to be key to the analysis of this model; see Subsection 4.5.
The infinite version of the line-of-sight model is in fact a lattice percola-
tion model. Although percolation is the study of a certain kind of random
graph, the results and methods tend to have a rather different flavour to
those of classical random graph theory. Nevertheless, there is a close connec-
tion between certain kinds of percolation model, in particular the spread-out
model of Penrose [160], branching processes, and the BJR model; see Bol-
lobás, Janson and Riordan [42].
A rather different type of random graph model was proposed by Derényi,
Palla and Vicsek [83], motivated by the study of ‘communities’ in real-world
networks: given some (usually random) starting graph G, and parameters
1 ≤  < k, let Gk, be the graph whose vertex set is all complete subgraphs
in G of order k, with an edge between two vertices if the corresponding
cliques share at least  vertices. Derényi, Palla and Vicsek [83, 158] consid-
ered in particular the special case where G is simply G(n, p), and studied
heuristically the emergence of a giant component in the derived graph Gk, .
A directed version of this model was introduced by Palla, Farkas, Pollner,
Derényi and Vicsek [159] (for an overview of these and related models, see
Palla, Ábel, Farkas, Pollner, Derényi and Vicsek [157]). The thresholds for
the appearance of the giant component in these models, and the asymptotic
size of any such component, were found rigorously in [52].
Random processes on inhomogeneous graphs have also received much
attention. For example, Berger, Borgs, Chayes and Saberi [24] studied
the spread of viruses on the LCD model. There are much older models of
random processes on graphs with structure, for example the ‘stepping stone’
model of Kimura [136], modelling the spread of genes between different
populations, studied by, among others, Cox and Durrett [81], Zähle, Cox
and Durrett [200] and Durrett and Restrepo [91].

3. The Phase Transition in G(n, p)

Although a key topic of this survey is inhomogeneous random graphs, the

starting point in understanding results about these is the homogeneous case,
Random Graphs and Branching Processes 35

and in particular the model G(n, p). We focus throughout on the case of
bounded average degree, i.e., consider limiting behaviour with np bounded
and n → ∞. Sometimes, to keep things simple, we shall take np constant,
i.e., we study G(n, c/n), with c constant.
In the questions and results we are interested in, thengraph G(n, m) is
2
equivalent to G(n, 2m/n ) or to G(n, m/N ), with N = 2 , so much of the
time we shall state the results for G(n, p) rather than G(n, m). To translate
a result from G(n, p) to G(n, m), set m = pN , say.

3.1. Historical remarks

As we have already mentioned, the greatest single result of Erdős and

Rényi about random graphs concerns the ‘phase transition’ in the compo-
nent structure of a random graph process or, equivalently, a random graph
G(n, c/n), as c grows from c < 1 to c > 1. In fact, in [94] this is not so much
a single theorem as an attractive formulation of the union of several results.
We shall spend some time on the rather peculiar history of this result and
the refinements that came many years later.
Let us start by quoting from the original paper of Erdős and Rényi [94],
changing only the notation (and translating from G(n, m) to G(n, p), as
promised).
Thus the situation
 can
 be summarized as follows: the largest
component of G n, p(n) is of order log n for np(n) ∼ c < 1, of
order n2/3 for np(n) ∼ 1 and of order n for np(n) ∼ c > 1. This
double “jump” of the size of the largest component when np(n) passes
the value 1 is one of the most striking facts concerning random graphs.

There is no doubt that this is a striking result indeed: it is not surprising

that for the next quarter of a century this amazing “double jump” was the
crown jewel of the theory of random graphs, with the emphasis on the double
jump, rather than the sudden growth of the largest component from Θ(log n)
to Θ(n), the birth of the unique ‘giant’ component. And, reading the
summary above, one has to agree.  In fact,
 this result describes the growth
of the largest component of G n, p(n) in the entire range p = Θ(1/n):
after all, for any such function p(n) there is a sequence n1 < n2 < . . . and
a constant c > 0 such that limi→∞ ni p(ni ) = c, i.e., p(ni ) ∼ c/ni . Briefly,
concerning the birth of the giant component, the Erdős–Rényi results tell
us everything we wish to know.
36 B. Bollobás and O. Riordan

In fact, this is not the case at all. It is regrettable that the above
formulation of the Erdős–Rényi result lulled interested combinatorialists
into the belief that there was essentially nothing left to prove about the
emergence of the giant component, and the field lay dormant for over two
decades.

As Littlewood said, a first proof is allowed to be unduly complicated,

and a major paper is allowed to have mistakes. There are few papers to
which this dictum is more applicable than the Erdős–Rényi paper [94] of
1960 on the evolution of random graphs. This major paper has opened
up new vistas, and has been the starting point of a rich theory of random
graphs that is still alive and well, but several of the theorems are incorrect,
and some of the proofs are a little cavalier.

In
 particular,
 the assumption that in our study of the largest component
of G n, p(n) , the relation np(n) ∼ c with c > 0 constant is just about the
same as np(n) = c happens to be correct for c = 1, but for c = 1 this
could hardly be further from the truth:
 knowing
 that np(n) ∼ 1 tells us
very little about the structure of G n, p(n) . For example, Theorem 7c
of [94] asserts that if np(n) ∼ 1 and ω(n) → ∞, then whp the largest
tree component of G n, p(n) has at least n2/3 /ω(n) and at most ω(n)n 
2/3

vertices. (And nowhere is it excluded that in this range G n, p(n) may

have even larger non-tree (cyclic) components.) In fact, in this range the
order of the largest tree component depends very strongly on the ‘error
term’ ε = ε(n)
 = p(n)n
 − 1. As we shall see in Subsection 3.4, setting
p = p(n) = 1 + ε(n) /n, with ε = −1/(log n)2 , say, the order of the largest
tree component in G(n, p) is about 2(log n)5 , if ε = −n−1/5 then it is about
4(log n)n2/5 /5, and if ε = +n−1/5 then about 2n4/5 .

Strangely, the fact that there are problems with the results concerning
the range p ∼ 1 or m ∼ n/2 went unnoticed for twenty-four years, until
1984, when Bollobás [29] took a closer look at the phenomena in this range.

As we shall see more precisely later, in [29] it was proved that if m(n) =
n/2 + s with n2/3 log n ≤ s = o(n) (corresponding to ε ≥ n−1/3  log n and

ε = o(1)), then whp the largest component in G(n, m) has 4 + o(1) s
vertices. On a more ‘philosophical’ level, the main discovery of [29] was
that for m = n/2 + s the order of the largest component of G(n, m) is a
rather pleasant function of m, provided |s| ≥ n2/3 log n, say. (As always,
we mean everything with high probability, whp.) Although within this
‘window’ |s| = |n/2 − m| = O(n2/3 log n) not much was said about the
Random Graphs and Branching Processes 37

order of the largest component, outside this window it was proved to be

concentrated.
Let us note that, contrary to a widely held belief, the size of the window
(about n2/3 ) has nothing to do with the maximal order of a component at
p = 1/n or m = n/2, which is also about n2/3 . Although Erdős and Rényi
were far from expecting any ‘window’, there is some evidence that they
expected that G(n, m) is especially like G(n, n/2) provided the difference
m − n/2 has order O(n1/2 ).
The logarithmic factor in the size of the window was due to the somewhat
crude bounds used in [29]; in those days, even these bounds were deemed
to be over-precise, as an estimate of the form n2/3+o(1) was considered
accurate enough. Later, in 1990, the unnecessary logarithmic factor was
duly removed by L  uczak [143], using a much more careful analysis. With
 uczak established that the true size of the window is Θ(n2/3 );
this result, L
equivalently, the window in G(n, p) has width Θ(n−4/3 ). Since then, many
papers have appeared describing the behaviour of G(n, m) or G(n, p) in and
around this window (see [59, 124, 146, 145, 153, 161, 162, 175, 186]): for
lack of space, we shall comment on only four of them.
First, L
 uczak and Wierman [146] corrected another false theorem in [94].
They proved that at the critical probability p = 1/n, whp every cycle in
G(n, p) is a ‘hole’, i.e., a cycle without a diagonal. A consequence of this is
that whp G(n, 1/n) has chromatic number 3.
Second, Janson, Knuth, L uczak and Pittel [124] wrote a monumental pa-
per in which they combined algebraic, analytic and combinatorial techniques
to study G(n, p) at its phase transition. This paper marked the re-entry of
generating functions into the mainstream theory of random graphs: start-
ing with generating functions for the exact number of certain subgraphs,
Janson, Knuth, L uczak and Pittel used intricate analysis to turn unenlight-
ening exact formulae into accessible asymptotic expressions. For example,
they showed that the probability that a random graph process G = (Gt )N
0
is such that no Gt has more than one component containing cycles tends
to 5π/18. With this paper, generating functions, discarded by Erdős and
Rényi, were returned to the arsenal of many practitioners of probabilistic
combinatorics.
Third, in 1999, L uczak, Pittel and Wierman [145] answered a question
asked and left open by Erdős and Rényi about the planarity of G(n, p).
They proved that G(n, p) is non-planar whp if and only if p is beyond the
window, i.e., p = 1/n + ω(n)n−4/3 , with ω(n) → ∞.
38 B. Bollobás and O. Riordan

Finally, in a difficult and very technical paper, Pittel [162] determined

the limiting distribution of the order of the largest component of G(n, p)
with p inside the critical window, i.e., when p = (1 + λn−1/3 )/n with λ a
constant or a bounded function of n. An immediate consequence of this
very impressive result is the following. Let a > 0 and λ be fixed. For
p = (1 + λn−1/3 )/n, the probability that G(n, p) has no component with
more than an2/3 vertices tends to
 3 
λ   λ log (λ log λ)
exp − − 1 + η(λ) ,
6 a

where η(λ) = O(1/ log λ) as λ → ∞. A point process describing the limiting

joint distribution of the sizes of the largest components of G(n, p) inside
the window was studied by Aldous [6] and Janson and Spencer [128], for
example.
Similar windows arise in many other contexts; a particularly closely
related one is random 2-SAT, where the width of the window was determined
by Bollobás, Borgs, Chayes, Kim and Wilson [37].
After this very brief review of the results, let us turn to the methods
that have been employed to study the phase transition in G(n, p). First,
before Erdős and Rényi wrote their series of papers, the accepted wisdom
had been that, to learn about G(n, p), generating functions have to be used:
these methods give exact results, but much of the time only for very small
values of n can one see what these results really mean. The Erdős–Rényi
revolution replaced this approach by a ‘static’ probabilistic point of view:
to study G(n, p), consider the probability space G (n, p); on this space our
favourite graph parameter (for example the number of triangles, or the
number of pairs of vertices not joined by a path of length at most three)
is a random variable, which we can study by approximating its mean and
variance, and perhaps its higher moments.
This ‘static’ approach still dominates today, although by now we have
much better tools, including martingale and correlation inequalities, to
bound the moments or to approximate the probabilities in other ways.
However, the formidable paper of Janson, Knuth, L  uczak and Pittel [124]
has put an end to the hegemony of the Erdős–Rényi approach, so that
by now generating functions are again important in the theory of random
graphs.
The use of branching processes is newer still. As one of the aims of this
review is to demonstrate their use in the theory of random graphs, later
Random Graphs and Branching Processes 39

in this section we shall give a rather simple proof of the phase transition
in G(n, p) we keep mentioning. (In fact, we shall give two proofs, one of a
very weak form of the result, and one of a strong form.) Our proof is very
different from the recent simple proof given by Nachmias and Peres [153],
which is based on martingale techniques.
One of the reasons why branching processes are so effective in studying
components of random graphs is that they allow us to bypass a number of
intricate combinatorial difficulties. To conclude this section, we shall give a
brief description of one of these problems and the efforts to solve it, at least
partially.
The difficulty alluded to above is the problem of bounding the number of
components of different types: trees, unicyclic components, and (k, k + )-
components, i.e., components with k vertices and k +  edges. If c is a
constant different from 1, then this method works very well for G(n, c/n),
since the expectation of the number of vertices on non-tree components with
at most n2/3 vertices, say, is bounded as n → ∞. However, near the point
of the phase transition, when c is no longer constant but is 1 ± ε(n) with
ε(n) tending to 0 at a suitable speed, we may have to deal with (k, k + )-
components for  = (k) not that small, so one needs a good bound for
C(k, k + ), the number of connected graphs with k labelled vertices and
k +  edges, which applies to a wide range of  = (k).
While Cayley’s classical formula tells us that C(k, k − 1) = k k−2 , i.e.,
there are k k−2 trees on k ≥ 1 labelled vertices (see, e.g., [33]), and there are
numerous proofs (see, e.g., [108, 132, 164, 10]) of the following formula for
the number of unicyclic graphs:

1  kj
k−3
C(k, k) = (k − 1)! ∼ (π/8)1/2 kk−1/2 ,
2 j!
j=0

for  = (k) there is no easy-to-use formula for C(k, k + ). [Here and
elsewhere the expression f ∼ g means  that the functions f and g are
asymptotically equal: f = 1 + o(1) g, as our variable – here k, but usually
n – tends to infinity.] In a series of papers, Wright [196, 197, 198, 199] gave
(more and more complicated) exact formulae for C(k, k + ) for general k
and successive , and the following asymptotic formula when  = o(k 1/3 ):
 
C(k, k + ) = γ k k−1/2+3/2 (1 + O 3/2 k −1/2 ),
40 B. Bollobás and O. Riordan

where γ depends only on . Here γ−1 = 1, γ0 = (π/8)1/2 , γ1 = 5/24 and

  e /2
γ ∼ 3/(4π)
12
as  → ∞. The constant in the last formula was first identified by Bagaev
and Dmitriev [12]. Some years later, Bender, Canfield and McKay [18] gave
extremely accurate estimates for C(k, k + ), for a general function  = (k),
but their complex formula seems to be difficult to work with.
In fact, much of the time, in our study of G(n, p) it suffices to use an
upper bound (rather than an asymptotic expression) for C(k, k + ). Such
an upper bound was given by Bollobás [29]: he gave a very simple proof of
the fact that the leading factors in Wright’s formula give an upper bound
in a wide range of . More precisely, there is a constant c > 0 such that
⎧ /2
⎪ c
⎪
⎨  kk−1/2+3/2 , 1 ≤  ≤ k,
C(k, k + ) ≤  
⎪
⎪ k
⎩(ck) ,
k+
−1 ≤  ≤ − k.
2

This bound was the starting point of the analysis in [29] of the structure of
G(n, p) near the critical probability, and was heavily used in other studies
as well, including those of L
 uczak [143] and Pittel [162].

3.2. Local behaviour

In the modern view of G(n, c/n), branching processes play a fundamental

role. Let X(λ) denote the branching process defined as follows: we start with
a single particle in generation 0. Each particle in generation t gives rise to a
Poisson number of children in generation t + 1 with mean λ, independently
of the other particles and of the history. This is one of the simplest examples
of a Galton–Watson branching process, introduced by Watson to answer a
question of Galton about the probability of surnames dying out [194]. For
the history see, for example, Kendall [134]; for basic results on branching
processes, see, for example, the books by Harris [119, 120] or Athreya and
Ney [9].
As we shall see, there is a very close connection between G(n, c/n)
and X(c). Indeed, the local connection is almost immediate. Let v be a
random vertex of G(n, p); more precisely, choose v uniformly form [n] and
Random Graphs and Branching Processes 41

independently of which edges are selected to form G(n, p). Let Cv denote
the component of G(n, p) containing v. When studying the distribution of
|Cv |, the number of vertices in Cv , we shall think of choosing v first, and
then revealing which edges of G(n, p) are present.
In fact, one tests edges one by one in what is known as the component
exploration process. This was applied by Karp [131] to study the size of
the giant component in random directed graphs; similar ideas had been
used earlier in other contexts, for example by Kendall [133] in the study of
epidemics. The application to undirected graphs is particularly simple.
The component exploration process is simply breadth first search: we
maintain two lists, one of ‘explored’ vertices, and one of ‘unexplored’ ver-
tices. The remaining vertices are ‘new’. Initially, v is ‘unexplored’, and all
other vertices are ‘new’. In step t of the exploration, we pick an unexplored
vertex vt , if there is one (thus v1 = v). Note that vt may be picked in
any way we like: this makes no difference; for definiteness, one often picks
the first unexplored vertex in some fixed order chosen in advance. Having
picked an unexplored vertex vt , we test all possible edges vt w from vt to
new vertices w to see which ones are present in G(n, p). If vt w is present,
we add w to our ‘unexplored’ list. At the end of step t we move vt from
the unexplored list to the explored list. We stop when there are no unex-
plored vertices. Clearly, at the end of this process, the set {v1 , v2 , · · · , vr }
of explored vertices is precisely the vertex set of Cv .
The key property of this process is that an edge may only be tested
once, so when we come to test an edge, the probability (conditional on the
history) that it is present is exactly p. For comparison with a branching
process, note that if there are nt ‘new’ vertices at the start of step t, then the
number Xt of new neighbours of vt that we find has the binomial distribution
Bi (nt , p). Specializing to G(n, c/n) with c constant, i.e., taking p = c/n,
when considering components of size k, we stop our exploration immediately
if at some point our explored and unexplored lists together contain more
than k vertices. Up to this point, at every step t, we have n − k ≤ nt ≤ n, so
(crudely) nt ∼ n and E(Xt ) ∼ np = c. Since p → 0, it follows that Xt may
be coupled with a Poisson random variable Yt with mean c so that Xt and
Yt agree with probability 1 − o(1). Hence, for k fixed, we may couple the
exploration above with the branching process X(c) so that the probability
of a mismatch before both processes have reached size k + 1 is o(1).
Let Nk (G) denote the number of vertices of a graph G in components
of order k. The trivial argument above gives the following lemma.
42 B. Bollobás and O. Riordan

Lemma 1. Let c > 0 and a positive integer k be fixed, and let Gn =

G(n, c/n). Then

1    
E Nk (Gn ) → P ( X(c) = k ),
n
 
where  X(c) denotes the total number of particles in all generations of X(c).

Proof. Choose v ∈ [n] randomly as above, and consider the coupling of

the initial stages of the branching process and the component exploration
started at v described above. The probability that our coupling fails before
uncovering k + 1 vertices is o(1). If we have seen k + 1 vertices in Cv , then
certainly |Cv | = k,
 and a similar
 comment
  applies to the branching process.
It follows that P |Cv | = k = P ( X(c) = k ) + o(1). Revealing the edges
   
of Gn before choosing v shows that P |Cv | = k = E Nk (Gn ) /n, so the
result follows.

Let Nk0 (G) denote the number of vertices of G that are in k-vertex
components that are trees. If we explore the component Cv as above and
find that it has size k, then we have so far revealed exactly k − 1 edges
within this component. We have also tested a certain number t ≥ 0 of other
potential edges within Cv and found them to be absent. Considering the
edges not so far tested, the conditional probability that Cv is not a tree is
thus
    
( k
)−(k−1)−t k k
1 − (1 − p) 2 ≤p − (k − 1) − t ≤ p ≤ pk2 /2.
2 2

If p = c/n with c > 0 constant and k is fixed, this probability is o(1), and
it follows from Lemma 1 that
1  0   
(4) E Nk (Gn ) → P ( X(c) = k ).
n

Although this distinction

 is not always made, there is in principle a big
difference between E Nk (Gn ) /n and what we would really like to study:
the fraction of vertices in components of order k. This fraction is a random
quantity, and it could a priori vary a lot, and so need not be close to its
expectation. However, there is a simple trick to show that this is not the
case here, requiring essentially no further work. This trick was used by
Bollobás, Janson and Riordan [41], for example.
Random Graphs and Branching Processes 43

Lemma 2. Let c > 0 be fixed, and let Gn = G(n, c/n). For each fixed k we
have
1    
E Nk (Gn )2 → P ( X(c) = k ) ,
2
n 2

and
1  
Nk (Gn ) → P ( X(c) = k ).
p
n
Proof. For the first statement we simply start two explorations at vertices
v and w chosen uniformly at random from [n], independently of each other
and of Gn . Provided we have so far uncovered at most k vertices in each
exploration, the chance of meeting at each step is O(k/n), so the total
chance of meeting by this point is O(k 2 /n). It follows that
   
P |Cv | = k, |Cw | = k = P ( X(c) = k ) + o(1).
2

   
Since E Nk (Gn )2 = n2 P |Cv | = k, |Cw | = k , the
 first statement follows.
Using Lemma 1, it follows that Var Nk (Gn )/n = o(1), and the second
statement follows.

Although it may not be obvious at first sight, the connection described

above is very useful. This is because the branching process, despite be-
ing potentially infinite, is mathematically a simpler object than G(n, c/n).
Indeed, if one can reduce a question about random graphs to one about
branching processes, one often thinks of it as (essentially) solved.

3.3. The giant component

The connection between G(n, c/n) and X(c) described above is essentially
trivial: it is not very surprising that G(n, c/n) looks locally like X(c).
On the other hand, what is surprising is that one can get much global
information about G(n, c/n) from the branching process. The reason why
this is surprising is that the connection between the exploration process and
the branching process must break down: eventually, in the graph we run
out of vertices, while the branching process (for c > 1) may well continue
forever. Moreover, for c > 1, the graph contains many cycles: in the
branching process we never see these. Depending on how close a connection
we want, the local approximation may well break down quite soon; certainly
we cannot push the coupling argument as far as k = Θ(n) without losing
44 B. Bollobás and O. Riordan

accuracy. Nonetheless, by less direct arguments we can understand even

the limiting size of the largest component.
The first concrete global question we consider is: how large is the largest
component of G(n, c/n)? Thus, writing C1 (G) for the maximum number of
vertices in a component of G, we should
 like to determine the distribution
of the random variable C1 G(n, c/n) , as a function of c and n.
Let ρ(λ) denote the survival probability of the Galton–Watson branching
process X(λ), i.e., the probability that there are particles in all generations.
It is essentially immediate from basic properties of Poisson processes that
ρ(λ) is given by the largest solution to

(5) ρ(λ) = 1 − e−λρ(λ) .

Furthermore, this equation has at most one non-zero solution, ρ(λ) is a

continuous function of λ, and ρ(λ) > 0 if and only if λ > 1.
Although, as we have seen in Subsection 3.1, the original viewpoint was
very different, we know by now that branching process analysis gives the
approximate size of the largest component of G(n, c/n). The simplest form
of this result is one of the first and best known results in the theory of
random graphs, due to Erdős and Rényi [94], although they did not state it
in quite this form.

Theorem 3. Let c > 0 be fixed. For any η > 0 the bounds

     
ρ(c) − η n ≤ C1 G(n, c/n) ≤ ρ(c) + η n

hold whp as n → ∞.

In other words, the normalized size of the giant component of G(n, c/n)
converges in probability to ρ(c). In fact, as remarked in Subsection 3.1,
Erdős and Rényi proved much more,
 e.g. that
 if c < 1 is constant then
there is an A = A(c) such that C1 G(n, c/n) ≤ A log n holds whp.

Proof. Let Gn = G(n, c/n). For each fixed k, from Lemma 2 we have
 
 
n Nk (Gn ) → P ( X(c) = k ). Defining N<ω (Gn ) and N≥ω (Gn ) in the
1 p

obvious way, it follows that there is some function ω = ω(n) tending to

Random Graphs and Branching Processes 45

infinity, which we may take to be o(n), so that

   
1       
 N≥ω (Gn ) − P ( X(c) ≥ ω ) =  1 N<ω (Gn ) − P ( X(c) < ω )
n  n 

 

ω−1
1    p
=  Nk (Gn ) − P ( X(c) = k ) → 0.
n 
k=1

Now
   
ρ(c) = P ( X(c) = ∞) = lim P ( X(c) ≥ k ).
k→∞
 
Since ω → ∞, it follows that P ( X(c) ≥ ω ) → ρ(c), so we have shown that

1 p
(6) N≥ω (Gn ) → ρ(c).
n

In other words, we have (about) the right number of vertices in ‘large’

components, where ‘large’ means containing more than ω(n) vertices.
The upper bound
 in Theorem
 3 follows immediately, using the fact that
C1 (Gn ) ≤ max ω, N≥ω (Gn ) , and recalling that ω = o(n). It remains to
prove the lower bound; for this we shall use the ‘sprinkling’ method of Erdős
and Rényi [94], which exploits the continuity of ρ(c) as a function of c.
We may of course assume that ρ(c) > 0. Given an η > 0, which we shall
take to be smaller than ρ(c)/2, choose c < c so that ρ(c ) ≥ ρ(c) −η/3.
  
Applying (6)to Gn = G(n, c /n), we see that whp at least ρ(c ) − η/3 n ≥
ρ(c) − 2η/3 n vertices of Gn are in large components, where ‘large’ means
of size at least some ω(n) → ∞. Let B denote the set of such vertices.
(Here B stands for ‘big’; we shall use L for something else shortly.)
We may construct Gn from Gn by adding to it each non-edge with a
certain probability r, independently of the others. The relevant probability
r is given by (1 − c /n)(1 − r) = 1 − c/n, but it is enough to note that
r ≥ (c − c )/n. Let us condition on Gn . We shall assume that |B| ≥
ρ(c) − 2η/3 n. It suffices to show that given any 
 Gn satisfying
 this
condition, the conditional probability that C1 (Gn ) ≥ ρ(c)−η n is 1−o(1).
   
If |B| ≥ ρ(c) − 2η/3 n and C1 (Gn ) ≤ ρ(c) − η n, then no component
of Gn contains more than |B| − ηn/3 vertices of |B|. It follows that there
is a cut B = B1 ∪ B2 with the following properties: B1 and B2 are disjoint,
|B1 |, |B2 | ≥ ηn/3, and no edge of Gn joins B1 to B2 . To complete the proof,
46 B. Bollobás and O. Riordan

it suffices to show that the expected number of such cuts is o(1); this is
precisely the method Erdős and Rényi used in [94].
Certainly, in any such cut, B1 must be a union of large components of
Gn .Since B contains at most |B|/ω(n) such components, the number of
possible choices for B1 is at most 2|B|/ω ≤ 2n/ω = 2o(n) . On the other hand,
for (B1 , B2 ) to form a cut, each of the
   
|B1 | |B2 | ≥ (ηn/3) |B| − ηn/3 ≥ (ηn/3) ρ(c) − η n = Θ(n2 )

possible cross edges must be absent from Gn \ Gn , an event of probability

at most
  Θ(n2 )
1 − (c − c )/n = e−Θ(n) .

Hence the expected number of cuts is eo(n)−Θ(n) = e−Θ(n) = o(1).

There are many other ways of writing the ‘sprinkling’ argument above.
For example, one can take the local coupling much further than we did here,
up to a size growing with n at a suitable rate, and then one needs many
fewer sprinkled edges, and very nearly all the ‘large’ components will join
up; we shall use this method in Subsection 3.6. The argument above has
the advantage of requiring only extremely weak bounds at each stage, which
means that it adjusts readily to other contexts. See, for example, Bollobás,
Janson and Riordan [41, 44], and Bollobás, Borgs, Chayes and Riordan [36].

3.4. Stronger results for G(n, p)

It is perhaps not so well known that much stronger results about the size
of the largest component of G(n, p) can also be proved (relatively) easily by
branching process methods. Here we ‘zoom in’ on the critical probability.
As mentioned in Subsection 3.1, results of this kind were first proved by
Bollobás [29], who was the first to notice that Erdős and Rényi had not in
fact completely settled this question.
Throughout this section we consider G(n, λ/n), where λ = 1 ± ε, with
ε = ε(n) > 0 and ε → 0 as n → ∞. We write L for ε3 n; we shall always
assume that L → ∞, i.e., that we are outside the critical window.
Bollobás [29, 31] studied the size of the giant component of G(n, λ/n)
under (essentially) these assumptions, concentrating on the more interesting
Random Graphs and Branching Processes 47

supercritical case. In our notation, he proved, among other related results,

Theorem 4 below. We have translated this from G(n, m) to G(n, p); as noted
in Subsection 3.1, results of this type have trivially equivalent formulations
for the two models. We write Ci (G) for the number of vertices in the ith
largest component of a graph G. Note that we are simplifying the results
for the presentation here.
Theorem 4. Let λ = 1+ε, where ε = ε(n) ≥ 4(log n)1/2 n−1/3 but ε = o(1).
Then
  ε + ε  √ 
C1 G(n, λ/n) = n + Op εn/ L
1+ε
 √ 
= 2εn + O(ε2 n) + Op εn/ L ,

where ε is the positive solution to


(7) (1 − ε )eε = (1 + ε)e−ε .

Also, if L ≥ nγ0 for some γ0 > 0 and ε = o(1/ log n), then
 
  −2 5 log log n
(8) C2 G(n, λ/n) = 2ε log L − + Op (1) .
2

Note that there are two typographical errors in the last line of Theorem
6.9 of [31], repeated when the results are quoted by L  uczak [143], for
example. That the errors are typographic may be seen by comparison
with Corollary 6.4 of [31], from which the formula is incorrectly copied.
Bollobás [29, 31] also gave results for the (less interesting) subcritical case.
L
 uczak [143] weakened the restrictions on ε in the result above to what
they ‘should be’. The case ε constant is much easier and had already been
treated by Erdős and Rényi [94], so there is no loss in assuming that ε → 0.
At the other extreme, if L = ε3 n does not tend to infinity, then we are
inside the critical window, and results of the type we state here do not
hold. Again we translate L  uczak’s result to our notation, and simplify. (He
gave the distribution of the Op (1) term in the first and last statements, and
considered the ith largest component for each fixed i.)
Theorem 5. Let λ = 1 − ε, where ε = ε(n) satisfies L = ε3 n → ∞ and
ε = o(1). Then
 
  −2 5 log log L
(9) C1 G(n, λ/n) = 2ε log L − + Op (1) .
2
48 B. Bollobás and O. Riordan

Let λ = 1 + ε, where ε = ε(n) satisfies L = ε3 n → ∞ and ε = o(1). Then

  ε + ε  √ 
C1 G(n, λ/n) = n + Op εn/ L ,
1+ε

where ε is defined by (7), and

 
  −2 5 log log L
(10) C2 G(n, λ/n) = 2ε log L − + Op (1) .
2

Unfortunately, as we shall see below, Theorem 5 is not correct as stated:

the condition ε = o(1/ log n) in the second part of Theorem 4 is no accident!
Here, we shall prove two results; we start with the supercritical case.

Theorem 6. Let λ = 1 + ε, where ε = ε(n) > 0 satisfies ε → 0 and

L = ε3 n → ∞. Then
   √ 
(11) C1 G(n, λ/n) = ρ(λ)n + Op εn/ L .

From (5) it is easy to check that

8 28 464 4
ρ(1 + ε) = 2ε − ε2 + ε3 − ε + O(ε5 ),
3 9 135

so (11) implies in particular that

 
C1 (G n, (1 + ε)/n ) p
→ 1.
2εn

Depending on how small εis, it may be more natural to express the error
term in Theorem 6 as Op ( n/ε); the form given above makes the relation-
ship of the error term to the main term clearer.
Theorem 6 is best
√ possible in some sense: the standard deviation is
indeed of order εn/ L. In fact, at least for some range of ε, more  precise
√ 
results have been proved, giving the limiting distribution of the Op εn/ L
term (see, for example, Stepanov [183, 184, 185]). We shall not go into this
much detail, as our aim here is to present simple proofs to illustrate the use
of branching process and tree counting methods.
In the subcritical case, we shall prove the following result.
Random Graphs and Branching Processes 49

Theorem 7. Let λ = 1 − ε, where ε = ε(n) > 0 satisfies ε → 0 and

L = ε3 n → ∞. Then
 
  5 log log L
(12) C1 G(n, λ/n) = δ −1 log L − + Op (1) ,
2

where

(13) δ = λ − 1 − log λ = −ε − log (1 − ε) = ε2 /2 + ε3 /3 + O(ε4 ).

This certainly implies that

 
C1 G(n, λ/n) p
  → 1.
δ −1 log L − (5/2) log log L

Theorem 7 is best possible, in the sense that for any constant M , when n is
large, the term written as Op (1) really does have probability bounded away
from 0 of exceeding M , or of being less than −M . In fact, just as in [31], by
considering all moments rather than just the first two, our proof gives the
limiting distribution of the Op (1) term; we shall not emphasize this. Note
that Theorems 6 and 7 remain correct if we replace ε by a constant; in this
case they become equivalent to Corollary 5.11 in [31].
As we shall see later, Theorem 7 gives a corresponding bound on
C2 G(n, λ/n) in the supercritical case; see Theorem 12.
Before turning to the proofs, let us comment on the relationship of
Theorems 6 and 7 to Theorem 4 and (the incorrect) Theorem 5. Firstly,
 +ε
the rather mysterious ratio ε1+ε appearing in the latter results is simply
ρ(1 + ε). To see this, for λ > 1, let λ be the ‘dual’ branching process
parameter defined by
 
λ = λ 1 − ρ(λ) .
From elementary  properties
 of Poisson processes, the distribution of X(λ)
 
conditioned on X(λ) < ∞ is simply that of X(λ ). Using (5), one can
check that λ is also the unique solution to

λ e−λ = λe−λ .


Thus ε as defined by (7) above is simply 1 − λ , λ = 1 + ε, and one does

indeed obtain that
ε + ε λ − λ
= = ρ(λ).
1+ε λ
50 B. Bollobás and O. Riordan

We hope the reader will agree that, with hindsight, ρ(λ) is the more natural
description of this constant!
A minor difference between the formulae in Theorem 4 and those in
Theorems 5 and 7 is the appearance of log log n in the former in place
of log log L; for the parameters considered, the difference is o(1), which is
absorbed into the error term.
A more significant difference is the appearance of δ −1 in our results, in
place of 2ε−2 in Theorems 4 and 5. From (13), we have
 
δ −1 = 2ε−2 1 + Θ(ε) .

If ε = o(1/ log n), as in the relevant part of Theorem 4, then it makes no

difference whether we write δ −1 or 2ε−2 in front of the bracket in (8), (9),
(10) or (12). Thus our results are consistent with, and extend, Theorem 4.
On the other hand, in Theorem 5, whenever ε log n is unbounded, the
difference is important, and to obtain a correct result one must replace
2ε−2 by δ −1 . (The mistake in L  uczak [143] is in his (2.1), where terms of
order ks3 /n3 are neglected incorrectly.) As we shall see, δ plays a natural
role in the branching process: it is the decay constant in the exponential
tail probability that the process has large (but finite) total size.
In the next two subsections we turn to our new proofs of Theorems 6
and 7, based on branching processes. We start with the latter, which is
simpler, although our initial remarks will apply with λ > 1 as well as λ < 1.

3.5. The subcritical case

As we have seen in Subsection 3.1, the original proofs of (versions of) the
results above were based on counting graphs with slightly more edges than
vertices, and were rather involved. Much simpler proofs of somewhat weaker
results were given by Nachmias and Peres [153] using martingale methods.
Here we give simple proofs of strong results based on direct comparison with
a branching process. It turns out that we still have to count trees; however,
this is the easy part. Branching process arguments will allow us to avoid
counting components containing cycles.
We shall prove Theorems 6 and 7 via a sequence of simple lemmas. We
shall need the basic fact from combinatorics mentioned in Subsection 3.1,
Cayley’s formula k k−2 for the number of trees on a given set of k (labelled)
vertices.
Random Graphs and Branching Processes 51

As before, let Gn = G(n, λ/n), where λ = λ(n). For the moment, we

consider the sub- and super-critical cases together, so we assume only that
λ(n) = Θ(1).
Our first aim is to prove a fairly precise formula for the expected number
of vertices in tree components of a given size. Let Tk (Gn ) denote the number
of components of Gn that are trees with k vertices, and, as before, let
Nk0 (Gn ) = kTk (Gn ) be the number of vertices in such components. A key
quantity that we shall study is
 
μn,k,λ = E Nk0 (Gn ) ,

where Gn = G(n, λ/n).

It is very easy to see that
   
n k−2 λ k−1 2
(14) μn,k,λ = k k (1 − λ/n)kn−k /2−3k/2+1 .
k n
Indeed, we must show that the same formula without the initial
 factor of k
gives the expectation of Tk (Gn ). To see this, note that nk is the number
of choices for the vertex set, k k−2 for which tree we have on this vertex set,
and then this particular tree T is present in Gn if and only if each of its
k − 1 edges is present, an event of probability pk−1 = (λ/n)k−1 . Finally, if
present, T is a tree component if and only if there are no edges from V (T )
to the remaining n − k vertices,
  and no other edges between the vertices
of T . In total, k(n − k) + k2 − (k − 1) = kn − k 2 /2 − 3k/2 + 1 other edges
must be ruled out, giving the final factor in (14).
Unfortunately, to obtain strong results we must calculate a little. This
is very easy but slightly messy. Since for |x| < 1 we have log (1 − x) =
−x − x2 /2 − x3 /3 · · · , if 0 < x ≤ 1/3 then

(15) −x − x2 < −x − x2 /2 − x3 /2 < log (1 − x) < −x − x2 /2 < −x,

and, although we shall not use this,

x − x2 /2 < log (1 + x) < x − x2 /3.

Writing n(k) for the falling factorial n(n − 1)(n − 2) · · · (n − k + 1) =

n!/(n − k)!, applying (15) to log (1 − i/n) and summing from i = 1 to
k − 1, it follows that for k ≤ n/3 we have
 2   2 
k k3 n(k) k k
(16) exp − − ≤ k ≤ exp − + .
2n 3n2 n 2n 2n
52 B. Bollobás and O. Riordan
√
If k ≤ n, then the +k/(2n) term on the left that we have ignored above
is larger than the −k3 /(3n2 ) term, and we have
   
k2 n(k) k2 k
(17) exp − ≤ k ≤ exp − + .
2n n 2n 2n

We shall assume throughout these arguments that k 3 /n2 → 0, which

implies that k/n → 0. From (16) or (17), in this range we have
 
n(k) = nk exp − k2 /(2n) + o(1) .
   −2

Writing (1 − λ/n)
n as exp log (1 − λ/n) = exp − λ/n − O(n ) and using
the fact that k = n(k) /k!, it follows after a little calculation that
 
k k−1 −1  −λ  k k2
(18) μn,k,λ =n λ λe exp (λ − 1) 3 2
+ O(k /n + k/n) .
k! 2n

Digressing for a moment, we may use (18) to tell us something about

the branching process. Let us write
 
ρk (λ) = P ( X(λ) = k )

for the probability that the branching process X(λ) consists of exactly k
particles in total, and
 
ρ≥k (λ) = P ( X(λ) ≥ k )

for the probability that it consists of at least k particles.

Lemma 8. For any λ > 0 and k ≥ 1 we have

  kk−1 −1  −λ  k
(19) ρk (λ) = P ( X(λ) = k ) = λ λe .
k!

Proof. This formula may be obtained by evaluating directly the probability

that the branching process yields each possible tree on k vertices; although
straightforward, this is a little tricky to write down. Alternatively, we may
simply apply Lemma 1, or rather its consequence (4). Indeed, from (4) we
have μn,k,λ /n = ρk (λ) + o(1) as n → ∞ with k and λ fixed. Since ρk (λ) does
not depend on n, we thus have ρk (λ) = limn→∞ μn,k,λ /n. Dividing (18) by
n and taking the limit immediately yields (19).
Random Graphs and Branching Processes 53

√ k → ∞, then using Stirling’s formula in the (rather crude) form

When
k! ∼ 2πk k+1/2 e−k , we may rewrite (19) as
 k
(20) ρk (λ) ∼ (2π)−1/2 k−3/2 λ−1 λe1−λ .

The quantity λe1−λ turns out to play a fundamental role in the analysis
of X(λ) or of G(n, λ/n). It is easily seen that λe1−λ is at most 1, so it is
convenient to consider the negative of its logarithm. Thus we set
 
(21) δ = δ(λ) = − log λe1−λ = λ − 1 − log λ.

In this notation, recalling that the approximation in Stirling’s formula is

correct within 10% for all k ≥ 1, we have

(22) ρk (λ) ≤ k−3/2 λ−1 e−δk

for all k and λ.

It is easy to check that as ε → 0, we have δ(1 ± ε) ∼ ε2 /2. If λ ∼ 1 and
k → ∞, then (20) reduces to

(23) ρk (λ) ∼ (2π)−1/2 k−3/2 e−δk .

In the case λ < 1, summing over k we obtain a tail bound on the branching
process.
Lemma 9. Suppose that λ → 1 and A → ∞. Then
 
(24) P (A/δ ≤  X(λ) < ∞) ∼ (2π)−1/2 A−3/2 e−A δ 1/2 .

If, in addition, λ < 1, then

 
(25) ρ≥A/δ (λ) = P ( X(λ) ≥ A/δ ) ∼ (2π)−1/2 A−3/2 e−A δ 1/2 .

Proof. To prove (24) we simply sum (23) over k ≥ k0 = A/δ. Let  f (k)

denote the right-hand side of (23). Since δ → 0, we have f (x) ∼ f x ,
say, and it easily follows that
 
 
P (A/δ ≤  X(λ) < ∞) = f (k) ∼ f (x) dx.
k≥k0 x≥k0

The integral above can be evaluated exactly in terms of the normal distrib-
ution function, but there is no need: simply note that x−3/2 hardly changes
54 B. Bollobás and O. Riordan
 
over the relevant range, namely A/δ ≤ x ≤ A + O(1) /δ, so the integral
is asymptotically f (k0 )/δ. The estimate (24) follows.
 
To deduce (25), simply note that either  X(λ) is finite, or it is infinite,
an event of probability ρ(λ). If λ < 1, then ρ(λ) = 0.

To compare the graph and the branching process, we first need to com-
pare the relevant binomial and Poisson distributions. Given two real-valued
random variables X and Y , we say that X stochastically dominates Y , and
write X  Y , if P (X ≥ x) ≥ P (Y ≥ x) for every x. Of course, this is
really a statement about the distributions of X and Y . It is a basic fact
of probability theory that X  Y if and only if there are random variables
X  and Y  with the same distributions as X and Y respectively, such that
X  ≥ Y  always holds, i.e., X  Y if and only if X and Y may be coupled
so that X ≥ Y .
We shall need an approximate version of this concept: X stochastically
dominates Y up to η if we may couple (the distributions of) X and Y so
that P (X ≥ Y ) ≥ 1 − η. The next lemma is essentially trivial, although to
prove it takes a little calculation.

Lemma 10. For any 0 < p < 1, the binomial distribution Bi (n, p) is sto-
chastically dominated by the Poisson distribution with mean −n log (1 − p)
= np + O(np2 ).
For any 0 < a < 3/2, the distribution Bi (n, a/n) stochastically
 dom-
−100
inates the Poisson distribution with mean b up to o n , where b =
a(1 − log n/n).

Proof. For any p we may couple the (rather simple) binomial (or Bernoulli)
distribution Bi (1, p) with the Poisson distribution with mean − log (1 − p)
so that the latter dominates. To see this, note that if X and Y have these
distributions, then P (X = 0) = 1 − p = P (Y = 0), while if Y = 0 then
Y ≥ 1 ≥ X. As pointed out by Svante Janson, the first statement of the
lemma follows by taking n independent copies of this coupling.
The second statement requires a little calculation. Let X have the
binomial distribution Bi (n, a/n), and Y the Poisson distribution Po (b).
Note first that b ≤ a ≤ 3/2, so

 r  (3/2)r   3e r  
−b b
P (Y ≥ log n) = e ≤ ≤ = o n−100 .
r! r! 2r
r≥log n r≥log n r≥log n
Random Graphs and Branching Processes 55

Assuming, as we may, that n ≥ 100, say, we claim that for 2 ≤ r ≤ log n we

have
P (X = r) > P (Y = r).

Indeed, setting ξ = log n/n, so b = a(1 − ξ), and using the lower bound in
(17), we have
   
P (X = r) n a r a n−r b
= 1− e r!/br
P (Y = r) r n n
 2   
r a a2
> exp − − + (n − r) + b (a/b)r
2n n n2
 2 
r a2 ar
> exp − −a− + + b (1 − ξ)−r
2n n n
 2 
r
> exp − − a + b erξ
2n
 2 
r 3
> exp − − ξ + rξ
2n 2
 2 
r ξ
> exp − + ≥ 1.
2n 2

Summing over r in the range s ≤ r ≤ log n, it follows that for s ≥ 2 we have

(26) P (X ≥ s) ≥ P (s ≤ X ≤ log n)
 
≥ P (s ≤ Y ≤ log n) = P (Y ≥ s) + o n−100 ,

where the error term is uniform in s. It remains only to note that P (X =

0) = (1 − a/n)n ≤ e−a , while P (Y = 0) = e−b ≥ e−a . Thus P (X ≥ 1) ≥
P (Y ≥ 1). Together with (26) this establishes the claimed approximate
stochastic domination.

When we explore a component of Gn = G(n, p), at each step the number

of new vertices reached has a Bi (r, p) distribution, for some r ≤ n − 1 ≤ n.
By Lemma 10, this distribution is stochastically dominated by a Po (λ+ )
distribution, where λ+ = −n log (1 − p). Thus we may assume that the
56 B. Bollobás and O. Riordan

component exploration is dominated by the branching process X(λ+ ). In-

deed, we have shown that if v is a random vertex of G(n, p), then
   
P |Cv | ≥ k ≤ P ( X(λ+ ) ≥ k).

Let us now specialize to the values we are interested in, treating the
subcritical case first: from now on we take

p = λ/n with λ = 1 − ε,

where ε = ε(n) > 0 is a function of n satisfying ε → 0 and L = ε3 n → ∞.

In this case we have

(27) λ+ = −n log (1 − p) = np + O(np2 ) = 1 − ε + O(n−1 ).

Let δ = δ(λ) and δ + = δ(λ+ ) be given by (21). Since

dδ(x)
= 1 − x−1 = O(1 − x)
dx

for x near 1, we have

   
(28) δ + = δ + O(εn−1 ) = δ 1 + O(ε−1 n−1 ) = δ 1 + o(L−1 ) ,

where the last relation follows from εn/(ε3 n) → ∞.

We shall prove the upper bound claimed in Theorem 7 in the next lemma.

Lemma 11. Suppose that p = (1 − ε)/n with ε → 0 and L = ε3 n → ∞.

Then for any K = K(n) → ∞ we have
   
C1 G(n, p) ≤ δ −1 log L − (5/2) log log L + K

whp, where δ = δ(1 − ε) is defined by (21):

δ = −ε − log (1 − ε) = ε2 /2 + O(ε3 ).

Proof. Let Gn = G(n, p), and let C1 = C1 (Gn ). Let λ+ = 1 − ε + O(1/n)

be given by (27). Given a real number k ≥ 1, if v is a random vertex of Gn
then
 
ρ≥k (λ+ ) ≥ P |Cv | ≥ k ≥ (k/n) P (C1 ≥ k),
Random Graphs and Branching Processes 57

with the first inequality from our stochastic domination argument, and the
second from considering choices of v in the largest component. Turning this
around,

(29) P (C1 ≥ k) ≤ (n/k)ρ≥k (λ+ ).

 
Set k = δ −1 log L − (5/2) log log L + K . Using (28), we see that δ + /δ =
1 + o(L−1 ), so we may write k as A/δ + , where

(30) A = log L − (5/2) log log L + K 

with K  ∼ K → ∞. From (29) we have P (C1 ≥ k) ≤ (δ + n/A)ρ≥k (λ+ ).

Applying Lemma 9, it follows that
   
P (C1 ≥ k) = O n(δ + ) A−5/2 e−A = O ε3 nA−5/2 e−A ,
3/2

recalling that δ + ∼ δ ∼ ε2 /2. From (30), the final estimate is asymptotically

L(log L)−5/2 L−1 (log L)5/2 e−K ∼ e−K → 0, and the result follows.
 

We shall use the second moment method to prove a lower bound for
C1 (Gn ), i.e., to prove that a component of at least a certain size exists whp;
in fact, we shall consider only tree components. Set L = ε3 n, as before.

Proof of Theorem 7. Let

(31) A1 = log L and A0 = log L − (5/2) log log L − K,

where K = K(n) → ∞ but K ≤ (1/3) log log L, say, so A0 → ∞. Note that

−5/2 −A0
(32) LA0 e ∼ eK → ∞,

while
−5/2 −A1
(33) LA1 e = (log L)−5/2 → 0.

For i = 0, 1, let ki = Ai /δ, where δ = δ(1 − ε) is defined by (21), and let


S+ = kTk
k0 ≤k≤k1
58 B. Bollobás and O. Riordan

be the number of vertices of Gn = G(n, (1 − ε)/n) in tree components of

order between k0 and k1 . The required lower bound follows if we can show
that S+ ≥ 1 whp.

1
Note that E(S+ ) = kk=k 0
μn,k,λ . Now k1 = δ −1 log L = O(ε−2 log L), so
   
k13 /n2 = O ε−6 (log L)3 n−2 = O (log L)3 /L2 = o(1).

Furthermore, (λ − 1)k12 /n = εk12 /n ∼ εε−4 (logL)2 /n 2

 = (log L) /L → 0, so
the exponential correction term in (18) is exp o(1) throughout the range
of our sum. Hence, comparing with (19), we have μn,k,λ ∼ nρk (λ) for k ≤ k1 .
The error term in this approximation is uniform, so it follows that
 
E(S+ ) ∼ n P (k0 ≤  X(λ) ≤ k1 ).

Applying Lemma 9 twice, it follows that

 −3/2 −3/2 
(34) E(S+ )/k0 ∼ (2π)−1/2 A0 e−A0 − A1 e−A1 nδ 3/2 /A0
−5/2 −A0
∼ (2π)−1/2 A0 e nδ 3/2 → ∞,

recalling that nδ 3/2 = Θ(nε3 ) = Θ(L) and using (32) and (33).
Our next aim is to approximate the second moment of S + . We may
write E(S+
2 ) as M + M , where M is the expected number of pairs (v, w)
1 2 1
in the same tree component whose order is between k0 and k1 , and M2 is
the expected number of pairs in distinct such components. We start by
examining M2 .
Let μn,k,,λ denote the expected number of ordered pairs (v, w) of vertices
of G(n, λ/n) that are in different tree components with |Cv | = k and
|Cw | = , so writing Tk for the number of tree components of order k in
G(n, λ/n),
μn,k,,λ = E(kTk T )
if k = , and
 
μn,k,k,λ = E kTk k(Tk − 1) .
It is easy to give a formula for μn,k,,λ by arguing as for (14). Instead, let
us note that
μn,k,,λ (n − k)()
= (1 − λ/n)−k .
μn,k,λ μn,,λ n()
Random Graphs and Branching Processes 59

Indeed, having chosen the tree component of order k containing v, we must

choose the  vertices for that containing w from the remaining n−k vertices.
Counting non-edges as in the proof of (14), the final factor above accounts
for the double counting of the k potential edges between the components.
For any 0 ≤ k,  ≤ n/3, say, and any r ≤ , we have
n−k−r k k k
=1− =1− = 1 − + O(k/n2 ).
n−r n−r n − O() n
Taking logarithms and summing, it follows that
 
(n − k)() k  
log = − + O (k 2  + k2 )/n2 .
n() n
 
Since (1 − λ/n)−k = exp λk/n + O(k/n2 ) , it follows that
 
μn,k,,λ k
(35) = exp (λ − 1) + O(k /n + k /n ) .
2 2 2 2
μn,k,λ μn,,λ n

As noted earlier, k13 /n2 and (λ − 1)k12 /n are both o(1). Hence, for
k,  ≤ k1 we have

(36) μn,k,,λ ∼ μn,k,λ μn,,λ .

Summing over k and  in the relevant range, it follows that E(M2 ) ∼

E (S+ )2 .
Turning to pairs of vertices in the same tree component, we have


k1
M1 = kμn,k,λ ≤ k1 E(S+ ) ∼ k0 E(S+ ).
k=k0

Hence,

E(S+
2
) M1 + M2 M1 k0
2 = 2 = 1 + o(1) + 2 ≤ 1 + o(1) + E(S ) = 1 + o(1),
E (S+ ) E (S+ ) E (S+ ) +

using (34) for the last step. It follows by Chebyshev’s inequality that S+ is
concentrated about its mean. In particular,
 
P (C1 ≤ k0 ) ≤ P (S+ = 0) ≤ P ( S+ − E(S+ ) ≥ E(S+ )) = o(1).

Together with Lemma 11, this completes the proof of Theorem 7.

60 B. Bollobás and O. Riordan

Although our main aim here is simplicity, let us note that with very
little extra work one can obtain the distribution of the Op (1) term in (12),
in a form analogous to that obtained by Bollobás [29, 31] or L  uczak [143].
The key observation is that, defining A0 as in (31)  but with K fixed, the
branching process upper bound on E N≥k0 (Gn ) /k0 matches the lower
bound in (34) coming from trees, up to an additive error of o(1). This
shows that whp there are no non-tree components of order at least k0 . It
is easy to show that the distribution of the number of tree components of
order at least k0 is asymptotically Poisson with the appropriate mean, as
done by Bollobás [29, 31] and L  uczak [143]. Note that for this, it is more
natural to count large trees, rather than vertices in such trees; for the direct
comparison with the branching process, it seems clearer to count vertices
as above.

3.6. The supercritical case

The aim of this subsection is to prove Theorem 6. Throughout we set

p = λ/n, where λ = 1 + ε, and ε = ε(n) > 0 satisfies ε → 0 and
L = ε3 n → ∞; as before, we write Gn for G(n, λ/n).
We have done much of the work towards proving
 Theorem
 6 already.
Indeed, for any k we have C1 = C1 (Gn ) ≤ max k, N≥k (Gn ) . Hence,
     
E(C1 ) ≤ k + E N≥k (Gn ) = k + n P |Cv | ≥ k ≤ k + n P ( X(λ+ ) ≥ k),

where λ+ is given by (27). Since λ+ is very close to λ, choosing

 k suitably,
one can show in this way that E(C1 ) ≤ nρ(λ) 1 + O(L−1 ) . We omit the
details since we shall use a slightly different argument.
 
Proof of Theorem 6. Set λ = λ 1 − ρ(λ) . Since ρ(λ) = 1 − e−λρ(λ) , it
follows that
λ e−λ = λe−λ ,


i.e., that δ(λ ) = δ(λ), where δ is the function defined by (21).

With A1 = log L as before, let k1 = A1 /δ, and let

S= kTk
1≤k≤k1

be the number of vertices in ‘small trees’, i.e., in tree components of size

at most k1 . Our first aim is to estimate the mean and variance of S. We
Random Graphs and Branching Processes 61
 
shall show that E(S) is close to 1 − ρ(λ) n, and that its variance is small.
This will give us an upper bound on the number of vertices in any giant
component that holds whp, not just in expectation.
For k ≤ k1 , the quantity η inside the exponential correction term in
(18) is
   
η = O εk12 /n + k13 /n2 + k1 /n = O ε(log L)2 ε−4 n−1 + ε−6 (log L)3 n−2
 
= O (log L)2 /L + (log L)3 /L2 = o(1).

Unfortunately, we have to estimate E(S) with a rather smaller relative error.

Nevertheless, the weak estimate above is useful. Firstly, comparing (18) and
(19), for k ≤ k1 we have

(37) μn,k,λ = nρk (λ) exp (η) ∼ nρk (λ).

Also, since η = o(1) uniformly in the relevant range, we may write exp (η)
as 1 + O(η). Thus, for k ≤ k1 , we have
  
k2 k3 k
Δk = μn,k,λ − nρk (λ) = O nρk (λ) ε + 2+ .
2n n n

Using (22), we have ρk (λ) = O(k −3/2 e−δk ), so taking out a factor of εn, we
have
  
k 1/2 k 3/2 k −1/2
Δk
= O e−δk + + .
εn n εn2 εn

We shall use repeatedly the observation that for any constant a > −1,
∞

(38) ka e−δk = O(δ −(a+1) ) = O(ε−2(a+1) ),
k=1

which is most easily seen by noting that the exponential factor ensures that
the contribution to the sum from terms with k ≥ δ −1 = Θ(ε−2 ) is at most
a constant times the sum over k ≤ δ −1 .

1
Setting Δ = kk=1 Δk , it follows that
     
Δ ε−3 ε−5 ε−1 1 1 1 1
=O + 2+ =O + 2 +ε =O ,
εn n εn εn L L L L
62 B. Bollobás and O. Riordan

so
Δ = O(εn/L).
From (19) we see that ρk (λ) = (λ /λ)ρk (λ ). Hence,


k1 
k1
 
E(S) − Δ = n ρk (λ) = n(λ /λ) ρk (λ ) = n(λ /λ) 1 − ρ≥k1 +1 (λ ) .
k=1 k=1

Recalling that λ < 1, and that δ(λ ) = δ(λ), since k1 = (log L)/δ, from
Lemma 9 we have
 
ρ≥k1 (λ ) = P ( X(λ ) ≥ (log L)/δ ) ≤ L−1 δ 1/2 = O(ε/L).

It follows that

(39) E(S)/n = λ /λ + O(ε/L) = 1 − ρ(λ) + O(ε/L),

 
recalling that λ = λ 1 − ρ(λ) .
We can use this to get an upper bound on the expected size of the giant
component, but we would like a bound in probability. To obtain this we
consider the second moment of S, the number of vertices in ‘small’ trees.
As before, we may write E(S 2 ) as M1 + M2 , where M1 is the expected
number pairs (v, w) in the same tree component with at most k1 vertices,
and M2 is the expected number of pairs in distinct such components.
Now, using (37) and (22),


k1 
k1 
k1 ∞

E(M1 ) = kμn,k,λ ∼ nkρk (λ) ≤ nk −1/2 −δk
e ≤ nk −1/2 e−δk .
k=1 k=1 k=1 k=1

By (38), the final sum is O(n/ε).

We bound E(M2 ) as in the subcritical case: let


k1 
k1
Δ2 = E(M2 ) − E (S) =2
μn,k,,λ − μn,k,λ μn,,λ .
k=1 =1

The estimate (35) applies in this range, and from (36), the ratio in (35) is
o(1). Writing exp (η) as 1 + O(η) as before, we thus have


k1 
k1  
k k 2  k2
Δ2 = μn,k,λ μn,,λ O ε + 2 + 2 .
n n n
k=1 =1
Random Graphs and Branching Processes 63

Using (37) and (22) again, this gives


k1 
k1  
2 −3/2 −kδ −3/2 −δ k k 2  k2
Δ2 = n k e  e O ε + 2 + 2 .
n n n
k=1 =1

Applying (38) to the inner sum,


k1  
2 −3/2 −kδ kε−1 k2 ε−1 kε−3
Δ2 = n k e O ε + + 2 .
n n2 n
k=1

Applying (38) again,

 
2ε−1 ε−1 ε−3 ε−1 ε−1 ε−3
Δ2 = n O ε + +
n n2 n2
 
1 ε−4
2
=n O + 2 = O(ε−1 n + ε−4 ).
εn n

Since ε−4 = L−1 ε−1 n, the first term dominates, and Δ2 = O(n/ε).
Recalling that E(M1 ) = O(n/ε), we find that

Var (S) = E(M1 ) + Δ2 = O(n/ε).

In other words, the standard deviation of S is of order

 √
ω = n/ε = εn/ L.

(With a little more care, comparing the sums to integrals, we can find the
asymptotic variance of S.) Note that this is the same order as the error
term claimed in Theorem 6. Indeed, from (39), the variance bounds and
Chebyshev’s inequality, we have
 
(40) S = 1 − ρ(λ) n + Op (ω).

Subtracting from n gives the upper bound on C1 claimed in Theorem 6.

To prove the lower bound, we could try to count vertices in small non-
tree components, but this would get us into the kind of complications
in earlier work on this subject discussed in Subsection 3.1. Fortunately,
using a branching process argument, we can avoid these complications. We
64 B. Bollobás and O. Riordan

use a more careful version of the very simple sprinkling argument used in
Subsection 3.3.
In order to be able to apply the sprinkling method of Erdős and
Rényi [94], we must first show that we have the right number
√ of vertices in
‘large’ components, now meaning of size at least 5ω = 5εn/ L.
Set
√
λ = 1 + ε − 10ε/ L,
and
√
λ− = 1 + ε − 20ε/ L,
so λ− ≤ λ ≤ λ.
First, we consider B  , the set of vertices of Gn = G(n, λ /n) in compo-
nents of order at least 5ω. Let v be a random vertex of Gn . We explore its
component Cv as before, except that we stop if we reach at least 5ω ver-
tices. Up to this point, the number of new vertices found at each√stage has
a binomial distribution Bi (r, λ /n) where r ≥ n − 5ω = n − 5εn/ L. Since
 √  √ √
rλ /n ≥ (1 − 5ω/n)λ = 1 − 5ε/ L λ ≥ λ − 6ε/ L = λ− + 4ε/ L,
√
and ε/ L is much bigger than log n/n, by the second part of Lemma 10 we
may couple our exploration with X(λ− ) so that, up to a totally negligible
error probability, the exploration dominates the branching process until we
have reached 5ω vertices.
It follows that
     
E |B  | = n P |Cv | ≥ 5ω ≥ nρ(λ− ) − o n−99

= nρ(λ) − O(λ− n − λ n) = nρ(λ) − O(ω).

Let B be the set of vertices of Gn = G(n, λ/n) in components of order at

least 5ω. Coupling G(n, λ /n) and G(n, λ/n) so that the latter contains the
former, we have B  ⊂ B, so
   
(41) E |B| ≥ E |B  | ≥ nρ(λ) − O(ω).

With k1 = log L/δ ∼ 2ε−2 log L as before, we have

5ω 5εnL−1/2 5LL−1/2
∼ −2 = → ∞.
k1 2ε log L 2 log L
Random Graphs and Branching Processes 65

Hence, for n large enough, our ‘small’ (order ≤ k1 ) tree components are defi-
nitely smaller than our ‘large’ (order at least 5ω) components. In particular,
recalling that S is the number of vertices of Gn in small tree components,

S + |B  | ≤ S + |B| ≤ n.
 
From (39) and (41) we have E S +|B  | = n+O(εn/L)+O(ω) = n+O(ω).
It follows that n − S − |B  | has expectation O(ω). Since n − S − |B  | ≥ 0,
we thus have n − S − |B  | = Op (ω). Using (40), it follows that

(42) |B  | = ρ(λ)n + Op (ω).

Similarly, since n − S − |B| = Op (ω), we have

(43) |B| = ρ(λ)n + Op (ω).

Let K = K(n) tend to infinity; our aim is to show that whp C1 (Gn ) is
within Kω of ρ(λ)n. In doing so, we may of course assume that K tends to
infinity as slowly as we like.
We now ‘sprinkle’ edges as in Subsection 3.3. Indeed, we may form
Gn = G(n, λ/n) from Gn = G(n, λ /n) by adding each non-edge with a
probability r given by (1 − λ /n)(1 − r) = (1 − λ/n). Note that
√
r ≥ (λ − λ )/n = 10εn−1 / L = 10ω/n2 .

We condition on Gn = G(n, λ /n), assuming that

(44) ρ(λ)n − Kω ≤ |B  | ≤ ρ(λ)n + Kω,

which holds whp by (42). Let Z1 ,√ . . . , Zs list all ‘large’ components of Gn ,
i.e., those with at least 5ω = 5εn/ L vertices.
The calculation we are about to present is simplest if the sizes of the Zi
are roughly equal. Fortunately, we may artificially achieve this situation: we
simply divide any Zi with more than 10ω vertices into parts with between 5ω
and 10ω vertices. In this way we obtain disjoint sets Z1 , . . . , Zs  of vertices
of Gn such that 5ω ≤ |Zi | ≤ 10ω for every i, with each Zi contained in some
Zj . Note that we do not assume that each Zi induces a connected subgraph
of Gn : all that matters is that all vertices of Zi are in the same component
of Gn .
66 B. Bollobás and O. Riordan

If K tends to infinity slowly enough, then the upper√bound in (44) is

at most 5εn, say, and it follows that s ≤ (5εn)/(5ω) = L.√ On the other
hand, the lower bound is at least εn, so s ≥ (εn)/(10ω) = L/10.
Given a pair {Zi , Zj }, i = j, the probability that there is at least one
sprinkled edge from Zi to Zj is

 
1 − (1 − r)|Zi | |Zj | ≥ 1 − (1 − r)25ω .
2

Since rω 2 = Θ(ω 3 /n2 ) which, as we shall now see, is o(1), the bound above
is asymptotically
√
25ω2 r ≥ 250ω 3 /n2 = 250ε3 n3 L−3/2 /n2 = 250/ L.

It follows easily that√ the probability that at least one such edge is present is
at least p1 = 240/ L, say. The presence of sprinkled edges between different
pairs {Zi , Zj } is independent. Let H be the graph on [s ] in which each
edge is present independently with probability p1 . We√have √shown that H
stochastically dominates G(s , p1 ). Since s p1 ≥ 250/ L L/10 = 25,

it follows that H stochastically dominates G(s , 25/s ). 
√
Since s ≥ L/10 → ∞, it follows from the very weak Theorem 3 that
whp at least 99% of the vertices of H are in a single component. More
crudely, simply by counting the expected number of partitions of the vertex
set into classes of size s /20 and 19s /20 with no
 edge from one class to
the other, it is trivial to show that C1 G(s , 25/s ) ≥ 19s /20 whp.
Hence, whp Gn = G(n, λ/n) has a single component containing all but
at most s /20 of the sets Zi , and hence all but at most
√
(s /20)(10ω) ≤ Lω/2 = εn/2
 
of the vertices in B  = i V (Zi ) = i V (Zi ). Using (44) again, we see that
whp Gn = G(n, λ/n) contains a component of order at least 1.4εn, say.
This bound is much weaker than the one we claimed, but nevertheless,
we are essentially home! Roughly speaking, deleting such a large component
from Gn will leave a subcritical graph, and by Theorem 7, whp this will
contain no components of size at least 5ω. In this case, C1 = |B|, and (43)
gives the required result.
To make this precise, suppose for a contradiction that with positive
probability the second largest component of Gn has size at least 5ω. As
shown above, whp there is a component of size at least 1.4εn, so with
Random Graphs and Branching Processes 67

positive probability there is a component X1 with |X1 | ≥ 1.4εn and a

different component X2 with |X2 | ≥ 5ω. Suppose that Gn has this form,
and we repeatedly pick a random vertex of Gn not yet uncovered, and then
uncover its component in Gn . We are likely to hit X1 before X2 , so in
this process, with positive probability we find a component of order at least
1.4εn, and then later find one of order at least 5ω.
This is impossible: stopping after we first uncover a component of order
at least 1.4εn, the distribution of the graph Gn on the unseen vertices
is exactly that of G(n , λ/n), where n ≤ (1 − 1.4ε)n is the number of
vertices not yet uncovered. Adding vertices
 if necessary, we may assume
 
that n ∼ (1 − 1.4ε)n. Since n λ/n ≤ 1 − 0.4ε + O(ε2 ) , we may apply
 −2
Theorem 7 to√ deduce that
−2
√ whp C1 (Gn ) ≤ 20ε log L. This is much smaller
than ω = εn/ L = ε L.

Once we know that whp there is a unique giant component, it follows

that, writing G− n for the graph formed from Gn = G(n, λ/n) by deleting
this component, conditional on n = n − C1 (Gn ) = |G− n |, the distribution
of G−n is essentially that of G(n , λ/n) = G n , (λn /n)/n ; these random

graphs may be coupled to agree whp. (The details of this argument are
spelled out in a more general context in [41, page 79].) In studying the
second (and third etc) largest components  in Gn , one can condition
√  on n ;
by Theorem 6 we may assume that n = 1 − ρ(λ) n + O εn/ L , which
 √ 
gives (λn /n) = λ + O ε/ L .

 Recalling
√  that δ(λ ) = δ(λ), since the relative correction to ε is
O 1/ L = o(1/ log L), the formula in Theorem 7 applies as is, to give
the following result.

Theorem 12. Let λ = 1 + ε, where ε = ε(n) > 0 satisfies ε → 0 and

L = ε3 n → ∞. Then
 
  −1 5 log log L
C2 G(n, λ/n) = δ log L − + Op (1) ,
2

where
δ = λ − 1 − log λ = ε2 /2 − ε3 /3 + O(ε4 ).

Just as for Theorem 7, this is consistent with, and extends the range
of, the results of Bollobás [29, 31]; the formula given by L
 uczak differs in
−1 −2
that δ is replaced by 2ε , which means his formula is wrong if ε log n is
68 B. Bollobás and O. Riordan

unbounded. Of course, just as for Theorem 7, one can obtain the distribu-
tion
 of the Op (1) term by considering trees, and indeed the distribution of
C3 G(n, p) etc.

4. The Phase Transition in Inhomogeneous Random

Graphs

The ‘phase transition’ in G(n, c/n) as c is varied, i.e., the radical difference in
structure for c < 1 constant and c > 1 constant, is one of the most important
properties of G(n, p). As we have seen, for G(n, p), starting with the paper
of Bollobás [29], the focus has moved to much more detailed questions, in
particular what happens if c is not constant but depends on n, with c → 1.
For more complicated graph models, the first question to ask is still whether
there is a phase transition, and where, i.e., for which parameters is there a
‘giant’ component containing Θ(n) vertices? As the model gets more and
more complicated, one is likely to have to content oneself with less and less
precise answers. Here we shall only consider results generalizing Theorem 3.
While this is a very weak result about G(n, p), the corresponding results for
other models can be rather difficult.

4.1. Graphs with a given degree sequence

For random graphs with a given degree sequence, i.e., for the configuration
model of Bollobás [25], in 1995, Molloy and Reed [151] gave a simple
necessary and sufficient condition for the existence of a giant component
whp. They showed that, under mild assumptions, there is a giant component
if and only if

d(d − 2)λd > 0,
d

where λd n is the (asymptotic) number of vertices with degree d. In [152] they

analyzed the size of the giant component when it exists: again under mild
assumptions, the asymptotic fraction of vertices in the giant component is

  d/2
2α
(45) σ =1− λd 1− ,
d≥1
d
Random Graphs and Branching Processes 69

where d = d dλd is the (asymptotic) average degree, and α is the smallest

positive solution to

  d/2
2α
(46) d − 2α − dλd 1− = 0.
d≥1
d

Although they did not take this point of view, σ is exactly the survival
probability of a certain branching processes associated to the exploration of
the neighbourhood of a random vertex; we return to this in Subsection 4.5.
Kang and Seierstad [130] initiated the study of the window of the phase
transition in the configuration model, and determined its width up to a
log factor. Using the properties of empirical distributions of independent
random variables, Janson and Luczak [126] gave a simpler proof of this
result, and removed the log factor.
One can also study percolation on a random graph G(n, d) with a given
degree sequence, first forming the random graph, and then deleting vertices
and or edges at random; the resulting graph can be expressed as G(n, d ),
with d an appropriate random sequence. This approach was used by
Fountoulakis [109] and Janson [122], for example.
Aiello, Chung and Lu [2] considered random graphs with a given degree
sequence, where the degree sequence follows a power law: there are (essen-
tially) A/kβ  vertices of each degree k ≥ 1, where β is a parameter and A
is a suitable normalizing factor. Unlike Molloy and Reed, they considered
a random multi-graph with this degree sequence; this is much simpler to
study. Writing β0 = 3.47875 . . . for the solution to ζ(β − 2) − 2ζ(β − 1) = 0,
they showed that for β < β0 there is a giant component, while for β > β0
there is not. These results fit with those of Molloy and Reed [151] for the
more complicated simple graph model, although (due to the restrictions on
the degree sequence there) the latter apply directly only if β > 8.
In the rest of this section we turn to analogues of this result, or of The-
orem 3, for the various new inhomogeneous models described in Section 2.
We shall briefly state some of the known results, and describe the connec-
tion to branching processes. Before doing so, let us explain one possible
motivation for studying this topic.
70 B. Bollobás and O. Riordan

4.2. Robustness of the BA or LCD model

One of the very first questions asked about the Barabási–Albert model was:
how robust are the graphs it generates under random failures? More con-
cretely, if we delete vertices or edges (or both), keeping each with probability
p independently of the others, to what extent does the graph remain con-
nected? In particular, how large is the largest component in what remains?
If one views these graphs as models of, for example, a communication net-
work, then these are clearly very important questions.
The cleanest mathematical formulation of the questions above is as
follows: given a graph G and a real number 0 < p < 1, let G(p) denote the
random subgraph of G obtained by deleting edges independently of each
other (and, if G is random, independently of G), keeping each edge with
probability p. Similarly, let G[p] be obtained by deleting vertices, keeping
(m)
each with probability p. Let Gn denote the n-vertex LCD graph in which
m edges are added per new vertex. (This was not the original notation
in [49], where we introduced this model; there we wrote Gnm ; the notation
(m)
Gn now seems more natural, and fits better with other notation here. In
any case, n is always the number of vertices, and m the number of edges
added per vertex.) With m ≥ 1 and 0 < p < 1 fixed, it is not hard to
(m)
convince oneself that the largest component of Gn [p], say, will contain
λm (p) + op (1) n vertices, for some constant λm (p). The key question is:
when is λm (p) non-zero? Also, when λm (p) is non-zero, how large is it?
This question, stated in a rather less precise form, was raised by Albert,
(m)
Jeong and Barabási [5], who studied Gn [p] (or rather the equivalent for
the imprecisely defined BA model) using simulations. They also asked a
(m)
related question about the vulnerability of Gn to malicious attack, i.e.,
the size of the largest component that remains after deleting pn vertices,
minimized over all choices of the vertices to delete. Heuristic answers to
these questions were given by Cohen, Erez, ben-Avraham and Havlin [73,
74], and Callaway, Newman, Strogatz and Watts [64] (see also [3]). In
contrast to the suggestions of [5], these heuristics suggested that there is
always a giant component.
Following these heuristic results, we proved in [48] that there is indeed
always a giant component, but that for small p, this giant component is
extremely small: for m ≥ 2 and 0 < p < 1 fixed,
 
C1 (G(m)
n [p]) = λm (p) + op (1) n
Random Graphs and Branching Processes 71

and
 
C1 (G(m)
n (p)) = λm (p)/p + op (1) n,

where the function λm (p) is positive for all m ≥ 2 and p > 0, and satisfies
   
(47) exp − Θ(p−2 ) ≤ λm (p) ≤ exp − Θ(p−1 )

as p → 0 with m fixed. The case m = 1 behaves completely differently: in

(m)
this case Gn is essentially a tree, and it turns out that λm (p) = 0 for all
p < 1; this radically different behaviour is not predicted by the heuristics
mentioned above.
Although this point of view was not explicitly expressed in [48], with
hindsight the arguments in [48] show that λm (p) is exactly the survival
probability of a certain (rather complicated) branching process. However,
this is far from the end of the story. Indeed, it was shown in [48] that λm (p)
may be written as
 1   m 
1 − L(α)
λm (p) = p 1−   m dα,
α=0 1 + R(α)

where L and R are the maximum solution to the simultaneous integral

equations
 α   m 
p 1 1 − L(β)
(48) L(α) = √ √ 1−   m+1 dβ.
2 α β=0 β 1 + R(β)

and
    m−1 
p 1
1 1 − L(β)
(49) R(α) = √ √ 1−  m dβ.
2 α β=α β 1 + R(β)

The bounds (47) were obtained by bounding the solutions to these

equations above and below. Later, Riordan [166] proved much better bounds
on the solutions to (48) and (49), obtaining the following result.
(m)
Theorem 13. Let m ≥ 2 and 0 < p < 1 be fixed, and let Gn [p]
(m)
be the graph formed from Gn by keeping vertices independently with
probability p. There is a constant λm (p) such that
 
C1 (G(m)
n [p]) = λm (p) + op (1) n
72 B. Bollobás and O. Riordan

as n → ∞. Furthermore, as p → 0 with m fixed we have

  1  1
m−1 2p m−1 2p
Ω p2 = λm (p) = O .
m+1 m+1

In particular, as p → 0 with m fixed,

   
cm   cm + o(1)
λm (p) = exp − + O log (1/p) = exp − ,
p p

where
 
1 m+1 1 1  
cm = log = + + O m−5 .
2 m−1 m 3m 3

In [48], corresponding results were proved concerning (a modified form

of) the vulnerability question of Albert, Jeong and Barabási [5]. Omitting
the details, the overall conclusion is that the LCD model is much more
robust than a homogeneous (Erdős–Rényi) random graph with the same
number of vertices and edges, but at the same time somewhat more vulner-
able to attack.
It is natural to ask whether this difference between the BA or LCD model
and G(n, p) is due to the growing nature of the model, or to preferential
attachment. To address this, we turn next to the growing m-out model.

4.3. The uniformly grown models and Dubins’ model

(m)
Let Hn denote the n-vertex graph obtained by starting with m vertices
and no edges, adding vertices one by one, and joining each new vertex to a
set of m earlier vertices chosen uniformly at random, and independently of
the history. Once again, this is not our original notation from [50], where
(n)
we wrote Hm .
It was suggested already in [48] that for m ≥ 2, the critical probability
(m) (m)
for the emergence of a giant component in Hn (p) or Hn [p] should be
given by
  
1 m−1 1 1  −3 
(50) pc (m) = 1− = + + O m .
2 m 4m 16m2
Random Graphs and Branching Processes 73

Before saying anything more about this, let us turn our attention to a closely
related, but much simpler model.
(m)
As noted in Section 2, for i < j the probability that ij is an edge of Hn
(m)
is essentially m/j, so the ‘mean-field’ version of Hn is the c = m case of
1/j
the uniformly grown random graph Gn (c), the graph on [n] in which edges
are present independently, and, for 1 ≤ i < j ≤ n, the probability that ij is
present is min {c/j, 1}.
1/j
If we start with Gn (c) and select edges with probability p, then, apart
1/j
from a small correction when j ≤ c, we obtain exactly Gn (pc). Thus,
in this case, just as for G(n, c/n), rather than starting with a given graph
and deleting edges with some probability, it makes sense simply to vary the
parameters of the model. Then, just as Erdős and Rényi did for G(n, c/n),
one studies the phase transition, asking for which parameters there is a giant
component, and how large it is. We shall take this point of view for most
of the models we consider here; the reader more interested in robustness
type questions should note that that is exactly what we are considering.
(While this equivalence applies exactly only for edge deletion, it turns out
that, almost always, the analysis for edge deletion and for vertex deletion
is essentially identical.)
In 2001, Callaway, Hopcroft, Kleinberg, Newman and Strogatz [63]
raised the question of determining the critical point of the phase transi-
1/j
tion in a model essentially equivalent to Gn (c). For a discussion of the
equivalence, see Durrett [89] or Bollobás, Janson and Riordan [40]; using
this equivalence, we shall state all results about either model in terms of
1/j
Gn (c).
Callaway, Hopcroft, Kleinberg, Newman and Strogatz [63] derived a
generating function for the limiting fraction of vertices in components of
each fixed small size in their model, and presented evidence that the critical
point is at c = 1/4 (δ = 1/8 in their formulation). Also, assuming for the
moment the existence of such a function, let f (ε) denote the asymptotic
fraction of vertices in the giant component when c = 1/4 + ε, ε > 0, so
   
C1 G1/j
n (1/4 + ε) = f (ε) + op (1) n.

Numerically integrating the differential equation for the generating function,

Callaway, Hopcroft, Kleinberg, Newman and Strogatz [63] suggested that
f (ε) ∼ exp (αε−β ), with β = 0.499 ± 0.001.
74 B. Bollobás and O. Riordan

In 1984, Dubins introduced an infinite random graph GD (c) that is ex-

1/j
actly the natural infinite version of the much later model Gn (c): in Dubins’
model, the vertex set is {1, 2, . . .}, different edges are present independently,
and the probability that ij is present is c/j for i < j. Dubins asked whether
this graph is almost surely (a.s.) connected when c = 1 (see [129, 176]).
Kalikow and Weiss [129] showed that for c < 1/4, the graph GD (c) is a.s.
disconnected, and that for c > 1 it is a.s. connected. Shepp [176] established
the critical point, showing that GD (c) is a.s. connected for c > 1/4, and also
that it is a.s. disconnected at c = 1/4.
Although this may seem unrelated to the CHKNS question, there is in
fact a very close connection: as pointed out independently by Bollobás,
Janson and Riordan [40] and Durrett [89], the results above immediately
imply that the critical value c0 for the uniformly grown graph is 1/4,
corresponding to δ = 1/8 for the CHKNS model.
Turning to the size of the giant component, analyzing the generating
function equation using more mathematical, but still not rigorous, methods,
Dorogovtsev, Mendes and Samukhin [88] obtained
 
π 1
f (ε) = 2c exp − √ ,
2 ε

with c = 0.295 . . . . (We have translated this from their CHKNS formula-
tion.) Presumably, they did not mean to assert exact equality! Durrett [89]
made a start on making their methods rigorous, in particular obtaining the
critical probability
 rigorously,
 and giving a rigorous lower bound on f (ε) of
the form exp − Θ(1/ε) .
Using path counting methods, Bollobás, Janson and Riordan [40] proved
that
 √ 
f (ε) = exp ( − Θ 1/ ε ).

In particular, for any k one has f (ε) = o(εk ) as ε → 0 from above: the
phase transition has infinite order, as suggested by the non-rigorous results
of [63] and [88].
Using branching processes, Riordan [166] gave much tighter bounds on
f (ε), finding the constant in the exponent. More precisely, he showed that
 
π 1  
(51) f (ε) = exp − √ + O log (1/ε) .
2 ε
Random Graphs and Branching Processes 75
(m)
Returning to Hn , building on the results above, the critical probability
was determined in [50], where it was proved that (50) is indeed correct.

Theorem 14. Let m ≥ 2 and 0 < p < 1 be fixed, and set

  
1 m−1 1 1  
pc (m) = 1− = + 2
+ O m−3 .
2 m 4m 16m

If p ≤ pc (m), then C1 (Hn (p)) = o(n) holds whp as n → ∞. If p =

(m)

pc (m) + ε, ε > 0, then

C1 (Hn(m) (p)) = fm (ε)n + o(n)

and C2 (Hn (p)) = o(n) hold whp as n → ∞ with ε fixed, where fm (ε) is
(m)

positive and satisfies

 √ 
fm (ε) = exp ( − Θ 1/ ε )

as ε → 0 with m fixed.

(m) 1/j
This result shows that Hn (p) and its ‘mean-field’ version Gn (pm)
are essentially different: with m constant, the phase transition is shifted
by a constant amount. In other words, the dependence between edges in
(m)
Hn (p) is enough to shift the critical point of phase transition. Despite
(m)
this, Riordan [166] managed to obtain bounds for Hn (p) corresponding to
(51), showing that
 
π 1  
f (ε) = exp −   1/4 √ + O log (1/ε) .
2 m(m − 1) ε

As mentioned in Subsection 2.4, there is a natural common generaliza-

1/j (m)
tion of Gn (c) and Hn , the uniformly grown Z-out random graph, intro-
duced in full generality in [166]. (A restricted version that does not include
1/j
Gn (c) was studied in [50].) Here Z is a distribution on the non-negative
integers satisfying some mild conditions, and for each vertex added, one
chooses a number Zi of edges to add, where the Zi are independent and
have (essentially – small deviations are allowed) the distribution Z. With
76 B. Bollobás and O. Riordan

Z Poisson with mean c, one can recover (exactly, using the flexibility of the
1/j (m)
full definition) the model Gn (c). With Z = m constant, one obtains Hn .
Riordan [166] proved a common generalization
 of the results above:
writing c = E(Z) and r = E Z(Z − 1) / E (Z)2 , and setting
√
1 1− r
c0 =
2 1−r

for r = 1, and c0 = 1/4 if r = 1 (as in the Poisson case), he showed that

under mild assumptions, there is a giant component if and only if c > c0 ,
and that its normalized size is
 
π 1  
exp − 1/4 √ + O log (1/ε) .
2r ε

if c = c0 + ε with ε → 0 from above and r fixed. For full details see [166,
Theorem 8.1].

4.4. Graphs with independence between edges

In the previous subsection, we considered one very important graph with

1/j
independence between edges, namely the finite version Gn (c) of Dubins’
model.
Another natural model with independence between edges is the mean-
field version of the Barabási–Albert or LCD model, i.e., the graph in which
√
the edge ij has probability c/ ij. In [166] it was shown that (like random
subgraphs of the m ≥ 2 LCD model), this has a giant component for any
c > 0. More importantly, its normalized size σ(c) is the survival probability
of a certain multi-type branching process, and satisfies
 
σ(c) ∼ 2e1−γ exp − 1/(2c)

as c → 0, where γ is Euler’s constant.

The phase transition in ‘random graphs with given expected degree se-
quences’ was studied by Chung and Lu [68, 70] and Norros and Reittu [155].
These models both have (conditional) independence between edges; as we
are not attempting to be comprehensive, we omit the precise details of these
models and results.
Random Graphs and Branching Processes 77

Turning to Turova’s ‘dynamical random graph’ described in Subsec-

tion 2.4, Turova [187, Corollary 4.1] and Söderberg [179] obtained (without
giving full technical details) a formula for the critical point for the phase
transition, which may be written in the form λc = γ(μ/γ)2 zγ/μ−1
2
/8, where
zν denotes the first positive zero of the Bessel function Jν ; see Bollobás,
Janson and Riordan [41, Section 16.5].

As noted earlier, the very general BJR model G(n, κ) of [41] includes
exactly the various sparse models with independence mentioned above, as
well as many others. In [41], the size of the giant component of G(n, κ)
is found, in terms of the survival probability ρ(κ) of a certain branching
process X(κ). Our next aim is to describe this result, simplifying somewhat
by considering a special case. We shall then comment on the relationship
of this result to the results for specific models mentioned above.

Simplifying the set-up in [41] somewhat, let κ be a kernel, i.e., a sym-

metric measurable function from [0, 1]2 to the non-negative reals. For the
moment, we assume in addition only that κ is integrable. Let X(κ) be the
multi-type Galton–Watson process defined as follows: we start with a sin-
gle particle in generation 0, whose type is chosen uniformly and randomly
from [0, 1]. Given generation t, each particle P in this generation gives rise
to children in the next generation whose types form a Poisson process on
[0, 1] with intensity measure κ(x, y) mathrmdy, where x is the type of P .
In other words, the number
 of particles with types in a measurable set A
is Poisson with mean y∈A κ(x, y) dy, and these numbers are independent
for disjoint sets A. The children of different particles are independent, and
independent of the history.

In the simplest special case, κ(x, y) = c is constant. In this case, or,

1
more generally, if 0 κ(x, y) dy = c for (almost) all x, then the number of
children of a particle of any type is Poisson with mean c, so the types are
irrelevant and X(κ) reduces to the single-type process X(c) considered in
Section 3.

The main theme of the results of Bollobás, Janson and Riordan [41] is
that X(κ) has the same relationship to G(n, κ) as X(c) does to G(n, c/n).
Not only does X(κ) provide a good approximation to the neighbourhood
exploration process in G(n, κ) (an unsurprising but also non-trivial fact),
but the normalized size of the giant component (if any) is given, under
suitable assumptions, by the survival probability ρ(κ) of X(κ). To state a
precise result, we need some further definitions.
78 B. Bollobás and O. Riordan

Let Tκ be the integral operator associated to the kernel κ, defined by

 1
(52) (Tκ f )(x) = κ(x, y)f (y) dy,
0

for any (measurable) function f such that this integral is defined (finite or
+∞) for almost every x; since κ is non-negative, Tκ f is defined for any
f ≥ 0.
Let
 
Tκ  = sup Tκ f 2 : f ≥ 0, f 2 ≤ 1 ,
noting that we may  have Tκ  = ∞; when finite, Tκ  is the norm of Tκ as
an operator on L2 [0, 1] .
We say that a kernel κ is reducible if there is a measurable set A ⊂ [0, 1]
with 0 < μ(A) < 1 such that κ = 0 almost everywhere on A × Ac , where
Ac = [0, 1] \ A; otherwise, κ is irreducible.
A special case of the main result of [41] is as follows; here we restrict
the way the ‘types’ of the vertices are chosen for simplicity.

Theorem 15. Let κ be a symmetric measurable function from [0, 1]2 to the
non-negative reals, with κ < ∞. Let Gn = G(n, κ) be the graph ob-
tained by first choosing x1 , . . . , xn independently and uniformly from [0, 1],
and then, given x1 , . . . , xn , selecting each possible edge ij with probability
 
pij = min 1, κ(xi , xj )/n ,

independently of the other edges.

(i) If Tκ  ≤ 1, then C1 (Gn ) = op (n), while if Tκ  > 1, then C1 (Gn ) =
Θ(n) whp.
(ii) For any ε > 0, whp we have

1
C1 (Gn ) ≤ ρ(κ) + ε.
n

(iii) If κ is irreducible, then

 
C1 (Gn ) = ρ(κ) + op (1) n.

For the fully general result, which involves several further definitions,
see [41, Theorem 3.1]. (In fact, there are some additional restrictions on κ
Random Graphs and Branching Processes 79

in [41], needed when the xi can be chosen in a more general way. As noted
in [44], these restrictions are not necessary when the xi are iid as above.)
The key points of the result above are: under suitable assumptions,
C1 (Gn ) is asymptotically ρ(κ)n, generalizing Theorem 3. Furthermore,
there is a giant component if and only if Tκ  > 1. Finally, as shown in [41,
1
Theorem 6.2], ρ(κ) may be written as 0 ϕ(x) dx, where ϕ is the maximum
solution ϕ : [0, 1] → [0, 1] to a certain non-linear equation, namely,
 
(53) ϕ(x) = 1 − exp − (Tκ ϕ)(x)

for every x. This generalizes the corresponding equation for a single-type

process, namely (5).
Although for lack of space we cannot go into details, let us mention
that one of the main ideas of the proofs in [41] is to reduce whenever
possible to the much simpler finite-type case. To perform this reduction,
frequent use is made of monotonicity and convergence results about (infinite-
type) branching processes. In turn, key to the proof of these is showing
that when κ is irreducible, there is at most one non-zero solution to (53).
This is not surprising, but not so easy to prove in general. Indeed, the
analogous statement for a general (non-Poisson) branching process is false!
See Bollobás, Janson and Riordan [44] for a brief discussion of this.
In one sense, the results of Bollobás, Janson and Riordan [41] generalize
many of the results mentioned earlier; however, in another sense, this is
far from the case. Without further work, Theorem 15 gives the critical
point and giant component size in a rather cryptic form. In fact, it is often
not too hard to determine Tκ , and thus whether or not there is a giant
component; see [41, Section 16] for a discussion of this in various cases. On
the other hand, obtaining precise results such as (51) from (53) requires a
great deal of further work. Nevertheless, Theorem 15 shows that no further
combinatorics is needed. It also makes very clearly the point that multi-
type branching processes are fundamental to the understanding of (sparse)
inhomogeneous random graphs.

4.5. Applications of non-Poisson branching processes

The single-type Poisson Galton–Watson process X(λ) used in Section 3 is

in some sense the canonical branching process. In the same way, X(κ) is
perhaps the canonical multi-type branching process. These processes arise
80 B. Bollobás and O. Riordan

more or less automatically in settings with independence between edges.

As we shall now see, these are not the only situations in which branching
processes can be used to study random graphs.
In the single-type setting, we have alluded to one non-Poisson example
already. Indeed, let us consider the configuration model of Bollobás in the
setting where, for each d, we have approximately λd n vertices of degree d
(we shall not state a precise result). When following a random edge in the
graph, the destination vertex will be chosen randomly but with probability
proportional to its degree. It follows that the neighbourhood exploration
process in the graph may be initially coupled with a single-type Galton–
Watson branching process in which

the probability that a particle has d − 1
children is dλd /d, where d = d dλd is the average degree. Corresponding
to starting at a random vertex, one must start with a random number N0
of particles (one for each edge), with P (N0 = d) = λd .
This process is supercritical if and only if

  
dλd
(d − 1) > 1,
d
d

i.e., if and only if d d(d − 2)λd > 0; furthermore, the survival probability
in this case is given by (45). Indeed, writing x for the probability that a
single particle survives, one has

  
dλd
1−x= (1 − x)d−1 ,
d
d

so, comparing with (46), (1 − x)2 is the quantity 1 − 2α/d considered by

Molloy and Reed [152]. The overall survival probability is

1− (1 − x)d λd ,
d

which coincides with (45). Although they did not take this viewpoint,
branching processes can thus be used to give an alternative proof of the
results of Molloy and Reed [151, 152].
The results of [48, 50, 166] concerning the phase transition in the LCD
model were also proved using (this time, multi-type) branching process
methods; for details, see [50] or [166]. As suggested by the equations (48),
Random Graphs and Branching Processes 81

(49) for the survival probability, this branching process is rather compli-
cated. For the LCD model, to get to the branching process formulation re-
quires transforming the definition of the model first to a static form (using
linearized chord diagrams), and then to one with conditional independence
between edges (using pairings of random points). These transformations
were discovered in [49], where the LCD model was introduced; see also [47].
A rather different application of non-Poisson branching processes was
given by Bollobás, Janson and Riordan [43], analyzing the ‘line-of-sight’
percolation model of Frieze, Kleinberg, Ravi and Debany [111]. In this, one
studies the graph Gn,ω,p formed from an n-by-n grid or torus by selecting
vertices independently with probability p, and joining two selected vertices
if they lie on a common horizontal or vertical line and are within distance
ω. For fixed ω, determining the critical value pc = pc (ω) of p above which a
giant component emerges is a hopeless task: already the case ω = 1 amounts
to finding the critical probability for site percolation on the square lattice,
a well-known open problem. (Indeed, it may well be that there is no simple
expression for the answer.) For this reason, Frieze, Kleinberg, Ravi and
Debany [111] asked for the limiting behaviour of ωpc (ω) as ω → ∞, and
proved, essentially, that the limit is between 1/(4e) and 1. Bollobás, Janson
and Riordan [43] settled this question, proving that

ωpc (ω) → log (3/2) = 0.4054 . . . ,

among other results. This limit arises as the critical point of a certain
branching process, associated to a non-traditional exploration of the com-
ponent of a vertex of Gn,ω,p . The proof involves studying a branching ran-
dom walk, where each particle of a branching process also has a position in
R2 , corresponding to the re-scaled location of the corresponding vertex in
the grid.
Turning briefly to another dependent context, namely the clique perco-
lation model of Derényi, Palla and Vicsek [83] mentioned in Subsection 2.7,
the critical point of this model was found rigorously in [52]. This time (un-
like in the example above), it is clear what the relevant branching process
should be: optimistically, when we look for new Kk s sharing a particular set
of  vertices with a Kk we have already discovered, we hope to find essen-
tially a Poisson number of such new Kk s, withacertain mean λ. Each new
Kk gives rise, in the simplest case, to exactly k − 1 new sets of  vertices
to explore from in the next step. Bollobás and Riordan [52] showed that the
critical point does indeed correspond to the point at which the correspond-
ing compound Poisson branching process has positive survival probability.
82 B. Bollobás and O. Riordan

The difficulty in this case is carrying out the local coupling: the problem
is that the distribution of edges in the graph on the Kk s is very far from
independent. Nevertheless, this problem can be overcome with careful use
of correlation arguments, exploring in a special way that separates positive
and negative correlation, and allows the positive correlation to be bounded.
See [52] for the details, which are rather involved.
We close this section by describing a perhaps more canonical example
of the use of non-Poisson branching processes to study a sparse random
graph with dependence between edges, namely the inhomogeneous cluster-
ing model, or inhomogeneous hypergraph model, of Bollobás, Janson and Ri-
ordan [44]. Recall that this graph is defined by starting with a kernel family
(κF ) consisting of one kernel for each isomorphism class of connected fi-
nite graphs; the kernel associated to a graph F with r vertices is a function
from [0, 1]r to the non-negative reals. Choosing x1 , . . . , xn independently
and uniformly from [0, 1], for each F and each r-tuple v1 , . . . , vr , one adds
a copy of F with the ith vertex mapped to vi with probability

κF (xv1 , . . . , xvr )
,
nr−1
 
all choices being independent. Note that the resulting graph G n, (κF )
certainly does not have independence between  edges. In particular, taking
κF non-zero only for F a triangle, say, G n, (κF ) is a union of Θ(n)
triangles, and almost all (in the multi-graph version, all) degrees are even.
For the questions we consider here, namely the existence and size of the
giant component, since adding any connected graph F on v1 , . . . , vr unites
the components containing these vertices, the detailed structure of F does
not matter, and one may as well replace F by Kr . In this way the set-up
simplifies slightly: one need only consider a hyperkernel κ = (κr ), consisting

of one symmetric measurable function κr on [0, 1]r for each r ≥ 2. We write
G(n, κ) for the random graph defined as above from κ, using each kernel κr

to describe the addition of copies of Kr .
There is a multi-type compound Poisson branching process X(κ) associ-
ated to κ in a natural way: we start with generation 0 consisting of a single

particle whose type is chosen uniformly from [0, 1]. A particle P of type x
gives rise to children in the next generation according to a two step process:
first, for each r ≥ 2, construct a Poisson process Zr on[0, 1]r−1 with in-
tensity rκr (x, x2 , . . . , xr ) dx2 · · · dxr . The points of Z = r≥2 Zr are called
the child cliques of P . There are r − 1 children of P for each child clique
Random Graphs and Branching Processes 83

(x2 , . . . , xr ) ∈ [0, 1]r−1 , one each of types x2 , . . . , xr . Thus the children of

P form a multiset on [0, 1], with a certain compound Poisson distribution
we have just described. As usual, the children of different particles are
independent of each other, and of the history.
A hyperkernel κ = (κr ) is integrable if

 
r κr < ∞,
r≥2 [0,1]r

corresponding to the sum of the orders of the cliques added to G(n, κ)

having expectation O(n), i.e., to a random vertex of G(n, κ) being in O(1)
corresponding
cliques on average. The hyperkernel κ is irreducible if the
edge kernel

 
κe (x, y) = r(r − 1) κ(x, y, x3 , x4 , . . . , xr ) dx3 · · · dxr
r≥2 [0,1]r−2

is irreducible as a kernel on [0, 1]2 . Note that κe captures (at least to

first order) the individual edge probabilities in G(n, κ), but ignores the
dependence between them.
One of the main results of Bollobás, Janson and Riordan [44] is the
following.

Theorem 16. Let κ = (κr ) be an irreducible, integrable hyperkernel. Then

 
C1 G(n, κ) = ρ(κ)n + op (n),

 
and C2 G(n, κ) = op (n).

Here C1 (G) and C2 (G) denote the number of vertices in the largest and
second largest components of a graph G, and ρ(κ) is the survival probability
of X(κ).

For Theorem 16 to be useful, one needs a handle on the survival prob-
ability ρ(κ); in particular, we would like to know when it is non-zero. In
) is given by integrating the solution to a certain equation
general, ρ(κ

 
(54) ϕ(x) = 1 − exp − (Sκ ϕ)(x) ,

84 B. Bollobás and O. Riordan

where Sκ is now a non-linear integral operator defined using κ (for details,

see [44]). The equation above is likely to be rather difficult to solve in any
particular case. Fortunately, however, there is a much simpler answer to
the question of when ρ(κ) is strictly positive: as shown in [44], if κ is an
ρ(κ) > 0 if and only if Tκ  > 1, where Tκ is
integrable hyperkernel, then e e
the integral operator defined from κe using (52).
As in [41], a key technique of the proofs in [44] is to use approximation
and monotonicity results. This involves proving that if κ is irreducible, then
(54) has at most one non-zero solution. For the rather general branching
processes under consideration, this is far from obvious; is not true for general
branching processes.
Many special cases of the model G(n, κ) have been studied before.
Firstly, if only κ2 is non-zero, then we obtain the sparse inhomogeneous
model G(n, κ) with, unfortunately, a slightly different normalization; for a
discussion of the normalization, see [44]. More simply, if exactly one of the
kernels κr is non-zero, and that kernel is constant, then we obtain the nat-
ural r-uniform hypergraph equivalent of G(n, c/n). The critical point for
this model was found by Schmidt-Pruzan and Shamir [174], and the size of
the giant component above the critical point by Coja-Oghlan, Moore and
Sanwalani [76]. Behrisch, Coja-Oghlan and Kang [16, 15] gave very pre-
cise results about the size of the giant component, proving central and local
limit theorems; see the references in [15] for earlier related work.

5. Branching Processes and Other Global Properties

In some sense, the size of the giant component is the global graph property
most obviously related to branching processes. However, branching process
methods can also be used to study many other properties.

5.1. Diameter

For the first example we consider the diameter. Given two vertices v, w of
a graph G, let d(v, w) = dG (v, w) be the graph distance between them, i.e.,
the minimum number of edges in a path in G starting at v and ending at w.
If there is no such path, i.e., if v and w are in different components of G,
then we set dG (v, w) = ∞.
Random Graphs and Branching Processes 85

The standard graph theoretic definition of the diameter of G is simply

 
diam (G) = max dG (v, w) : v, w ∈ V (G) ,

so diam (G) is finite if and only if G is connected. This is perhaps the most
natural version of the definition: for example, asserting that diam (G) ≤ d
is the same as asserting that any two vertices are joined by a path of length
at most d. This usual graph diameter is not our topic here, so we shall say
almost nothing about the long history of its study. Let us mention in passing
only three results. For random r-regular graphs, Bollobás and Fernandez
de la Vega [39] gave tight bounds on the diameter,
 showing
 in particular
that if r ≥ 3 is fixed, then the diameter is 1 + op (1) log n/ log (r − 1).
Bollobás and Chung [38] showed in 1988 that the diameter of a cycle plus a
random matching is whp logarithmic in n. This phenomenon of (at most)
logarithmic diameter in a very wide variety of random graphs was very well
known before the 1998 observations of Watts and Strogatz [195] on the
‘surprising’ small-world phenomenon; indeed, the Watts–Strogatz model is
extremely close to that studied by Bollobás and Chung.
Some of the first studies of the diameter of the new inhomogeneous
models were the computer experiments presented by Barabási, Albert and
Jeong [4, 14] and heuristic arguments given by Newman, Strogatz and
(m)
Watts [154], suggesting that the LCD model Gn should have diameter
of the form A + B log n. (These experiments and results were stated for
the imprecisely defined Barabási–Albert model; it is not clear in which way
it was made definite enough for simulations to be possible!) In fact, as
(m)  
shown in [49], for m ≥ 2 the diameter of Gn is 1 + op (1) log n/ log log n.
In addition to the proof in [49], a simple heuristic related to branching
processes was given in [49], and later by Cohen and Havlin [75].
Turning to the various inhomogeneous models generalizing G(n, c/n),
since these graphs are very unlikely to be connected, one works with a
slightly different definition of the diameter diam (G) of G:
 
diam (G) = max dG (v, w) : v, w ∈ V (G), dG (v, w) < ∞ ,

i.e., the maximum diameter of any component of G. If G has a giant

component, then it is likely that diam (G) will be the diameter of the giant
component, although this of course needs to be proved in any given context.
Perhaps surprisingly, the diameter of G(n, c/n) received little attention
until fairly recently. For c < 1 L
 uczak [144] obtained very detailed results,
86 B. Bollobás and O. Riordan

treating the case where c approaches 1 at a suitable rate as well as that

of c constant. Chung and Lu [66] gave partial results for c > 1, but did
not obtain the correct asymptotic form. Indeed, they conjectured that,
under suitable conditions, the diameter is approximately log n/ log c. This
is what one might naively expect from the branching process. However, this
turns out not to be correct. The correct asymptotic form of the diameter
of G(n, c/n) was in fact first found as a special case of results for two much
more general models.
Fernholz and Ramachandran [104] studied a version of the configuration
model with a given asymptotic degree sequence. In other words, they
assumed that the numbers ad (n) of degree-d vertices in the n vertex graph
Gn are given, and that the numbers ad (n)/n converge in a suitable sense to
a distribution Λ = (λd )d≥0 satisfying some mild assumptions. Then Gn is
chosen uniformly at random from among all graphs with the given degree
sequence. Using branching process analysis of the 2-core, they showed that
(under mild technical conditions),
 
diam (Gn ) = b + op (1) log n,

where b = b(Λ) is a constant defined in terms of the distribution Λ. Their

results apply to G(n, c/n) by conditioning on the degrees, taking Λ to be
the Poisson distribution with mean c, and they obtained in this case that
 
diam G(n, c/n) 1 2
(55) = + + op (1)
log n log c log (1/c )

for c > 1 constant, 

 where c is the dual branching process parameter to c,
defined by c = c 1 − ρ(c) , or by c < 1 and c e−c = ce−c .
 

Independently, Bollobás, Janson and Riordan [41] obtained a corre-

sponding result for the finite-type case of their general model, showing that
in the subcritical case, when Tκ  < 1, one has
 
diam G(n, κ) p 1
→ ,
log n log Tκ −1

and that in the irreducible, supercritical case,

 
diam G(n, κ) p 1 2
→ + ,
log n log Tκ  log Tκ̂ −1
Random Graphs and Branching Processes 87

where κ̂ is the kernel associated to the ‘dual’ branching process to X(κ).

This of course coincides with (55) if κ is constant.
The proofs of the results above are rather involved, due to the generality
of the models considered. It is thus natural to hope for a much simpler
proof for G(n, c/n) itself. Such a proof has been obtained by Riordan and
Wormald [169], again using branching process arguments. In fact, they
proved much more.

Theorem 17. Let c > 1 be constant. Then

  log n 2 log n
diam G(n, c/n) = + + Op (1).
log c log (1/c )

This result is best possible, in that the diameter does vary by more than
any given constant with probability bounded away from zero. In fact, they
obtained a (rather involved) description of the distribution of the error term.
Riordan and Wormald [169] also obtained a result for G(n, λ/n), where
λ = 1 + ε with ε = ε(n) > 0 tending to zero at a suitable rate. Assuming
that L = ε3 n is at least some large power of log n, they showed that

  log (ε3 n) 2 log (ε3 n)

diam G(n, λ/n) = + + Op (1/ε).
log λ log (1/λ )

Again, the error term is best possible.

Results related to those described in this section have been proved by
many people, often concerning the ‘typical’ distance between vertices, rather
than the diameter. Examples include the results of Chung and Lu [67, 69],
and van den Esker, van der Hofstad, Hooghiemstra, van Mieghem and
Znamenski [192, 191, 193]; for a discussion of related work see [192], for
example.

5.2. The k-core

In this subsection we consider one final example of a global graph property

that, for many random graphs, can be studied using branching processes.
Given a graph G and an integer k ≥ 2, the k-core ck (G) of G is the
subgraph formed by repeatedly deleting any vertices of degree less than k
until none remain. Equivalently, ck (G) is the maximal subgraph of G with
minimal degree k; note that this may be empty. This concept was introduced
88 B. Bollobás and O. Riordan

by Bollobás [30], in the context of finding large k-connected subgraphs of

random graphs. As edges are added to a graph G, its k-core grows. In
particular, up to some point there is no k-core, and beyond some point
there is. The question of when the k-core emerges in a random graph (or
random graph process) arose in the context of finding the chromatic number
of sparse random graphs; Chvátal [72] used the fact that a graph with no
3-core is 3-colourable to show that G(n, 2.88/n) is whp 3-colourable.
A basic question about the emergence of the k-core is as follows: if k ≥ 3
is fixed, what is the critical value λc = λc (k) of λ above which a (non-empty)
k-core first appears whp in G(n, λ/n)? Also, above this point, how large is
the k-core as a function of λ? It turns out that there is a natural ‘guess’ as
to the answer, given in terms of our favourite branching process X(λ): let
Bk be the event that the initial particle x0 has at least k children each of
which has at least k − 1 children each of which has at least k − 1 children
each of which . . . , i.e., that x0 is in a k-regular tree contained in X(λ). Let
βk (λ) be the probability that X(λ) has the property Bk . (The notation βk+
is used in [167].) It is natural to believe that, up to a small error term, a
vertex is in the k-core if and only if its neighbourhoods up to a suitable
distance have a property corresponding to Bk , and thus that the fraction
of vertices of G(n, λ/n) in the k-core is βk (λ) + op (1); this turns out to be
the case.
Pittel, Spencer and Wormald [163] showed that, except at the critical
point, the number of vertices in the k-core of G(n, λ/n), λ constant, is
indeed βk (λ)n + op (n). In particular, the
 threshold λc for the emergence
of the k-core is λc = inf λ : βk (λ) > 0 . The result they actually proved
is rather more precise; they also showed that the k-core emerges extremely
suddenly: for k ≥ 3, adding edges one by one, i.e., considering a random
whp the very first time the k-core is non-empty, it already
graph process G,
has size Θ(n).
Recently, simpler proofs of this result have been given, as well as gen-
eralizations to various other contexts; see, for example, Cooper [77], Mol-
loy [150], Fernholz and Ramachandran [103], Cain and Wormald [62], Jan-
son and Luczak [125] and Darling and Norris [82]. Although the branching
process heuristic mentioned above was described already in the original pa-
per of Pittel, Spencer and Wormald [163], none of these proofs works by
directly coupling the neighbourhoods of a vertex of the graph with the
branching process. Indeed, it turns out to be very difficult to relate the in-
escapably global property of lying in the k-core to a simple local property.
(To see that being in the k-core is a genuinely global property, note that
Random Graphs and Branching Processes 89

adding one single edge to the graph can suddenly cause Θ(n) vertices to
have this property, whereas before adding the edge, no vertices did.)
A direct branching process derivation of the size of the k-core was finally
given by Riordan [167]; the arguments are rather involved. This branching
process approach has the advantage that it generalizes to other models, in
particular, under mild assumptions, to the general inhomogeneous model
of Bollobás, Janson and Riordan [41]: the fraction of vertices in the k-
core turns out to be the probability of the (natural analogue) of the event
Bk in the multi-type branching process. (Related heuristic results were
given by Goltsev, Dorogovtsev and Mendes [117, 86]; in the special case of
G(n, c/n), their arguments reduce to the original heuristic of Pittel, Spencer
and Wormald [163]. Of course, the whole point of [163] was to prove that
this heuristic in fact gives the correct answer.)
One special case treated in [167] is of particular interest. This is the
mean-field version of the Barabási–Albert or LCD model, which we regard
as the BJR model G(n, cκ0 ), where κ0 is the kernel defined by κ0 (x, y) =
√
1/ xy. In this result, βk (cκ0 ) denotes the probability that the multi-type
branching process X(cκ0 ) has the property Bk described above.
√
Theorem 18. Let κ0 (x, y) = 1/ xy. For c > 0, let G(n, cκ0 ) be the graph
on [n] = {1, 2, . . . , n} in which edges are present independently,
 and the
√
probability that i and j are joined is min c/ ij, 1 . For each k ≥ 2 we
have
 
|ck G(n, cκ0 ) | = βk (cκ0 )n + op (n).
If k ≥ 3, then βk (cκ0 ) = 0 for c ≤ (k − 2)/2, while

(k − 1)!2/(k−2) 2/(k−2)
(56) βk (cκ0 ) ∼ ε
(k − 1)(k − 2)

when c = (1 + ε)(k − 2)/2 and ε → 0 from above.

If k = 2, then βk (cκ0 ) > 0 for every c > 0, and

1 2−2γ
βk (cκ0 ) ∼ e exp (−1/c)
2c
as c → 0.

For k = 2, this result states that there is always a ‘giant’ 2-core. How-
ever, the main interest is the case k ≥ 3, where the k-core emerges at a
90 B. Bollobás and O. Riordan

positive threshold, c = (k − 2)/2. Moreover, it emerges slowly: in the ter-

minology of [41], for k ≥ 3 the emergence of the k-core exhibits a phase
transition of exponent 2/(k − 2), where this is the exponent of ε appearing
in (56). This in very sharp contrast to G(n, λ/n), where the k-core emerges
suddenly: the function βk (λ) has a discontinuity at the critical value.

6. Appropriateness of Models: Metrics on Graphs

In this final section we turn to a topic that is often neglected, but is

clearly fundamental in any field where one is using a mathematical model
of something observed in the real world: how to decide whether a particular
model is appropriate, i.e., how to measure how well a certain model fits
the real-world data. In statistics, this is of course a well-developed field.
The situation here is rather different, however. Although our mathematical
model is almost always random, it may well be that there is only one real-
world instance of the network we are trying to model (the internet, for
example). Thus our aim is not to measure how close some set of examples
is to the distribution given by our model, but rather to measure the degree
of similarity between a ‘typical’ graph produced by the model and a single
real-world example. In other words, we would like an appropriate metric on
graphs, giving some kind of distance d(G, H) between two graphs G and H.

In some sense, this question of similarity isn’t really a mathematical

one: the model is appropriate if whatever parameters of the graph one is
interested in are similar in the model and the real network. However, one
can still ask: What if we are interested in the whole graph structure? It may
happen (in very rare cases) that the model is in some sense an (almost) exact
fit to reality: maybe one can find real examples that do grow in a certain
uniformly random manner, so G(n, p), for example, is an (almost) exactly
appropriate model. We should like to be able to recognize this situation if
it should arise and, in general, to measure how far we are from it.

Although our main interest here is the sparse case, we shall start by
discussing the dense case, concerning graphs with n vertices and Θ(n2 )
edges; this is much better understood. There are many sensible metrics to
consider in the dense case: we start with some of the ‘strongest’, i.e., those
where very few pairs of graphs are close in the metric.
Random Graphs and Branching Processes 91

6.1. The edit distance(s): dense case

Recall that we almost always take our graphs to have vertex set [n] =
{1, 2, . . . , n}. The strongest possible notion of similarity between two graphs
G, H on [n] is that they be identical, i.e., that ij ∈ E(G) if and only if
ij ∈ E(H) for all 1 ≤ i < j ≤ n. There is a correspondingly strong notion
of similarity, given by the number of edges present in G but not  in H, or vice

versa. The corresponding (unnormalized) distance is simply  E(G)E(H) ,
where  denotes symmetric difference. In this dense context, the sensible
normalization is to divide by the maximum possible distance, giving the
metric
 
  n
(57) 
dedit1 (G, H) = E(G)E(H) /  .
2

This is the normalized form of the number of changes (edge additions or

deletions) that must be made to one of G or H to turn it into the other,
and is sometimes known as  the edit distance; viewing a graph on [n] as a
point in {0, 1}N , N = n2 , this version of the edit distance is simply the
normalized form of the Hamming distance.
In some contexts, the above notion is appropriate. In particular, when
comparing networks with the same set of nodes, we may wish to know
whether, for the most part, the same pairs of nodes are connected in each
network. Often, however, our labelling of the nodes in at least one network
is totally arbitrary: for example, in the real-world network the nodes might
be the set of people in some community, while the mathematical model uses
[n] as the vertex set. In this context, two graphs are considered the same if
they are isomorphic, i.e., identical up to relabelling the nodes. For graphs
G and H on [n], we have G ∼ = H if and only if there is a permutation σ
of [n] such that ij ∈ E(G) if and only if σ(i)σ(j) ∈ E(H). There is a
corresponding notion of similarity, sometimes called the edit distance. As
before, we normalize, so
 −1
n  
(58) dedit2 (G, H) = min  E(G)E(H  ) .
2  ∼
H =H

Unfortunately, the term ‘edit distance’ is not consistently used: it may refer
to either of the notions in (57) or (58), and is also often unnormalized. The
(unnormalized) version of dedit2 seems to have been first defined explicitly
by Axenovich, Kézdy and Martin [11], although implicitly the notion had
92 B. Bollobás and O. Riordan

been used much earlier, e.g., by Erdős [92] and Simonovits [178] in 1966,
and in many subsequent papers.
Note that this notion of distance is already not so easy to compute in
practice. Indeed, the ‘isomorphism problem’, to determine whether two
given graphs are isomorphic, is a well known ‘hard’ problem. (Interestingly,
it is not known to be NP-complete; nevertheless, it is widely believed not
to lie in P.)
It might seem that (58) answers the question of how close two dense
graphs are. To show that this is not the appropriate measure in this context,
we consider the example of the Erdős–Rényi graph G(n, 1/2). For the
motivation, consider the hypothetical situation that the real-world network
we are studying is produced by some (physical, say) process that really
does produce connections independently and at random, so our real-world
network H is simply an instance of G(n, 1/2). Our aim is to somehow
recognize this.
Now G(n, 1/2) is obtained by choosing any graph on [n], with all choices
equally likely, so it seems absurd to speak of a typical instance of G(n, 1/2).
Nevertheless, there is an intuitive sense in which G(n, 1/2) is hardly ran-
dom at all! Indeed, as mentioned in Subsection 2.1, for almost any ‘sensible’
numerical property one can measure (for example, the number of triangles,
or the variance of the degree distribution, or the size of the largest cut,
or the approximate size of the giant component), it turns out that for n
large, with very high probability, the value of this property falls in a rather
small range. Furthermore, for almost all natural yes/no properties, either
G(n, 1/2) almost certainly has this property, or it almost certainly does not;
these observations go back to Erdős and Rényi. It is not easy (indeed, per-
haps impossible) to make this ‘sameness principle’ mathematically precise,
but anyone who has worked with G(n, p) will recognize it. A consequence of
this is that, if we are working with graphs similar to G(n, 1/2), our metric
d should have the property that the distance d(G, H) between two indepen-
dent copies of G(n, 1/2) is almost always very small, whatever that means.
Unfortunately, in the edit distance, two independent instances G, H of
G(n, 1/2) are almost always almost as far apart as possible: we include the
details of this simple fact.

Theorem 19. Let G and H be independent instances of G(n, 1/2). Then,

for any 0 < ε < 1/2, we have dedit2 (G, H) ≥ 1/2 − ε whp.
Suppose  G0 and H0 are any two graphs with between
nthat η > 0 and that
n
(1/2 − η) 2 and (1/2 + η) 2 edges. Then dedit2 (G0 , H0 ) ≤ 1/2 + 2η 2 .
Random Graphs and Branching Processes 93

Note that for any η > 0, G and H satisfy the conditions of the second
part whp, so G and H are indeed almost as far apart as possible given their
numbers of edges.
Proof. Given a permutation σ of [n] and a graph H on [n], let H σ be
the graph with ij ∈ E(H σ ) if and only if σ(i)σ(j) ∈ E(H). With σ
fixed and H an instance G(n, 1/2), H σ has the same distribution as H.
Taking G ∈ G(n, 1/2) independent of H, for each ij the events ij ∈ E(G)
and ij ∈ E(H σ ) are independent and have probability 1/2, so P ij ∈
E(G)E(H σ ) = 1/2.  Furthermore, these events are independent for
σ  σ 
  so X = E(G)E(H ) has a binomial
different edges,   distribution with
mean μ = 2 2 . Let ε > 0 be fixed, and set N = n2 . From the Chernoff
1 n

bounds, for example, (see [34]), one has

 
P |X σ − μ| ≥ εN ≤ e−2ε N .
2

Since dedit2 (G, H) ≤ 1/2 − ε if and only if there is some σ such that
dedit1 (G, H σ ) ≤ 1/2 − ε, i.e., such that X σ ≤ (1/2 − ε)N , it follows that
 
P dedit2 (G, H) ≤ 1/2 − ε ≤ n!e−2ε N = o(1).
2

Thus dedit2 (G, H) ≥ 1/2 − ε whp, proving the first part of the theorem.
For the second statement, let G0 and H0 be any two graphs on [n], and
define 0 ≤ a, b ≤ 1 by e(G0 ) = aN and e(H0 ) = bN . Then
  1   
N dedit2 (G0 , H0 ) = min  E(G0 )E(H0σ ) ≤ E(G0 )E(H0σ )
σ n! σ
  
= P ij ∈ E(G0 )E(H0σ ) ,
ij

where in the final formula we choose σ uniformly at random from among

all n! permutations of [n]. Thinking of ij as also random, the random edges
ij and σ(i)σ(j)
 are independent.
  Thus, even though
 G0 and H0 are  fixed,
the events ij ∈ E(G0 ) and ij ∈ E(H0 ) = σ(i)σ(j) ∈ E(H0 ) are
σ

independent. Consequently,
 
P ij ∈ E(G0 )E(H0σ ) = a(1 − b) + b(1 − a) = a + b − 2ab.

Hence, for any two graphs G0 and H0 with aN and bN edges respectively,
we have
dedit2 (G0 , H0 ) ≤ a + b − 2ab.
94 B. Bollobás and O. Riordan

For a, b ∈ [1/2 − η, 1/2 + η] this expression is maximized when {a, b} =

{1/2 − η, 1/2 + η}, in which case a + b − 2ab = 1/2 + 2η 2 .

Having seen that (either version of) the edit distance is a totally inappro-
priate in the present context, we next turn to metrics that are appropriate.

6.2. The subgraph distance: dense case

Some of the most basic questions one can ask about a graph are: How
many edges does it have? How many triangles does it contain? and so on.
Taken together, the answers to these questions provide a complete ‘local’
description of the graph, and it makes sense to consider two graphs ‘close’
if the answers to these questions are similar for the two graphs.
More formally, thinking of F as a small ‘fixed’ graph, and G as some large
graph, let XF (G) denote the number of subgraphs of G that are isomorphic
to F . The same information is captured by the number of embeddings
of F into G, i.e., the number of injections ϕ : V (F ) → V (G) such that
ϕ(x)ϕ(y) ∈ E(G) whenever xy ∈ E(F ). Indeed, one has

emb (F, G) = aut (F )XF (G),

where aut (F ) is the number of automorphisms of F . If |F | = k and |G| = n,

then the natural normalization is to divide by n(k) = emb (F, Kn ), the
maximum possible number of embeddings of F into an n-vertex graph. This
gives the subgraph density of F in G:

emb (F, G) XF (G)

s(F, G) = = ∈ [0, 1],
n(k) XF (Kn )

generalizing the edge density, which is the case F = K2 .

Letting F run over one representative Fi of each isomorphism
 class
 ∞of
finite graphs, we may map a given graph G to a point s(G) = s(Fi , G) i=1
in [0, 1]N . Taking any metric d on [0, 1]N giving rise to the product topology,
we may then define the subgraph distance between two graphs G and H as
 
dsub (G, H) = d s(G), s(H) .
Random Graphs and Branching Processes 95

More concretely, we may take, for example,

∞
 2 
dsub (G, H) = 2−|Fi |  s(Fi , G) − s(Fi , H) .
i=1

Such a concrete definition is likely to be more useful in practice, but,

mathematically, usually the main property of a metric one is interested in is
the resulting topology, i.e., for which sequences (Gn ), (Hn ) does d(Gn , Hn )
tend to 0. For this, the choice of metric on [0, 1]∞ is an irrelevant distraction.
The key property of the subgraph distance is that a sequence (Gn ) is
Cauchy in dsub if and only if s(F, Gn ) converges for every F . Such sequences
are sometimes called convergent, although using this term without first
defining the corresponding limit points conflicts with the usual usage in
topology.
It is not hard to check that dsub (G, H) = 0 if and only if G ∼= H, so dsub
is a genuine metric on the set F of isomorphism classes of finite graphs. It
is intuitively clear that dsub is a ‘reasonable’ metric to use when deciding
whether two dense graphs have similar local properties. We shall return to
global properties shortly; we first note some very nice properties of dsub .
It is immediate from the definition of dsub that the completion of
(F , dsub ) is a compact metric space. What is interesting is that the ex-
tra points in this space, corresponding to limits of sequences (Gn ) with
|Gn | → ∞, have a very nice alternative description, found by Lovász and
Szegedy [141]. Note that we follow here the precise definitions and notation
of [51]: Lovász and Szegedy worked instead with normalized homomorphism
counts. This makes some formulae a tiny bit cleaner, but also gives only a
pseudo-metric, not a metric.
By a graphon we shall mean a symmetric measurable function W from
[0, 1]2 to [0, 1]. Given a graph F with vertex set k and a graphon W , one
can define the subgraph density of F in W by


k
s(F, W ) = W (xi , xj ) dxi .
[0,1]k ij∈E(F ) i=1

(The same quantity is often denoted t(F, W ).)

Lovász and Szegedy [141] proved (essentially) the following result.
96 B. Bollobás and O. Riordan

Theorem 20. Let (Gn ) be a Cauchy sequence in (F , dsub ). Then either

(Gn ) is eventually constant, or there is a graphon W such that s(F, Gn ) →
s(F, W ) for every F .

Except that they are bounded, graphons are essentially the same as
the kernels discussed in Subsection 2.5. Recall from Subsection 2.5 that
Lovász and Szegedy [141] defined a natural dense random graph G(n, W ) =
G1 (n, W ) associated to a given graphon W ; they called this random graph a
W -random graph. To define G(n, W ), first choose x1 , . . . , xn independently
and uniformly from [0, 1], and then, given x1 , . . . , xn , include each possible
edge ij independently with probability W (xi , xj ). As noted earlier, apart
from the normalization, this is exactly the same as (a special case of)
the definition of the model G(n, κ) = G1/n (n, κ) in Bollobás, Janson and
Riordan [41].
It is very easy to check that the subgraph density of F in G(n, W ) =
G1 (n, W ) is concentrated
 about s(F, W ) as n → ∞, so the random sequence
(Gn ) = G(n, W ) will be Cauchy in dsub with probability 1. The force
of Theorem 20 is now apparent: it states that the random graph models
G(n, W ) are universal with respect to dsub : any Cauchy sequenceis ‘close to’

(would remain Cauchy if interleaved with) a random sequence G(n, W ) .
This is further evidence that dsub is a natural metric.

6.3. The cut metric: dense case

In the previous subsection we considered only ‘local’ properties. What if

we would like our graphs G and H also to be similar in terms of ‘global’
properties, such as the size of the maximum cut? For this, one uses the cut
metric defined by Borgs, Chayes, Lovász, Sós and Vesztergombi [57], based
on a norm used by Frieze and Kannan [110].
Given a bounded measurable function W from [0, 1]2 to R, its cut norm
W cut is defined by
 
 
W cut = sup  W (x, y) dx dy ,

S,T ⊂[0,1] S×T

where the supremum is over all pairs of measurable subsets of [0, 1].
 WG associated
Given a graph G with n vertices, there is a graphon 
to G in a natural way: WG is constant on each square (i − 1)/n, i/n ×
Random Graphs and Branching Processes 97
 
(j − 1)/n, j/n , taking the value 1 on this square if ij ∈ E(G) and 0
otherwise. One could use the cut-norm to measure the distance between
graphs G and H by considering WG − WH cut , but this takes no account
of relabelling; in particular, isomorphic graphs will not be at zero distance.
Given a measure-preserving map ϕ from [0, 1] to itself and a graphon

W , let W (ϕ) (ϕ)
be the graphon defined by W (x, y) = W ϕ(x), ϕ(y) . If ϕ
is a bijection (or a bijection between subsets of measure 1 – in this context
sets of measure 0 do not matter), then we call W (ϕ) a rearrangement of W .
We write W1 ≈ W2 if W1 is a rearrangement of W2 . The cut metric dcut
of Borgs, Chayes, Lovász, Sós and Vesztergombi [57] may be defined for
(signed) graphons by

dcut (W1 , W2 ) = inf W1 − W2 cut ,

W2 ≈W2

and for graphs simply by dcut (G, H) = dcut (WG , WH ). It is easy to see that
if G and H are isomorphic, then dcut (G, H) = 0.
The cut metric is not quite a metric on (isomorphism) classes of graphs:
it is possible for graphs with different numbers of vertices to give rise to the
same graphon. Nonetheless, it is very useful. Although not quite immediate
from the definition, it does turn out that if graphs G and H are close in
dcut then they do have maximum cuts of almost the same size, for example.
Indeed, much more is true: after relabelling, any cut in G has almost the
same size as the corresponding cut in H (see [57]).
It may seem that by considering subgraph and cut metrics we are mul-
tiplying possibilities. Remarkably, this is not the case. One of the main
results (Theorem 2.6) of Borgs, Chayes, Lovász, Sós and Vesztergombi [57]
is that the metrics dsub (defined in a slightly different way than we have
done here) and dcut are equivalent: (Gn ) is a Cauchy sequence for dsub if
and only if it is a Cauchy sequence for dcut .
In a series of papers, Borgs, Chayes, Lovász, Sós, Szegedy and Veszter-
gombi [56, 55, 141, 142, 57, 58] have taken this much further, showing that
this notion of convergence is equivalent to various other natural notions.
These results show that in the dense case, although a priori it is not clear
what ‘similar’ should mean for two graphs, there are many different notions
that turn out to be equivalent. Moreover, Cauchy sequences correspond in a
very nice way to inhomogeneous random graphs. This is a very satisfactory
answer to the dense case of the rather vague question posed at the start of
this section.
98 B. Bollobás and O. Riordan

6.4. The sparse case

In the sparse case, there is unfortunately no simple analogue of the results

described above. Many of the definitions make sense, at least in various
density ranges, but the nice equivalence breaks down. We discuss this
phenomenon at great length in [51], presenting many (in some cases partial)
results, conjectures as to when equivalence may hold, and examples where
it does not. In this subsection we shall briefly describe the picture painted
in [51].
In the dense case, one of the key tools in the analysis of general graphs is
Szemerédi’s Lemma, and the accompanying embedding or counting lemmas.
Together these say that any graph G may be partitioned into not too many
parts, so that for any given ‘small’ graph F , the number of copies of F in G
can be calculated approximately simply from the edge densities between the
various parts of G. This is extremely important when relating the cut and
count metrics; in the weak form due to Frieze and Kannan [110], Szemerédi’s
Lemma gives exactly a simple graphon that is close in the cut metric to the
original graph, and the counting lemma implies that the subgraph densities
in the graph and the graphon are similar.
In the sparse case, concerning graphs with o(n2 ) but many more than n
edges, there are various forms of Szemerédi’s Lemma that (with some mild
assumptions) provide a satisfactory approximation in the (appropriately
normalized) cut metric, but in general there is no satisfactory counting
lemma. Indeed, one of our main aims in [51] was to prove such a counting
lemma for a certain class of subgraphs, extending a result of Chung and
Graham [65]. In general, it is not clear what one can expect from a counting
lemma, or how to define the analogue of dsub . For some conjectured possible
relationships between the cut metric and suitably defined versions of dsub ,
see [51].
The case of graphs with O(n) edges (called the ‘extremely sparse case’
in [51]) seems to be even more difficult, and at the same time especially
interesting. Even a survey of the relevant sections (6 and 7) of [51] would
take many pages, so we shall content ourselves with a very brief and informal
overview of some of the relevant issues. We should emphasize that this is
a very recent area of investigation; despite the current shortage of results,
there are many interesting avenues to explore.
For graphs with O(n) edges, Szemerédi’s Lemma breaks down com-
pletely: there seems to be no natural definition of regularity such that
Random Graphs and Branching Processes 99

non-trivial regular pairs exist at all! Correspondingly, defining dcut in the

natural way (i.e., as above, but normalizing so that the distance between
an empty graph and one with n edges is 1), it turns out that a sequence
(Gn ) in which Gn has n vertices is Cauchy in dcut only in the trivial case
when e(Gn ) = o(n); see Theorem 6.2 of [51]. Also, dcut and a suitably nor-
malized edit distance turn out to be essentially equivalent, and both have
the undesirable property that two different instances of G(n, 2/n), say, are
far apart in this metric.
Turning to metrics defined by ‘local’ properties, the natural normaliza-
   graph F in graphs Gn with n
tion when counting copies of a connected
vertices and density p (i.e., with Θ p n2 edges) is to consider

XF (Gn ) emb (F, Gn )

(59) sp (F, Gn ) =   = .
E (XF G(n, p) ) n(|F |) pe(F )

For this count to be well behaved, the denominator should tend to infinity.
(We don’t want to say that two graphs are far apart simply because one
contains a single triangle, while the other does not.) In the extremely
sparse case, where p = 1/n, this happens only when F is a tree. One
can define a tree subgraph distance this way, as in [51]: writing s(T, Gn ) for
emb (T, Gn )/n, to obtain a well behaved metric, it is convenient to restrict to
graphs Gn in which the tree counts s(T, Gn ) are bounded. Fixing constants
cT > 0, we consider only graphs Gn for which s(T, Gn ) ≤ cT for every T .
Picking one representative Ti of each isomorphism
 class
 ∞of finite
 trees, one
may then map each graph Gn to s(Gn ) = s(Ti , Gn ) i=1 ∈ i [0, ci ], and
proceed as before.
Counting only subtrees is particularly appropriate in graphs with few
short cycles, such as the graphs given by the BJR model G(n, κ) =
G1/n (n, κ). Indeed, as discussed in [51], for such ‘locally acyclic graphs’,
the tree counts capture essentially the distribution of local structure. Writ-
ing Γ≤h (v) for the subgraph of a given graph Gn induced by all vertices
within graph distance h of v, viewed as a rooted graph with root v, it turns
out that (under mild assumptions), the tree counts in Gn capture, for each
rooted tree T of height at most h, the probability that Γ≤h (v) is isomorphic
to T when v is chosen randomly from V (Gn ); for the details, see [51]. If
Gn = G(n, κ), then this information captures exactly the branching process
X(κ) but seen without types.
For sparse graphs which do contain many short cycles, there is a different
natural normalization in (59): since Gn has Θ(n) edges, it will presumably
100 B. Bollobás and O. Riordan

contain O(n) copies of any given graph F , so we simply consider

s̃(F, Gn ) = emb (F, Gn )/n,

and use these (now differently normalized) subgraph counts to define a

metric dloc , by first mapping Gn to
 
s̃(Gn ) = s̃(F, Gn ) F ∈F ∈ [0, ∞)F .

Once again, it turns out that this is simply a different way of describing
local information: given a rooted graph F in which every vertex is within
distance h of the root, let
 
p(F, Gn ) = ph (F, Gn ) = P Γ≤h (v) ∼
=F ,

when v is a random vertex of Gn . Defining a corresponding metric, a se-

quence (Gn ) is Cauchy if and only if there are numbers ph (F ) such that
ph (F, Gn ) → ph (F ) for every F . The numbers ph (F ) may be combined to
form a probability distribution on infinite locally finite rooted graphs F .
The notion of limit one arrives at in this way is extremely natural: it was
used by Benjamini and Schramm [19] (for a random rather than determin-
istic sequence Gn ) to define a ‘distributional limit’ of certain random planar
graphs. The same notion in slightly different generality was studied by Al-
dous and Steele [8], under the name ‘local weak limit’, and Aldous and
Lyons [7], who used the term ‘random weak limit’. The definition leaves
open many interesting questions, in particular, the question of which limit-
ing distributions can arise in this way. This question was posed by Aldous
and Lyons [7]; see also [51].
Returning to global properties, consider the question of how to dis-
tinguish the random graph G(n, 10/n) from the random bipartite graph
G(n/2, n/2; 20/n) with vertex classes of size n/2 (n even), in which each
possible edge is present independently with probability 20/n. The kernels
corresponding to these graphs give rise to the same branching process, and
the graphs have the same local structure. However, one is bipartite, while
in the other any cut into parts of size n/2 spans Θ(n) edges. The cut met-
ric would distinguish them, but as we have seen, it also distinguishes two
different instances of G(n, 10/n). For this reason it seems to be better to
consider partitions.
The following definition is from [51]. Here Gn denotes a graph with
n vertices, and p = p(n) a normalizing density. Thus, for example, the
Random Graphs and Branching Processes 101

normalized density of edges

 dp (A, B) between sets A and B of vertices of
Gn is simply eGn (A, B)/ p|A| |B| .
Let k ≥ 2 be fixed. For n ≥ k and Π = (P1 , . . . , Pk ) a partition
of V (Gn ) into k non-empty parts, let MΠ (Gn ) be the k-by-k matrix
whose ijth entry is dp (Pi , Pj ), the normalized density of edges from
Pi to Pj . Since MΠ (Gn ) is symmetric, we may think of this matrix as
an element of Rk(k+1)/2 . Set

Mk (Gn ) = {MΠ (Gn )} ⊂ Rk(k+1)/2 ,

where Π runs over all balanced partitions of V (Gn ) into k parts, i.e.,
all partitions (P1 , . . . , Pk ) with |Pi − Pj | ≤ 1.
As usual, we always assume that Gn has O(pn2 ) edges. For
definiteness, let us assume that e(Gn ) ≤ M pn2 . Since each part
of a balanced partition has size at least n/(2k), the entries of any
MΠ (Gn ) ∈ Mk (Gn ) are bounded by Mk = (2k)2 M , say. Thus,
k(k+1)/2
Mk (Gn ) is a subset of the compact space Mk = [0, Mk ] .
Let C (Mk ) denote the set of non-empty compact subsets of Mk ,
and let dH be the Hausdorff metric on C (Mk ), defined with respect to
the ∞ distance, say. Note that (C (Mk ), dH ) is compact, since Mk is
compact. To ensure that the metric we are about to define is a genuine
metric, it is convenient to add the empty set to C (Mk ), extending dH
so that the empty set is an isolated point.

Let C = k≥2 C (Mk ), and let M : F → C be the map defined by
 ∞
M (Gn ) = Mk (Gn )
k=2

for every graph Gn on n vertices, noting that Mk (Gn ) is empty if

k > n. Then we may define the partition metric dpart by
 
dpart (G, G ) = d M (G), M (G ) ,

where d is any metric on C giving rise to the product topology. Consid-

ering the partition of an n vertex graph into n parts shows that dpart is
a metric on the set F of isomorphism classes of finite graphs. Recall-
ing that each space (C (Mk ), dH ) is compact, the key property of the
partition metric is that (Gn ) is Cauchy with respect to dpart if and only
 
if there are compact sets Yk ⊂ Mk such that dH Mk (Gn ), Yk → 0
for each k. In particular, convergence in dpart is equivalent to conver-
gence of the set of partition matrices for each fixed k. Thus we may
always think of k as fixed and n as much larger than k.
102 B. Bollobás and O. Riordan

An analogous metric was introduced independently by Borgs, Chayes,

Lovász, Sós and Vesztergombi [58] in the dense case; apart from normaliza-
tion, the only difference is that in [58], all partitions into k parts are con-
sidered, rather than just balanced partitions. In the dense case, as shown
in [51, Subsection 6.4.1], the partition and cut metrics are essentially equiv-
alent. In fact, this applies also to sparse graphs, as long as np → ∞. In
contrast, in the extremely sparse case we are interested here, dcut and dpart
behave very differently. Indeed, crucially, different instances of the BJR
model G(n, κ) are close in dpart at least if they have the same number of
vertices. More precisely, the following result is proved in [51].

Theorem 21. Let κ be a bounded kernel, let k ≥ 2 be fixed, and let

Gn = G(n, κ) = G1/n (n, κ). There is a sequence (Yn )n≥1 of sets Yn ∈ Mk
 
such that dH Mk (Gn ), Yn converges to 0 in probability.

Unfortunately, it is not known that the sets Yn cannot ‘jump around’

as n varies, although this is extremely unlikely. If they do not, i.e., if we
may take for all Yn the same set Y = Y (κ), then the following conjecture
from [51] holds.

Conjecture 22. For any bounded kernel κ, the random sequence G1/n (n, κ)
is Cauchy with respect to dpart with probability 1.

If this holds, then the partition metric is likely to be a good metric for
distinguishing between graphs arising from the BJR model G(n, κ) with dif-
ferent kernels κ. Note, however, that this is a much more delicate question
than one might think. There are ‘obviously’ different kernels which nonethe-
less give rise to essentially the same random graph model; see the discussion
in [51, Section 6.5].

6.5. New metrics and models

In the very sparse case, considering graphs with Θ(n) edges, it seems likely
that the partition metric described above does a fairly good job of capturing
global information. On the other hand, dloc or the notion of local weak limit
captures local information. One can combine these notions, into the coloured
neighbourhood metric. Again we quote from [51].
Let Gn be a graph with n vertices, and k ≥ 1 an integer. We shall
think of Gn as having Θ(n) edges, though this is only relevant when
Random Graphs and Branching Processes 103

we come to sequences (Gn ). Let Π = (P1 , . . . , Pk ) be a partition of the

vertex set of Gn , which we may think of as a (not necessarily proper)
k-colouring of Gn . This time, for variety, we do not insist that the
parts have almost equal sizes; this makes essentially no difference. Let
r
Gk,t be the set of isomorphism classes of k-coloured connected rooted
r
graphs with radius at most t. For each F ∈ Gk,t , let pk,t (Gn , Π)(F )
be the probability that the t-neighbourhood of a random vertex of the
coloured graph (Gn , Π) is isomorphic to F as a coloured rooted graph,
r
so pk,t (Gn , Π) is a probability distribution on Gk,t . Finally, let

Mk,t (Gn ) = {pk,t (Gn , Π)},

where Π runs over all k-partitions of V (Gn ). Thus Mk,t (Gn ) is a finite
r r
subset of the space P(Gk,t ) of probability distributions on Gk,t .
...
r r
The space P(Gk,t ) of probability distributions on Gk,t is naturally
viewed as a metric space, with the total variation distance dTV between
r
two distributions as the metric. In other words, regarding P(Gk,t ) as
r
a subset of the unit ball of 1 in RGk,t , we simply take the 1 -metric
on this set. Let dH denote the Hausdorff distance between compact
r
subsets of P(Gk,t ), defined with respect to dTV . Then we may define
the coloured neighbourhood metric dcn by
   
dcn (G, G ) = 2−k−t dH Mk,t (G), Mk,t (G ) ,
k≥1 t≥1

 r
say. (As before, we can instead use any metric on t,k P(Gk,t ) giving
rise to the product topology.) If we restrict our attention to graphs
with maximum degree at most some constant Δ, then the correspond-
r r
ing sets Gk,t are finite, so each P(Gk,t ) is compact, and any sequence
(Gn ) has a subsequence that is Cauchy with respect to dcn , and in fact
r
converges to a limit point consisting of one compact subset of P(Gk,t )
for each k, t. In fact, it is not hard to check that whenever (Gn ) has
bounded tree counts (i.e., contains O(n) copies of any fixed tree T ), it
has a convergent subsequence with respect to dcn .

Taking k = 1, one sees that if the sequence (Gn ) is Cauchy in the

coloured neighbourhood metric, then it has a local weak limit. On the
other hand, considering only 1-neighbourhoods, one can check that (Gn ) is
Cauchy with respect to the partition metric, so dcn provides a strong notion
of similarity between two graphs. Unlike the cut metric, it likely that this
notion is not too strong, and in particular that sequences Gn = G(n, κ)
generated by the model of Bollobás, Janson and Riordan [41] are Cauchy
in dcn .
104 B. Bollobás and O. Riordan

As commented above, in the dense case, the W -random graph models

G(n, W ) = G1 (n, W ) are in some sense universal with respect to many
equivalent metrics, including dcut , dsub and the dense version of dpart . One
would like to know whether there are similar random graph models which
have the same relationship to dpart or dcn in the extremely sparse case. The
following slightly more concrete form of this question is from [51].

Question 23. Given a metric d, can we find a ‘natural’ family of random

graph models with the following two properties: (i) for each model, the
sequence of random graphs (Gn ) generated by the model is Cauchy with
respect to d with probability 1, and (ii) for any sequence (Gn ) with |Gn | = n
that is Cauchy with respect to d, there is a model from the family such that,
if we interleave (Gn ) with a sequence of random graphs from the model, the
resulting sequence is still Cauchy with probability 1?

The clustering model of Bollobás, Janson and Riordan [44] is much more
general than the model in [41], and does produce graphs with interesting
local cyclic structure. However, it is still far from an answer to the (very
ambitious) program suggested by Question 23. For example, the degree
distributions generated by the clustering model are approximately mixtures
of compound Poisson distributions. Considering only subgraph counts, a
positive answer to Question 23 for dloc would involve a model producing,
for example, graphs consisting almost entirely of edge-disjoint triangles,
with almost all degrees equal to 6.
For lack of space we cannot go into the many examples of sparse metrics
and models discussed in [51], but we hope that we have managed to whet
the reader’s appetite to take a closer look at the host of intriguing questions
discussed at greater length there.

References

[1] W. Aiello, F. Chung, and L. Lu, A random graph model for massive graphs, in: Pro-
ceedings of the Thirty-Second Annual ACM Symposium on Theory of Computing,
pages 171–180 (electronic), New York, 2000. ACM.
[2] W. Aiello, F. Chung, and L. Lu, A random graph model for power law graphs,
Experiment. Math., 10 (2001), 53–66.
[3] R. Albert and A.-L. Barabási, Statistical mechanics of complex networks, Rev.
Mod. Phys., 74 (2002), 47–97.
Random Graphs and Branching Processes 105

[4] R. Albert, H. Jeong, and A.-L. Barabási, Diameter of the world-wide web, Nature,
401 (1999), 130–131.
[5] R. Albert, H. Jeong, and A.-L. Barabási, Error and attack tolerance of complex
networks, Nature, 406 (2000), 378–382.
[6] D. Aldous, Brownian excursions, critical random graphs and the multiplicative
coalescent, Ann. Probab., 25 (1997), 812–854.
[7] D. Aldous and R. Lyons, Processes on unimodular random networks, Electron.
J. Probab., 12(54) (2007), 1454–1508 (electronic).
[8] D. Aldous and J. M. Steele, The objective method: probabilistic combinatorial
optimization and local weak convergence, in: Probability on discrete structures,
volume 110 of Encyclopaedia Math. Sci., pages 1–72. Springer, Berlin, 2004.
[9] K. B. Athreya and P. E. Ney, Branching processes, Springer-Verlag, New York,
1972, Die Grundlehren der mathematischen Wissenschaften, Band 196.
[10] T. L. Austin, R. E. Fagen, W. F. Penney, and J. Riordan, The number of compo-
nents in random linear graphs, Ann. Math. Statist, 30 (1959), 747–754.
[11] M. Axenovich, A. Kézdy, and R. Martin, On the editing distance of graphs J. Graph
Theory, 58 (2008), 123–138.
[12] G. N. Bagaev and E. F. Dmitriev, Enumeration of connected labeled bipartite
graphs, Dokl. Akad. Nauk BSSR, 28 (1984), 1061–1063, 1148.
[13] A.-L. Barabási and R. Albert, Emergence of scaling in random networks, Science,
286 (1999), 509–512.
[14] A.-L. Barabási, R. Albert, and H. Jeong, Scale-free characteristics of random
networks: the topology of the world-wide web, Physica A, 281 (2000), 69–77.
[15] M. Behrisch, A. Coja-Oghlan, and M. Kang, Local limit theorems and number of
connected hypergraphs, Preprint available from arXiv:0706.0497, 2007.
[16] M. Behrisch, A. Coja-Oghlan, and M. Kang, The order of the giant component of
random hypergraphs, Preprint available from arXiv:0706.0496, 2007.
[17] E. A. Bender and E. R. Canfield, The asymptotic number of labeled graphs with
given degree sequences, J. Combinatorial Theory Ser. A, 24 (1978), 296–307.
[18] E. A. Bender, E. R. Canfield, and B. D. McKay, The asymptotic number of labeled
connected graphs with a given number of vertices and edges, Random Structures
Algorithms, 1 (1990), 127–169.
[19] I. Benjamini and O. Schramm, Recurrence of distributional limits of finite planar
graphs, Electron. J. Probab., 6, article no. 23 (2001), 13 pp. (electronic).
[20] N. Berger, B. Bollobás, C. Borgs, J. Chayes, and O. Riordan, Degree distribution
of the FKP network model, in: Automata, Languages and Programming, volume
2719 of Lecture Notes in Comput. Sci., pages 725–738. Springer, Berlin, 2003.
[21] N. Berger, B. Bollobás, C. Borgs, J. Chayes, and O. Riordan, Degree distribution
of the FKP network model, Theor. Comput. Sci., 379 (2007), 306–316.
106 B. Bollobás and O. Riordan

[22] N. Berger, C. Borgs, J. T. Chayes, R. M. D’Souza, and R. D. Kleinberg, Competi-

tion-induced preferential attachment, in: Automata, Languages and Programming,
volume 3142 of Lecture Notes in Comput. Sci., pages 208–221. Springer, Berlin,
2004.
[23] N. Berger, C. Borgs, J. T. Chayes, R. M. D’Souza, and R. D. Kleinberg, Degree dis-
tribution of competition-induced preferential attachment graphs, Combin. Probab.
Comput., 14 (2005), 697–721.
[24] N. Berger, C. Borgs, J. T. Chayes, and A. Saberi, On the spread of viruses on
the internet, in: Proceedings of the Sixteenth Annual ACM-SIAM Symposium on
Discrete Algorithms, pages 301–310 (electronic), New York, 2005. ACM.
[25] B. Bollobás, A probabilistic proof of an asymptotic formula for the number of
labelled regular graphs, European J. Combin., 1 (1980), 311–316.
[26] B. Bollobás, The independence ratio of regular graphs, Proc. Amer. Math. Soc.,
83 (1981), 433–436.
[27] B. Bollobás, The asymptotic number of unlabelled regular graphs, J. London Math.
Soc. (2), 26 (1982), 201–206.
[28] B. Bollobás, Almost all regular graphs are Hamiltonian, European J. Combin., 4
(1983), 97–106.
[29] B. Bollobás, The evolution of random graphs, Trans. Amer. Math. Soc., 286 (1984),
257–274.
[30] B. Bollobás, The evolution of sparse graphs, in: Graph Theory and Combinatorics
(Cambridge, 1983), pages 35–57. Academic Press, London, 1984.
[31] B. Bollobás, Random Graphs, Academic Press Inc. [Harcourt Brace Jovanovich
Publishers], London, 1985.
[32] B. Bollobás, The isoperimetric number of random regular graphs, European J. Com-
bin., 9 (1988), 241–244.
[33] B. Bollobás, Modern Graph Theory, volume 184 of Graduate Texts in Mathematics,
Springer-Verlag, New York, 1998.
[34] B. Bollobás, Random Graphs, volume 73 of Cambridge Studies in Advanced Math-
ematics, Cambridge University Press, Cambridge, second edition, 2001.
[35] B. Bollobás, C. Borgs, J. Chayes, and O. Riordan, Directed scale-free graphs, in:
Proceedings of the Fourteenth Annual ACM-SIAM Symposium on Discrete Algo-
rithms (Baltimore, MD, 2003), pages 132–139, New York, 2003. ACM.
[36] B. Bollobás, C. Borgs, J. Chayes, and O. Riordan, Percolation on dense graph
sequences, Ann. Probab., to appear.
[37] B. Bollobás, C. Borgs, J. T. Chayes, J. H. Kim, and D. B. Wilson, The scaling
window of the 2-SAT transition, Random Structures Algorithms, 18 (2001), 201–
256.
[38] B. Bollobás and F. R. K. Chung, The diameter of a cycle plus a random matching,
SIAM J. Discrete Math., 1 (1988), 328–333.
[39] B. Bollobás and W. Fernandez de la Vega, The diameter of random regular graphs,
Combinatorica, 2 (1982), 125–134.
Random Graphs and Branching Processes 107

[40] B. Bollobás, S. Janson, and O. Riordan, The phase transition in the uniformly
grown random graph has infinite order, Random Structures Algorithms, 26 (2005),
1–36.
[41] B. Bollobás, S. Janson, and O. Riordan, The phase transition in inhomogeneous
random graphs, Random Structures Algorithms, 31 (2007), 3–122.
[42] B. Bollobás, S. Janson, and O. Riordan, Spread-out percolation in Rd , Random
Structures Algorithms, 31 (2007), 239–246.
[43] B. Bollobás, S. Janson, and O. Riordan, Line-of-sight percolation, Combinatorics
Probability and Computing, 18 (2009), 83–106.
[44] B. Bollobás, S. Janson, and O. Riordan, Sparse random graphs with clustering,
Preprint available from arXiv:0807.2040, 2008.
[45] B. Bollobás and B. D. McKay, The number of matchings in random regular graphs
and bipartite graphs, J. Combin. Theory Ser. B, 41 (1986), 80–91.
[46] B. Bollobás and O. Riordan, Constrained graph processes, Electron. J. Combin.,
7, Research Paper 18 (2000), 20 pp. (electronic).
[47] B. Bollobás and O. Riordan, Mathematical results on scale-free random graphs, in:
Handbook of Graphs and Networks, pages 1–34. Wiley-VCH, Weinheim, 2003.
[48] B. Bollobás and O. Riordan, Robustness and vulnerability of scale-free random
graphs, Internet Math., 1 (2003), 1–35.
[49] B. Bollobás and O. Riordan, The diameter of a scale-free random graph, Combi-
natorica, 24 (2004), 5–34.
[50] B. Bollobás and O. Riordan, Slow emergence of the giant component in the growing
m-out graph, Random Structures Algorithms, 27 (2005), 1–24.
[51] B. Bollobás and O. Riordan, Sparse graphs: metrics and random models, Preprint
available from arXiv:0708.1919, 2007.
[52] B. Bollobás and O. Riordan, Clique percolation, Preprint available from
arXiv:0804.0867, 2008.
[53] B. Bollobás, O. Riordan, J. Spencer, and G. Tusnády, The degree sequence of a
scale-free random graph process, Random Structures Algorithms, 18 (2001), 279–
290.
[54] B. Bollobás, A. Saito, and N. C. Wormald, Regular factors of regular graphs,
J. Graph Theory, 9 (1985), 97–103.
[55] C. Borgs, J. Chayes, L. Lovász, V. T. Sós, B. Szegedy, and K. Vesztergombi, Graph
limits and parameter testing, in: STOC’06: Proceedings of the 38th Annual ACM
Symposium on Theory of Computing, pages 261–270, New York, 2006. ACM.
[56] C. Borgs, J. Chayes, L. Lovász, V. T. Sós, and K. Vesztergombi, Counting graph
homomorphisms, in: Topics in Discrete Mathematics, volume 26 of Algorithms
Combin., pages 315–371. Springer, Berlin, 2006.
[57] C. Borgs, J. Chayes, L. Lovász, V. T. Sós, and K. Vesztergombi, Convergent
sequences of dense graphs I: Subgraph frequencies, metric properties and testing,
Adv. Math., 219 (2008), 1801–1851.
108 B. Bollobás and O. Riordan

[58] C. Borgs, J. Chayes, L. Lovász, V. T. Sós, and K. Vesztergombi, Convergent

sequences of dense graphs II: Multiway cuts and statistical physics, Preprint, 2007.
[59] V. E. Britikov, The structure of a random graph near a critical point, Diskret.
Mat., 1 (1989), 121–128.
[60] T. Britton, M. Deijfen, and A. Martin-Löf, Generating simple random graphs with
prescribed degree distribution, J. Stat. Phys., 124 (2006), 1377–1397.
[61] P. G. Buckley and D. Osthus, Popularity based random graph models leading to a
scale-free degree sequence, Discrete Math., 282 (2004), 53–68.
[62] J. Cain and N. Wormald, Encores on cores, Electron. J. Combin., 13, R81 (2006),
13 pp. (electronic).
[63] D. S. Callaway, J. E. Hopcroft, J. M. Kleinberg, M. E. J. Newman, and S. H.
Strogatz, Are randomly grown graphs really random? Phys. Rev. E., 64, 041902
(Sep. 2001).
[64] D. S. Callaway, M. E. J. Newman, S. H. Strogatz, and D. J. Watts, Network
robustness and fragility: Percolation on random graphs, Phys. Rev. Lett., 85 (Dec.
2000), 5468–5471.
[65] F. Chung and R. Graham, Sparse quasi-random graphs, Combinatorica, 22 (2002),
217–244.
[66] F. Chung and L. Lu, The diameter of sparse random graphs, Adv. in Appl. Math.,
26 (2001), 257–279.
[67] F. Chung and L. Lu, The average distances in random graphs with given expected
degrees, Proc. Natl. Acad. Sci. USA, 99 (2002), 15879–15882 (electronic).
[68] F. Chung and L. Lu, Connected components in random graphs with given expected
degree sequences, Ann. Comb., 6 (2002), 125–145.
[69] F. Chung and L. Lu, The average distance in a random graph with given expected
degrees, Internet Math., 1 (2003), 91–113.
[70] F. Chung and L. Lu, The volume of the giant component of a random graph with
given expected degrees, SIAM J. Discrete Math., 20 (2006), 395–411 (electronic).
[71] F. Chung, L. Lu, and V. Vu, Spectra of random graphs with given expected degrees,
Proc. Natl. Acad. Sci. USA, 100 (2003), 6313–6318 (electronic).
[72] V. Chvátal, Almost all graphs with 1.44n edges are 3-colorable, Random Structures
Algorithms, 2 (1991), 11–28.
[73] R. Cohen, K. Erez, D. ben Avraham, and S. Havlin, Resilience of the internet to
random breakdowns, Phys. Rev. Lett., 85 (Nov. 2000), 4626–4628.
[74] R. Cohen, K. Erez, D. ben Avraham, and S. Havlin, Breakdown of the internet
under intentional attack, Phys. Rev. Lett., 86 (Apr. 2001), 3682–3685.
[75] R. Cohen and S. Havlin, Scale-free networks are ultrasmall, Phys. Rev. Lett., 90
(Feb. 2003), 058701.
[76] A. Coja-Oghlan, C. Moore, and V. Sanwalani, Counting connected graphs and hy-
pergraphs via the probabilistic method, Random Structures Algorithms, 31 (2007),
288–329.
Random Graphs and Branching Processes 109

[77] C. Cooper, The cores of random hypergraphs with a given degree sequence, Random
Structures Algorithms, 25 (2004), 353–375.
[78] C. Cooper and A. Frieze, A general model of web graphs, Random Structures
Algorithms, 22 (2003), 311–335.
[79] C. Cooper, A. Frieze, and B. Reed, Random regular graphs of non-constant degree:
connectivity and Hamiltonicity, Combin. Probab. Comput., 11 (2002), 249–261.
[80] C. Cooper, A. Frieze, B. Reed, and O. Riordan, Random regular graphs of non-
constant degree: independence and chromatic number, Combin. Probab. Comput.,
11 (2002), 323–341.
[81] J. T. Cox and R. Durrett, The stepping stone model: new formulas expose old
myths, Ann. Appl. Probab., 12 (2002), 1348–1377.
[82] R. Darling and J. Norris, Differential equation approximations for markov chains,
Probab. Surv., 5 (2008), 37–79.
[83] I. Derényi, G. Palla, and T. Vicsek, Clique percolation in random networks, Physical
Review Letters, 94 (2005), 160202.
[84] L. Devroye, C. McDiarmid, and B. Reed, Giant components for two expanding
graph processes, in: Mathematics and Computer Science, II (Versailles, 2002),
Trends Math., pages 161–173. Birkhäuser, Basel, 2002.
[85] S. Dorogovtsev and J. Mendes, Evolution of networks, Adv. Phys., 51 (2002), 1079–
1187.
[86] S. N. Dorogovtsev, A. V. Goltsev, and J. F. F. Mendes, k-core architecture and
k-core percolation on complex networks, Phys. D, 224 (2006), 7–19.
[87] S. N. Dorogovtsev and J. F. F. Mendes, Evolution of networks: From biological nets
to the Internet and WWW, Oxford University Press, Oxford, 2003.
[88] S. N. Dorogovtsev, J. F. F. Mendes, and A. N. Samukhin, Anomalous percolation
properties of growing networks, Phys. Rev. E., 64 (Nov. 2001), 066110.
[89] R. Durrett, Rigorous result for the CHKNS random graph model, in: Discrete
Random Walks (Paris, 2003), Discrete Math. Theor. Comput. Sci. Proc., AC, pages
95–104 (electronic). Assoc. Discrete Math. Theor. Comput. Sci., Nancy, 2003.
[90] R. Durrett and H. Kesten, The critical parameter for connectedness of some ran-
dom graphs, in: A tribute to Paul Erdős, pages 161–176. Cambridge Univ. Press,
Cambridge, 1990.
[91] R. Durrett and M. Restrepo, One-dimensional stepping stone models, sardine
genetics and Brownian local time, Ann. Appl. Probab., 18 (2008), 334–358.
[92] P. Erdős, On some new inequalities concerning extremal properties of graphs, in:
Theory of Graphs (Proc. Colloq., Tihany, 1966), pages 77–81. Academic Press,
New York, 1968.
[93] P. Erdős and A. Rényi, On random graphs. I, Publ. Math. Debrecen, 6 (1959),
290–297.
[94] P. Erdős and A. Rényi, On the evolution of random graphs, Magyar Tud. Akad.
Mat. Kutató Int. Közl., 5 (1960), 17–61.
110 B. Bollobás and O. Riordan

[95] P. Erdős and A. Rényi, On the strength of connectedness of a random graph, Acta
Math. Acad. Sci. Hungar., 12 (1961), 261–267.
[96] P. Erdős and A. Rényi, On random matrices, Magyar Tud. Akad. Mat. Kutató Int.
Közl, 8 (1964), 455–461 (1964).
[97] P. Erdős and A. Rényi, On the existence of a factor of degree one of a connected
random graph, Acta Math. Acad. Sci. Hungar., 17 (1966), 359–368.
[98] P. Erdős and A. Rényi, On random matrices. II, Studia Sci. Math. Hungar., 3
(1968), 459–464.
[99] P. Erdős, S. Suen, and P. Winkler, On the size of a random maximal graph, Random
Structures Algorithms, 6 (1995), 309–318.
[100] A. Fabrikant, E. Koutsoupias, and C. H. Papadimitriou, Heuristically optimized
trade-offs: A new paradigm for power laws in the internet, in: ICALP ’02: Proceed-
ings of the 29th International Colloquium on Automata, Languages and Program-
ming, pages 110–122, London, UK, 2002. Springer-Verlag.
[101] M. Faloutsos, P. Faloutsos, and C. Faloutsos, On power-law relationships of the in-
ternet topology, in: SIGCOMM ’99: Proceedings of the conference on Applications,
technologies, architectures, and protocols for computer communication, pages 251–
262, New York, NY, USA, 1999. ACM.
[102] T. I. Fenner and A. M. Frieze, Hamiltonian cycles in random regular graphs,
J. Combin. Theory Ser. B, 37 (1984), 103–112.
[103] D. Fernholz and V. Ramachandran, Cores and connectivity in sparse random
graphs, Technical Report UTCS TR04–13, Department of Computer Science, Uni-
versity of Texas at Austin, 2004.
[104] D. Fernholz and V. Ramachandran, The diameter of sparse random graphs, Ran-
dom Structures Algorithms, 31 (2007), 482–516.
[105] G. W. Ford, R. Z. Norman, and G. E. Uhlenbeck, Combinatorial problems in the
theory of graphs. II, Proc. Nat. Acad. Sci. U. S. A., 42 (1956), 203–208.
[106] G. W. Ford and G. E. Uhlenbeck, Combinatorial problems in the theory of graphs. I,
Proc. Nat. Acad. Sci. U. S. A., 42 (1956), 122–128.
[107] G. W. Ford and G. E. Uhlenbeck, Combinatorial problems in the theory of
graphs. III, Proc. Nat. Acad. Sci. U.S.A., 42 (1956), 529–535.
[108] G. W. Ford and G. E. Uhlenbeck, Combinatorial problems in the theory of
graphs. IV, Proc. Nat. Acad. Sci. U.S.A., 43 (1957), 163–167.
[109] N. Fountoulakis, Percolation on sparse random graphs with given degree sequence,
Preprint available from arXiv:math/0703269, 2007.
[110] A. Frieze and R. Kannan, Quick approximation to matrices and applications,
Combinatorica, 19 (1999), 175–220.
[111] A. Frieze, J. Kleinberg, R. Ravi, and W. Debany, Line-of-sight networks, in: Proc.
18th ACM-SIAM Symposium on Discrete Algorithms, pages 968–977, 2007.
[112] A. Frieze, M. Krivelevich, and C. Smyth, On the chromatic number of random
graphs with a fixed degree sequence, Combin. Probab. Comput., 16 (2007), 733–
746.
Random Graphs and Branching Processes 111

[113] A. M. Frieze, Finding Hamilton cycles in sparse random graphs, J. Combin. Theory
Ser. B, 44 (1988), 230–250.
[114] A. M. Frieze, Random regular graphs of non-constant degree, Technical Report,
Department of Mathematical Sciences, Carnegie Mellon University, 1988.
[115] E. N. Gilbert, Random graphs, Ann. Math. Statist., 30 (1959), 1141–1144.
[116] N. Gilbert, A simulation of the structure of academic science, Sociological Research
Online, 2, 1997.
[117] A. V. Goltsev, S. N. Dorogovtsev, and J. F. F. Mendes, k-core (bootstrap) percola-
tion on complex networks: critical phenomena and nonlocal effects, Phys. Rev. E.,
73, 10 (2006), 056101.
[118] C. Greenhill, A. Ruciński, and N. C. Wormald, Random hypergraph processes with
degree restrictions, Graphs Combin., 20 (2004), 319–332.
[119] T. E. Harris, The theory of branching processes, Die Grundlehren der Mathematis-
chen Wissenschaften, Bd. 119. Springer-Verlag, Berlin, 1963.
[120] T. E. Harris, The theory of branching processes, Dover Phoenix Editions. Dover
Publications Inc., Mineola, NY, 2002, Corrected reprint of the 1963 original
[Springer, Berlin].
[121] S. Janson, On a random graph related to quantum theory, Combin. Probab. Com-
put., 16 (2007), 757–766.
[122] S. Janson, On percolation in random graphs with given vertex degrees, Electron.
J. Probab., 14 (2009), 87–118.
[123] S. Janson, Standard representation of multivariate functions on a general proba-
bility space, Preprint available from arXiv:0801.0196, 2008.
[124] S. Janson, D. E. Knuth, T. L
 uczak, and B. Pittel, The birth of the giant component,
Random Structures Algorithms, 4 (1993), 231–358. With an introduction by the
editors.
[125] S. Janson and M. J. Luczak, A simple solution to the k-core problem, Random
Structures Algorithms, 30 (2007), 50–62.
[126] S. Janson and M. J. Luczak, A new approach to the giant component problem,
Random Structures Algorithms, 34 (2009), 197–216.
[127] S. Janson, T. L
 uczak, and A. Rucinski, Random graphs, Wiley-Interscience Series
in Discrete Mathematics and Optimization. Wiley-Interscience, New York, 2000.
[128] S. Janson and J. Spencer, A point process describing the component sizes in the
critical window of the random graph evolution, Combin. Probab. Comput., 16
(2007), 631–658.
[129] S. Kalikow and B. Weiss, When are random graphs connected, Israel J. Math., 62
(1988), 257–268.
[130] M. Kang and T. G. Seierstad, The critical phase for random graphs with a given
degree sequence, Combin. Probab. Comput., 17 (2008), 67–86.
[131] R. M. Karp, The transitive closure of a random digraph, Random Structures
Algorithms, 1 (1990), 73–93.
112 B. Bollobás and O. Riordan

[132] L. Katz, Probability of indecomposability of a random mapping function, Ann.

Math. Statist., 26 (1955), 512–517.
[133] D. G. Kendall, Deterministic and stochastic epidemics in closed populations, in:
Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Prob-
ability, 1954–1955, vol. IV, pages 149–165, Berkeley and Los Angeles, 1956. Uni-
versity of California Press.
[134] D. G. Kendall, Branching processes since 1873, J. London Math. Soc., 41 (1966),
385–406.
[135] V. I. Khokhlov and V. F. Kolchin, On the structure of a random graph with nonuni-
form distribution, in: New trends in probability and statistics, Vol. 1 (Bakuriani,
1990), pages 445–456. VSP, Utrecht, 1991.
[136] M. Kimura, ”stepping stone” model of population, Ann. Rep. Nat. Inst. Genet.
Japan, 3 (1953), 62–63.
[137] V. F. Kolchin and V. I. Khokhlov, On the number of cycles in a random non-
equiprobable graph, Diskret. Mat., 2 (1990), 137–145.
[138] M. Krivelevich, B. Sudakov, V. H. Vu, and N. C. Wormald, Random regular graphs
of high degree, Random Structures Algorithms, 18 (2001), 346–363.
[139] R. Kumar, P. Raghavan, S. Rajagopalan, D. Sivakumar, A. Tomkins, and E. Upfal,
Stochastic models for the web graph, in: 41st Annual Symposium on Foundations
of Computer Science (Redondo Beach, CA, 2000), pages 57–65. IEEE Comput.
Soc. Press, Los Alamitos, CA, 2000.
[140] A. Lotka, The frequency distribution of scientific productivity, J. Washington Acad.
Sci., 16 (1926), 317.
[141] L. Lovász and B. Szegedy, Limits of dense graph sequences, J. Combin. Theory
Ser. B, 96 (2006), 933–957.
[142] L. Lovász and V. T. Sós, Generalized quasirandom graphs, J. Combin. Theory
Ser. B, 98 (2008), 146–163.
[143] T. L
 uczak, Component behavior near the critical point of the random graph
process, Random Structures Algorithms, 1 (1990), 287–310.
[144] T. L
 uczak, Random trees and random graphs, Random Structures Algorithms, 13
(1998), 485–500.
[145] T. L
 uczak, B. Pittel, and J. C. Wierman, The structure of a random graph at the
point of the phase transition, Trans. Amer. Math. Soc., 341 (1994), 721–748.
[146] T. L
 uczak and J. C. Wierman, The chromatic number of random graphs at the
double-jump threshold, Combinatorica, 9 (1989), 39–49.
[147] B. D. McKay, Asymptotics for 0–1 matrices with prescribed line sums, in: Enumer-
ation and design (Waterloo, Ont., 1982), pages 225–238. Academic Press, Toronto,
ON, 1984.
[148] B. D. McKay and N. C. Wormald, Asymptotic enumeration by degree sequence of
graphs of high degree, European J. Combin., 11 (1990), 565–580.
[149] B. D. McKay and N. C. Wormald, Uniform generation of random regular graphs
of moderate degree, J. Algorithms, 11 (1990), 52–67.
Random Graphs and Branching Processes 113

[150] M. Molloy, Cores in random hypergraphs and Boolean formulas, Random Structures
Algorithms, 27 (2005), 124–135.
[151] M. Molloy and B. Reed, A critical point for random graphs with a given degree
sequence, Random Structures Algorithms, 6 (1995), 161–179.
[152] M. Molloy and B. Reed, The size of the giant component of a random graph with
a given degree sequence, Combin. Probab. Comput., 7 (1998), 295–305.
[153] A. Nachmias and Y. Peres, Component sizes of the random graph outside the scal-
ing window, ALEA Lat. Am. J. Probab. Math. Stat., 3 (2007), 133–142 (electronic).
[154] M. E. J. Newman, S. H. Strogatz, and D. J. Watts, Random graphs with arbitrary
degree distributions and their applications, Phys. Rev. E., 64 (Jul. 2001), 026118.
[155] I. Norros and H. Reittu, On a conditionally Poissonian graph process, Adv. in Appl.
Probab., 38 (2006), 59–75.
[156] D. Osthus and A. Taraz, Random maximal H-free graphs, Random Structures
Algorithms, 18 (2001), 61–82.
[157] G. Palla, D. Ábel, I. Farkas, P. Pollner, I. Derényi, and T. Vicsek, k-clique perco-
lation and clustering, this volume (2009), 369–408.
[158] G. Palla, I. Derényi, and T. Vicsek, The critical point of k-clique percolation in
the Erdős-Rényi graph, J. Stat. Phys., 128 (2007), 219–227.
[159] G. Palla, I. Farkas, P. Pollner, I. Derényi, and T. Vicsek, Directed network modules,
New Journal of Physics, 9 (2007), 186 (21 pages).
[160] M. D. Penrose, On the spread-out limit for bond and continuum percolation, Ann.
Appl. Probab., 3 (1993), 253–276.
[161] B. Pittel, On tree census and the giant component in sparse random graphs,
Random Structures Algorithms, 1 (1990), 311–342.
[162] B. Pittel, On the largest component of the random graph at a nearcritical stage,
J. Combin. Theory Ser. B, 82 (2001), 237–269.
[163] B. Pittel, J. Spencer, and N. Wormald, Sudden emergence of a giant k-core in a
random graph, J. Combin. Theory Ser. B, 67 (1996), 111–151.
[164] A. Rényi, On connected graphs. I, Magyar Tud. Akad. Mat. Kutató Int. Közl., 4
(1959), 385–388.
[165] R. J. Riddell, Jr. and G. E. Uhlenbeck, On the theory of the virial development of
the equation of state of mono-atomic gases, J. Chem. Phys., 21 (1953), 2056–2064.
[166] O. Riordan, The small giant component in scale-free random graphs, Combin.
Probab. Comput., 14 (2005), 897–938.
[167] O. Riordan, The k-core and branching processes, Combin. Probab. Comput., 17
(2008), 111–136.
[168] O. Riordan, The mathematics of the Barabási–Albert network model, Lecture Notes
of the Institute of Mathematical Sciences, University of Singapore, to appear, 2008.
[169] O. Riordan and N. Wormald, The diameter of G(n, p), Preprint available from
arXiv:0808.4067, 2008.
114 B. Bollobás and O. Riordan

[170] R. W. Robinson and N. C. Wormald, Almost all regular graphs are Hamiltonian,
Random Structures Algorithms, 5 (1994), 363–374.
[171] A. Ruciński and N. C. Wormald, Random graph processes with degree restrictions,
Combin. Probab. Comput., 1 (1992), 169–180.
[172] A. Ruciński and N. C. Wormald, Random graph processes with maximum degree
2, Ann. Appl. Probab., 7 (1997), 183–199.
[173] A. Ruciński and N. C. Wormald, Connectedness of graphs generated by a random
d-process, J. Aust. Math. Soc., 72 (2002), 67–85.
[174] J. Schmidt-Pruzan and E. Shamir, Component structure in the evolution of random
hypergraphs, Combinatorica, 5 (1985), 81–94.
[175] A. D. Scott and G. B. Sorkin, Solving sparse random instances of Max Cut and
Max 2-CSP in linear expected time, Combin. Probab. Comput., 15 (2006), 281–315.
[176] L. A. Shepp, Connectedness of certain random graphs, Israel J. Math., 67 (1989),
23–33.
[177] H. Simon, On a class of skew distribution functions, Biometrika, 42 (1955), 425–
440.
[178] M. Simonovits, A method for solving extremal problems in graph theory, stability
problems, in: Theory of Graphs (Proc. Colloq., Tihany, 1966), pages 279–319.
Academic Press, New York, 1968.
[179] B. Söderberg, General formalism for inhomogeneous random graphs, Phys. Rev. E.,
66, 6 (2002), 066121.
[180] B. Söderberg, Random graphs with hidden color, Phys. Rev. E., 68(R) (2003),
015102.
[181] B. Söderberg, Properties of random graphs with hidden color, Phys. Rev. E., 68,
12 (2003), 026107.
[182] B. Söderberg, Random graph models with hidden color, Acta Physica Polonica B.,
34 (2003), 5085–5102.
[183] V. E. Stepanov, The probability of the connectedness of a random graph Gm (t),
Teor. Verojatnost. i Primenen., 15 (1970), 58–68.
[184] V. E. Stepanov, Phase transitions in random graphs, Teor. Verojatnost. i Prime-
nen., 15 (1970), 200–216.
[185] V. E. Stepanov, Structure of the random graphs Gm (xh), Teor. Verojatnost.
i Primenen., 17 (1972), 238–252.
[186] V. E. Stepanov, Some features of the structure of a random graph near a critical
point, Teor. Veroyatnost. i Primenen., 32 (1987), 633–657.
[187] T. S. Turova, Dynamical random graphs with memory, Phys. Rev. E., 65, 9 (2002),
066102.
[188] T. S. Turova, Long paths and cycles in dynamical graphs, J. Statist. Phys., 110
(2003), 385–417.
[189] T. S. Turova, Phase transitions in dynamical random graphs, J. Stat. Phys., 123
(2006), 1007–1032.
Random Graphs and Branching Processes 115

[190] T. S. Turova, Continuity of the percolation threshold in randomly grown graphs,

Electron. J. Probab., 12 (2007), 1036–1047 (electronic).
[191] H. van den Esker, R. van der Hofstad, G. Hooghiemstra, and D. Znamenski,
Distances in random graphs with infinite mean degrees, Extremes, 8 (2005), 111–
141 (2006).
[192] R. van der Hofstad, G. Hooghiemstra, and P. Van Mieghem, Distances in random
graphs with finite variance degrees, Random Structures Algorithms, 27 (2005), 76–
123.
[193] R. van der Hofstad, G. Hooghiemstra, and D. Znamenski, Distances in random
graphs with finite mean and infinite variance degrees, Electron. J. Probab., 12
(2007), 703–766 (electronic).
[194] H. Watson and F. Galton, On the probability of the extinction of families, Journal
of the Anthropological Institute of Great Britain, 4 (1875), 138–144.
[195] D. Watts and S. Strogatz, Collective dynamics of ‘small-world’ networks, Nature,
393 (1998), 440–442.
[196] E. M. Wright, The number of connected sparsely edged graphs, J. Graph Theory,
1 (1977), 317–330.
[197] E. M. Wright, The number of connected sparsely edged graphs. II. Smooth graphs
and blocks, J. Graph Theory, 2 (1978), 299–305.
[198] E. M. Wright, The number of connected sparsely edged graphs. III. Asymptotic
results, J. Graph Theory, 4 (1980), 393–407.
[199] E. M. Wright, The number of connected sparsely edged graphs. IV. Large nonsep-
arable graphs, J. Graph Theory, 7 (1983), 219–229.
[200] I. Zähle, J. T. Cox, and R. Durrett, The stepping stone model. II. Genealogies and
the infinite sites model, Ann. Appl. Probab., 15 (2005), 671–699.
[201] G. Zipf, Human behavior and the principle of least effort, Hafner, New York, 1949.

Béla Bollobás Oliver Riordan

Department of Pure Mathematics and Mathematical Institute
Mathematical Statistics University of Oxford
University of Cambridge 24–29 St Giles’
Cambridge CB3 0WB, UK Oxford OX1 3LB, UK
and e-mail: [email protected]
Department of Mathematical Sciences
University of Memphis
Memphis TN 38152, USA
e-mail:
[email protected]
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 117–142.

Chapter 2
Percolation, Connectivity, Coverage and
Colouring of Random Geometric Graphs

PAUL BALISTER, BÉLA BOLLOBÁS∗ and AMITES SARKAR

In this review paper, we shall discuss some recent results concerning several
models of random geometric graphs, including the Gilbert disc model Gr , the
k-nearest neighbour model Gnn k and the Voronoi model GP . Many of the results
concern finite versions of these models. In passing, we shall mention some of the
applications to engineering and biology.

1. Introduction

Place a million points uniformly at random in a large square and connect

every point to the six points closest to it. What can we say about the re-
sulting graph? Is it connected, and, if not, does it contain a connected
component with at least a hundred thousand vertices? In this paper, we
consider such questions for some of the most natural models of a random
geometric graph, including the one above. From a practical point of view,
these graphs are excellent models for ad-hoc wireless networks, in which
some radio transceivers lie scattered over a large region, and where each
transceiver can only communicate with a few others nearby. From a more
mathematical standpoint, the models act as a bridge between the theory
∗
Research supported in part by NSF grants DMS-0906634, CNS-0721983 and CCF-
0728928, and ARO grant W911NF-06-1-0076.
118 P. Balister, B. Bollobás and A. Sarkar

of classical random graphs [17] and that of percolation [18], and our study
of them will draw inspiration and use tools from both these more estab-
lished fields. (It is tempting to write “older” fields, but in fact Gilbert’s
pioneering papers [31, 32] appeared only shortly after those of Broadbent
and Hammersley on percolation, and those of Erdős and Rényi on random
graphs.)
For all the models, we will take the vertex set to be a unit intensity
Poisson process P in R2 , or the restriction of P to a square, so it is
convenient to make some remarks about this at the outset. All our finite
results will carry over to the case of uniformly distributed points, but for
their proofs, and for the statements of our infinite results, it is far better to
use a Poisson process.
There are several ways of defining such a process, of which the following
is perhaps the easiest to describe. First recall that a Poisson random
variable of mean λ is a discrete random variable X for which

λk
P(X = k) = e−λ .
k!

As is usual, we will denote this by X ∼ Po (λ). Now tessellate R2 with

unit squares, and consider a family (Xi ) of independent random variables,
indexed by the squares, where each Xi ∼ Po (1). We place Xi points
uniformly at random in the square i. The result is a unit intensity Poisson
process P.
One of the key features of this model is its independence: the number
of points of P in a measurable region A ⊂ R2 is a Poisson random variable
with mean |A| (the Lebesgue measure of A), regardless of what is happening
outside A. Moreover, conditioned on there being k points in A, their
distribution is uniform. See [18] for more background.
Throughout the paper, the phrase “with high probability” will mean
“with probability tending to one as n → ∞”. Also, all logarithms in this
paper are to the base e.

2. The Gilbert Disc Model

Our first model was introduced and studied by E. N. Gilbert in 1961 [31],
and has since become known as the disc model or Gilbert model. To define
Percolation, Connectivity, Coverage and Colouring 119

it, fix r > 0, let P be a Poisson process of intensity one in the plane R2 ,

and connect two points of P by an edge if the distance between them is less
than r. Denote the resulting infinite graph by Gr .

2.1. Percolation

Although Gilbert’s main focus was the study of communications networks,

he noted that Gr could also model the spread of a contagious disease. For the
second application, and perhaps also for the first, one is primarily interested
in percolation, in the following sense. Let us suppose that, without loss of
generality, the origin is one of the points of P. Writing a = πr2 , and θ(a) for
the probability that the origin belongs to an infinite connected component
of Gr , Gilbert defined the critical area ac as
 
ac = sup a : θ(a) = 0 .

In other words, for a > ac , there is a non-zero probability that the disease
spreads, or that communication is possible to some arbitrarily distant nodes
of the network. In this case we say that the model percolates. Since θ(a)
is clearly monotone, θ(a) = 0 if a < ac . Currently the best known bounds,
due to Hall [36], are
2.184 ≤ ac ≤ 10.588,
although in a recent paper Balister, Bollobás and Walters [13] used 1-inde-
pendent percolation to show that, with confidence 99.99%,

4.508 ≤ ac ≤ 4.515,

which is consistent with the non-rigorous bounds

4.51218 ≤ ac ≤ 4.51228

obtained by Quintanilla, Torquato and Ziff [50].

We can consider the same problem in d dimensions, where we use balls of
volume v rather than discs, and define θ d (v) and vcd in the obvious manner.
Penrose [47] proved the following.

Theorem 1.
vcd → 1 as d → ∞.
120 P. Balister, B. Bollobás and A. Sarkar

Most of the volume in a high-dimensional ball is close to the boundary,

and hence one might expect that the same conclusion holds for a two-
dimensional annulus where the ratio of the inner and outer radii tends to 1.
This is indeed true, and was proved independently by Balister, Bollobás
and Walters [12] and Franceschetti, Booth, Cook, Meester and Bruck [28].
However, as shown in [12], the corresponding result for square annuli is false.
A general condition under which the critical area tends to 1 is given in [14].

2.2. Connectivity

Penrose [48, 49] considered the following finite version of Gr . For this, we
only consider points of P lying in a fixed square Sn of area n, again joining
two points if the distance between them is less than r. Penrose proved the
following result on the connectivity of the resulting model Gr (n).

Theorem 2. If πr2 = log n + α then

 −α
P Gr (n) is connected) → e−e .

In particular, the above probability tends to 1 iff α → ∞.

This result has an exact analogue in the theory of classical random

graphs [17]. Indeed, in both cases the obstructions to connectivity are
isolated vertices. In fact, for both models, it is not hard to calculate the
expected number of isolated vertices, and then to show that their number
has a distribution that is approximately Poisson. The hard part is to show
that there are no other obstructions. For Gr (n), Penrose first shows that
the obstructions must be small, that is, of area at most C log n (with our
normalization). This he achieves by discretization. Since many of the proofs
of the theorems we will discuss use this technique, we give a brief account of
it, for this case. The basic
√ idea is to tessellate our large square with smaller
squares of side length r/ 5. Any component in Gr (n) must be surrounded
by a connected path consisting of, say, l vacant squares, none of which can
contain any points of P. Even though the number of such paths of squares is
exponential in l, if the component is large (so that l ≥ K for some absolute
constant K), it is not hard to show that the probability of such a vacant
path existing anywhere in Sn tends to zero. Thus, with high probability, if
Gr (n) is disconnected, it contains a small component. Penrose completes his
proof with a delicate local argument, showing that, for the relevant range
Percolation, Connectivity, Coverage and Colouring 121

of values of r, this small component is, with high probability, an isolated

vertex.
As shown by Penrose, the story for k-connectivity also mirrors that for
classical random graphs, in that the principal obstructions are vertices of
degree exactly k − 1. For detailed statements and proofs, the reader is
referred to [49].
There are various ways to generalize this model. One, treated thor-
oughly in [43], is to choose the disc radii to be independent and identically
distributed random variables. Another possibility is to keep the radii fixed
at r, but vary the intensity of the underlying Poisson process. In one such
model, suggested by Etherington, Hoge and Parkes [24], the intensity ρ(x)
of P at distance x from the origin is given by a gaussian distribution, so
that
n
ρ(x) = e−x .
2

π
This model GGauss
r (n) was analyzed in detail by Balister, Bollobás, Sarkar
and Walters [8], who determined the threshold for connectivity.
Theorem 3. If
 1
2r log n = log log n − log log log n + f (n),
2
then, with high probability, GGauss
r (n) is connected if f (n) → ∞ and
disconnected if f (n) → −∞.

2.3. Coverage

For most of the remainder of this section, we will imagine that the points
of our Poisson process P are sensors designed to monitor a large square
region Sn of area n. Such monitoring is feasible if the sensing discs Dr (p)
cover Sn , so that

Sn ⊂ Dr (p).
p∈P

How large should we make r = r(n) so that this occurs with high probabil-
ity?
Before turning to recent results, we consider the original application
of Moran and Fazekas de St Groth [45]. They considered the problem of
covering the surface of a sphere with circular caps, and write:
122 P. Balister, B. Bollobás and A. Sarkar

This problem arises in practice in the study of the theory of the manner
in which antibodies prevent virus particles from attacking cells. Thus an
influenza virus may be considered to be a sphere of radius about 40 mμ.
Antibodies are supposed to be cigar-shaped molecules of length about
27 mμ and of a thickness which will be neglected. The antibodies are
assumed to attach themselves at their ends to the virus particle, standing
up rigidly on the surface and thus shielding a circular area on the virus
from possible contact with the surface of a cell.

Also noteworthy is their simulation method:

. . . an experiment was carried out using table tennis balls. These had a
mean diameter of 37.2mm. with a standard deviation around this mean
of 0.02mm. One hundred holes of diameter 29.9mm. were punched in an
aluminium sheet forming one side of a flat box. The balls were held firmly
against the holes by a foam rubber pad, and sprayed with a duco paint.
After drying they were removed and replaced at random by hand. Forty
sprayings were done in each of three sets of 100 balls.

This was also one of the problems considered by Gilbert [32], who
performed his simulations on an IBM 7094 computer. His paper contains
the following critical observation, which we will state in the context of our
original formulation of the problem. For the (open) discs Dr (p) to cover Sn ,
it is not only necessary but also sufficient that the following three conditions
hold:
• Every intersection of 2 disc boundaries inside Sn is covered by a third
disc
• Every intersection of a disc boundary with ∂Sn is covered by a second
disc
• There is at least one such intersection (of either type)
Hall [35] used this observation to establish the following criterion.

Theorem 4. If
πr2 = log n + log log n + f (n),
then a necessary and sufficient condition for the discs Dr (p) to cover Sn
with high probability is that f (n) → ∞.

The proof proceeds by showing that if r is as in the statement of The-

orem 4, then the expected number of uncovered intersections is asymptot-
ically 4e−f (n) . Thus if f (n) → ∞, by Gilbert’s observation, we obtain
Percolation, Connectivity, Coverage and Colouring 123

coverage with high probability. For the other direction, Hall applies the
second moment method (to the uncovered area).
Slightly later, Janson [38] obtained very general results on the probabil-
ity of coverage. For our case, his result is as follows.

Theorem 5. If
πr2 = log n + log log n + x,

then as n → ∞
−x
P(Sn is covered) → e−e .

Recently, a shorter proof of Theorem 5, with bounds on the error term,

was obtained by Balister, Bollobás, Sarkar and Walters [11]. The idea is
that, while the uncovered intersections occur in groups, these groups consist
simply of the intersections bordering the uncovered regions, which are small
(area C/ log n), and essentially form their own Poisson process of intensity
e−f (n) . (It is very unlikely that two such uncovered regions are close, because
the discs bordering them are, on their scale, almost half-planes. Moreover,
the expected number of sides of an uncovered region is the same as that
of any other “atomic” region, namely 4. To make these heuristics rigorous,
one can use the Stein–Chen method [2].)

2.4. Colouring

Both Hall [35] and Janson [38] considered not only the case of coverage, but
also that of k-coverage. Our square Sn is said to be k-covered by the discs
Dr (p) if every point of Sn is contained in at least k discs. This property is
useful for sensor networks, since it allows for the possibility that up to k − 1
sensors in a small region might simultaneously fail. Now, in our model, for
a fixed instance of P, suppose that we increase r until Sn is covered. It
turns out that just a small additional increase in r ensures k-coverage.

Theorem 6 [11, 38]. For any fixed k ≥ 1, if

πr2 = log n + k log log n + x,

then as n → ∞
−x /(k−1)!
P(Sn is k-covered) → e−e .
124 P. Balister, B. Bollobás and A. Sarkar

However, suppose we are more optimistic and instead request the fol-
lowing. We would like to devise a rota system so that each sensor can sleep
for most of the time, for example, to extend battery life. A natural way of
doing this would be to colour the set of sensors with k colours, and arrange
that only the sensors with colour  are active in the th time slot. After k
time slots have expired, we repeat the process. In order to detect an event
occurring anywhere and at any time, it is necessary that the sensors in each
colour class themselves form a single cover of Sn . Thus our question be-
comes: for fixed k, how large should r be to ensure that the sensors can be
partitioned into k groups, each of which covers the sensing region? We call
this the problem of sentry selection, since each of the groups is a group of
sentries keeping watch over the region while the others are sleeping.
It is important to note that a k-cover of an arbitrary set cannot always
be partitioned into k single covers. For instance, let S be the set of all
subsets of A = {1, 2, . . . , n} of size k. The n sets Si = {B ∈ S : i ∈ B},
1 ≤ i ≤ n, form a k-cover of S which cannot even be partitioned into two
single covers if n ≥ 2k − 1. This example shows that a solution to our
problem must make some use of its geometric setting. Also, even restricting
ourselves to discs of equal radii, it is possible to construct k-covers of the
plane that are not (2k + 2)/3-partitionable. Thus we must also make use
of the probabilistic setting.
Let n, r ∈ R. For k ∈ N, write Erk for the event that the discs Dr (p) form
a k-cover of Sn , and Frk for the event that they may be partitioned into k
single covers of Sn . Balister, Bollobás, Sarkar and Walters [11] proved that
most random k-covers are in fact k-partitionable.

Theorem 7. With notation as above,

  ck
P Erk \ Frk ≤ ,
log n

for some constant ck .

They also proved that this is sharp, up to the value of the constant ck .
Two hitting time versions of Theorem 7 are also obtained: if we fix n and
slowly increase r, or if we fix r and add points uniformly at random to a given
area, then with high probability, k-partitionability occurs as soon as we have
k-coverage. In particular, Theorem 6 holds also for k-partitionability.
Let us suppose that πr2 ≥ log n + (k − 12 ) log log n and attempt to prove
Theorem 7. For this range of values of r, a typical point in Sn is covered
Percolation, Connectivity, Coverage and Colouring 125

at least log n times. Intuitively, in most of Sn , we can simply colour the

discs randomly, and the probability of failure, that is, of a point x ∈ Sn
not being covered by discs of every colour, will be negligible. Indeed, if the
level of coverage is at least 3k log log n everywhere, we can apply the Lovász
local lemma to prove that a suitable colouring exists. However, there will
be many regions in Sn which are covered less than 3k log log n times. Call
these thinly covered regions. It turns out that, with high probability, such
regions occur in small, well-separated clusters. At the scale of the clusters,
the curvature of the discs is negligible, so that they behave like half-planes.
Let us examine one such cluster. We will probably find some very thinly
covered regions, which are covered less than 3k times. These turn out to
have a very useful property (with high probability): some set of k − 1
discs D1 , . . . , Dk−1 covers all of them. This facilitates the following simple
deterministic colouring method. Suppose that all the discs are actually
half-planes. Remove the Di , and suppose that we still have a cover of Sn
(otherwise, we did not have a k-cover to begin with). By Helly’s theorem,
we can find three of the half-planes which cover Sn , which we colour with
colour k and remove. Now restore Dk−1 and repeat the procedure, colouring
three half-planes with colour k − 1 before removing them. Because of the
property mentioned above, we can repeat the process k − 2 times, using
all the colours, and the level of coverage in the cluster will never drop to
zero until we have finished. We do this for every cluster, and, outside the
clusters, we complete the colouring using the Lovász local lemma, as before.
The actual proofs require somewhat more detailed estimates than the
above sketch suggests. As a by-product, we can identify the principal
obstructions to k-partitionability in a k-cover as small non-partitionable
k-covered configurations which are covered by k − 2 common discs. Since
these configurations are very small, the curvature of the discs forming them
is negligible, so that our obstructions are essentially 2-covers with half-
planes which cannot be partitioned into two single covers. It is therefore
of interest to classify such configurations. Such a classification is presented
in [11].

2.5. Thin strips

Suppose that instead of examining points of a Poisson process P inside a

large square, we instead consider the restriction of P to a thin strip T . As
before, we will join two points of P at distance less than r. Such a model
126 P. Balister, B. Bollobás and A. Sarkar

was suggested by the engineering problem of building an electronic “fence”

surrounding a large region. The points of P are sensors, and the fence con-
sists of a thin strip of sensors bordering the region, which has the ability
to detect intruders if there is no continuous path crossing it, no point of
which lies within distance r/2 of any sensor. Note that this is a different
condition from both connectivity of the underlying graph Gr [T ] of sensors
and coverage of the sensing regions (of radius r/2). Indeed, if T is a long
thin rectangle, our requirement
 is weaker than both connectivity of Gr [T ]
and coverage of T by p∈P∩T Dr/2 (p). However, it is not hard to see that
the new condition is both necessary and sufficient not only for the ability
to detect intruders, but also for the ability to relay information longitudi-
nally across T , assuming that the transmission range of the sensors is r.
Informally, if there is a continuous crossing path γ avoiding all the sensing
regions, then the sensors on one side of γ will be unable to communicate
with those on the other side.
To fix ideas, let Th = R × [0, h], and construct the infinite graph Gr [Th ].
Define a separating path to be a continuous simple path in Th starting at
some point on the line y = h, ending at some point on the line y = 0, and not
passing strictly within distance r/2 of any point of P ∩ Th . This path would
be a feasible path for an intruder to take in order to avoid detection. It
also identifies a communication breakdown in the information transmission
problem. We wish to estimate the frequency with which these paths occur
along Th , but some care is needed with the definition of when two such paths
are essentially the same. To this end, we say that a component of Gr [Th ]
√
3
is good if it contains a vertex strictly within distance 2 r of the top of Th ,
√
3
and also a vertex strictly within distance
√ 2 r of the bottom of Th . The
3
significance of the factor 2 is that the good components can be ordered
along Th , since no good component can “jump over” another. Now we may
define a break in Gr [Th ] to be a partition of the good components into two
classes: those on the left of the break and those on the right. It is not hard
to see that any separating path defines a break, and conversely that, given
a break, there exists a separating path which separates the components
on each side of the break. However, two separating paths γ1 and γ2 may
correspond to the same break. The point of this definition is that the breaks
count separating paths that are essentially different.
Horizontal translation is an ergodic transformation on the probability
space of this model, and consequently it is possible to define the intensity
Ih,r of breaks along Th . (In fact, this can also be seen directly.) Loosely
Percolation, Connectivity, Coverage and Colouring 127

Fig. 1. A break between two good components. Figure taken from [4]

speaking, a long section of Th of length  will contain approximately Ih,r

breaks. Our problems thus reduce to the single problem of estimating Ih,r .
This was done by Balister, Bollobás, Kumar and Sarkar [4, 6].

Theorem 8. The intensity of breaks I(h, r) is given by

   
Ih,r = r1/3 ε hr−1/3 exp ( − hr + O hr−5/3 ),

where
log ε(z) = αz + β + oz (1).

Numerical simulations give α ≈ 1.12794 and β ≈ −1.05116. The proof is

long and complicated, so we shall content ourselves with a very brief outline.
First, a discretization argument shows that, for moderately large values of
r and h, the breaks are typically narrow, that is, they tend to cut straight
across√ Th and have width Θ(r). Next, a version of Theorem 8 is obtained for
h ≤ 23 r. This involves, among other things, an area-preserving rescaling
of Th near a break, which approximates disc boundaries by parabolas and
enables us to replace the two parameters r and h by the single parameter
z = hr−1/3 . (We also make use of a quantitative version of Perron’s theorem
for eigenvalues of strictly positive matrices.) To extend this to larger values
of h, we require two additional lemmas. The first is a technical lemma on
the typical shape of a break: loosely speaking we require that most breaks
are “rectangular”. The second lemma states that Ih,r is approximately
multiplicative, in the sense that if r ≥ 6 and h = h1 + h2 with h1 , h2 ≥ r,
then there is some c > 0 such that

cr−1 Ih1 ,r Ih2 ,r ≤ Ih,r ≤ 50hIh1 ,r Ih2 ,r .

Naturally, the proof of this begins by splitting the strip Th into two strips
Th1 and Th2 , but many difficulties arise at the interface, and we also require
our bound on the expected width of a break established earlier.
128 P. Balister, B. Bollobás and A. Sarkar

For many applications, it is useful to know information about the distrib-

ution of the breaks, rather than simply their expected number. It is possible
to show, using the Stein–Chen method [2], that for r ≥ 6 and x > 0, the
probability that Gr [Th ] restricted to [0, x/Ih,r ] contains exactly k breaks
tends to e−x xk /k! as h → ∞. For this, we need to know that, for large
values of h, the good components are typically wide, and this necessitates
a somewhat elaborate discretization argument, owing to complications aris-
ing from tiles of our discretization intersecting previously examined regions.
For details, see [4].

3. The k-nearest Neighbour Model

Our second model is very similar to the first. As before, we begin with a
Poisson process P of intensity one in the plane R2 . This time, however,
we join each point p ∈ P to its k nearest neighbours: those points of P
which are the closest, in the usual euclidean norm, to p. (With probability
one, there are no ties.) Initially, this creates a directed graph with out-
degree k: however, we convert this into an undirected graph by removing
the orientations. Note that while the maximum degree of the resulting
graph Gnn
k may be significantly more than k, the average degree is certainly
between k and 2k, and it is not hard to see (from elementary properties
  of
the Poisson distribution) that, as k → ∞, the average degree is 1 + o(1) k.
The maximum degree is at most 6k [44].
This is also a very natural model for a transceiver network: one can
imagine, for instance, that each transceiver can initiate a connection with
at most k others. Indeed, this was the original application, and as such was
studied in a series of papers in the engineering literature (see [62] and the
references therein).

3.1. Percolation

Percolation in this model is defined as for the Gilbert model, one difference
being that k is an integer, so that there is some hope of determining the
percolation threshold exactly. To be precise, suppose that the origin is one
of the points of P, write θnn (k) for the probability that the origin belongs
Percolation, Connectivity, Coverage and Colouring 129

to an infinite connected component of Gnn

k , and define kc by the formula
 
kc = min k : θ(k) > 0 .

Simulations [7] suggest that θnn (1) = θnn (2) = 0 and that θnn (3) ≈ 0.985,
so that kc = 3, but proving this is another matter. The best published
bounds are due to Teng and Yao [53], and Bagchi and Bansal [3], who show
that
2 ≤ kc ≤ 188,
although in a paper to be published, Balister, Bollobás and Walters [15]
used a certain oriented 1-independent percolation model to prove that

kc ≤ 11,

and that kc = 3 with confidence 99.99%.

As for the Gilbert model, we can consider the same problem in d dimen-
sions. This was done by Häggström and Meester [34]. Writing kc (d) for the
d-dimensional analogue of kc , they proved that there exists a d0 such that

kc (d) = 2 for all d ≥ d0 ,

and carried out Monte Carlo simulations which suggest that


3 for d = 2
kc (d) =
2 for d ≥ 3.

3.2. Connectivity

Since all transceiver networks are finite, it is natural to consider finite

versions of the model Gnn
k . With this in mind, we restrict attention to points
of P within a fixed square Sn of area n, and ask questions about the graph
Gn,k formed by joining each point of P within Sn to its k nearest neighbours
within Sn . Note that this is different from the subgraph of Gnnk induced by
the vertices of P within Sn . One can now ask for an analogue of Penrose’s
theorem. In other words, how large should we make k = k(n) so as to make
Gn,k connected with high probability? The obstructions to connectivity
cannot be isolated vertices, since there are no isolated vertices in our new
130 P. Balister, B. Bollobás and A. Sarkar

model: the minimum degree of Gn,k is at least k. On the other hand, it is not
hard to see that, for connectivity, we should look at the range k = Θ(log n).
To see this, imagine tessellating the square Sn with small squares Qi of area
about log n. Then the probability that a small square contains no points
of the process is about e− log n = n−1 , so that, with high probability, every
small square contains at least  one point. A short
 calculation now shows
that, if k ≥ 50 log n, then P Po (5π log n) > k = o(n−1 ), so that, again
with high probability, every point of Gn,k contained in a square Qi is joined
to every other point in Qi , and also to every point in every adjacent square.
This is enough to make Gn,k connected. For a lower bound, imagine a small
cluster of k + 1 points surrounded by a large annulus containing no points
of P. These points will form a component of Gn,k if the thickness of the
annulus is greater than the (euclidean) diameter of the cluster it encloses,
and if each point outside the annulus has all its k nearest neighbours outside
the annulus. It is easy to exhibit an example of such a configuration which
occurs in a specified location with probability e−ck : the constant c depends
on the exact specifications of the configuration. It is now a simple matter

to show that if e−ck ≥ n−c for some c < 1 (i.e. if k ≤ c log n for some
c < 1/c), such a configuration will, with high probability, occur somewhere
in Sn , disconnecting Gn,k .
Define cl and cu by

cl = sup {c : P(Gn,c log n is connected) → 0},

and
cu = inf {c : P(Gn,c log n is connected) → 1}.
Xue and Kumar [62] were the first to publish bounds on cl and cu : they
obtained cl ≥ 0.074 and cu ≤ 5.1774, although a bound of cu ≤ 3.8597 can be
read out of earlier work of Gonzáles-Barrios and Quiroz [33]. Subsequently,
Wan and Yi [60] showed that cu ≤ e and Xue and Kumar [63] improved
their bound to cu ≤ 1/ log 2. The best bounds to date are due to Balister,
Bollobás, Sarkar and Walters [7], who proved that cl ≥ 0.3043 and cu ≤
1/ log 7 ≈ 0.5139.
In some sense, the lower bound comes from optimizing the shape (and
other characteristics) of the cluster of k + 1 points alluded to above. The
details are far from straightforward, however, and most of the work consists
of optimizing the region outside the “empty” annulus. For the upper bound
in [7], it is important to show first that the obstructions to connectivity
are small (of area C log n). For this, in turn, one first needs to observe
Percolation, Connectivity, Coverage and Colouring 131

that no two edges belonging to different components to Gn,k may cross, and
indeed that, for k = Θ(log n), any two edges belonging to different compo-
nents of Gn,k are, with high probability, separated by a certain minimum
distance (which depends on k). One can then mimic Penrose’s discretiza-
tion argument to prohibit the existence of two large components, with high
probability. The remainder of the proof is very different in character and
we will not discuss it here.
The natural conjecture that cl = cu = c was made in [7] and proved
in [10]. More precisely, we have the following theorem.

Theorem 9. There exists a constant ccrit such that if c < ccrit and k =
c log n then P(Gn,k is connected) → 0 as n → ∞, and if c > ccrit and
k = c log n then P(Gn,k is connected) → 1 as n → ∞.

One of the ideas in the proof of Theorem 9 is that the essentials of a

small component in Gn,k can be captured “up to ε” by a sufficiently fine
discretization (depending on ε but not on k), which can then be scaled
for different values of k. The details, however, are complicated. The proof
suggests that c = “0.3043” (where “0.3043” refers to the bound on cl from [7]
mentioned above). To some extent, this is backed up by simulations [7].
From the above results, it follows from the theorems in [9] that also cl =
cu = c for the problem of s-connectivity, for any fixed s. For information on
the directed model Dn,k , related coverage problems, and several conjectures,
see the papers [7, 9].

3.3. Sharp thresholds

We have seen that, for the Gilbert model, very precise results are known
about the nature of the transition from non-connectivity to connectivity.
For the k-nearest neighbour model, the picture is much less clear, since the
obstructions to connectivity are only conjectural. Writing

p(n, k) = P(Gn,k is connected),

let us fix n and focus on the case k ≈ c log n, where c is the critical constant
from the previous section. We would like to know how quickly p(n, k)
changes from almost 0 to almost 1 as k increases. Specifically, write
 
kn (p) = min k : p(n, k) ≥ p .
132 P. Balister, B. Bollobás and A. Sarkar

It seems very likely that, for any 0 < ε < 1, there exists C(ε) such that, for
all sufficiently large n,

(1) kn (1 − ε) < C(ε) + kn (ε).

However, this is not known. What is known is that, for fixed k, p(n, k)
decreases sharply from almost 1 to almost 0 as n increases. (One has to
increase n by a multiplicative factor to make p(n, k) go from 1 − ε to ε,
but that is only to be expected since k ≈ c log n.) Ignoring problems at the
boundary, the basic idea is that if p(n, k) = 1 − ε, say, then we can consider
the square SM 2 n as the union of M 2 copies of Sn , each of which contains a
small disconnecting component with probability about ε. Consequently,
2
p(M 2 n, k) ≈ (1 − ε)M < ε,

for a suitable multiplier M = M (ε). In [9], a weak form of (1) is derived

from this result via a complicated double-counting argument.

4. Random Tessellations

Random tessellations of R3 were introduced into the study of rock forma-

tions by Delesse [22] 160 years ago, and in recent years they have been used
to study a great variety of problems from kinetics to polymers, ecological
systems and DNA replication (see, among others, Evans [25], Fanfoni and
Tomellini [26], [27], Ramos, Rikvold and Novotny [51], Tomellini, Fanfoni
and Volpe [54], [55], and Pacchiarotti, Fanfoni and Tomellini [46]). In this
section we shall concentrate on planar tessellations and give a brief review
of the results concerning percolation on the two most frequently studied
models, the so called Voronoi and Johnson–Mehl tessellations.
Strictly speaking, it would suffice to discuss the Johnson–Mehl tessel-
lations only, since a Voronoi tessellation is just a special Johnson–Mehl
tessellation. Nevertheless, as Voronoi tessellations have been studied for
much longer and are much more basic than Johnson–Mehl tessellations, we
shall discuss them in a separate subsection.
In fact, first we shall describe a rather general tessellation in Rd , a trivial
extension of the one defined by Johnson and Mehl. Suppose that ‘particles’
(also called ‘nucleation centres’) arrive at certain times according to some
Percolation, Connectivity, Coverage and Colouring 133

spatial process, which may be deterministic or random. The moment a

particle arrives, it starts to grow a ‘crystal’ at a certain pace, which may be
constant or varying, either deterministically, or in a random way, occupying
the ‘unoccupied’ space around it as it grows. Whatever space a particle
occupies belongs to that particle or crystal forever.
In this very general model, the crystal of a ‘fast’ particle may well
overtake and surround the crystal formed by an earlier, but ‘slow’ particle,
and the crystal of a particle may well consist of an infinite number of
components. Not surprisingly, such a general model does not seem to be of
much use. Needless to say, it is easy to define even more general models of
crystals: e.g., we may use different norms rather than the Euclidean.
In the Johnson–Mehl model, all particles have the same constant speed,
so the crystals are simply connected regions and a particle arriving in the
crystal of another particle does not even start to form any crystal of its own,
so may be ignored. In a Voronoi tessellation the particles not only have the
same speed, but also arrive at the same time.

4.1. Random Voronoi Percolation

Let us start with a slightly different definition of a Voronoi tessellation.

Let Z be a set of points in Rd . (In our terminology above, Z is the set
of ‘particles’ or ‘nucleation centres’ that grow into ‘crystals’ or ‘tiles’.) For
z ∈ Z, let Vz be the closed ‘cell’ consisting of those points of Rd that
are at most as far from z as from any other point of Z. In all cases of
interest, Z is taken to be a countable set without accumulation points;
also, Z is not too ‘lop-sided’: its convex hull is the entire space Rd . In
particular, each cell Vz is the intersection of finitely many closed half-
spaces, so is a convex polyhedron with finitely many faces. Trivially, each
Vz is the closure of its interior Uz ; also the total boundary of the cells,

z=z  ∈Z Vz ∩ Vz  , has measure 0. The tessellation or tiling of R into the
d

‘cells’ or ‘tiles’ Vz is the Voronoi tessellation associated with Z. As we have

already remarked, these tessellations were first introduced by Delesse [22] to
study the formation of rocks; their mathematical study was initiated only
a little later by Dirichlet [23] in connection with quadratic forms, and their
detailed study was started by Voronoi [59] about sixty years later. Today,
in mathematics they tend to be called Voronoi tessellations (or tilings),
although occasionally they are named after Dirichlet.
134 P. Balister, B. Bollobás and A. Sarkar

In a random Voronoi tessellation the set Z used to define the Voronoi

cells is taken to be a homogeneous Poisson process P on Rd , of intensity 1,
say. The choice of the points z ensures that, with probability 1, the tes-
sellation has no ‘pathologies’ (in fact, is as ‘regular’ as possible): any two
cells Vz , Vz  are either disjoint or share a full (d − 1)-dimensional face, and
in every vertex of a cell precisely d + 1 cells meet.

Fig. 2. Part of a random Voronoi tiling in R2 . The dots are the points of a Poisson
process. Figure adapted from [19]

Having defined the cells Vz associated with the points z ∈ P, we define

a graph GP with vertex set P by joining two vertices by an edge if their
cells share a (d − 1)-dimensional face. Now, a random Voronoi percolation
in Rd is simply a site percolation on GP , where GP itself depends on the
random set P. To spell it out, let 0 < p < 1 be a parameter, and assign one
of two states to each vertex of GP , open or closed, such that, given P, each
vertex is open with probability p, and the state of a vertex is independent
from the set of states assigned to the other vertices. As always, our system
is said to percolate if the graph GP contains an infinite path all of whose
vertices are open. Equivalently, we may colour a cell black with probability
p, independently of the colours of the other cells, and colour a point of Rd
black if it belongs to a black cell: then percolation means that the set of
black points has an unbounded component.
There is a more user-friendly way of defining random Voronoi percola-
tion: in this approach we take two independent Poisson processes on Rd ,
P + and P − , with intensities p and 1 − p, respectively. Then P = P + ∪ P −
is a Poisson process of intensity 1, P + is the set of black (open) points,
Percolation, Connectivity, Coverage and Colouring 135

and P − is the set of white (open) points that are used to define Voronoi
cells. Define a graph on P + by joining two of its points z and z  if there
is a path in Rd from z to z  which does not go nearer to another point of
P = P + ∪ P − than to the nearer of z and z  . We have percolation if this
graph has an infinite component.
By making use of Kolmogorov’s 0-1 law one can show that, for each
0 < p < 1, the probability of percolation is either 0 or 1. In the first
instance, we are interested in the critical probability pc = pc (d) such that
for p < pc the probability of percolation is 0, and for p > pc it is 1.
Unlike in the case of the classical bond and site percolations on lattices,
it is not entirely immediate that this critical probability pc (d) is non-trivial,
i.e., 0 < pc (d) < 1. A way of showing this is to use (P + , P − ) to define
appropriate 1-independent percolations on Zd that imply bounds on pc (d).
However, in order to prove better bounds for pc (d), we have to work rather
hard.
For large d, Balister, Bollobás and Quas [5] have proved the following
bounds on pc (d). The proof of the lower bound is fairly easy, but that of
the upper bound is more difficult.

Theorem 10. If d is sufficiently large then the critical probability pc (d) for
random Voronoi percolation on Rd satisfies
√
2−d (9d log d)−1 ≤ pc (d) ≤ C2−d d log d,

where C is an absolute constant.

Not surprisingly, most of the interest in random Voronoi percolation

centres round percolation in the plane. In fact, in one of the early papers
on percolation, Frisch and Hammersley [30] challenged mathematicians to
work on problems of this kind. From the late 1970s, much numerical work
was done on random Voronoi percolation in the plane (see, e.g., Winterfeld,
Scriven and Davis [61], Jerauld, Hatfield, Scriven and Davis [39], and Jer-
auld, Scriven and Davis [40]). In particular, Winterfeld, Scriven and Davis
estimated that the critical probability for random Voronoi percolation in
the plane is 0.500 ± 0.010. In spite of this, it was not even proved that the
critical probability pc (2) is strictly between 0 and 1.
The 1990s brought about substantial mathematical work on random
Voronoi percolation, notably by Vahidi-Asl and Wierman [56, 57, 58], Zvav-
itch [64], Aizenman [1], Benjamini and Schramm [16] and Freedman [29].
136 P. Balister, B. Bollobás and A. Sarkar

Of these papers, only [64] is about the critical probability: in this paper
Zvavitch proved that pc (2) ≥ 1/2.
Even without computer experiments, it is difficult not to guess that the
critical probability pc (2) is exactly 1/2, but a guess like this is very far from
a mathematical proof. Such a proof was given by Bollobás and Riordan [19]
in 2006.

Theorem 11. The critical probability for random Voronoi percolation in

the plane is 1/2.

Very crudely, the ‘reason why’ the critical probability is 1/2 is ‘self-
duality’. For any rectangle R, either there is a ‘black crossing’ from top to
bottom or a ‘white crossing’ from left to right. In particular, if p = 1/2
then the probability that for a given square S there is a black crossing from
top to bottom is precisely 1/2. All this is very well, but there are major
difficulties in piecing together such crossings to form appropriate paths.
In fact, ‘self-duality’ is the reason why the critical probability for bond
percolation in the plane is 1/2, but after Harris’s proof [37] ten years passed
before Kesten [41] could prove the matching upper bound. By now there
are numerous elegant and simple proofs of this fundamental Harris–Kesten
theorem (see Bollobás and Riordan [20, 18]), but it seems that there is no
easy way of adapting any of these proofs to random Voronoi percolation,
as the technical problems of overcoming ‘singularities’ are constantly in the
way. Indeed, in order to prove Theorem 11, Bollobás and Riordan [19] had
to find a much more involved and delicate argument than those used to
tackle percolation on lattices.
To conclude this subsection, let us mention an important question con-
cerning random Voronoi percolation in the plane: is it conformally invari-
ant? (Rather than explaining what this question means, we refer the reader
to Benjamini and Schramm [16] and to Chapter 8 of Bollobás and Rior-
dan [18].) Let us just add that in 1994 Aizenman, Langlands, Pouliot and
Saint-Aubin [42] made the famous conjecture that under rather weak condi-
tions percolation in the plane is conformally invariant. This has been proved
for site percolation in the triangular lattice by Smirnov [52]. Since random
Voronoi percolation has much more built-in symmetry than percolation on
lattices, like the triangular lattice, it would not be unreasonable to expect
that conformal invariance is easiest to prove in this case. Unfortunately, so
far this expectation has not been justified.
Percolation, Connectivity, Coverage and Colouring 137

4.2. Random Johnson–Mehl Percolation

This time we shall consider only Johnson–Mehl percolation in the plane. Let
us recall the definition in the simplest case. ‘Particles’ or ‘nucleation centres’
arrive randomly on the plane at random times, according to a homogeneous
Poisson process P on R2 × [0, ∞), of intensity 1, say. Thus, if z = (w, t) ∈ P
then at time t a nucleation centre arrives in the point w ∈ R2 . As soon as
this nucleation centre arrives, it starts to grow at speed 1, say, so that by
time t + u it reaches every point x within distance u of w, and claims it for
its crystal, provided it had not been claimed by another nucleation centre.
A little more formally, if a nucleation centre w ∈ R2 arrives at time t then
its crystal Vz = V(w,t) consists of all points x such that

d2 (x, w) + t ≤ d2 (x, w ) + t

for every point z  = (w , t ) ∈ P. (Here d2 (x, x ) is the Euclidean distance

of x and x . In defining a cell, we may safely ignore what happens at the
boundary: if a point may be claimed by several particles, we may assign it
at random to any one of them.)
In yet another description of this random tessellation, we keep the points
z = (w, t) ∈ P themselves, grow them in the space R3 (rather than the
plane), and then slice this tessellation with the plane R2 ⊂ R3 . To spell this
out, define the Johnson–Mehl norm  · JM as the 1 -sum of the 2 -norms
on R2 and R:
!
(x1 , x2 , t) JM = x21 + x22 + |t| = (x1 , x2 ) 2 + |t|,

and write d = dJM for the corresponding distance. Then the crystal Vz =
V (w, t) of the nucleation centre w that arrived at time t is
   
(2) Vz = {x ∈ R2 : d (x, 0), z = inf

d (x, 0), z  }.
z ∈P

Putting it in this way, we see that Johnson–Mehl tessellations of R2 corre-

spond to two-dimensional slices of Voronoi tessellations of R3 with respect
to the somewhat unusual sum-metric dJM .
To define percolation on a random Johnson–Mehl tessellation, we pro-
ceed as in the case of Voronoi tessellations: we assign black and white (or
open and closed) states to the cells, and look for an unbounded component
in the union of black cells.
138 P. Balister, B. Bollobás and A. Sarkar

Fig. 3. Part of a random Johnson–Mehl tessellation of R2 . The dots are the projections
onto R2 of those points z of a Poisson process in R2 × [0, ∞) for which the corresponding
cell Vz is non-empty. Figure taken from [21]

By adapting their proof of Theorem 11 to the case of random Johnson–

Mehl tessellations, Bollobás and Riordan [21] determined the critical prob-
ability in this case as well.

Theorem 12. The critical probability for random Johnson–Mehl percola-

tion in the plane is 1/2.

Once again, this result is not too surprising, but what is surprising
is that although the Johnson–Mehl model is more complicated than the
Voronoi model, the proof of this result is actually simpler than that of
Theorem 11. This seeming contradiction is explained by the fact that in
proving Theorem 12 we can make use of the third dimension in the last
representation.

5. Outlook

In this brief review we have seen that although in the past fifty years much
work has been done on properties of random geometric graphs, including
percolation on them, the subject is still in its infancy. We very much
hope that the host of beautiful open problems in the area will attract some
beautiful solutions.
Percolation, Connectivity, Coverage and Colouring 139

References

[1] M. Aizenman, Scaling limit for the incipient spanning clusters, in Mathematics
of multiscale materials (Minneapolis, MN, 1995–1996), IMA Vol. Math. Appl., 99
(1998), 1–24.
[2] R. Arratia, L. Goldstein and L. Gordon, Two moments suffice for Poisson approx-
imations: The Chen-Stein method, Ann. Probab., 17 (1989), 9–25.
[3] A. Bagchi and S. Bansal, On the metric distortion of nearest-neighbour graphs on
random point sets, available at http://arxiv.org/abs/0804.3784.
[4] P. Balister, B. Bollobás, S. Kumar and A. Sarkar, Reliable density estimates for
deployment of sensors in thin strips (detailed proofs), Technical Report, University
of Memphis, 2007.
Available at http://umdrive.memphis.edu/pbalistr/public/ThinStripComplete.pdf
[5] P. Balister, B. Bollobás and A. Quas, Percolation in Voronoi tilings, Random
Structures and Algorithms, 26 (2005), 310–318.
[6] P. Balister, B. Bollobás, A. Sarkar and S. Kumar, Reliable density estimates for
coverage and connectivity in thin strips of finite length, ACM MobiCom, Montréal,
Canada (2007), 75–86.
[7] P. Balister, B. Bollobás, A. Sarkar and M. Walters, Connectivity of random k-
nearest neighbour graphs, Advances in Applied Probability, 37 (2005), 1–24.
[8] P. Balister, B. Bollobás, A. Sarkar and M. Walters, Connectivity of a gaussian
network, International Journal of Ad-Hoc and Ubiquitous Computing, 3 (2008),
204–213.
[9] P. Balister, B. Bollobás, A. Sarkar and M. Walters, Highly connected random
geometric graphs, Discrete Applied Mathematics, 157 (2009), 309–320.
[10] P. Balister, B. Bollobás, A. Sarkar and M. Walters, A critical constant for the
k-nearest neighbour model, Advances in Applied Probability, 41 (2009), 1–12.
[11] P. Balister, B. Bollobás, A. Sarkar and M. Walters, Sentry selection in wireless
networks, submitted.
[12] P. Balister, B. Bollobás and M. Walters, Continuum percolation with steps in an
annulus, Annals of Applied Probability, 14 (2004), 1869–1879.
[13] P. Balister, B. Bollobás and M. Walters, Continuum percolation with steps in the
square or the disc, Random Structures and Algorithms, 26 (2005), 392–403.
[14] P. Balister, B. Bollobás and M. Walters, Random transceiver networks, to appear
in Advances in Applied Probability.
[15] P. Balister, B. Bollobás and M. Walters, Percolation in the k-nearest neighbour
model, submitted.
[16] I. Benjamini and O. Schramm, Conformal invariance of Voronoi percolation, Com-
mun. Math. Phys., 197 (1998), 75–107.
[17] B. Bollobás, Random Graphs, second edition, Cambridge University Press, 2001.
140 P. Balister, B. Bollobás and A. Sarkar

[18] B. Bollobás and O. M. Riordan, Percolation, Cambridge University Press, 2006,

x + 323pp.
[19] B. Bollobás and O. M. Riordan, The critical probability for random Voronoi per-
colation in the plane is 1/2, Probability Theory and Related Fields, 136 (2006),
417–468.
[20] B. Bollobás and O. M. Riordan, A short proof of the Harris–Kesten Theorem, Bull.
London Math. Soc., 38 (2006), 470–484.
[21] B. Bollobás and O. M. Riordan, Percolation on random Johnson–Mehl tessellations
and related models, Probability Theory and Related Fields, 140 (2008), 319–343.
[22] A. Delesse, Procédé méchanique pour déterminer la composition des roches, Ann.
des Mines (4th Ser.), 13 (1848), 379–388.
[23] G. L. Dirichlet, Über die Reduktion der positiven quadratischen Formen mit drei
unbestimmten ganzen Zahlen, Journal für die Reine und Angewandte Mathematik,
40 (1850), 209–227.
[24] D. W. Etherington, C. K. Hoge and A. J. Parkes, Global surrogates, manuscript,
2003.
[25] J. W. Evans, Random and cooperative adsorption, Rev. Mod. Phys., 65 (1993),
1281–1329.
[26] M. Fanfoni and M. Tomellini, The Johnson–Mehl–Avrami–Kolmogorov model – a
brief review, Nuovo Cimento della Societa Italiana di Fisica. D, 20 (7-8) (1998),
1171–1182.
[27] M. Fanfoni and M. Tomellini, Film growth viewed as stochastic dot processes,
J. Phys.: Condens. Matter, 17 (2005), R571–R605.
[28] M. Franceschetti, L. Booth, M. Cook, R. Meester and J. Bruck, Continuum perco-
lation with unreliable and spread-out connections, Journal of Statistical Physics,
118 (2005), 721–734.
[29] M. H. Freedman, Percolation on the projective plane, Math. Res. Lett., 4 (1997),
889–894.
[30] H. L. Frisch and J. M. Hammersley, Percolation processes and related topics, J. Soc.
Indust. Appl. Math., 11 (1963), 894–918.
[31] E.N. Gilbert, Random plane networks, J. Soc. Indust. Appl. Math., 9 (1961), 533–
543.
[32] E. N. Gilbert, The probability of covering a sphere with N circular caps, Bio-
metrika, 56 (1965), 323–330.
[33] J. M. Gonzáles-Barrios and A. J. Quiroz, A clustering procedure based on the
comparison between the k nearest neighbors graph and the minimal spanning tree,
Statistics and Probability Letters, 62 (2003), 23–34.
[34] O. Häggström and R. Meester, Nearest neighbor and hard sphere models in con-
tinuum percolation, Random Structures and Algorithms, 9 (1996), 295–315.
[35] P. Hall, On the coverage of k-dimensional space by k-dimensional spheres, Annals
of Probability, 13 (1985), 991–1002.
Percolation, Connectivity, Coverage and Colouring 141

[36] P. Hall, On continuum percolation, Annals of Probability, 13 (1985), 1250–1266.

[37] T. E. Harris, A lower bound for the critical probability in a certain percolation
process, Proc. Cam. Philos. Soc., 56 (1960), 13–20.
[38] S. Janson, Random coverings in several dimensions, Acta Mathematica, 156 (1986),
83–118.
[39] G. R. Jerauld, J. C. Hatfield, L. E. Scriven and H. T. Davis, Percolation and con-
duction on Voronoi and triangular networks: a case study in topological disorder,
J. Physics C: Solid State Physics, 17 (1984), 1519–1529.
[40] G. R. Jerauld, L. E. Scriven and H. T. Davis, Percolation and conduction on the
3D Voronoi and regular networks: a second case study in topological disorder,
J. Physics C: Solid State Physics, 17 (1984), 3429–3439.
[41] H. Kesten, The critical probability of bond percolation on the square lattice equals
1/2, Comm. Math. Phys., 74 (1980), 41–59.
[42] R. Langlands, P. Pouliot and Y. Saint-Aubin, Conformal invariance in two-dimen-
sional percolation, Bull. Amer. Math. Soc., 30 (1994), 1–61.
[43] R. Meester and R. Roy, Continuum Percolation, Cambridge University Press, 1996.
[44] G. L. Miller, S. H. Teng and S. A. Vavasis, An unified geometric approach to graph
separators, in IEEE 32nd Annual Symposium on Foundations of Computer Science,
1991, 538–547.
[45] P. A. P. Moran and S. Fazekas de St Groth, Random circles on a sphere, Biometrika,
49 (1962), 389–396.
[46] B. Pacchiarotti, M. Fanfoni and M. Tomellini, Roughness in the Kolmogorov–
Johnson–Mehl–Avrami framework: extension to (2+1)D of the Trofimov–Park
model, Physica A, 358 (2005), 379–392.
[47] M. D. Penrose, Continuum percolation and Euclidean minimal spanning trees in
high dimensions, Annals of Applied Probability, 6 (1996), 528–544.
[48] M. D. Penrose, The longest edge of the random minimal spanning tree, Annals of
Applied Probability, 7 (1997), 340–361.
[49] M. D. Penrose, Random Geometric Graphs, Oxford University Press, 2003.
[50] J. Quintanilla, S. Torquato and R. M. Ziff, Efficient measurement of the percolation
threshold for fully penetrable discs, J. Phys. A, 33 (42): L399–L407 (2000).
[51] R. A. Ramos, P. A. Rikvold and M. A. Novotny, Test of the Kolmogorov–Johnson–
Mehl–Avrami picture of metastable decay in a model with microscopic dynamics,
Phys. Rev. B, 59 (1999), 9053–9069.
[52] S. Smirnov, Critical percolation in the plane: conformal invariance, Cardy’s for-
mula, scaling limits, Comptes Rendus de l’Académie des Sciences. Série I. Mathé-
matique, 333 (2001), 239–244.
[53] S. Teng and F. Yao, k-nearest-neighbor clustering and percolation theory, Algo-
rithmica, 49 (2007), 192–211.
[54] M. Tomellini, M. Fanfoni and M. Volpe, Spatially correlated nuclei: How the
Johnson–Mehl–Avrami–Kolmogorov formula is modified in the case of simulta-
neous nucleation, Phys. Rev. B, 62 (2000), 11300–11303.
142 P. Balister, B. Bollobás and A. Sarkar

[55] M. Tomellini, M. Fanfoni and M. Volpe, Phase transition kinetics in the case of
nonrandom nucleation, Phys. Rev. B, 65 (2002), 140301-1 – 140301-4.
[56] M. Q. Vahidi-Asl and J. C. Wierman, First-passage percolation on the Voronoi
tessellation and Delaunay triangulation, in Random graphs ’87 (Poznań, 1987),
Wiley, Chichester (1990), pp. 341–359.
[57] M. Q. Vahidi-Asl and J. C. Wierman, A shape result for first-passage percolation
on the Voronoi tessellation and Delaunay triangulation, in: Random graphs, Vol. 2
(Poznań, 1989), Wiley-Intersci. Publ., Wiley, New York (1992), pp. 247–262.
[58] M. Q. Vahidi-Asl and J. C. Wierman, Upper and lower bounds for the route length
of first-passage percolation in Voronoi tessellations, Bull. Iranian Math. Soc., 19
(1993), 15–28.
[59] G. Voronoi, Nouvelles applications des paramètres continus à la théorie des formes
quadratiques, Journal für die Reine und Angewandte Mathematik, 133 (1908), 97–
178.
[60] P. Wan and C.W. Yi, Asymtotic critical transmission radius and critical neighbor
for k-connectivity in wireless ad hoc networks, ACM MobiHoc, Roppongi, Japan
(2004).
[61] P. H. Winterfeld, L. E. Scriven and H. T. Davis, Percolation and conductivity of
random two-dimensional composites, J. Physics C, 14 (1981), 2361–2376.
[62] F. Xue and P. R. Kumar, The number of neighbors needed for connectivity of
wireless networks, Wireless Networks, 10 (2004), 169–181.
[63] F. Xue and P. R. Kumar, On the theta-coverage and connectivity of large random
networks, IEEE Transactions on Information Theory, 52 (2006), 2289–2399.
[64] A. Zvavitch, The critical probability for Voronoi percolation, MSc. thesis, Weiz-
mann Institute of Science (1996).

Paul Balister Béla Bollobás

Department of Mathematical Department of Pure Mathematics and
Sciences, Mathematical Statistics,
University of Memphis, University of Cambridge,
Memphis, TN 38152, Cambridge CB3 0WB,
U.S.A. UK
e-mail: [email protected] and
Department of Mathematical
Sciences,
University of Memphis,
Amites Sarkar
Memphis, TN 38152,
Department of Mathematics, U.S.A.
Western Washington University,
Bellingham, e-mail:
WA 98225, [email protected]
U.S.A.
e-mail: [email protected]
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 143–169.

Chapter 3
Scaling Properties of Complex Networks and
Spanning Trees

REUVEN COHEN and SHLOMO HAVLIN

We present a relation between three properties of networks: the fractal properties

of the percolation cluster at criticality, the optimal path between vertices in the
network under strong disorder (i.e., a broad distribution of edge weights) and the
minimum spanning tree. Based on properties of the percolation cluster we show
that the distance between vertices under strong disorder and on the minimum
spanning tree behaves as N 1/3 for the N vertex complete graph and for Erdős–
Rényi random graphs, as well as for scale free networks with exponent γ > 4.
For scale free networks with 3 < γ < 4 the distance behaves as N (γ−3)/(γ−1) . For
2 < γ < 3, our numerical results indicate that the distance scales as lnγ−1 N .
We also discuss a fractal property of some real world networks. These networks
present self similarity and a finite fractal dimension when measured using the box
covering method.

1. Random Graphs and Complex Networks

About 50 years ago a model for random networks was developed, combining
ideas from graph theory with ideas from probability theory. The model was
presented and its properties studied in a series of seminal papers by Erdős
and Rényi [24, 25, 26] (A somewhat similar model was discussed also by
Rapoport [38]). In this model, graphs consist of N vertices and M edges
randomly selected between them. That is, in this model all graphs contain-
ing N labeled vertices and M edges are equiprobable. This model has come
144 R. Cohen and S. Havlin

to be known as the Erdős–Rényi (ER) model. An alternative and closely

related model [28] is obtained
  when considering N labeled vertices selecting
edges from the possible N2 edges independently with some probability p.
This model has the advantage of avoiding correlations between edges that
are present in the original ER model. However, both models behave simi-
larly when the number of edges in not extremely small or extremely large
(See, e.g., [7]). We will refer to both models loosely as the “ER model”.
The ER model was presented and studied as an abstract mathematical
object due to its simplicity and elegance. It was also later applied by re-
searchers in different disciplines to describe various real world network mod-
els. In recent years it has become apparent that many of the networks in the
real world are not well described by the ER model. Watts and Strogatz [49]
observed that in many networks edges are not distributed completely ran-
domly between vertices, but rather, edges tend to cluster, i.e., form more
triangles than expected from a completely random distribution. Barabási
and Albert [3] observed that in many real world networks the degree distri-
bution is not Poissonian, as obtained in the ER model, but rather a broad,
power law, distribution. Faloutsos, Faloutsos and Faloutsos [27] observed
the same phenomenon in the Internet router network. For some more re-
cent reviews on the subject see [2, 32, 35, 23]. Several models have been
suggested, trying to better describe the nature of real world networks. Most
of the models studied today attempting to describe real world networks fall
into two main classes. One class of models is based on the Watts–Strogatz
small-world model [49], focusing on the small distance versus high cluster-
ing occurring in real networks by interpolating between a regular lattice
and a random graph. The second class is that of the Barabási–Albert scale
free model [3], focusing on the power law degree distribution observed in
real world networks, as opposed to the Poisson distribution occurring in ER
graphs.
Here we will focus mainly on ER networks and on scale free networks,
having degree distribution

(1) P (k) = ck −γ ,

where k is the degree (number of connections) of a vertex, c is a normaliza-

tion factor, and γ is some exponent (usually 2 < γ < 3). The BA model and
its variants lead to this kind of degree distributions. However, we will focus
on the class of equilibrium scale free networks, obtained by the Bollobás
configuration model, described below (Sect. 2).
Scaling Properties of Complex Networks and Spanning Trees 145

It should be noted that the results supplied here are based on analytical
methods providing insight into the problems. They do not constitute rigor-
ous proofs of the presented results, but mostly heuristic arguments. When
a rigorous proof exists for a certain result we attempt to supply a reference
to the proof. For some of these results no full proof has been obtained, and
they may be seen as open challenges for the mathematical community.

2. The Bollobás Configuration Model

The Bollobás configuration model [6] is a model for random graphs with a
prescribed degree sequence. Given a degree sequence, all graphs having this
degree sequence are equiprobable in this model. In [6] it was shown that an
equiprobable distribution on all graphs with a given degree sequence may be
obtained by starting with a set of N vertices and assigning to each vertex its
degree from the sequence1 . The vertex is then equipped with this number
of “stubs” (i.e., links, currently leading nowhere). Random pairs of stubs
are then connected to each other, forming an edge between the respective
vertices. This process continues until no stubs are left. It should be noted
that this process actually may produce a multi-graph, which is a graph with
self loops and multiple links between some pair of vertices, in which case
one may discard the graph and restart the process. However, for a sparse
graph with no vertices having degrees of order N , the expected number of
self loops and multiple edges is of lower order than the node degrees, and
thus discarding these leads to a graph with degree sequence very close to
the prescribed.
This model leads to a graph having the prescribed degree distribution
and no other correlations. That is, it is maximally random for the given
degree distribution. Therefore, it can be seen as an “equilibrium” or “maxi-
mal entropy” model of random graphs with a prescribed degree distribution.
This model can also be seen as an expansion of the ER model, by considering
a configuration model network with a Poisson degree distribution,

ak
(2) P (k) = exp (−a) ,
k!
1
Assuming the degree sequence is also random, the degrees may be assumed to be
assigned using the desired degree distribution (such as Eq. (1)). In case the sum of all
degrees is odd and a graph can not be constructed, the degrees should be reassigned from
the distribution.
146 R. Cohen and S. Havlin

where a = k is the average degree. While this model does not repro-
duce exactly the ER model probability distribution, it is very close in its
properties.

3. Percolation on Random Graphs

One of the simplest and most common models for phase transitions is the
percolation model [11, 44, 31]. It is based only on the topological properties
of the underlying network and on a single externally tunable parameter. In
this model each vertex (for site percolation) or edge (for bond percolation) is
occupied with some probability p and vacant otherwise. Alternatively, it can
be thought of as the process of removing vertices or links with probability
q = 1 − p.
When only a small fraction of the vertices (or links) is occupied the net-
work is, with high probability, composed of a large number of very small
components, unreachable from each other through occupied vertices (or
links). However, when the occupied fraction becomes large, a giant com-
ponent emerges, connecting a finite fraction of the vertices in the network.
This giant component usually appears at some critical concentration, pc .
The percolation transition is usually second order, i.e., the size of the gi-
ant component varies continuously from zero below and at pc to some finite
value above pc .
The subject of percolation in uncorrelated random networks has been
studied thoroughly (See, e.g. [16, 12]). Here we present a derivation of the
percolation threshold and some properties of the percolating network.
Consider a random network with some degree distribution P (k). The
degree distribution

of a randomly selected vertex is P (k), and its average
degree is k = kP (k). However, when following a link to reach a vertex,
the probability of reaching a vertex with degree k is proportional to its
number of links, i.e., to its degree. The distribution of degrees of vertices
reached by following a random link is therefore φ(k) =
kP (k)/k and the
average degree of a vertex reached this way is κ ≡ kφ(k) = k2 /k.
Since links are randomly connected, uncorrelated random graphs tend, with
high probability, to include almost no small loops, and are locally tree-like.
When considering the process of exploring the graph as a branching process,
each reached vertex has one link through which it has been arrived at, and
k − 1 outgoing link leading to new vertices (until most of the graph is
Scaling Properties of Complex Networks and Spanning Trees 147

explored and some links may lead back to known vertices). When only a
fraction p of the vertices (or links) are occupied, on average only (k − 1)p
of the “descendants” of the
vertex will be reachable, and on average, per
an explored vertex, k̃ = (k − 1)pφ(k) = p(κ − 1) descendants will be
reachable. When k̃ < 1 the branching process will die with probability 1
after a finite number of explored vertices, while when k̃ > 1 the process may
continue indefinitely and a giant component will exist. This leads to the
following value for the critical threshold [16],

1
(3) pc = .
κ−1

It should be noted that if the second moment of the degree distribution,

k 2 , diverges, pc → 0, i.e., the network contains a giant component when
any finite fraction of the vertices or links are removed [16]. This result has
been proven rigorously for the LCD model in [8]. This is the case for scale
free degree distributions (Eq. (1) with γ ≤ 3), which is a common distrib-
ution for real networks. Therefore, it is expected that these networks will
be very resilient to random breakdown of vertices or links. The vanishing
percolation threshold also indicates that epidemic diseases and viruses can
propagate in the network with no critical threshold. See also [34].

3.1. Generating functions

A different approach for calculating the critical percolation threshold and

other important percolation properties is by utilizing the generating func-
tions approach [33, 12].
Denote by G0 (x) the generating function for the vertex degree distribu-
tion, P (k), i.e., the formal power series,

(4) G0 (x) = P (k)xk .
k

Note that G0 (1) = 1 by normalization, and G0 (1) = k. The generating
function for the out degrees of vertices reached by following a link is then

  kP (k) G0 (x)

(5) G1 (x) = φ(k)xk−1 = xk−1 = .
k G0 (0)
k k
148 R. Cohen and S. Havlin

Note that G1 (1) = 1 also by normalization (assuming k is finite), and that
G1 (1) = κ − 1, the average outgoing degree of a vertex reached through a
link.
Consider the following process in the network: Start exploring the net-
work by picking a link and following it to one of its vertices. Then all other
links emanating from this vertex are explored and so on. If no vertex is
reached twice during this exploration we may refer to the explored region of
the network as a “branch”, and view this process as a probabilistic branch-
ing process. This process may die out after a finite number of vertices are
reached, or it may continue indefinitely.2 A generating function can be con-
structed for the sizes of branches defined as above by noticing that this is a
branching process where (almost) each link explored leads to a new branch
with the same distribution as the original one. The number of branches
at each point is distributed according to φ(k) (or G1 (x)). Assuming the
concentration of links is p,3 the generating function for branch sizes is,
 
(6) H1 (x) = (1 − p) + pxG1 H1 (x) .

Note that it is no longer necessarily true that H1 (1) = 1. In fact, H1 (1) gives
the probability that a branch has a finite size. See Fig. 1 for illustration.

Fig. 1. An illustration of the recursive branch definition

Starting from a random vertex and studying the size distribution of the
component to which it belongs, is similar to selecting a degree k, using the
distribution P (k), and then summing the sizes of the k branches to which
2
Naturally, in a finite network the whole network will be explored in a finite time.
However, when the size of the network, N → ∞, it can be assumed that as long as the
branch size is of o(N ), the fraction of edges linking back to explored vertices is of o(1),
and therefore the branching process is a good approximation. When N → ∞ the process
can continue indefinitely (up to O(N )) before loops significantly affect the behavior of the
branching process.
3
From here on we consider link, and not vertex, percolation. Vertex percolation can
be handled in a very similar manner. See, e.g, [12] for details.
Scaling Properties of Complex Networks and Spanning Trees 149

it leads. Therefore, to find the distribution of component sizes one can

construct the following generating function

 
(7) H0 (x) = xG0 H1 (x) .

Again, H0 (1) does not necessarily equal 1, and gives the probability of a
vertex to belong to a finite component. Thus, the probability of a vertex
to belong to the infinite component, which is the relative size of the infinite
component is,

(8) P∞ = 1 − H0 (1).

3.2. Critical exponents

The generating functions presented in Section 3.1 can be used to find the
properties of percolating networks near and at the percolation transition
point. The behavior near a physical phase transition point is known, both
experimentally and using heuristic arguments, to be universal [11, 44], i.e.,
to depend only on the dimensionality of the physical space in which it occurs,
and on the symmetries of the order parameter. It is also known, however,
that heterogeneity in space may break the universality and lead to non-
universal behavior. Below we present a case in which universality is broken
also by heterogeneity in the degrees. We closely follow [15].
Consider percolation in a network. As shown in Section 3, the critical
point can be found at pc = (κ − 1)−1  (Eq. (3)). P∞ can by found using
Eq. (8) and substituting H0 (1) = G0 H1 (1) . H1 (1) is to be found using
Eq. (6). This can be done numerically and for some distributions even
analytically. However, near the critical point one can find the leading order
of the behavior of H1 (1) for a general distribution. At p = pc + δ, with
δ → 0, the size of the giant component is still very small and the probability
of belonging to it is close to zero. Therefore, u ≡ H1 (1) = 1 − ε. Expanding
Eq. (6) one obtains

∞
(pc + δ) 
(9) 1 − ε = 1 − pc − δ + kP (k)(1 − ε)k−1 .
k
k=0
150 R. Cohen and S. Havlin

The sum in Eq. (9) can be expanded in powers of ε

∞
 " # 1" #
(10) kP (k)uk−1 ∼ k − k(k − 1) ε + k(k − 1)(k − 2) ε2 + · · · .
2
k=0

Using this expansion, Eq. (9) leads to ε ∼ δ, to first order approximation,

implying that P∞ ∼ (p−pc ). This is in accordance with the known universal
behavior for percolation in high dimension, d ≥ dc = 6, where the size of
the giant component (or “spanning cluster”) grows linearly with p − pc near
the transition point.
For scale free networks, however, the behavior is different.
" When the #
degree distribution is given by Eq. (1) with γ < 4, the term k(k−1)(k−2)
diverges and Eq. (10) no longer holds. The behavior of the sum near u = 1
may be determined using Abelian methods (See, e.g., [50]). In this case
∞
 " #
(11) kP (k)uk−1 ∼ k − k(k − 1) ε + cΓ(2 − γ)εγ−2 + · · · ,
k=0

where Γ denotes the Gamma function. Thus, Eq. (6) leads to εγ−3 ∼ δ and
therefore,

1
(12) P∞ ∼ (p − pc )β , β= .
γ−3

This result is already non-universal, in the sense that the critical exponent
for scale free networks with γ < 4 is different from that obtained for lattices
in high dimensions, for Cayley trees [11], and for ER networks (or scale free
networks with γ > 4).
The distribution of component sizes in the network can also be deter-
mined from the coefficients of the expansion of H0 (x) as a power series in
x. The coefficient of xs in this expansion gives P (s), the probability that a
vertex belongs to a component of size s. The number of components of size
s is denoted by ns = N P (s)/s. In mean-field percolation (i.e., percolation
above the critical dimension), it is known (See, e.g., [11, 44, 33]) that near
the threshold
∗
(13) ns ∼ s−τ e−s/s ,

where τ = 2.5 and s∗ diverges exactly at the critical point, leading to a

pure power law distribution. To find P (s) one can study Eqs. (7) and (6)
Scaling Properties of Complex Networks and Spanning Trees 151

exactly at p = pc . Letting x = 1 − ε and denoting φ(ε) = 1 − H1 (1 − ε),

Eq. (6) leads to

$ " #
φ k(k − 1)(k − 2) 2
(14) − φ = −pc + (1 − ε)pc 1 − + φ
pc 2k
%
Γ(2 − γ) γ−2
+ ··· + c φ .
k

Again, the analytical terms dominate for ER networks and scale free net-
works with γ > 4 and the non analytical term dominates for scale free
networks with γ < 4. Using Tauberian theorems, linking between the ana-
lytical properties of a function and its power series expansion, one obtains
that P (s) ∼ s−τ +1 with τ = 2.5 for ER networks and scale free networks
with γ > 4 and τ = 2γ−3
γ−2 for γ < 4.
To find the size of the largest component in the network at criticality,
one may consider the number of components of size s, ns . The extreme value
statistics on the largest component size, S may be estimated by taking the
integral over the tail of the distribution to equal 1/N , as this signifies that
approximately one component will have this size. Thus,
 ∞  ∞
1
(15) = ns ds = c1 s−τ ds = c2 s−τ +1 .
N S S

It follows that the size of the largest component scales as

(16) S ∼ N 1/(τ −1) ,

where, as above, τ = 2.5 for ER networks and τ = 2γ−3 γ−2 for scale free
networks with γ < 4. For ER networks and for scale free networks with
γ > 4 at the transition point this leads to the well known result S ∼ N 2/3 [7].

3.3. Fractal dimensions

Several, not necessarily equivalent, definitions exist to the concept of a

dimension. One of the common definitions of the dimension of a graph (as
well as of continuous objects) is based on the dependence of the number of
vertices (or “mass”) as a function of the distance from some initial vertex,
i.e., the size of the l-neighorhood of a random vertex. If the number of
152 R. Cohen and S. Havlin

vertices, Ml , up to some large distance, l, from some initial vertex, scales

as Ml ∼ ld , then the dimension of the graph is considered to be d. This is
true for regular lattices in all dimensions, as well as for many other graphs
embedded in finite dimensional space. In Configuration Model random
graphs, the number of vertices at a distance l from a vertex usually (for high
enough average degree) grows as Ml ∼ bl with some b.4 This exponential
growth is faster than any power law, and therefore random graphs are
usually considered to be infinite dimensional.
At the percolation critical point, however, the network becomes very
diluted, and the growth of M becomes slower. To find the behavior of Ml
we use the following consideration (See [18]). Consider Nl , the generating
function for the distribution of the number of vertices at a distance l along
some branch, i.e., l hops from a random vertex arrived by following some
link. At the end of a followed link there is always one vertex. Therefore
N0 (x) = x. The distribution of the number of this vertex’s neighbors
(excluding the link through which this vertex was reached to) is G1 (x).
In general, each such branch consists of a vertex whose degree distribution
is represented by the generating function G1 , and each of whose links leads
to a new branch. The total number of neighbors at distance l from the
vertex consists of the total number of l−1-distance neighbors of this vertex’s
neighbors. Thus, the generating function of this distribution can be found
using the recursive equation
 
(17) Nl (x) = G1 Nl−1 (x) .

Since at criticality the branching factor, pc (κ − 1) is exactly 1, the average

number of vertices at distance l+1 is exactly the average number at distance
l. However, we are only interested in branches that survive at least l layers.
That is, branches in which at least one vertex is at a distance l from the
origin. Since Nl (x) is the generating function for the number of vertices at
distance l, the coefficient of x0 gives the probability of dying before or at
the lth layer. Therefore, the probability to die before the lth layer is given
by Nl (0). The average number of vertices at the lth layer for surviving
branches, Al is thus given by the average for all branches divided by the
survival probability
1
(18) Al = .
1 − Nl (0)
4
For equilibrium random graphs with finite κ, b = κ − 1, and when κ diverges, M
grows even faster. See, e.g., [7, 33, 17, 48].
Scaling Properties of Complex Networks and Spanning Trees 153

To find the behavior of Nl (0) for large l, one should expand Eq. (17) at
the critical point. Assuming κ − 1 = 1,5 one obtains

(19) & " # '

1 " # k(k − 1)(k − 2) 2
G1 (1 − ε) = 1 − k − k(k − 1) ε + ε + ··· .
k 2

Letting Nl (1 − ε) = 1 − εl , Eq. (17) leads to

& " # '

1 " # k(k − 1)(k − 2) 2
(20) 1−εl+1 = 1− k − k(k − 1) εl + εl + · · · .
k 2

Guessing a solution of the form εl ≈ Bl−h leads to

(21)
" #
−h −h −h−1 −h k(k − 1)(k − 2)
B(l + 1) = B(l − hl ) + · · · = Bl − Bl−2h ,
2k

and thus, h = 1. From Eq. (18) follows Al ∼ l. Therefore,6


l
(22) Ml ≈ Al ∼ lh+1 ∼ l2 ,
=1

and the fractal dimension is dl = 2.7 For scale free networks with 3 < γ < 4
similar considerations lead to dl = (γ − 2)/(γ − 3) [18].

5
For simplicity of notation we assume here that the original graph is at criticality,
rather than arriving at criticality through a dilution of the links or vertices. See [29] for
a more complete treatment.
6
Note that Al is the average number of vertices in the lth layer provided the branch
survived at least up to the lth layer. This is not the same as the average number of
vertices in the lth layer provided the branch survived l + l layers. However, Al gives a
lower bound for this quantity and Al+l gives the upper bound.
7
Notice that there is no real embedding space here, and the dimension is based on
the shortest distance metric on the graph itself. This is actually known in physics as the
“chemical dimension”, whereas the fractal dimension depends on the embedding space
and for random embedding it is twice the chemical dimension [11].
154 R. Cohen and S. Havlin

4. Weighted Networks

In the following we consider a network in which every link is associated

with a weight – a positive number representing some property of the link.
This property can be the cost, time or capacity of the link. In physical
systems this weight is usually associated with the energy of the bond. We
assume that the weights are randomly selected from some distribution. We
will mainly concentrate on the behavior of the “optimal path” between
vertices, i.e., the path with minimal total weight connecting the two vertices.
We begin by considering distances between vertices in networks with no
disorder.

4.1. Shortest paths in networks

Since networks are usually not considered to be embedded in real space, no

a-priori notion of the distance between vertices exists. Therefore, distances
should be defined based only on the topology of the network. The most
natural definition of a distance between vertices is the “hop distance”, i.e.,
the minimal number of links that need to be transversed to reach one vertex
from the other. This is analogous to assigning a weight of 1 to each link
and considering the minimal weight path between vertices.
Considering a two dimensional lattice, the distance between vertices
using the above definition is the “Manhattan distance”, i.e., the sum of the
absolute difference between the x and y coordinates. Considering an L × L
lattice, with N = L2 vertices, the average distance between two randomly
chosen vertices is L/3 in the x coordinate ad L/3 in the y coordinate.
Therefore, the average total distance is l = 2L/3, which scales as l ∼ L ∼
√
N . This behavior, N ∼ ld or l ∼ N 1/d implies that the network has
dimension d. In this case d = 2. This shows that the dimension of a lattice
can be defined even if no a-priori assumptions about an embedding space is
made.
For a configuration model random graph, the scaling of the distance is
quite different. As stated above, the random graph has no small loops and
therefore is locally tree-like. Thus, it behaves locally as a branching process
with average branching factor κ − 1. The average number of vertices at a
distance l scales as (κ − 1)l . As this is an exponential growth process (for
κ > 2), the number of vertices up to distance l is proportional to that of
Scaling Properties of Complex Networks and Spanning Trees 155

the lth layer. This implies that the average distance between vertices scales
logarithmically with the size of the network l ∼ log N/ log(κ−1) (See, e.g.,
[7, 46] for a more detailed account). For scale free networks with γ < 3, κ
diverges, and therefore the distances are even shorter than logarithmic. For
2 < γ < 3 the distances behave as [19, 17, 22, 48]
 
(23) l ∼ ln ln N/ ln(γ − 2) .

4.2. Strong and Weak Disorder

Assume now that random weights are associated to the links. The weights
are drawn from some distribution P (w). The total weight of a path is the
sum of the weights of the links along the path. The “optimal path” between
vertices is the path of minimal total weight between these vertices. Two
classes of behavior are possible, with a crossover regime between them [13,
14]. If the weights are drawn from a relatively narrow distribution, the
weight of a path will be closely related to its hop number, as every link
will contribute a similar weight. Thus, the length of the optimal path is
expected to be proportional to the length of the shortest path. This case is
known as “weak disorder”. If the distribution of weights is broad enough,
such that, e.g., each weight is at least twice as large as the next highest
weight, the total weight of a path is determined by the highest weight along
the path, and is almost independent of all other weights. In this case, paths
can be compared by the highest weight on them. If they share the highest
weight links, the lower weight path is determined by comparing the highest
weight between the non-shared links. This is termed “strong disorder”.
While in weak disorder the behavior of the optimal path length is very
similar to that of shortest paths, in the strong disorder regime the behavior
is quite different. In strong disorder the optimal paths attempts to avoid
high weight links whenever possible. This implies that the optimal path
may follow a very long distance to avoid passing through nearby high weight
links. In lattices it is well-known [13, 14, 37] that optimal paths in strong
disorder are fractal, i.e., have dimension higher than 1.8 On the other hand,
in weak disorder the optimal paths are only self-affine, meaning they still
8
This implies that if the shortest hop distance between two vertices is l, the optimal
path length will scale as lopt ∼ ldopt with some dopt > 1.
156 R. Cohen and S. Havlin

scale similarly to the shortest paths, but allow some small deviations to find
a lower weight path.9
Strong disorder also appears in cases where a path is determined by the
highest weight link without summation. An example of such a case is circuit
allocation in a communication network, where the bandwidth of a path is
determined by the minimum bandwidth of a link along the paths. In this
case only one link determines the bandwidth of the path. However, it may
happen that one wishes to optimize also the bandwidth of the next lowest
bandwidth links to prevent congestion. In this case the optimal between
two paths will be determined by comparing the minimal bandwidth between
their non-shared links. An illustration is presented in Fig. 2.

4.3. Minimum Spanning Trees

Spanning trees are trees that span the network. That is, trees containing
a subset of the links in the network, while still connecting all vertices in
the network. Minimum spanning trees (MSTs) are the spanning trees that
have the lowest total weight of all spanning trees. When all weights are
equal, all trees are minimal, as they all have the same weight (since all
trees on N vertices have N − 1 links). This model, where all spanning trees
are equiprobable, is called uniform spanning tree (UST). When weights are
drawn from a continuous distribution, ties between the weights of different
links become statistically insignificant, as their probability approaches zero,
and the MST becomes unique.
An interesting feature of the minimum spanning tree is that it induces
global, rather than local, optimization. The tree is selected to minimize
the total weight of links under the constraint that all vertices must be
connected. This implies that paths between vertices on the MST are not
chosen to minimize the distance between vertices, but rather are the result
of the global optimization of the tree weight. Therefore, these paths may be
very long compared to the shortest distance between vertices in the original
network. Furthermore, paths between vertices in the tree are unique, so for
the paths between three vertices, A, B and C, it must hold that one of these
paths is the sum of the other two. This structure is, in some sense, “super-
critical”, as it represents a backbone that has been diluted to the maximum
9
Formally, self-affinity implies that the width of the path (i.e., the deviations from the
shortest path) scales as W ∼ lα with some α < 1. Therefore, the optimal path length is
still proportional to l with some higher order corrections [4, 10].
Scaling Properties of Complex Networks and Spanning Trees 157

Fig. 2. Searching for the optimal path between vertices A and B in a network with
strong disorder. The length of a path is calculated using lexicographic comparison
rather than the sum of weights. The first (shortest) path found is (8, 10). However, the
subsequent paths found (8, 7, 6) and (8, 7, 4, 3) have lower weight, although they are
longer. After [9]

possible level while maintaining connectivity. The structure of the minimum

spanning tree in a network is thus very similar to the optimal paths in strong
disorder. However, due to their globally optimized nature, MSTs maintain
this behavior for every distribution of weights, even a narrow one.
Two widespread algorithms exist to find the MST: The Prim algorithm
and the Kruskal algorithm (See, e.g., [20]). The Kruskal algorithm operates
by starting from a forest of N vertices and no edges and adding each time the
minimal weight edge that does not close a loop. The Prim algorithm starts
from a vertex, and at each step adds to the tree the lowest weight adjacent
link that does not close a loop. The order at which the edges are added
is random, since the weights were chosen randomly and independently.
Hence, the Kruskal algorithm resembles percolation and the Prim algorithm
resembles invasion percolation [36]. The main difference is that both are
“guarded” percolation, i.e., percolation with the modification that only
the removal of links that do not disconnect the network is allowed. The
percolation process ends when no loops are left and the remaining network
is a tree [21].
158 R. Cohen and S. Havlin

4.4. Fractal properties of optimal paths

To study the behavior of optimal paths and MSTs we use the following
algorithm (which may be seen as the inverse of the Kruskal algorithm for
MSTs): We begin with the full network, and then start removing the highest
weight links. Whenever the removal of a link will break a component into
two smaller components, the link will be “guarded” and will not be removed.
Since the weights are randomly and independently chosen, the order of link
removal is random, and there is no need to actually draw the weights, as
their numerical value plays no role in the model. This algorithm is naturally
justified for the MST, as discussed above. For strong disorder, it is known
that the weight of the highest weight link is larger than the sum of all weights
below it, so, as long as an alternative path with lower weights exists, it will
always be preferable over the higher weight link.
In configuration model random graphs, the analogy between optimal
paths and percolation becomes more exact [9]. Unlike finite dimensional
percolating lattices, in random graphs the components at criticality are tree
like and contain almost no loops. Thus, the percolation components at and
below criticality are subgraphs of the MST. We expect the properties of the
MST or optimal path tree in strong disorder to be similar to that of the
percolation components at criticality.
As discussed in Section 3.2, at pc the size of the largest component at
the percolation threshold scales as S ∼ N 1/(τ −1) . The distances between
vertices in this component scale as lopt ∼ S 1/dl , where dl is the fractal dimen-
sion. Most of the path is along the largest percolation components [30, 1].
It follows that the distances scale as

(24) lopt ∼ N 1/(dl (τ −1)) .

This leads to the conclusion that in ER networks optimal paths scale as [9]

(25) lopt ∼ N 1/3 .

This scaling also holds when starting from the complete graph, KN , as ER
graphs are obtained by randomly removing links from the complete graph.
On the other hand, distances in uniform spanning trees of the complete
graph and ER graphs are known to scale as l ∼ N 1/2 [45]. It therefore follows
that MSTs are more compact than USTs, due to the global constraint on
MSTs, forcing them to be drawn from a different distribution than the
uniform one.
Scaling Properties of Complex Networks and Spanning Trees 159

In scale free networks with γ > 4 the behavior, as for the other critical
exponents, is similar to that of ER networks. When 3 < γ < 4 the behavior
is [9], according to Eq. (24),

(26) lopt ∼ N (γ−3)/(γ−1) .

When γ < 3 the analogy to percolation gives only little insight, as percola-
tion is only achieved in the limit of zero concentration. However, simulation
results indicate that the optimal path lengths are polylogarithmic. Thus,
the dimension of the MST is still infinite. However, these distances are
exponentially larger than the hop distances, which scale as log log N , see
Eq. (23).
To complete the investigation of the length of the optimal path, one
should establish that the percolation components connect between them in
a compact way through the guarded links. This would establish the above
estimation of lopt also as an upper bound and therefore establish that the
scaling is correct. An investigation of the properties of optimal paths and
their partition to percolation components and guarded links is presented
in [51].
A recent paper [1] finally establishes a tight bound on the optimal path
length for MSTs on the complete graph and ER networks. To establish
the compactness of the connections between components of the network,
the links are divided to three regimes. Up to the critical concentration
all links that do not form loops are added, leading to a critical network
with O(N 1/3 ) path length as presented above, Eq. (25). Then, a series of a
few steps in which the size of the largest component grows from O(N 2/3 )
to O(N/ ln N ). In every such step the length of the optimal path does
not increase too much. Eventually, after reaching a component of size
O(N/ ln N ), all other components are considerably smaller and then connect
to the largest component through a short sequence of small components. For
full details see [1].
It is well known that in most physical models weak disorder does not
affect the scaling of the optimal path (See, e.g, [11]. See also [47] for
a rigorous result.). Therefore, when disorder is weak, the optimal path
lengths are expected to scale similarly to the shortest paths. Simulations
confirm this expectation. Simulation results [9] indicate that the distances
in both ER and scale free networks with γ > 3 behave logarithmically with
the network size, similarly to shortest paths. For scale free networks with
2 < γ < 3 simulation results lead to the conjecture that the optimal path
160 R. Cohen and S. Havlin

length also scales as powers of log N , which is different than the log log N
scaling of the shortest path length. We have no explanation or analytical
confirmation for these results.

5. Fractal Networks

As discussed earlier, the dimension of scale free networks and mean field
random graph models, such as ER graphs, can be considered to be infinite,
as the growth of the number of vertices at a distance l from an arbitrary
node grows exponentially with l. This seems to stand in contradiction with
the notion that the “scale free” nature of a network implies some fractal
properties of this network. However, as shown below, a recent study [42]
indicates that actually these notions can be reconciled. To explain how, we
first present two methods of finding the fractal dimension.10

5.1. The cluster growing method

In the cluster growing method one begins with an arbitrary vertex of the
network, the distance l neighborhood of the vertex (i.e., all vertices at a
distance at most l from it) is explored, and the number of vertices at a
distance at most l, Ml , is plotted. Results obtained by starting from many
initial vertices are then averaged, and the growth of Ml determines the
fractal dimension. That is, if Ml ∼ ld for some d then d is the fractal
dimension.

5.2. The box covering method

In the box covering method, the fractal is covered by a minimal number of

boxes of some (Euclidean) linear size l. Since the fractal does not cover the
full space, the smaller the boxes, the more holes can be left uncovered. The
dependence of the number of boxes on l determines the fractal dimension
of the network. That is, if the number of boxes of linear size l needed to
cover the fractal is NB (l) ∼ N l−dB , dB being the fractal dimension of the
10
It should be noted that several models have been presented to scale free networks
with fractal structure. See, e.g., [39].
Scaling Properties of Complex Networks and Spanning Trees 161

network. Since the networks discussed here are not embedded in Euclidean
space a analogous algorithm is covering the network with “boxes”, each of
which contains the nodes within a hop distance L from some starting node.

5.3. The fractal nature of scale free networks

In a recent paper [42, 43], Song et al investigated the fractal properties

of several real world networks. As discussed above it was widely believed
that due to the exponential growth of the number of vertices at a distance
l there can be no fractal properties to scale free networks. Song et al
introduce the following idea: While the cluster growing method always leads
to exponential behavior, the box covering method may lead to a different
behavior. To uncover this behavior one should notice the following point:
A box with linear dimension l covers several vertices within a distance l of
some initial node. More boxes are added to cover the rest of the network,
where a covered vertex is not to be covered again. This last requirement
stems from the fact that high degree nodes have a large number of nodes in
their neighborhood. Allowing their repeated use would lead to a large boxes
containing the same nodes with their sizes growing exponentially with l.
The large number of vertices in the l-distance neighborhood of high
degree vertices leads to the covering of high degree vertices quickly in the
process. This effectively lowers the number of high degree vertices for the
rest of the network; thus reducing the expansion of the network. This allows
the growth to become slower than exponential despite the fast growth of the
cluster growing method.
Empirical results show that the application of the optimal box covering
method to networks such as the WWW and protein interaction networks
results in a power law

(27) N (l) ∼ l−dB ,

where dB is the fractal dimension obtained by the box covering method. For
the WWW, dB ≈ 4.1 and for protein interaction networks, dB ≈ 2.3. For
the Internet network, as well as for different models for network formation
no fractality is found, and no power law can be fitted to NB (l).
The fractal nature of networks is also seen by applying the renormal-
ization technique. In the renormalization process each box is replaced by a
“super-vertex” and two such super-vertices are connected if there is a link
162 R. Cohen and S. Havlin

between vertices in the respective boxes. The process of renormalization

produces a new degree distribution of the super-vertices P (k ) whose tail is
invariant under the renormalization, P (k  ) ∼ P (k) ∼ (k )−γ .
From empirical results, the number of links in each box is proportional
to the number of links in the highest degree vertex in the box, with some
proportionality factor depending on the box size l. This implies

(28) k → k = s(l)k .

The proportionality factor also behaves as a power law in l satisfying

s(l) ∼ l−dk for some dk . For the WWW dk ≈ 2.5. Consider now the
transformation k → k . Since n(k)dk = n (k )dk  , where n(k) = N P (k) and
n (k  ) = N  P  (k  ) are the respective number of vertices of degree k (before
the renormalization transformation) and k (after the renormalization trans-
formation) respectively. Using P (k) ∼ k −γ , P (k ) ∼ (k  )−γ and Eq. (28), it
follows that N  = N sγ−1 . Since N  = NB (l), the number of boxes of size l
needed to cover the network, it follows that NB (l) ∼ N sγ−1 = N l−dk (γ−1) .
By definition NB (l) ∼ N l−dB . Therefore, there exists a relation between
the fractal dimensions and the degree distribution exponents,

(29) γ = 1 + dB /dk .

The box covering and renormalization methods are illustrated in Fig. 3.

Fig. 3. The box covering method and the renormalization of the network for different
box sizes. After [42]
Scaling Properties of Complex Networks and Spanning Trees 163

6. Fractal Properties of Network Boundaries

Most work on distances in networks focused on the average, or typical

distance between vertices. In a recent work [40, 41], however, the properties
of the vertices far from a given vertex were investigated. It was found that
the number of vertices at a large distance from an arbitrary vertex follows
a power law distribution.
Consider an N -vertex network with some degree distribution P (k). Start
from some arbitrary vertex and observe the vertices at distance l from this
vertex. For ER networks and small l, the growth with l is approximately
exponential. The average hop distance between vertices is approximately
l ∼ log N/ log (κ − 1) when κ is finite. In the following, we study the
structure of layers with l > l. That is, we study the properties of the
vertices at a distance l from an arbitrary vertex, where l is larger than the
average distance in the network.

Fig. 4. Shell and cluster (component) structure of the boundary of a network. After [41]
164 R. Cohen and S. Havlin

6.1. Distribution of outer shells population

We now follow [41] to prove that the distribution of the number of vertices
in the nth shell for large n follows p(yn ) ∼ yn−2 . When exploring the
network, the probability of reaching a vertex with k outgoing

links through
a link is p̃(k) = (k + 1)P (k + 1)/k. Define g(k) =

p̃(k)xk , and k̃ =

k p̃(k). The generating function for the number
  of vertices at the nth
shell of a branch is therefore Gn (x) = Gn−1 g(x) . We denote by p̃n (Kn )
the probability of finding Kn vertices at the nth shell. p̃n (Kn ) are the
coefficients of the Taylor expansion of Gn (x).
For high shell numbers, by the law of large numbers, we expect the
number of vertices to increase by a factor of k̃ ≡ κ−1 at every consecutive
shell. (For ER networks k̃ = k.)  Hence, we can conclude that Gn (x)
n
converges to a function of the form f˜ (1 − x)k̃ for large n.
The solution of g(f∞ ) = f∞ gives the probability of a vertex not to con-
nect to infinity, i.e., not to belong to the giant component. Near criticality
we can guess a solution of the form f (y) = f∞ + ay −α . Expanding g we
obtain,

(30) g(f∞ + ay −α ) = g(f∞ ) + g  (f∞ )ay −α + o(y −α ).

 
Solving g f (y) = k̃f (y) we obtain α = − ln g  (f∞ )/ lnk̃. In [5], the
behavior of supercritical branching processes was studied. It was proved
that the tail of the distribution of layer sizes follows a power law, pn (Kn ) ∼
Knμ , with μ = α − 1.

yn be the
Let
k
fraction of vertices not connected after the nth shell. Then
yn = p̃(k)yn−1 , and thus yn = g(yn−1 ). Accordingly, the relation between
any two shells, m and n, is given by
 n−m 
(31) yn = f (1 − ym )k̃ = f∞ + a(1 − ym )−γ .

Using the same reasoning as above we obtain γ = ln g /f∞ lnk̃.

For a large m, ym = 1 − m l=1 Kl , where Kl is the fraction of vertices in

the lth shell. For large enough l, by the law of large numbers Kl+1 = k̃Kl .
Thus, the sum is a geometric series equal to k̃Km /(k̃ − 1). Therefore,
  −γ
yn ∼ (k̃Km / k̃ − 1 ) .
μ
As shown above P (Km ) ∼ Km . Using p(yn )dyn = p(Km )dKm , we
obtain p(yn ) ∼ yn−β , where β = 1 + (μ + 1)/γ. For ER networks μ + 1 = γ,
and thus β = 2, leading to p(yn ) ∼ yn−2 .
Scaling Properties of Complex Networks and Spanning Trees 165

6.2. Component size distribution

We now turn to study the distribution of the component sizes left when ver-
tices up to distance l from a given vertex are removed from the network [41].
The component size distribution in percolation at some concentration p is
determined using the formula
 
(32) Pp (s > S) ∼ S −τ +1 exp − S|p − pc |−1/σ .

This distribution can be approximated by considering the exponent to in-

troduce a sharp cutoff,

S −τ +1 , S < |p − pc |−1/σ ,
(33) Pp (s > S) ∼
0, S > |p − pc |−1/σ .

When links are uncovered one by one as the percolation threshold is ap-
proached in a uniform way, so for any a, P (p − pc < a) ∼ a. Therefore, it
follows that
 
(34) P (s > S) = Pp (s > S)P S < |p − pc |−1/σ = S −τ +1 S −σ .

Therefore, the component size distribution follows ns ∼ s−(τ +σ) . For ER

networks σ = 1/2 and τ = 5/2 and thus ns ∼ s−3 .

7. Summary

Complex networks present several different types of scaling behavior. A well-

studied property is the scale free nature of the degree distribution of numer-
ous real networks. The Internet, both at the AS and router levels, and the
WWW present a power law degree distribution for several orders of mag-
nitude. Several other natural, technological and social networks are also
believed to posses a power law degree distribution.
In physics, it is well known that systems at a critical point exhibit
fractal, scale invariant, properties. We discussed the critical exponents
associated with the percolation phase transition in general random networks
and in particular in ER and scale free networks. We surveyed the behavior
in different power law regimes and showed that scale free networks have
166 R. Cohen and S. Havlin

anomalous percolation properties even when the critical concentration is

finite. We have shown that the structure of the percolation components
at the phase transition point is fractal, and discussed how to calculate the
fractal dimension. We also calculated the size and typical length of the
largest components at the critical point and showed that it is a fractional
power of the network size.
We have also shown that despite an apparent contradiction between the
infinite dimensional nature of random networks and the notion of fractality,
many naturally occurring scale free networks have fractal properties even
far from criticality. This property is not present in most equilibrium and
growth models of scale free networks. However, it is observed in many
biological and socio-technological networks in the real world. This seems to
hint some scale free properties of the underlying mechanism for the network
formation.
Finally, it was shown that the distribution of the number of vertices at
a large distance from an arbitrary vertex follows a power law. The sizes of
components left after the removal of the l closest layers of a certain vertex
for large l also follow a power law distribution. Thus, even equilibrium
random networks far from criticality exhibit scaling properties in some of
their features.

References

[1] L. Addario-Berry, N. Broutin and B. Reed, The Diameter of the Minimum Spanning
Tree of a Complete Graph, in: Proc. of the Fourth Colloquium on Mathematics and
Computer Science, DMTCS, Nancy, France, (2006), pages 237–248.
[2] R. Albert and A.-L. Barabási, Statistical mechanics of complex networks, Rev. of
Mod. Phys., 74 (2002), 47.
[3] A.-L. Barabási and R. Albert, Emergence of scaling in random networks, Science,
286 (1999), 509–512.
[4] A.-L. Barabási and H. E. Stanley, Fractal Concepts in Surface Growth, Cambridge
University Press, 1995.
[5] N. H. Bingham, On the limit of a supercritical branching process, Journal of Applied
Probability, 25 (1988), 215–228.
[6] B. Bollobás, A probabilistic proof of an asymptotic formula for the number of
labeled regular graphs, European Journal of Combinatorics, 1 (1980), 311–316.
[7] B. Bollobás, Random Graphs, Academic Press, London (1985).
Scaling Properties of Complex Networks and Spanning Trees 167

[8] B. Bollobás and O. Riordan, Robustness and Vulnerability of Scale-Free Random

Graphs, Internet Mathematics, 1 (2003), 1–35.
[9] L. A. Braunstein, S. V. Buldyrev, R. Cohen, S. Havlin and H. E. Stanley, Optimal
paths in disordered complex networks, Physical Review Letters, 91 (2003), 168701.
[10] S. V. Buldyrev, S. Havlin and H. E. Stanley, Optimal paths in strong and weak
disorder: A unified approach, Phys. Rev. E., 73 (2006), 036128.
[11] A. Bunde and S. Havlin, editors, Fractals and Disordered System, Springer (1996).
[12] D. S. Callaway, M. E. J. Newman, S. H. Strogatz and D. J. Watts, Network
robustness and fragility: percolation on random graphs, Phys. Rev. Lett., 85 (2000),
5468–5471.
[13] M. Cieplak, A. Maritan and J. R. Banavar, Optimal paths and domain walls in the
strong disorder limit, Phys. Rev. Lett., 72 (1996), 2320–2323.
[14] M. Cieplak, A. Maritan and J. R. Banavar, Invasion Percolation and Eden Growth:
Geometry and Universality, Phys. Rev. Lett., 76 (1996), 3754–3757.
[15] R. Cohen, D. ben-Avraham and S. Havlin, Percolation Critical Exponents in Scale
Free Networks, Phys. Rev. E., 66 (2002), 036113.
[16] R. Cohen, K. Erez, D. ben-Avraham and S. Havlin, Resilience of the Internet to
Random Breakdown, Physical Review Letters, 85 (2000), 4626–4628.
[17] R. Cohen and S. Havlin, Scale free networks are ultrasmall, Phys. Rev. Lett., 90
(2003), 058701.
[18] R. Cohen and S. Havlin, Fractal dimensions of percolating networks, Physica A,
336 (2004), 6–13.
[19] R. Cohen, S. Havlin and D. ben Avraham, Structural properties of scale free
networks, in: S. Bornholdt and H. G. Schuster, editors, Handbook of graphs and
networks, chapter 4. Wiley-VCH, Berlin (2002).
[20] T. H. Cormen, C. Leiserson and R. Rivest, Introduction to algorithms, MIT Press
(1990).
[21] R. Dobrin and P. M. Duxbury, Minimum spanning trees on random networks, Phys.
Rev. Lett., 86 (2001), 5076–5079.
[22] S. N. Dorogovtsev, J. F. F. Mendes and A. N. Samukhin, Metric structure of
random networks, Nucl. Phys. B., 653 (2003), 307–338.
[23] S. N. Dorogovtsev, and J. F. F. Mendes, Evolution of Networks: From Biological
Nets to the Internet and WWW, Oxford University Press (2003).
[24] P. Erdős and A. Rényi, On random graphs, Publicationes Mathematicae, 6 (1959),
290–297.
[25] P. Erdős and A. Rényi, On the evolution of random graphs, Publications of the
Mathematical Institute of the Hungarian Academy of Sciences, 5 (1960), 17–61.
[26] P. Erdős and A. Rényi, On the strength of connectedness of a random graph, Acta
Mathematica Scientia Hungary, 12 (1961), 261–267.
[27] M. Faloutsos, P. Faloutsos and C. Faloutsos, On Power-Law Relationships of the
Internet Topology, Proceedings of the ACM SIGCOMM, 1999.
168 R. Cohen and S. Havlin

[28] E. N. Gilbert, Random Graphs, Annals of Mathematical Statistics, 30 (1959), 1141–

1144.
[29] T. Kalisky and R. Cohen, Width of percolation transition in complex networks,
Phys. Rev. E., 73 (2006), 035101.
[30] T. Kalisky, S. Sreenivasan, L. A. Braunstein, S. V. Buldyrev, S. Havlin and H. E.
Stanley, Scale-free networks emerging from weighted random graphs, Phys. Rev. E.,
73 (2006), 025103.
[31] S. Kirkpatrick, Percolation and conduction, Rev. Mod. Phys., 45 (1973), 574–588.
[32] M. E. J. Newman, Structure and function of complex networks, SIAM Review, 45
(2002), 167–256.
[33] M. E. J. Newman, S. H. Strogatz and D. J. Watts, Random graphs with arbitrary
degree distributions and their applications, Phys. Rev. E, 64 (2001), 026118.
[34] R. Pastor-Satorras and A. Vespignani, Epidemic spreading in scale-free networks,
Phys. Rev. Lett., 86 (2001), 3200–3203.
[35] R. Pastor-Satorras and A. Vespignani, Evolution and Structure of the Internet:
A Statistical Physics Approach, Cambridge University Press (2003).
[36] M. Porto, S. Havlin, S. Schwarzer and A. Bunde, Optimal path in strong disorder
and shortest path in invasion percolation with trapping, Phys. Rev. Lett., 79 (1997),
4060–4062.
[37] M. Porto, N. Schwartz, S. Havlin and A. Bunde, Optimal paths in disordered media:
Scaling of the crossover from self-similar to self-affine behavior, Phys. Rev. E., 60
(1999), R2448–R2451.
[38] A. Rapoport, A contribution to the theory of random and biased nets, Bulletin of
Mathematical Biophysic, 19 (1957), 257–271.
[39] H. D. Rozenfeld, S. Havlin and D. ben-Avraham, Fractal and transfractal recursive
scale-free nets, New Journal of Physics, 9 (2007), 175.
[40] J. Shao, S. V. Buldyrev, L. A. Braunstein, S. Havlin and H. E. Stanley, Structure
of shells in complex networks, arXiv:0903.2070.
[41] J. Shao, S. V. Buldyrev, R. Cohen, M. Kitsak, S. Havlin and H. E. Stanley, Fractal
boundaries of complex networks, EPL, 84 (2008), 48004.
[42] C. M. Song, S. Havlin and H. A. Makse, Self-similarity of complex networks, Nature,
433 (2005), 392.
[43] C. M. Song, S. Havlin and H. A. Makse, Origins of fractality in the growth of
complex networks, Nature Physics, 2 (2006), 275.
[44] D. Stauffer and A. Aharony, Introduction to Percolation Theory, Taylor and Fran-
cis, 2nd edition (1994).
[45] G. Szekeres, Distribution of labelled trees by diameter, Lect. Notes in Math., 1036
(1983), 392–397.
[46] R. van der Hofstad, G. Hooghiemstra and P. Van Mieghem, Distances in random
graphs with finite mean and finite variance degrees, Random Structures and Algo-
rithms, 26 (2006), 76–123.
Scaling Properties of Complex Networks and Spanning Trees 169

[47] R. van der Hofstad, G. Hooghiemstra and P. van Mieghem, First passage per-
colation on the random graph, Probability in the Engineering and Informational
Sciences, 15 (2001), 225–237.
[48] R. van der Hofstad, G. Hooghiemstra and D. Znamenski, Distances in random
graphs with finite mean and infinite variance degrees, Electronic Journal of Prob-
ability, 12 (2007), 703–766.
[49] D. J. Watts and S. H. Strogatz, Collective dynamics of “small world” networks,
Nature, 393 (1998), 440–442.
[50] G. H. Weiss, Aspects and Applications of the Random Walk, North-Holland (1994).
[51] Z. Wu, L. A. Braunstein, S. Havlin and H. E. Stanley, Transport in weighted
networks: partition into superhighways and roads, Phys. Rev. Lett., 96 (2006),
148702.

Reuven Cohen Shlomo Havlin

Department of Mathematics Department of Physics
Bar-Ilan University Bar-Ilan University
Ramat-Gan 52900 Ramat-Gan 52900
Israel Israel
e-mail: [email protected] e-mail: [email protected]
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 171–202.

Chapter 4
Random Tree Growth
with Branching Processes – a Survey

ANNA RUDAS and BÁLINT TÓTH

We investigate the asymptotic properties of a random tree growth model which

generalizes the basic concept of preferential attachment. The Barabási–Albert
random graph model is based on the idea that the popularity of a vertex in the
graph (the probability that a new vertex will be attached to it) is proportional to
its current degree. The dependency on the degree, the so-called weight function,
is linear in this model. We give results which are valid for a much wider class
of weight functions. This generalized model has been introduced by Krapivsky
and Redner in the physics literature. The method of re-phrasing the model in
a continuous-time setting makes it possible to connect the problem to the well-
developed theory of branching processes. We give local results, concerning the
neighborhood of a “typical” vertex in the tree, and also global ones, about the
repartition of mass between subtrees under fixed vertices.

1. Random Tree Model

A natural concept of randomly growing trees is the class of models where

at each discrete time step, a new vertex appears, and is attached to an
already existing vertex randomly, the distribution of the choice depending
on the actual degrees of the vertices currently apparent in the tree. This
dependence on the degree structure is characterised by a weight function
w : N → R+ .
172 A. Rudas and B. Tóth

The concept of preferential attachment generally means that w is an in-

creasing function. One of the realizations of this idea is the Barabási–Albert
graph [2], which model reproduces certain phenomena observed in real-world
networks, the power-law decay of the degree sequence, for example. This
was proved in a mathematically precise way in [3] and, independently, in
[15]. The Barabási–Albert model, for trees, corresponds to the special case
of the model considered in this survey, when w is chosen to be linear. In [3]
and [15], the techniques used strongly depend on martingales that are ap-
parent in the system only in the linear case. For a survey on random graph
models that produce scale-free behavior, see Bollobás and Riordan [4] and
Chapter 4 of Durrett [7].

General weight functions are considered in the model of Krapivsky and

Redner [12], [13]. There w(k) ∼ kγ , and non-rigorous results are obtained,
showing the different behavior for γ > 1 and γ ≤ 1. In the first regime
the limiting object does not have a non-trivial degree sequence: a single
dominant vertex appears which is linked to almost every other vertex, the
others having only finite degree. This statement is made precise and proved
rigorously in [17]. See also [5] for a related model. In this survey, following
[20], we consider weight functions for which this does not happen, our class
includes the second regime γ ≤ 1 mentioned above.

Certain similar random recursive trees and random plane-oriented trees

have been studied before, see, for example, [21].

1.1. Notation

We consider rooted ordered trees, which are also called family trees or rooted
planar trees in the literature.

In order to refer to these trees it is convenient to use genealogical

phrasing. The tree is thus regarded as the coding of the evolution of a
population stemming from one individual (the root of the tree), whose
“children” form the “first generation” (these are the vertices connected
directly to the root). In general, the edges of the tree represent parent-
child relations, the parent always being the one closer to the root. The
birth order between brothers is also taken into account, this is represented
by the tree being an ordered tree (planar tree). The vertices are labelled by
Random Tree Growth with Branching Processes – a Survey 173

the set
∞

N = Zn+ , where Z+ := {1, 2, . . . }, Z0+ := {∅},
n=0

as follows. ∅ denotes the root of the tree, its firstborn child is labeled
by 1, the second one by 2, etc., all the vertices in the first generation are
thus labeled with the elements of Z. Similarly, in general, the children
of x = (i1 , i2 , . . . , ik ) are labeled by (i1 , i2 , . . . , ik , 1), (i1 , i2 , . . . , ik , 2), etc.
Thus, if a vertex has label x = (i1 , i2 , . . . , ik ) ∈ N , then it is the ith k
child of its parent, which is the ith k−1 child of its own parent and so on. If
x = (i1 , i2 , . . . , ik ) and y = (j1 , j2 , . . . , jl ) then we will use the shorthand
notation xy for the concatenation (i1 , i2 , . . . , ik , j1 , j2 , . . . , jl ), and with a
slight abuse of notation for n ∈ Z, we use xn for (i1 , i2 , . . . , ik , n).
There is a natural partial ordering  on N , namely, x  z if x is ancestor
of z, so if ∃y ∈ N such that z = xy. If we want to exclude equality, we use
x ≺ z, which thus means that x  z but x = z.
We can identify a rooted ordered tree with the set of labels of the vertices,
since this set already identifies the set of edges in the tree. It is clear that
a subset G ⊂ N may represent a rooted ordered tree iff ∅ ∈ G, and for
each (i1 , i2 , . . . , ik ) ∈ G we have (i1 , i2 , . . . , ik − 1) ∈ G if ik > 1, and
(i1 , i2 , . . . , ik−1 ) ∈ G if ik = 1.
G will denote the set of all finite, rooted ordered trees. The degree of
vertex x ∈ G will denote the number of its children in G:

deg (x, G) := max {n ∈ Z+ : xn ∈ G}.

Thus degree in this paper is one less than the usual graph theoretic degree,
except for the root, where it is the same. The subtree rooted at a vertex
x ∈ G is:
G↓x := {y : xy ∈ G},
this is just the progeny of x viewed as a rooted ordered tree.

1.2. The Model

First we describe the growth model in discrete time. Then we present a

continuous-time model which, at certain random stopping times, is equiv-
alent to the previous one. The method of investigating the discrete time
174 A. Rudas and B. Tóth

growth model by introducing the continuous-time setting described below

appears in [17] and in [20], independently of each other.

1.2.1. Discrete Time Model. Given the weight function w : N → R+ , let

us define the following discrete time Markov chain Υd on the countable state
space G , with initial state Υd (0) = {∅}. If for n ≥ 0 we have Υd (n) = G,
then for a vertex x ∈ G let k := deg (x, G) + 1. Using this notation, let the
transition probabilities be
 
w deg (x, G)
P(Υ (n + 1) = G ∪ {xk} | Υ (n) = G) =

d d  .
y∈G w deg (y, G)

In other words, at each time step a new vertex appears, and attaches to
exactly one already existing vertex. If the tree at the appropriate time is
G, then
 the probability
 of choosing vertex x in the tree G is proportional
to w deg (x, G) .

1.2.2. Continuous Time Model. Given the weight function w : N → R+ ,

let X(t) be a Markovian pure birth process with X(0) = 0 and birth rates
  
P X(t + dt) = k + 1  X(t) = k = w(k)dt + o(dt).

Let  : [0, ∞) → (0, ∞] be the density of the point process corresponding to

the pure birth process X(t), namely let
 
(1) (t) = lim ε−1 P (t, t + ε) contains a point from X .
ε→0

We denote the (formal) Laplace transform of  by (: (0, ∞) → (0, ∞]:

 ∞ ∞ n−1

−λt w(i)
(2) ((λ) := e (t) dt = .
0 n=1 i=0
λ + w(i)

The rightmost expression of ((λ) is easily computed, using the fact that
the intervals between successive jumps of X(t) are independent exponen-
tially distributed random variables of parameters w(0), w(1), w(2), . . . , re-
spectively. Let
 
(3) λ := inf λ : ((λ) < ∞ .
Random Tree Growth with Branching Processes – a Survey 175

Throughout this paper we impose the following condition on the weight

function w:

(M) lim ((λ) > 1.

λλ

We are now ready to define our randomly growing tree Υ(t) which will be
a continuous time, time-homogeneous Markov chain on the countable state
space G , with initial state Υ(0) = {∅}.
The jump rates are the following: if for a t ≥ 0 we haveΥ(t) = G then
the process may jump to G ∪ {xk} with rate w deg (x, G) where x ∈ G
 existing vertex x ∈ Υ(t) ‘gives
and k = deg (x, G) + 1. This means that each
birth to a child’ with rate w( deg x, Υ(t) ), independently of the others.
Note that condition (M) implies
∞
 1
(4) =∞
w(k)
k=0

and hence it follows that the Markov chain Υ(t) is well defined for t ∈ [0, ∞),
it does not blow up in finite time. If, on the other hand, (4) does not
hold, then λ is not finite. A rigorous proof of this statement follows from
the connection with general branching processes (see Sect. 3) for which the
related statement is derived in [9].
We define the total weight of a tree G ∈ G as
  
W (G) := w deg (x, G) .
x∈G

Described in other words, the Markov chain Υ(t) evolves as follows: assum-
ing Υ(t−) = G, at time t a new vertex is added to it with total rate W (G),
and it is attached with an edge to exactly one already existing vertex, which
is x ∈ G with probability
 
w deg (x, G)

 .
y∈G w deg (y, G)

Therefore, if we only look at our continuous time process at the stopping

times when a new vertex is just added to the randomly growing tree:
 
(5) Tn := inf {t :  Υ(t) = n + 1}
176 A. Rudas and B. Tóth

then we get the discrete time model: Υ(Tn ) has the same distribution as
Υd (n), the discrete time model at time n.

1.3. Some Additional Notation

We are ready to introduce the notations τx and σx , as follows. Let τx be

the birth time of vertex x,
 
(6) τx := inf t > 0 : x ∈ Υ(t) .

Let σx be the time we have to wait for the appearance of vertex x, starting
from the moment that its birth is actually possible (e.g. when no other
vertex is obliged to be born before it). Namely, let
(a) σ∅ := 0,
(b) σy1 := τy1 − τy , for any y ∈ N ,
(c) and σyk := τyk − τy(k−1) , for each y ∈ N and k ≥ 2.
Also, we will need the concept of historical orderings of the vertices in
a finite tree, as follows.
Consider a G ∈ G . An ordering s = (s0 , s1 , . . . , s|G|−1 ) of the elements
of G is called historical if it gives a possible ’birth order’ of the vertices in
G, formally if for each 0 ≤ i ≤ |G| − 1 we have {s0 , s1 , . . . , si } ∈ G . The
set of all historical orderings of G ∈ G will be denoted S (G). For a fixed
s ∈ S (G) the rooted ordered trees

(7) G(s, i) := {s0 , s1 , . . . , si } ⊂ G

give the evolution of G in this historical ordering s.

Let G ∈ G and one of its historical orderings s = (s0 , s1 , . . . , s|G|−1 ) ∈
S (G) be fixed. The historical sequence of total weights are defined as
 
(8) W (G, s, i) := W G(s, i)

for 0 ≤ i ≤ |G| − 1 while the respective weights of the appearing vertices

are defined as
 
(9) w(G, s, i) := w( deg (si )1 , G(s, i − 1) ),
Random Tree Growth with Branching Processes – a Survey 177
1 
for 1 ≤ i ≤ |G| − 1, here (si ) denotes the parent of si . Since deg (si )1 ,
G(s, i − 1) is the degree of si ’s parent just before si appeared, w(G, s, i)
is the rate with which our random tree process jumps from G(s, i − 1) to
G(s, i).

2. Local Properties

We ask questions about the neighborhood of the “typical” vertex (e.g.

sampled uniformly randomly from the tree), after a long time. We state
theorems regarding the limit distribution of the number of children, and of
the subtree under the uniformly selected vertex. We then comment on the
case of linear weight functions, where the formulae gain a simpler explicit
form. After that we state a more general version of these theorems, which
turns out to be the analogue of an important theorem in the theory of
general branching processes (see Sect. 3). At the end of the section we give
an argument on the convergence in Theorem 2, different from that of the
analoguous theorem discussed in [16]. Although this approach does not
give almost sure convergence, it is elementary and instructive, and gives
convergence in probability.

2.1. Statement of Results

Note that from condition (M) it follows that the equation

∞ n−1
 w(i)
(10) ρ((λ) = =1
λ + w(i)
n=1 i=0

has a unique root λ∗ > 0, called the Malthusian parameter. Now we are
ready to state our theorems, quoted from [20].

Theorem 1. Consider a weight function w which satisfies condition (M)

and let λ∗ be defined by (10). For any t ≥ 0 let ζt denote a random
vertex which is, once Υ(t) is given, selected uniformly from Υ(t). Then the
following limits hold almost surely:
178 A. Rudas and B. Tóth

(a) For any fixed k ∈ N

 
lim P(deg ζt , Υ(t) = k)
t→∞
 
|{x ∈ Υ(t) : deg x, Υ(t) = k }|
 
= lim  Υ(t) = pw (k),
t→∞

where the limit degree distribution pw on N is given by

λ∗ w(i)
k−1
pw (k) := ∗
.
λ + w(k) λ∗ + w(i)
i=0

(b) For any fixed G ∈ G

 
  | x ∈ Υ(t) : Υ(t)↓x = G
 
|
lim P Υ(t)↓ζt = G = lim  Υ(t) = πw (G),
t→∞ t→∞

where the limit subtree distribution πw on G is given by

|G|−2
 λ∗ w(G, s, i + 1)
πw (G) := ∗
.
λ + W (G) λ∗ + W (G, s, i)
s∈S (G) i=0

These results are direct consequences of the following theorem, as shown

in detail in [20], with the choices of ϕ : G → R as

ϕ(H) = 11{deg (∅,H)=k} ,

or, respectively,
ϕ(H) = 11G (H).

Theorem 2. Consider a weight function w satisfying condition (M) and let

λ∗ be defined by equation (10). Consider a bounded function ϕ : G → R.
Then the following limit holds almost surely:
   ∞
1  −λ∗ t
 

lim 
t→∞ Υ(t)
 ϕ Υ(t) ↓x = λ∗
e E( ϕ Υ(t) ) dt.
x∈Υ(t) 0

Remark. Based on Theorem 2 it is also possible to “look back” from

the uniform random vertex, and ask local questions regarding its parent,
Random Tree Growth with Branching Processes – a Survey 179

grandparent, etc. This Theorem is an application of a result of Nerman, see

the remark after Theorem 4, and for the detailed proof see [20]. The related
general concept of fringe distributions can be found in Aldous [1].

2.2. Linear Weight Function

In the linear case w(k) = αk + β (α, β > 0), all computations regarding
the distributions pw and πw in Theorem 1 are rather explicit. In [3] and
[15], the degree distribution is given, for any fixed finite number of vertices.
Our computations in the linear case (see below), reproduce the asymptotic
degree distribution p, as the size of the tree tends to infinity.
For sake of completeness, we perform these (explicit and straightfor-
ward) computations for the linear case. Multiplying the rate function with
a positive constant only means the rescaling of time in our model thus it is
enough to consider w(k) = k + β (with β > 0). In this case it is straight-
β
forward to compute that condition (M) holds, ((λ) = λ−1 , λ = 1 and
∗
λ = 1 + β. Thus both Theorems 2 and 1 hold.
For the asymptotic degree distribution we get

(k − 1 + β)k
p(k) = (1 + β) ,
(k + 1 + 2β)k+1

where we used the shorthand notation

k−1
Γ(x + 1)
(x)k := (x − i) = , k = 0, 1, 2, . . . .
Γ(x − k + 1)
i=0

For the calculation of π(G) first we show that the sum which defines it
contains identical elements. In order to avoid heavy notation, during the
following computations we will use n := |G| − 1 and deg (x) instead of
deg (x, G).
Clearly, for any s ∈ S (G)

n−1  deg
(x)−1   
w(G, s, i + 1) = w(j) = deg (x) − 1 + β deg (x)
.
i=0 x∈G j=0 x∈G
180 A. Rudas and B. Tóth

(Actually, the first equality holds for every weight function w.) It is also
easy to see that for any G ∈ G
 
W (G) = deg (x) + β = |G|(1 + β) − 1,
x∈G

thus for any s ∈ S (G)

λ∗
n−1
1 1
=   .
∗
λ + W (G)
i=0
∗
λ + W (G, s, i) (1 + β) n + 2 − (1 + β)−1 n+1
n

Therefore
  
  x∈G deg (x) − 1 + β deg (x)
π(G) =  S (G)   .
(1 + β)n n + 2 − (1 + β)−1 n+1

In the β = 1 case (i.e. if we consider random tree proposed in [2]) the

previous calculations give

4
p(k) =
(k + 1)(k + 2)(k + 3)

and  
2 S (G)
π(G) =   deg (x)!
2|G| + 1 !! x∈G
 
Although the value of  S (G) cannot be written as the function of

degrees of G only, one can compute it using the values G↓x  for x ∈ G, for
the detalis see [20].

2.3. Proof of Convergence in Probability

In this section we present a proof of the convergence in probability of certain

ratios of variables, see Theorem 3. The result formulated here exists in a
stronger form, namely, the convergence holds in the almost sure sense, as
stated by Theorem 4 in Sect. 3. The proof presented here, though, uses
more elementary methods and it is probabilistically instructive.
In order to simplify technicalities, we restrict the class of weight func-
tions from those satisfying condition (M) to a somewhat smaller, but still
Random Tree Growth with Branching Processes – a Survey 181

very wide class. We demand in this section that w(k) → ∞ as k → ∞, with

the weight function varying regularly.

(11) w(k) = k γ + v(k)

at some 0 < γ ≤ 1 and v(k) = o(kγ ) as k → ∞. (We do not need

monotonicity for w.)
Let us fix w(0) = 1, which can be done without loss of generality, since
multiplying all w(k) by a constant just corresponds to rescaling time in the
continuous time model.
Let λ∗ be the constant defined by (10), and with the letters ϕ and ψ we
denote positive bounded functions ϕ, ψ : G → R+ .
Define   
Ztϕ := ϕ Υ(t)↓x
x∈Υ(t)

(the analogous definition in Sect. 3 is (24)). We use the notation

 ∞

(12) 
κ := −∂λ ρ((λ) λ=λ∗ =
∗
te−λ t ρ(t) dt < ∞.
0

We also introduce the notation

 ∞
 
(13) (
ϕ(λ) := e−λs E(ϕ Υ(s) ) ds.
0

Theorem 3. Let w satisfy the conditions described in the beginning of

Sect. 2.3. Then
Ztϕ ( ∗)
ϕ(λ
→
Zψ ( ∗)
ψ(λ t

in probability, as t → ∞.
To prove Theorem 3 we need Lemmas 1, 2 and 3 below.
Lemma 1.
∗ 1
(14) E(e−λ t Ztϕ ) → ( ∗ ) =: dϕ ,
ϕ(λ as t → ∞.
κ

Remark. See Sect. 3 and Theorem 5 therein for the analogous, more general
∗
result. There we see that e−λ t Ztϕ itself converges almost surely to a random
variable with the appropriate expectation.
182 A. Rudas and B. Tóth

Proof. The key observation is the so-called basic decomposition, namely

that
   ϕj
(15) Ztϕ = ϕ Υ(t) + Zt ,
j∈N

where j runs over the children of the root, ϕj (G) := ϕ(G↓j ), and recall that
τj is the birth time of vertex j.
The advantage of this formula is due to the fact that given the sequence
ϕ ϕ
(τj )j∈Z+ , Zt j has the same conditional distribution as Zt−τj
.
At this point observe that if Ztϕ is of some exponential order eλt , then
λ must be the one defined by equation (10). This can be seen if we take
expectation of both sides in equation (15), supposing that limt→∞ e−λt Ztϕ
exists almost surely, we can write
    
E lim e−λt Ztϕ = E e−λτj lim e−λ(t−τj ) Zt j
ϕ
t→∞ t→∞
j∈N

    
 
−λτj
  −λ(t−τj ) ϕj −λτj
= E e E lim e Zt =E e E lim e−λt Ztϕ ,
t→∞ t→∞
j∈N j∈N

ϕ d
since limt→∞ e−λ(t−τj ) Zt j = limt→∞ e−λt Ztϕ .
So if the limit exists almost surely, and is non-zero and finite, then
 
−λτj
E e = ρ((λ) = 1
j∈N

must hold (compare with (10)).

For the convergence itself, using the notation

(16) mϕ ϕ
t := E(Zt ),

taking expectation on both sides of (15) in two steps (first conditionally on

(τj )j∈Z+ , then taking expectation regarding (τj )j∈Z+ ), we get



 t
(17) mϕ
t = E(ϕ Υ(t) ) + mϕ
t−s (s) ds.
0
Random Tree Growth with Branching Processes – a Survey 183

Taking the Laplace transform of both sides, we have

(
m(λ) (
= ϕ(λ) (
+ m(λ)(
ρ(λ),

so formally
(
ϕ(λ)
(
m(λ) = .
1 − ρ((λ)
From condition (M) it follows that there is an interval  of positive
 length
below λ∗ where the Laplace transform is finite, so 1/ 1 − ρ((λ) has a simple
pole at λ∗ (it is easy to check that ρ( (λ∗ ) < 0 and ρ( (λ∗ ) > 0). Taking series
expansion and inverse Laplace transform results that

1 ∗ ∗
mϕ
t = ( ∗ ) eλ t + o(eλ t ),
ϕ(λ
κ
so, indeed, the statement of the lemma holds.

Recall the notation in Sect. 1.2.2, the birth times of the vertices in the
first generation of the tree, (τj )j>0 , constitute the point process X. The
density function  has already been introduced, see (1). Similarly, let us
denote the second correlation function by 2 , namely, for u = s, let

2 (u, s) := lim (εδ)−1 P (u, u + ε)
ε,δ→0

and (s, s + δ) both contain a point from X ,

and we define it to be 0 if u = s.
The following estimates are needed.

Lemma 2. Suppose that w satisfies the conditions described in the begin-

ning of Sect. 2.3. Then
(a)
 ∞
∗
(18) C1 := e−2λ s (s) ds < 1,
0

(b)
 ∞ ∞
∗ (u+s)
(19) C2 := e−λ 2 (u, s) du ds < ∞.
0 0
184 A. Rudas and B. Tóth

Proof. The first statement is obvious, considering that

 ∞
∗
e−2λ s (s) ds = ((2λ∗ ) < ((λ∗ ) < 1,
0

since ( strictly decreases and λ∗ > 0.

As for statement (19), write C2 as follows:
  ∗ 
C2 = 2 E e−λ (τi +τj ) .
1≤i<j

Since for any i < j, τj can be decomposed as the sum of the two independent
variables τj = τi + (τj − τi ), it can be seen that
 ∗  ∗ ∗  ∗  ∗
E e−λ τj = E(e−λ τi e−λ (τj −τi ) ) = E e−λ τi E(e−λ (τj −τi ) ).

It now follows that

 ∗ 
      E e−λ τj
−2λ∗ τ −λ∗ (τ −2λ∗ τ
C2 = 2 E e i
E(e j −τi )
)=2 E e i  .
1≤i<j 1≤i<j
E e−λ∗ τi

From here we get the estimate

 
 E e−2λ∗ τi  ∗ 
C2 = 2  ∗
 E e−λ τj
E e −λ τ i
1≤i<j
  ∗   
E e−2λ τi   
  −λ∗ τj
≤2 E e ,
i
E e−λ∗ τi j

where the second sum is just ((λ∗ ) = 1, while the first is

∞ n−1
 λ∗ + w(k) 
∞ n−1
 
λ∗
= 1 − .
2λ∗ + w(k) 2λ∗ + w(k)
n=1 k=0 n=1 k=0

So far this was all general, but under the specific assumptions (see (11))
onw, the final expression is finite. If γ = 1 then the logarithm of the product
is − 1 + o(1) λ∗ log n + O(1), so for any ε > 0, the sum can be bounded

∗
by O(1) + n−(λ +ε) . This
is finite, since λ∗ > λ = 1. If 0 < γ < 1, then
the sum can be bounded by exp (−cn1−γ ) for some c > 0.
This completes the proof of Lemma 2.
Random Tree Growth with Branching Processes – a Survey 185

Lemma 3.
∗
E(e−2λ t Ztϕ Ztψ ) → C ϕ(λ ( ∗)
( ∗ )ψ(λ
for some constant C > 0. (This constant depends on the weight function w.)

Proof. According to the basic decomposition, we can write

   
Ztϕ Ztψ = ϕ Υ(t) ψ Υ(t)
∞ ∞
  ψ   ϕ
+ ϕ Υ(t) Zt j + ψ Υ(t) Zt j
j=1 j=1

∞
 ∞

ϕ ψ ψ
+ Zt j Zt j + Ztϕi Zt j .
j=1 i=j

Taking expectation yields

   
(20) mϕ,ψ
t := E(Ztϕ Ztψ ) = E(ϕ Υ(t) ψ Υ(t) )
 ∞ ∞ 
  ψ   ϕ
+ E ϕ Υ(t) Zt j + ψ Υ(t) Zt j
j=1 j=1
 t  t t
+ mϕ,ψ
t−s ρ(s) ds + mϕ ψ
t−u mt−s ρ2 (u, s) du ds,
0 0 0

recall the notation mϕ ϕ

t = E(Zt ) from (16).
∗
After multiplying the equation by e−2λ t , we can easily identify the limit
(as t → ∞) of the first, second and fourth terms in (20), as follows.
First term: since ϕ and ψ are bounded,
∗    
lim e−2λ t E(ϕ Υ(t) ψ Υ(t) )
t→∞

is trivially 0.
Second term: let ϕ be bounded by the constant D < ∞, then
 ∞  
−2λ∗ t
  ψ −2λ∗ t
t
e E ϕ Υ(t) Zt j ≤ De mψ
t−s ρ(s) ds
j=1 0

∗t  
= De−2λ (mψt − E(ψ Υ(t) ),
186 A. Rudas and B. Tóth

∗
the limit of which is 0 since mψ t is of order e
λ t (see (17)), and since ψ is

bounded.
−λ∗ t ϕ ∗
Fourth term: Let us introduce m̃ϕ t := e mt , and m̃ϕ,ψ
t := e−2λ t mϕ,ψ
t .
ϕ ψ
By Lemma 1, limt→∞ m̃t = dϕ and limt→∞ m̃t = dψ . With these, and using
Lemma 2,
 t t
−λ∗ (u+s)
lim m̃ϕ ψ
t−s m̃t−u e 2 (u, s) du ds = C2 dϕ dψ ,
t→∞ 0 0

by dominated convergence.
This way we see
 t
ϕ,ψ −2λ∗ s
(21) m̃t = m̃ϕ,ψ
t−s e (s) ds + C2 dϕ dψ + εt ,
0

where εt → 0 as t → ∞.
Now let us assume for a moment that the limit dϕ,ψ := limt→∞ m̃ϕ,ψ
t
does exist. In this case dominated convergence could also be used in the
third (normalized) term of (20), and the following would be true:
 t  t
−2λ∗ t ∗
−2λ s
lim e mϕ,ψ
t−s (s) ds = lim m̃ϕ,ψ
t−s e (s) ds = C1 dϕ,ψ ,
t→∞ 0 t→∞ 0

recall the notation and result in Lemma 2.

This way if m̃ϕ,ψ
t was convergent, then its limit could only be

C2
dϕ,ψ = dϕ dψ ,
1 − C1
recall that C1 < 1, by Lemma 2.
To show that the limit really exists, first note that m̃ϕ,ψ
t is bounded.
This is true since with Mt := sups<t m̃s and M := sups>0 m̃ϕ
ϕ,ψ ϕ,ψ ϕ
s , we get

Mtϕ,ψ ≤ E + Mtϕ,ψ C1 + M ϕ M ψ C2 ,

where E is an upper bound for εt . This way Mtϕ,ψ is bounded by a constant

independent of t (again, C1 < 1), thus m̃ϕ,ψ
t is bounded.
Let us introduce the difference of m̃ϕ,ψ
t and its supposed limit,

C2
(22) nt := m̃ϕ,ψ − dϕ dψ
t
1 − C1
Random Tree Growth with Branching Processes – a Survey 187

and rearrange equation (21),

 t
∗
nt = nt−s e−2λ s (s) ds + ε̄t ,
0

where ε̄t → 0 as t → ∞.
Since we have shown that m̃ϕ,ψ t is bounded, so is nt . Let Nt :=
sups≥t |ns |, Ēt := sups≥t |ε̄s |, and fix arbitrarily 0 < u < t0 . For these
and for all t > t0
 
 u 
|nt | ≤ |ε̄t | +  ns e−2λ∗ (t−s)
(t − s) ds
0
 t 
 
+  (t − s) ds.
−2λ ∗ (t−s)
ns e
u

∞ ∗ ∞ ∗
Recall that 0 e−λ t (t) dt = ((λ∗ ) = 1 and 0 e−2λ t (t) dt = ((2λ∗ ) =
C1 , and thus
∗ (t−u)
|nt | ≤ Ēt0 + e−λ N0 + Nu C1 .

This way
∗ (t
Nt0 ≤ Ēt0 + e−λ 0 −u)
N0 + Nu C1 .

Letting t0 → ∞ with u remaining fixed,

N∞ ≤ Nu C1 ,

and now letting u → ∞

N∞ ≤ N∞ C1 .

Since C1 < 1 this means that N∞ = 0, so m̃ϕ,ψt is convergent and its limit
is
∗ C2
lim m̃ϕ,ψ = lim e−2λ t E(Ztϕ Ztψ ) = dϕ dψ ,
t→∞ t t→∞ 1 − C1
as stated by the lemma.

Now we are ready to prove Theorem 3.

188 A. Rudas and B. Tóth

∗ ∗
Proof of Theorem 3. Let At := e−λ t Ztϕ and Bt := e−λ t Ztψ . Denote
the limits of their expectations a := limt→∞ E(At ) and b := limt→∞ E(Bt ).
From Lemma 3 we see that E(At Bt ) → Cab, and also E(A2t ) → Ca2 and
E(Bt2 ) → Cb2 . This implies that
 
E (bAt − aBt )2 → 0

so (bAt − aBt ) → 0 in L2 and thus in probability, too.

Now fix any positive δ, η > 0, then
  
 At a 
P  −  > δ
Bt b
)  *  )  * 
 At a   At a 
   
=P  Bt − b  > δ ∩ {Bt ≥ η} + P  Bt − b  > δ ∩ {Bt < η}
 
≤ P |bAt − aBt | > bδη + P(Bt < η).

Since the first term tends to 0 by the previous observation, it remains to

show that in the limit, Bt does not have a positive mass at 0, and then the
statement of the theorem is true.
But since (Bt )t>0 is tight, in every subsequence there is a sub-subse-
quence (tn )n>0 along which Btn converges weakly to some random vari-
able Y . By (15) for this variable, in distribution,
∞
 ∗τ
Y = e−λ j
Yj ,
j=1

where the Yj are iid with the same distribution as Y .

This means that
 k
P(Y = 0) = P(Yj = 0 for all j) = lim P(Y = 0) .
k→∞

It follows that if Y had a positive mass at 0, then Y would be a random

variable that is almost surely 0. Since we know that its expectation tends
to a positive limit, this could only happen if E(Bt2 ) converged to ∞, but
in fact it converges to a finite positive limit, according to Lemma 3. Thus,
Y does not have a positive mass at 0, so the statement of Theorem 3 holds.
Random Tree Growth with Branching Processes – a Survey 189

3. Branching Processes

The Random Tree Model, defined in continuous time, has the big advantage
that it fits into the framework of the well-established theory of branching
processes. We give a brief introduction to the fundamentals and state the
theorems that we rely on. We do not give a broad survey on the most
general types of branching processes here, we choose to focus on the results
which may be applied to our process. For more details see the monograph
[9] or the papers [10], [16], [18] and the references therein. For a survey on
branching processes, trees and superprocesses, see [14].
In the case of a general branching process, there is a population in which
each individual reproduces at ages according to i.i.d. copies of a random
point process ξ on [0, ∞). We denote by ξ(t) the ξ-measure of [0, t], this the
random number of children an individual has up to time t. (In the case of
our model, ξ is the Markovian pure birth process X.)
The individuals in the population are labelled with the elements of N ,
the same way as described in Sect. 1.1. The basic probability space is

(Ω, A , P ) = (Ωx , Ax , Px ),
x∈N

where (Ωx , Ax , Px ) are identical spaces on which ξx are distributed like ξ.

For each x ∈ N there is a .↓x shift defined on Ω by
 
ω↓x y
= ωxy ,

in plain words, ω↓x is the life of the progeny of x, regarding x as the ancestor.
The birth times τx of the individuals are defined in the obvious way:
τ0 = 0 and if x = xn with n ∈ Z+ then
 
(23) τx = τx + inf t : ξx (t) ≥ n ,

just like in the random tree model, see (6).

The branching process is often counted by a random characteristic, this
can be any real-valued process {Φ : R × Ω → R}. For each individual x, Φx
is defined by
Φx (t, ω) = Φ(τx + t, ω↓x ),
190 A. Rudas and B. Tóth

in plain words Φx (t) denotes the value of Φ evaluated on the progeny of x,

regarding x as the ancestor, at the time when x is of age t. We can think
about Φx (t) as a ‘score’ given to x when its age is t. With this,

(24) ZtΦ := Φx (t − τx )
x∈N

is the branching process counted by the random characteristic Φ (the ‘total

score’ of the population at time t).
For our applications we only consider random characteristics which are
0 for t < 0 and equal to a bounded deterministic function of the rooted tree
for t ≥ 0.
This means that only those individuals contribute to ZtΦ which are born
up to time t and their contribution is a deterministic function of their
progeny tree. (Random characteristics may be defined in a more general
way, see e.g. [9], [10].) One of the important examples is Φ(t) = 11{t ≥ 0}
when ZtΦ is just the total number of individuals born up to time t.
The Laplace-transform of dξ(t) is of great importance, we denote this
random variable by:
 ∞
(25) ( :=
ξ(λ) e−λt dξ(t).
0

We are interested in supercritical, Malthusian processes, meaning that there

exists a finite 0 < λ∗ < ∞ (the so-called Malthusian parameter) for which

(26) ( ∗ ) = 1,
E ξ(λ

and also
 ∞
 
(27) κ = −∂λ (
E ξ(λ) |λ=λ∗ = E
∗
te−λ t dξ(t) < ∞.
0

(The last property means that the process is Malthusian and the first
means that it is supercritical.)
Also, we require the reproduction to be non-lattice, which means that
the jumps of ξ(t) cannot be supported by any lattice {0, d, 2d, . . . }, d > 0
with probability one.
We quote here a weaker form of Theorem 6.3 from [16], using its exten-
sion which appears in Section 7 of the same paper. This way the conditions
of the original theorem are fulfilled automatically.
Random Tree Growth with Branching Processes – a Survey 191

Theorem 4 (Nerman, [16]). Consider a supercritical, Malthusian branch-

ing process with Malthusian parameter λ∗ , counted by two random charac-
teristics Φ(t) and Ψ(t) which have the properties described above (i.e. they
are 0 for t < 0 and a deterministic bounded function of the progeny tree for
t ≥ 0). Suppose that there exists a λ < λ∗ for which
( < ∞.
E ξ(λ)

Then, almost surely

ZtΦ ( ∗)
Φ(λ
→ as t → ∞,
ZtΨ ( ∗)
Ψ(λ
∞  
(
where Φ(λ) = exp (−λt)E Φ(t) dt.
0

Remark. Clearly, the time-evolution of the population has the same dis-
tribution as the evolution of the continuous time Random Tree Model cor-
responding to the weight function w. The vertices are the respective indi-
viduals and edges are the parent-child relations. It is also not hard to see
(
that the function E ξ(λ) for the branching process is the same as ((λ), and
the two definitions (12) and (27) for κ agree. This means that by condi-
tion (M) we may apply Theorem 4 with appropriate random characteristics.
Given any bounded
 function ϕ : G → R, setting the characteristics Φ, Ψ as
Φ(t) := ϕ Υ(t) 11{t ≥ 0} and Ψ(t) := 11{t ≥ 0} we get exactly the state-
ment of Theorem 2.
We have already seen in Sect. 2.3, Lemma 1, that
 ∗  1
(28) E e−λ t ZtΦ → Φ̂(λ∗ ).
κ
∗
Thus we need to divide ZtΦ by eλ t to get something non-trivial. Let us
quote a weaker form of Theorem 5.4 of [16].

Theorem 5 (Nerman, [16]). Consider a supercritical, Malthusian branching

process with Malthusian parameter λ∗ . Suppose that condition (M) holds
and Φ is a random characteristic with properties described before. Then
almost surely
∗ 1( ∗
(29) e−λ t ZtΦ → Φ(λ )W, as t → ∞,
κ
where W is a random variable that does not depend on Φ.
192 A. Rudas and B. Tóth

The necessary and sufficient condition for the random variable W to be

a.s. positive is the so-called x log x property of the reproduction process ξ:
 
(L) ( ∗ ) log+ ξ(λ
E ξ(λ ( ∗ ) < ∞.

We quote Theorem 5.3 of [10].

Theorem 6 (Jagers–Nerman, [10]). Consider a supercritical, Malthusian
branching process with Malthusian parameter λ∗ . If condition (L) holds
then W > 0 a.s. and E(W ) = 1; otherwise W = 0 a.s.

Remark. This theorem is the generalization of the Kesten–Stigum theorem,

which states this fact for Galton–Watson processes (see [11]).

4. Global Properties

The questions discussed in Sect. 2 investigated local properties of the ran-

dom tree: after a long time evolution, we asked about the neighborhood of
a typical (uniformly selected) vertex.
When we want to look “too far away” from the random vertex though,
the general theorem (Theorem 2) is of no use. What is the probability, for
example, that the random vertex is descendant of the first, second, or k th
child of the root? These are the types of questions that we address.
The results in this section are from [19], we give sketches of the basic
ideas and proofs.

4.1. Notation and Question

 
As we have seen in Sect. 3, the normalized size of the tree, exp (−λ∗ t) Υ(t) ,
converges almost surely to a random variable. Throughout this section, we
denote it by
∗  
(30) Θ := lim e−λ t  Υ(t) .
t→∞

For every x ∈ N , we introduce the variables Θx , corresponding to the

growth of the subtree under x, analogously to Θ,
 
Θx := lim e−λ (t−τx )  Υ↓x (t) .
∗

t→∞
Random Tree Growth with Branching Processes – a Survey 193

The letter Θ refers to the variable corresponding to the root, but when we
want to emphasize this, we sometimes write Θ∅ instead.
The most important relation between the different Θx variables in the
tree is the so-called basic decomposition (according to the first generation
in the tree):
 ∞
 
   
Θ∅ = lim e −λ∗ t  
Υ(t) = lim e −λ∗ t
1+  Υ↓k (t) 
t→∞ t→∞
k=1

∞
 
−λ∗ (t−τk ) 
  
∞
= −λ∗ τk
e lim e 
Υ↓k (t) =
∗
e−λ τk Θk ,
t→∞
k=1 k=1

or similarly for any x,

∞
 ∗ (τ −τ )
(31) Θx = e−λ xk x
Θxk .
k=1

Now we turn our attention to another random variable, which we denote

by Δx . In the introduction of this section, we asked about the probability
that for a fixed vertex x, after a long time, a randomly chosen vertex is
descendant of x. This probability tends to an almost sure limit as t → ∞,
   
 Υ↓x (t) ∗τ e−λ (t−τx )  Υ↓x (t)
∗ ∗
e−λ τx Θx
−λ
(32) Δx := lim   =e x
lim   = .
t→∞ Υ(t) t→∞ e−λ∗ t  Υ(t) Θ∅

Note that the value of Θ, obviously apparent in the continuous-time model,

seems to disappear when we investigate the discrete-time evolution of the
tree. The questions we naturally pose concern those properties which are
observable also in the discrete-time setting. The variable Δx is obviously
one of these observables since it can be written as the limit of a sequence of
random variables from the discrete time model:
 d 
 Υ ↓x (n)
Δx = lim  d  .
n→∞ Υ (n)

The question is, does Θ really “disappear” in the problems concerning the
discrete-time evolution? Recall that in Sect. 2, the main theorem discusses
the behavior of ratios of the form Ztϕ /Ztψ , and since both the numerator and
the denominator converges to constant times the common global variable,
194 A. Rudas and B. Tóth

Θ itself cancels out. In this section we are interested whether Θ is merely

an artificial side-product of the continuous-time embedding, or there exist
properties observable in the discrete-time model, where Θ itself appears.

4.2. Markov Property

Definition 1. We say that a system of random variables (Yx )x∈N consti-

tutes a tree-indexed Markov field if for any x ∈ N , the distribution of the
collection of variables (Yy : x ≺ y), and that of (Yz : x  z), are conditionally
independent, given Yx .

We state the following Lemma (recall that σx is the time we have to

wait before vertex x is born, as counted from the birth time of his youngest
brother, see Sect. 1.3).

Lemma 4. For each x ∈ N let Vx denote the vector Vx := (σx , Θx ). Then

the collection of variables Ax := (Vy : x ≺ y) and Bx := (Vz : x  z; σx ) are
conditionally independent, given Θx .

Proof. Recall (31), the decomposition of Θx according to the first genera-

tion of the subtree under x,
∞
 ∞

∗ (τ −τ ) ∗ (σ +σ +...+σ )
(33) Θx = e−λ xj x
Θxj = e−λ x1 x2 xj
Θxj .
j=1 j=1

Decompose the Θxj variables similarly, and so on, to arrive at the conclusion
that Θx , as well as the whole collection Ax , is in fact a function of the
variables (σy : x ≺ y).
By the same argument, the collection (Vz : x  z; σx ) is a function of
the collection of variables (σy : x ⊀ y; Θx ).
Now since for all x = y, σx and σy are independent, it is clear that
(σy : x ≺ y) and (σy : x ⊀ y) are independent sets of variables. Given Θx ,
the two collections Cx := (σy : x ≺ y; Θx ), and Dx := (σy : x ⊀ y; Θx ) are
conditionally independent. Since the variables in A are functions of C nad
similarly, B is function of D, the statement of the lemma follows.

Corollary 1. The variables (Θx )x∈N constitute a tree-indexed Markov

field.
Random Tree Growth with Branching Processes – a Survey 195

Proof. Direct consequence of Lemma 4, since Vx = (σx , Θx ).

It is clear that if x  x and x  x, then Θx and Θx are independent.

Using Lemma 4, any moment can be computed in the non-independent case.
We give formula for the covariance here for example:

Corollary 2. Let x = xy for some y ∈ N , then

∗τ
Cov (Θx , Θx ) = E(e−λ y
) Var (Θ).

Proof. With the notation |y| = n,

 

−λ∗ (τ xz −τx )
E(Θx Θxy ) = E Θxy e Θxz
z : |z|=n

∗ (τ −τ )
 ∗ (τ
= E(e−λ xy x
)E(Θ2xy ) + E(Θxy ) E(e−λ xz −τx )
)E(Θxz )
z : |z|=n, z=y

∗τ  2  ∗ 
= E(e−λ y
)E(Θ2 ) + E(Θ) (1 − E e−λ τy )
 2  ∗ 
= E(Θ) + E e−λ τy Var (Θ),

since by the results in Lemma 4, Θxz is independent of (τxz − τx ). Since

 2
E(Θx )E(Θx ) = E(Θ) , the formula for the covariance follows.

Let us introduce the following variables, indexed by N . For the root let
R∅ := 1 and for any other vertex y  which has a parent y, so for any y  = yk
with k ∈ N+ , let
 
 Υ↓yk (t) ∗
e−λ (τyk −τy ) Θyk Δyk
(34) Ryk := lim   = = .
t→∞ Υ↓y (t) Θy Δy

Notice that for x = (i1 i2 . . . in ), Δx is a telescopic product,

Δi1 i2 Δi1 i2 i3 Δi1 ...in

(35) Δx = Δi1 ... = Ri1 Ri1 i2 Ri1 i2 i3 . . . Ri1 ...in .
Δi1 Δi1 i2 Δi1 ...in−1

This decomposition is of interest due to the following Theorem.

196 A. Rudas and B. Tóth

Theorem 7. With the variables Rx defined as above, let Ux := (Rx , Θx ).

Given any sequence of positive integers (in )∞ n=1 , the sequence of variables
U∅ , Ui1 , Ui1 i2 , Ui1 i2 i3 , . . . constitutes a Markov chain, which is homogeneous:
the transition probabilities from Uy to Uyk depend on k, but not on y.

Proof. Let y, x, z be vertices in a progeny line, so let x be parent of z,

and y be parent of x. Given Θx , then, from Corollary 1, Θy and Θz are
conditionally independent. We show that so are Ry and Rz .
Consider that
∗
e−λ (τz −τx ) Θz
Rz = ,
Θx
so Rz is a function of Ax (recall the notation in Lemma 4). At the same
time,
∗ (τ −τ )
e−λ y y Θy
Ry = ,
Θy 
where y  is the parent of y, so Ry is a function of Θy and the collection
(σv : x ⊀ v), which implies that Ry is a function of Bx .
According to Lemma 4, Ax and Bx are conditionally independent, given
Θx , thus the proof is complete.

4.3. Fragmentation of Mass

The Δx variables can be thought of as relative “masses” which the limiting

tree “gives” to the individual vertices. For any fixed vertex x, Δx is the
limit of the probability that a uniformly selected vertex, is descendant of x.
If we thus look at all the x vertices in any fixed generation of the tree, the
sum of the respective Δx values is obviously 1.
How can we describe this fragmentation of mass 1, from the root to gen-
eration one, and then to generation two, etc? We investigate this question
in the following, in different cases of the choice of the weight function.

4.3.1. Linear Weight Function. Let us consider the case when the weight
function is linear. That is, for some α ≥ 0 and β > 0, for all k ∈ N+ , let
the weight of a vertex of degree k be

w(k) = αk + β.
Random Tree Growth with Branching Processes – a Survey 197

The corresponding Malthusian parameter (solution of (10)), is now λ∗ =

α + β. For the sake of computational simplicity it is convenient to re-scale
w so that λ∗ = 1, thus α = 1 − β,

(36) w(k) = (1 − β)k + β,

where 0 < β ≤ 1. Note that β = 1 is allowed, and means that the weight
function is constant, which corresponds to the Yule-tree model (see [22]).
m
Also, for any integer m > 1, we allow the choice of β = m−1 . In this case
w linearly decreases, and it hits level zero at m, meaning that each vertex
can have at most m children.
When a new vertex is added to the system, the sum of the weights in the
tree increases by two terms, with 1 − β because of the parent, and with β
because of the new vertex. Thus, each time a new vertex is added, the total
growth rate increases by 1, independently of the choice of the parent. This
intuitively explains why the size of the tree grows exponentially in time,
with parameter λ∗ = 1.
 
The previous observation means that Nt :=  Υ(t) is a Markov process,
which, at time t, increases by one with rate Nt −1+β. Thus it is straightfor-
ward to set up a partial differential equation for f (u, t) := E(e−uNt ), which
can be solved explicitly. By taking the limit limt→∞ f (ue−t , t), one arrives
at the conclusion that Θ has Gamma (1, β) distribution.
The fact that the growth rate is independent of the structure of the tree,
implies that anything that can be computed from the discrete time model,
is independent of Θ. This is in accordance with the distribution of Θ being
Gamma. To see this connection, consider for example that
∗ ∗
e−λ τ1 Θ1 e−λ τ1 Θ1 Θ1
(37) Δ1 = =
∞ −λ∗ τ =
∞ −λ ,
Θ∅ k=1 e
kΘ
k Θ1 + k=2 e ∗ (τk −τ1 ) Θk
X
which shows that Δ1 is a random variable of the form X+Y , where X and Y
are independent. For the ratio to be independent of the denominator (thus
Δ1 to be independent of Θ∅ ), X has to be of a Gamma distribution. This
result is in accordance with the above considerations.
This all implies that in the linear case, Δx is the product of independent
variables (see (35)), since now the Markov chain in Theorem 7 consists of
independent elements.
From this observation it follows that Θ, according to the first generation,
splits into the vector  −λ∗ τ 
e k
Θk k∈N+ ,
198 A. Rudas and B. Tóth

which is of a Poisson–Dirichlet distribution. For a precise formulation of

this fact, see [6].

4.3.2. Binary Tree. We now consider weight functions which ensure that
each vertex can have two children at maximum. First let us investigate
the distribution of Θ in this case, then we construct the system of the Δx
variables.
Let w(0) = a > 0, w(1) = 1, and w(k) = 0 for k ≥ 2. (We fix w(1) = 1
for the sake of computational simplicity, this is a different scaling from the
one used in the linear case). The Malthusian parameter λ∗ is now the
positive solution of the equation

(38) (λ∗ )2 + λ∗ − a = 0.

Consider the basic decomposition of Θ,

∗σ ∗ (σ +σ )
(39) Θ = e−λ 1
Θ1 + e−λ 1 2
Θ2 .

Let the moment generating function be ϕ(u) := E(e−uΘ ). Using (39),

computing ϕ(u) in two steps, by first taking conditional expectation with
σ1 and σ2 remaining fixed, then expectation with regards to Θ1 and Θ2 ,
yields the integral equation
 ∞ ∞
 ∗   ∗ ∗ 
(40) ϕ(u) = ϕ ue−λ x ϕ ue−λ x e−λ y ae−ax e−y dy dx.
0 0

Now, with two changes of variables, and twice differentiating the equation,
one arrives at the differential equation
    2
 ϕ(u)2 − ϕ(u) ϕ (u) ϕ (u)
(41) ϕ (u) = c − c + ,
u2 u ϕ(u)

where we introduced the shorthand notation

λ∗ + 1
(42) c := .
λ∗

The boundary values are ϕ(0) = 1 and ϕ (0) = −E(Θ) = − λ1∗ κ , where κ is
easily computed (recall (12)),
 
∂ a a 1 
(43) κ=− +  .
∂λ λ + a λ + a λ + 1 λ=λ∗
Random Tree Growth with Branching Processes – a Survey 199

Introducing g(u) = log ϕ(u), equation (41) is equivalent to

(44) g (u) = cu−2 (eg(u) − 1) − cu−1 g  (u).

We could have computed the moments of Θ already from (39), but (44)
offers a simple method. From the series expansion of eg(u) ,

eg(u) − 1 − g  (u)u c  2
g  (u) = c 2
→ ( − g  (0) + g (0) ), as u → 0.
u 2
  2  
This way g (0) = 2+c
c
g (0) , and so E(Θ2 ) = (1 + 2+c c
) E(Θ) 2 .
As for other moments of Θ, one can find a simple recursive formula for
the derivatives of g. The derivation of this recursion is not illuminating, so
we just state the result, namely,

(eg )(k)   g (k) g (k+1)

g (k+2) = c − ck + k(k − 1) − (c + 2k) , (∀ k ≥ 1),
u2 u2 u

from which all the values of g (k) (0) can be computed.

Now we turn our attention to the structure of the Δx system, meaning
that we want to construct the Δx variables step by step, from generation
to generation. For the rest of this section we treat the distribution of Θ as
known.
Recall that σ1 , σ2 , Θ1 and Θ2 are independent. Thus, from
∗σ  ∗σ 
Θ = e−λ 1
Θ1 + e−λ 2
Θ2 ,

the conditional distribution of σ1 , given Θ, is straightforward to calculate.

Then, given Θ and σ1 , the conditional distribution of Θ1 follows. After
this, given now Θ, σ1 and Θ1 , the conditional distribution of σ2 can be
determined. Finally, if we know Θ, σ1 , Θ1 and σ2 , then Θ2 is a deterministic
function of these.
Based on these considerations and on Theorem 7, now we can construct
the system of the Δx variables in the following steps.
1. Pick Θ∅ at random, according to its distribution.
2. First generation
a. Pick (σ1 , Θ1 ) according to their conditional distribution, given
∗
Θ∅ . These three numbers define Δ1 = R1 = exp (−λΘ σ1 )Θ1 .
200 A. Rudas and B. Tóth

b. Pick (σ2 , Θ2 ) similarly, according to their conditional distrib-

ution, given Θ∅ and (σ1 , Θ1 ). At this point we can compute
exp (−λ∗ (σ1 +σ2 ))Θ2
Δ 2 = R2 = Θ∅ .

3. Second generation

a. Repeat the steps seen before for the progeny of vertex 1, to get
R11 and R12 . Using the Markov property described in Theorem
7, this is done only using the information carried by Θ1 , con-
ditionally independently of Θ and also of Θ2 . Since we already
know R1 , we can now compute the values Δ11 = R1 R11 and
Δ12 = R1 R12 .
b. Independently of the previous step, split Θ2 to get R21 and R22 ,
thus also Δ21 and Δ22 .

4. Subsequent generations are constructed similarly.

Note that the special choice of a = 2 corresponds to the negative linear,

binary case of Section 4.3.1 (take m = 2 there). Then λ∗ = 1, κ = 12 , and
E(Θ) = 2. It can be checked that in this case, equation (41) is solved by
ϕ(u) = (1 + u)−2 , so indeed, if a = 2, then Θ is of Gamma (1, 2) distribution.
Also, by checking that no other function of the form (1 + u)−α satisfies (41),
it is verified that no other choice of a allows Θ to be of a Gamma distribution,
so in all a = 2 cases the structure of the tree is indeed dependent on Θ.
These results are extendable, in a very similar way, to the case where
the weight function allows the vertices to have at most m children for some
finite m > 2. Then, an mth order differential equation can be derived for the
moment generating function, and again, any moment of Θ can be computed.
The structure of variables Δx is also constructed analogously.
In [8], certain scaling exponents are derived explicitly, characterizing the
decay of the distribution of Δx .

Acknowledgements. We thank the anonymous referee for the thorough

work, and especially for his/her idea which shortened the proof of Lemma 2
considerably. We also thank Balázs Ráth and Imre Péter Tóth for stimulat-
ing discussions. The research work of the authors is partially supported by
the following OTKA (Hungarian National Research Fund) grants: K 60708
(for A. R. and B. T.), TS 49835 (for A. R.).
Random Tree Growth with Branching Processes – a Survey 201

References

[1] D. Aldous, Asymptotic fringe distributions for general families of random trees,
Ann. Appl. Probab., 1(2) (1991), 228–266.
[2] A.-L. Barabási and R. Albert, Emergence of scaling in random networks, Science,
286(5439) (1999), 509–512.
[3] B. Bollobás, O. Riordan, J. Spencer and G. Tusnády, The degree sequence of a
scale-free random graph process, Random Structures Algorithms, 18(3) (2001),
279–290.
[4] B. Bollobás and O. M. Riordan, Mathematical results on scale-free random graphs,
in: Handbook of graphs and networks, Wiley-VCH, Weinheim (2003), pp. 1–34.
[5] F. Chung, S. Handjani and D. Jungreis, Generalizations of Polya’s urn problem,
Ann. Comb., 7(2) (2003), 141–153.
[6] R. Dong, C. Goldschmidt and J. B. Martin, Coagulation-fragmentation duality,
Poisson–Dirichlet distributions and random recursive trees, Ann. Appl. Probab.,
16(4) (2006), 1733–1750.
[7] R. Durrett, Random Graph Dynamics, Cambridge University Press, Cambridge
(2007), Cambridge Series in Statistical and Probabilistic Mathematics.
[8] E. G. François David, Philippe Di Francesco and T. Jonsson, Mass distribution
exponents for growing trees, Journal of Statistical Mechanics: Theory and Experi-
ment, 2007(02) (2007), P02011.
[9] P. Jagers, Branching processes with biological applications, Wiley-Interscience [John
Wiley & Sons], London (1975), Wiley Series in Probability and Mathematical
Statistics – Applied Probability and Statistics.
[10] P. Jagers and O. Nerman, The growth and composition of branching populations,
Adv. in Appl. Probab., 16(2) (1984), 221–259.
[11] H. Kesten and B. P. Stigum, A limit theorem for multidimensional Galton–Watson
processes, Ann. Math. Statist., 37 (1966), 1211–1223.
[12] P. L. Krapivsky and S. Redner, Organization of growing random networks, Phys.
Rev. E, 63(6) (May 2001), 066123.
[13] P. L. Krapivsky, S. Redner and F. Leyvraz, Connectivity of growing random
networks, Phys. Rev. Lett., 85(21) (Nov. 2000), 4629–4632.
[14] J.-F. Le Gall, Processus de branchement, arbres et superprocessus, in: Development
of mathematics 1950–2000, Birkhäuser, Basel (2000), pp. 763–793.
[15] T. F. Móri, On random trees, Studia Sci. Math. Hungar., 39(1–2) (2002), 143–155.
[16] O. Nerman, On the convergence of supercritical general (C-M-J) branching
processes, Z. Wahrsch. Verw. Gebiete, 57(3) (1981), 365–395.
[17] R. Oliveira and J. Spencer, Connectivity transitions in networks with super-linear
preferential attachment, Internet Math., 2(2) (2005), 121–163.
[18] P. Olofsson, The x log x condition for general branching processes, J. Appl. Probab.,
35(3) (1998), 537–544.
202 A. Rudas and B. Tóth

[19] A. Rudas, Global properties of a randomly growing tree, Work in progress.

[20] A. Rudas, B. Tóth and B. Valkó, Random trees and general branching processes,
Random Struct. Algorithms, 31(2) (2007), 186–202.
[21] R. T. Smythe and H. M. Mahmoud, A survey of recursive trees, Teor. Probab. Mat.
Statist., (51) (1994), 1–29.
[22] G. Udny Yule, A mathematical theory of evolution, based on the conclusions
of Dr. J. C. Willis, F.R.S, Royal Society of London Philosophical Transactions
Series B, 213 (1925), 21–87.

Anna Rudas Bálint Tóth

Institute of Mathematics Institute of Mathematics
Budapest University of Technology Budapest University of Technology
e-mail: [email protected] e-mail: [email protected]
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 203–237.

Chapter 5
Reaction-diffusion Processes in Scale-free
Networks

MICHELE CATANZARO, MARIÁN BOGUÑÁ and

ROMUALDO PASTOR-SATORRAS

In this chapter we provide a review of the main results recently obtained in the
modeling of binary fermionic reaction-diffusion processes on scale-free networks.
We show how to derive rate equations within the heterogeneous mean-field for-
malism, and how information can be obtained from them both for finite networks
in the diffusion-limited regime and in the infinite network size lime. By means of
extensive numerical simulations, we check the mean field predictions and explore
other aspects of the reaction-diffusion dynamics, such as density correlations and
the effects of the minimum degree or a tree-like topology.

1. Introduction

In recent years, the theory of complex networks has proved to be an ex-

tremely useful tool for the study and characterization of the structure and
function of many complex systems [1, 2, 3]. In fact, many natural and man-
made systems have heterogeneous patterns of connexions and interactions
that can be properly described as a network or graph [4]. The statisti-
cal analysis of their topological organization has shown that most of them
seem to share some typical features, the most relevant being the small-world
property [5] and a large heterogeneity in the number of contacts per vertex
(or degree), which lacks any typical degree-scale [6].
204 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

The small-world property refers to the fact that the average distance 
between any two vertices – measured as the smallest number of connections
(or edges) on a path between one and the other – is very small, scaling
logarithmically or even slower with the network size N . This is to be
compared to the power-law scaling  ∼ N 1/d in a d-dimensional lattice.
Since the logarithm grows slower than any power-law function, even if d is
very large, small-world networks can be thought of as highly compact objects
of infinite dimensionality. As a consequence, the metric properties of this
class of systems are nearly absent, every vertex being roughly at the same
distance to any other vertex of the network, and fluctuations associated to
“spatial” effects can be neglected. This accounts for the accuracy of mean-
field theories applied to the understanding of dynamical processes taking
place in small-world networks [7].
On the other hand, real complex networks are not just small-worlds but
also scale-free (SF) networks. They are typically characterized by a degree
distribution P (k) – the probability that a randomly selected vertex has
degree k – that decreases as a power-law,

(1) P (k) ∼ k −γ ,

where γ is a characteristic degree exponent, usually in the range 2 < γ ≤ 3.

Within this domain the topological fluctuations of the degree, characterized
by its second moment k2 , diverge in the limit of large network sizes.
It is the conjunction of the small-world and SF properties that makes
complex networks radically different from more regular structures embedded
in metric spaces.
Initially, the interest in networked systems was concentrated on the de-
scription and modeling of their structural properties [1, 2]. More recently,
research efforts have placed the focus on the effects that these peculiar topo-
logical properties may have on the behavior of dynamical systems running
on top of networks [7]. Besides their obvious theoretical interest, these
efforts can have practical implications in real world systems, e.g., under-
standing traffic behavior in technological systems such as the Internet [8]
or epidemic spreading of sexually transmitted diseases in human popula-
tions [9]. Interest in these dynamics was triggered by the observation that
the heterogeneous connectivity pattern observed in SF networks with di-
verging degree fluctuations can lead to very surprising outcomes, such as
an extreme weakness in the face of targeted attacks aimed at destroying
the most connected vertices [10, 11], or the ease of propagation of infective
Reaction-diffusion Processes in Scale-free Networks 205

agents [12, 13]. These properties are due to the critical interplay between
topology and dynamics in heterogeneous networks and are absent in their
homogeneous counterparts. After those initial discoveries, a real avalanche
of new results have been put forward; for a summary of them we refer the
reader to Ref. [7].
The importance of dynamical processes on complex networks justifies the
pursuit of common frameworks to describe and analyze them in a general
way. One of them is the theory of reaction-diffusion (RD) processes, which
can be used to represent a wide spectrum of different dynamics. In very
general terms, RD processes are dynamic systems that involve particles of
different “species” (Ai , i = 1, . . . n) that diffuse stochastically and interact
among themselves according to a fixed set of reaction rules. In this kind of
processes, the interest is usually focused on the time evolution and steady
states of the densities of the different species ρAi (t), and on the possible
presence of phase transitions between those states [15].
Epidemic models, such as the Susceptible-Infected-Susceptible (SIS)
[16], represent classical examples of RD dynamics. The SIS model, for
instance, corresponds to a RD process with two species of particles (indi-
viduals), infected (I) and susceptible (S), that interact through the reactions
[16, 17]

δ
S + I −→ 2I,
(2) μ
I −→ S.

The first reaction in Eq. (2) corresponds to the infection of a susceptible

individual by contact with an infected one, with a probability per unit of
time (rate) δ. The second reaction stands for the spontaneous healing of
infected individuals at rate μ. The behavior of this epidemic model is ruled
by the ratio λ = δ/μ, the so-called spreading rate. The main prediction
of the model is the existence of an epidemic threshold λc , above which
the dynamics reaches an endemic state, with a nonzero density of infected
individuals ρI [17]. Below the threshold, on the other hand, any epidemics
dies out in the long term, and the systems is disease-free.
Much is known about the behavior of RD processes on regular homo-
geneous lattices. In particular, theoretical formalisms have been proposed
that allow for general descriptions of the process in terms of field theories
[18, 19, 20, 21] which are then susceptible of analysis by means of the renor-
malization group technique [22]. For example, for the simplest RD process,
206 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

the diffusion-annihilation process [23]

λ
(3) A + A −→ ∅

in regular lattices of Euclidean dimension d, it is well known that the local

density of A particles, ρ(x, t), is ruled by a Langevin equation [24],

∂ρ(x, t)
(4) = D∇2 ρ(x, t) − 2λρ(x, t)2 + η(x, t),
∂t

where η(x, t) is a Gaussian white noise, with correlations

" #
(5) η(x, t)η(x , t ) = −2λρ(x, t)2 δ d (x − x )δ(t − t ).

Dynamical renormalization
" group
# arguments show that the average density
of A particles, ρ(t) = ρ(x, t) , behaves in the long term as

1 1
(6) − ∼ tα ,
ρ(t) ρ0

where ρ0 is the initial particle density, and the exponent α takes the values
α = d/dc for d ≤ dc and α = 1 for d > dc , where dc = 2 is the critical
dimension of the process. For d > dc one thus recovers the homogeneous
mean-field solution

(7) ρ(t) ∼ t−1 ,

obtained from Eq. (4) by setting the diffusion coefficient D and the noise
term η(x, t) equal to zero.
When the substrate is not a regular lattice but a complex network with
the properties described above, the powerful machinery of the renormaliza-
tion group becomes incapable of solving the problem. The reason is rooted
in the small-world property and the lack of a metric structure that renders
the very concept of renormalization meaningless. We can then say that, in
comparison, the study of the interplay of RD processes with the heteroge-
neous topology of a complex network is still in its infancy.
RD processes in complex networks are usually defined within a “fermi-
onic” point of view, in which each vertex can hold at most one particle.
Diffusion and dynamics are thus coupled, in the sense that particles per-
form random jumps between adjacent vertices and react upon landing on
Reaction-diffusion Processes in Scale-free Networks 207

an already occupied vertex. This formalism has been applied, both theoret-
ically and numerically, to the case of binary diffusion-annihilation processes
with a single species A + A → ∅ [25, 26] and with two species A + B → ∅
[25, 27]. From the study of those systems, it has been already possible to
shed some light on RD dynamics, which turns out to have an unexpected
and interesting behavior when the substrate network is SF with an expo-
nent 2 < γ ≤ 3. The long term dynamics, for instance, is always faster
than in lattices, with the exponent α > 1. Besides, the dynamics evolves
in a hierarchical fashion, with the concentration of particles first decaying
at low degree vertices, then at higher degree vertices and so on until the
concentration of particles in the vertices with the maximum degree starts
decreasing. Another striking difference as compared to lattices is the ab-
sence of depletion zones in the A+A → ∅ process and the lack of segregation
between species in the A + B → ∅ process which, again, is a consequence of
the small-world property.
In this chapter we provide a review of the main results obtained in the
modeling of binary fermionic RD processes in SF networks. The rest of
the review is organized as follows: in section 2 we summarize the stan-
dard analytical technique used to deal with dynamical processes on com-
plex networks, namely heterogeneous mean-field (MF) theory. Applied to
the case of RD processes, we focus on a general process that encompasses
both A + A → ∅ and A + B → ∅ as particular cases. We show the gen-
eral procedure to obtain a MF rate equation for the particle density, and
how to solve it both in finite networks (within the so-called diffusion-limited
regime) and in the infinite network size limit, by means of a continuous de-
gree approximation. Section 3 is devoted to show the results than can be
obtained by means of numerical Monte Carlo simulations. After checking
the analytical predictions of MF theory, we consider other aspects numeri-
cally observed in this kind of systems: the lack of depletion and segregation
effects, which can be related to the density correlations of particles; the role
of the degree in the annihilation dynamics; the effects of the minimum de-
gree of the network substrate; and the effects of global constraints, such as
the absence of loops (tree topologies). Finally, in section 4 we present an
outlook of the future directions of research in the field of RD processes in
complex networks.
208 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

2. Analytical Description of Diffusion-annihilation

Processes in Complex Networks

Let us consider the most general two species diffusion-annihilation process,

given by particles of two species A and B, and that is defined by the reaction
rules

A + A → ∅ with probability paa

(8) A + B → ∅ with probability pab
B + B → ∅ with probability pbb .

The processes take place on an arbitrary complex network of size N , each

vertex of which can host at most one particle. The dynamics of the process
is defined as follows: At rate λ, an A particle chooses one of its nearest
neighbors. If it is empty, the particle fills it, leaving the first vertex empty.
If this nearest neighbor is occupied by another A particle, the two particles
annihilate with probability paa , leaving both vertices empty. If it is occupied
by a B particle, both particles annihilate with probability pab . If these
annihilation events do not take place, all particles remain in their original
positions. The same dynamics applies to B particles if we replace paa by
pbb . We also assume that the diffusion rate λ is the same for both types of
particles. It is easy to see that, with this formulation, paa = pbb = pab = 1
corresponds to the A + A → ∅ process since, in this case, there is no way
to distinguish between A and B particles. On the other hand, setting
paa = pbb = 0 and pab = 1, one recovers the A + B → ∅ process.

2.1. Heterogeneous mean-field formalism

The inherent randomness in the topology of complex networks forces us to

describe them using a statistical framework. In this chapter, we consider
the simplest statistical description of a network in terms of the properties of
single vertices (degrees) and correlations between pairs of such vertices, the
so-called degree-degree correlations. Single vertex statistical properties are
encoded in the degree distribution P (k) and degree-degree correlations are
described by the conditional probability P (k  |k) that a vertex of degree
k is connected to a vertex of degree k  [28]. We should note that all
the results presented here are correct for networks which are maximally
Reaction-diffusion Processes in Scale-free Networks 209

random – in the sense of maximizing the entropy of the network – under

the constraints of having given functions P (k) and P (k |k). Once this
assumption is made, we can get some analytical insights into the behavior of
diffusion-annihilation processes by applying the heterogeneous MF theory –
a mean-field description of the dynamics that discriminates between degree
classes [26, 27].
To study analytically this process within this approximation, we are
forced to consider the partial densities ρak (t) and ρbk (t), representing the
densities of A and B particles in vertices of degree k, or, in other words,
the probabilities that a vertex of degree k contains an A or B particle at
time t [12, 29]. From these partial densities, the total densities of A and B
particles are given as

 
(9) ρa (t) = P (k)ρak (t) and ρb (t) = P (k)ρbk (t).
k k

While it is possible to obtain rate equations for the densities ρak (t) and ρbk (t)
by means of intuitive arguments [12, 30], in what follows we will pursue a
more microscopical approach [26], which can be generalized to tackle other
kind of problems.
Let us consider the process defined by Eq. (8) on a network of N ver-
tices which is fully defined by its adjacency matrix aij , taking the values
aij = 1 if vertices i and j are connected by an edge and 0 otherwise. Let
nai (t) be a dichotomous random variable taking value 1 whenever vertex i
is occupied by an A particle and 0 otherwise. Analogously, let nbi (t) be
a dichotomous random variable taking value 1 whenever vertex i is oc-
cupied by a B particle and 0 otherwise. Notice that with the previous
definition, the random variable 1 − nai (t) − nbi (t) takes the value 1 when
vertex i is empty and is zero in every other situation, so that it defines
the complementary event of being occupied by an A or B particle. This
property assures the correct fermionic description of the dynamics. The
state of the
 system at time t iscompletely defined
 b by the state vectors 
na (t) = na1 (t), na2 (t), · · · , naN (t)
 a and nb (t) =
 n 1 (t), nb (t), · · · , nb (t) ,
2 N
denoted for simplicity n(t) ≡ n (t), nb (t) . Assuming that diffusion and
annihilation events of the particles follow independent Poisson processes
[14], the evolution of n(t) after a time increment dt can be expressed as
+ 
(10) nai (t + dt) = nai (t)η(dt) + 1 − nai (t) − nbi (t) ξ(dt),
210 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

with an analogous equation for the occupancy of B particles, nbi . Notice

that the two terms in the r. h. s. of Eq. (10) cannot be different from zero
simultaneously since they represent events that exclude each other – either
vertex i holds an A particle (the first term) or it does not (the second one).
In this way, the evolution at time t + dt is tied to the state of the system at
the previous time t.
The variables η(dt) and ξ(dt) in Eq. (10) are dichotomous random vari-
ables taking values

(11)
⎧ & + 
⎪  aij (1 − paa )naj (t) + (1 − pab )nbj (t)
⎪
⎪ 0 with prob. λ dt 1 −
⎪
⎪
⎪
⎪ ki
⎪
⎨
j
+ '
η(dt) =  aij paa naj (t) + pab nbj (t) ,
⎪
⎪ +
⎪
⎪ kj
⎪
⎪
⎪
⎪
j
⎩
1 otherwise

and
⎧  aij naj (t)
⎪
⎪
⎨1 with prob. λ dt
kj
(12) ξ(dt) = j ,
⎪
⎪
⎩
0 otherwise

where λ is the jumping rate that, without loss of generality, we set to 1.

The first term in Eq. (10) stands for an event in which vertex i is occupied
by an A particle and, during the time interval (t, t + dt), it becomes empty
either because the particle in it decides to move to another empty vertex or
because it annihilates with one of its nearby A or B particles. The second
term corresponds to the case in which vertex i is empty and an A particle
in a neighboring vertex of i decides to move to that vertex1 . Taking the
average of Eq. (10), we obtain
" a #
(13) ni (t + dt) | n(t) = nai (t)
& + 
 aij (1 − paa )nai (t)naj (t) + (1 − pab )nai (t)nbj (t)
− −
ki
j
1
Notice that the random variables η(dt) and ξ(dt) are not independent, since both
involve some common random movements. This fact, however, does not affect the mean-
field analysis, which do not involve cross correlations between them.
Reaction-diffusion Processes in Scale-free Networks 211
+ '
 aij paa nai (t)naj (t) + pab nai (t)nbj (t)
+ nai (t) + dt
kj
j

+   aij naj (t)

+ 1 − nai (t) − nbi (t) dt,
kj
j

equation that describes the average evolution of the system, conditioned to

the knowledge of its state at the previous time step. Then, after multiplying
Eq. (13) by the probability to find the system at state n at time t, and
summing for all possible configurations, we are led to

dρai (t)
(14) = −ρai (t)
dt
+ 
 aij (1 − paa )ρaa ij (t) + (1 − pab )ρij (t)
ab
+
ki
j
+  + 
 aij paa ρaa ab
ij (t) + pab ρij (t)
 aij ρaj (t) − ρaa
ij (t) − ρij (t)
ba
− + ,
kj kj
j j

where we have introduced the notation

(15) ρai (t) ≡ nai (t) , ρbi (t) ≡ nbi (t),

ij (t) ≡ ni (t)nj (t)

ρaa ij (t) ≡ ni (t)nj (t).
a a
(16) and ρab a b

ij (t) = ρij (t).

Notice that ρab ba

The derivation presented so far is exact, but difficult to deal with. How-
ever, it is possible to obtain useful information if we restrict our analysis
to the class of random networks with given degree distribution and degree-
degree correlations but maximally random at all other respects. For this
class of networks, vertices of the same degree can be considered as statisti-
cally equivalent. In mathematical terms, this means that [30]

(17) ρai (t) ≡ ρak (t) ∀i ∈ V(k),

(18) ρbi (t) ≡ ρbk (t) ∀i ∈ V(k),

(19) ij (t) ≡ ρkk  (t)

ρaa aa
∀i ∈ V(k), j ∈ V(k )
212 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

(20) ij (t) ≡ ρkk  (t)

ρab ab
∀i ∈ V(k), j ∈ V(k ),

where V(k) is the set of vertices of degree k. Besides, the small world
property present in this class of networks makes them objects of infinite di-
mensionality, well described by a MF theory. In such situation, correlations
between elements can be neglected and, consequently, we can approximate
two point correlations functions as ρkk (t) ρk (t)ρk (t). Using these ideas in
Eq. (14) we can write the following closed equation for the degree-dependent
densities

dρak (t)  + 
(21) = − ρak (t) + ρak (t) P (k |k) (1 − paa )ρak (t) + (1 − pab )ρbk (t)
dt  k

 kP (k  |k) + 
− ρak (t) paa ρak (t) + pab ρbk (t)
k
k

   kP (k  |k) a
+ 1 − ρak (t) − ρbk (t) ρk (t),

k
k

where we have made use of the identity [30]

1  
(22) aij = kP (k |k).
N P (k)
i∈V(k) j∈V(k )

The equation for the density of B particles can be obtained from Eq. (21)
by swapping indices a and b.
Eq. (21) can be further simplified if we assume that A and B particles
react among them with the same probability, that is, paa = pbb . In this case
(and assuming also the same concentration of A and B particles at t = 0),
the total density of particles in vertices of degree k, ρk (t) = ρak (t) + ρbk (t),
can be written as

dρk (t) 
(23) = −ρk (t) + (1 − μ)ρk (t) P (k  |k)ρk (t)
dt  k

+   kP (k  |k)
+ 1 − (1 + μ)ρk (t) ρk (t),

k
k
Reaction-diffusion Processes in Scale-free Networks 213

while the total density ρ(t) = k P (k)ρk (t) fulfills the differential equation

dρ(t) 
(24) = −2μ P (k)ρk (t)Θk (t),
dt
k

where we have used the degree detailed balance condition [28]

(25) kP (k)P (k  |k) = k  P (k )P (k|k  ).

In Eq. (24) we have defined


(26) Θk (t) = P (k  |k)ρk (t)
k

as the probability that a randomly chosen edge in a vertex of degree k points

to a vertex occupied by an A or B particle, while the parameter μ ∈ [0, 1] is
defined as μ = (paa + pab )/2. In this way, by setting μ = 1/2 we recover the
equation describing the A + B → ∅ process [27], whereas μ = 1 describes
the A + A → ∅ one [26]. It is interesting that, in this case, the model
is described by a single parameter μ, even if, originally, there were two of
them, paa and pab . This implies that there is a whole set of different models
which are governed by the same dynamical equation.

2.2. Finite networks: Diffusion-limited regime

In the case of networks with a general pattern of degree correlations, the

solution of Eq. (23) depends on the nature of the conditional probability
P (k |k) and can be a rather demanding task [31]. General statements on
the total density of particles can be made, however, in the limit of very long
time and very small particle density, when the concentration of particles is so
low that the RD process is driven essentially by diffusion. In this diffusion-
limited regime, it is possible to estimate the behavior of ρ(t), which turns
out to be independent of the correlation pattern of the network.
If we consider the limit k 0 in Eq. (23) (i.e. at large times), the linear
terms dominate, and we can consider the simplified linear equation

dρk (t)  kP (k  |k)

(27) −ρk (t) + ρk (t),
dt k
k
214 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

so that the density behaves as in a pure diffusion problem. The time scale
for the diffusion of the particles is much smaller than the time scale for two
consecutive reaction events. Therefore, the partial density can relax to the
stationary state of Eq. (27) and is well approximated by a pure diffusion of
particles [32, 33]

k
(28) ρk (t) ρ(t),
k

proportional to the degree k and the total instantaneous concentration of

particles, and independent of degree correlations. Inserting this quasi-static
approximation into the general Eq. (24), we obtain for long times and finite
size networks the equation

dρ(t) k 2 
(29) −2μρ(t)2 ,
dt k2

where k2  = k k2 P (k) is the second moment of the degree distribution.

The solution of this equation is

1 1 k 2 
(30) − 2μ t,
ρ(t) ρ0 k2

that is, linear in t with a prefactor depending on the fluctuations of the

degree. For homogeneous networks with a bounded degree distribution,
k2  is finite, and so is the density prefactor. In SF networks, the prefactor
depends on the cutoff – or maximum degree in the network – kc (N ) [34],
with

kc (N )3−γ for γ < 3
(31) k  ∼
2
,
ln kc (N ) for γ = 3

which is an increasing function of the network size N . The specific functional

form of kc on N depends, in general, on the particular model under consider-
ation. For uncorrelated SF networks, such as those created with the uncorre-
lated configuration model (UCM) [35], we have kc (N ) ∼ N 1/2 , the so-called
structural cutoff [36]. For correlated networks created with the configura-
tion model (CM) [37], we have instead a natural2 cutoff kc (N ) ∼ N 1/(γ−1)
2
The name “structural cutoff” refers to the fact that it is the maximum degree allowing
us to built uncorrelated networks. Beyond this limit, structural degree correlations appear
Reaction-diffusion Processes in Scale-free Networks 215

[34]. Therefore, the behavior of the particle density in the diffusion-limited

regime can be summarized as
⎧
⎪
⎪ N (3−γ)/2 t UCM, γ < 3
⎪
⎪
⎪
⎨N (3−γ)/(γ−1) t CM, γ < 3
1
(32) ∼ .
ρ(t) ⎪ ⎪ln N t γ=3
⎪
⎪
⎪
⎩
t γ>3

This result implies that, for the CM model, the diffusion-limited regime
is reached at lower particle concentrations as compared to the UCM. This
result has been observed in numerical simulations in Ref. [38].

2.3. Infinite networks: Continuous degree approximation

As we mentioned before, the solution of the general Eqs. (23) and (24) can
be very difficult to obtain in networks with general correlations P (k |k).
A completely analytical solution in the limit of infinite size networks can,
however, be obtained in the case of uncorrelated networks, in which the
conditional probability takes the simple form P (k |k) = k  P (k  )/k [30].
For this class of networks, the rate equation is

dρk (t) +  k
(33) = −ρk (t) + (1 − μ)Θ(t) + 1 − (1 + μ)ρk (t) ρ(t),
dt k

and
dρ(t)
(34) = −2μρ(t)Θ(t),
dt
where we have defined
1 
(35) Θ(t) = kP (k)ρk (t).
k
k

In order to solve Eq. (33), we perform a quasi-static approximation.

From the homogeneous MF solution of the A + A → ∅ and A + B → ∅
as a consequence of the closure of the network. Instead, “natural cutoff” refers to the
expected maximum degree when we generate N independent random trials from the degree
distribution P (k) without actually closing the network.
216 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

processes, we expect ρ(t) to be a decreasing function with a power-law-like

behavior. In this case, for large enough times, the time derivative of ρ(t) will
be much smaller that the density proper, that is, ∂t ρ(t) ! ρ(t). Extending
this argument to the partial densities ρk (t), at long times we can neglect
the left-hand-side term in Eq. (33), and solve for ρk (t) as a function of the
density, obtaining

kρ(t)
k
(36) ρk (t) = .
kρ(t)
1 + (1 + μ) − (1 − μ)Θ(t)
k

Substituting this approximation into the expression for Θ(t), we get

1  P (k)k2 ρ(t)
(37) Θ(t) = 2+  .
k 1 − (1 − μ)Θ(t) k 1 + 1+μ kρ(t)
1 − (1 − μ)Θ(t) k

Solving this self-consistent equation, we can find Θ(t) as a function of

ρ(t), and then proceed to solve Eq. (34). For heterogeneous networks with
a diverging second moment, as in the case of SF networks, we have to
consider carefully the solution of Eq. (37). If we consider a continuous
degree approximation, uncorrelated SF networks in the infinite size limit
are completely determined by the normalized degree distribution

(38) P (k) = (γ − 1)mγ−1 k−γ ,

where m is the minimum degree in the network, and we are approximating k

as a continuous variable. The average degree is thus k = m(γ − 1)/(γ − 2).
Within this approximation, Eq. (37) can be written as
 ∞
(γ − 1)mγ−1 ρ(t) k2−γ
(39) Θ(t) = +  dk
k2 1 − (1 − μ)Θ(t) m
1+
1+μ kρ(t)
1 − (1 − μ)Θ(t) k
& + '
1 k 1 − (1 − μ)Θ(t)
(40) = F 1, γ − 2, γ − 1, −
1+μ m(1 + μ)ρ(t)

where F [a, b, c, z] is the Gauss hypergeometric function [39]. Assuming

both Θ(t) and ρ(t) small (i.e. for sufficiently long times), we can use the
Reaction-diffusion Processes in Scale-free Networks 217

asymptotic expansion of the Gauss hypergeometric function

⎧
⎪
⎪z γ−2 γ<3
⎨
(41) F [1, γ − 2, γ − 1, −1/z] ∼ −z ln z γ=3, for z → 0,
⎪
⎪
⎩
z γ>3

to obtain and explicit expression of Θ(t) as a function of ρ(t). Inserting

it into the rate equation for ρ(t) and integrating, we obtain the explicit
solutions in the long time limit for infinite size uncorrelated SF networks
⎧
⎪
⎪ t1/(γ−2) γ<3
1 ⎨
(42) ∼ t ln t γ=3.
ρ(t) ⎪
⎪
⎩
t γ>3

That is, apart from irrelevant prefactors, the leading solution is independent
of μ, and therefore the same for both A + A → ∅ and A + B → ∅ processes
[26, 27].
The analytical exponents derived above are exact for infinite size net-
works. However, they may be difficult to observe in real computer simula-
tions performed on networks of finite size. We can see this fact from the
quasi-static approximation Eq. (36), in which, for the sake of simplicity, we
will focus in the case μ = 1 (A + A → ∅ process). Indeed, for a power-law
degree distribution, the largest weight in the sum in Eq. (37) is carried by
the large k values. If the network is composed of a finite number of ver-
tices, N , as it always happens in numerical simulations, it has a cutoff or
maximum degree kc (N ), Thus, there exists a cross-over time tc , defined by

2kc (N )ρ(tc )
(43) ∼ 1,
k

such that, for t > tc the particle density is so small that we can approximate

k
(44) ρk (t) ρ(t),
k

which corresponds to the diffusion-limited regime discussed in Section 2.2.

Therefore, for t > tc , we should expect to observe a linear behavior on
1/ρ(t), instead of the power-law predicted in infinite networks for the con-
tinuous degree approximation, while the region for the asymptotic infinite
218 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

size behavior should be observed for t < tc . From Eqs. (43) and (42) we can
predict

(45) tc (N ) ∼ kc (N )γ−2 ∼ N (γ−2)/2

for uncorrelated networks. This is an increasing function of N for γ < 3

and so one should expect that the region in which the infinite size behavior
is observed must be increasing with N . However, the width of this region
must be properly compared with the total surviving time of the process in
a finite network. Assuming that at long times the process is dominated by
the diffusion-limited regime, Eq. (32), we can estimate the total duration of
the process as the time td at which only two particles remain, that is,

2
(46) ρ(td ) ∼ .
N
From this definition, we can estimate

(47) td (N ) ∼ N (γ−1)/2 .

The ratio of the crossover time to the total duration of the process is thus

tc (N )
(48) ∼ N −1/2 ,
td (N )

that is, a decreasing function of N . Therefore, the RD dynamics is domi-

nated by its diffusion-limited regime. Consequently, even for very large sys-
tems, the asymptotic infinite size behavior will be very difficult to observe
whereas the diffusion-limited regime will span almost all the observation
time.

3. Numerical Simulations

The general process described by the reactions (8) can be easily implemented
in numerical simulations using a sequential updating algorithm [15]. An
initial fraction ρ0 < 1 of vertices in the networks are chosen and randomly
occupied by a ρ0 N/2 particles of species A and ρ0 N/2 particles of species B.
At time t in the simulation, a vertex is randomly chosen among the n(t) =
nA (t) + nB (t) vertices that host an A or B particle at that time. One of
Reaction-diffusion Processes in Scale-free Networks 219

its neighbors is selected also at random. If it is empty, the particle moves

and occupies it. If it contains a particle, both particles react according to
the rules in Eq. (8) and the particle numbers nA (t) and nB (t) are updated
according to the result of the reaction step. In any case, time is updated as
t → t + 1/n(t).
RD simulations are run on SF networks generated using the uncorrelated
configuration model (UCM) [35], which is defined as follows. 1) We first
assign to each vertex i – in a set of N initially disconnected vertices –
a degree ki extracted from the probability distribution

P (k) ∼ k −γ , and
subject to the constraints m ≤ ki ≤ N 1/2
and i ki even. 2) We
then
construct the network by randomly connecting the vertices with i ki /2
edges, respecting the preassigned degrees and avoiding multiple and self-
connections. Using this algorithm, it is possible to create SF networks
whose cutoff scales as kc (N ) ∼ N 1/2 for any degree exponent, and which
are completely uncorrelated. If instead of bounding the maximum degree
by N 1/2 , we leave degrees unbounded (i.e., m ≤ ki ≤ N , like in the
configuration model [40, 41, 42, 43]) and proceed to assemble the network
from step 2), avoiding the creation of multiple connections and self-loops,
the presence of very high degree vertices for γ < 3 leads to a cutoff scaling
as kc (N ) ∼ N 1/(γ−2) and to the emergence of unavoidable structural degree-
degree correlations [44, 45, 36].
From the theoretical formalism developed in Section 2, it is easy to see
that the second term in the right hand side of Eq. (23) describes diffusion
events that cannot take place because neighboring vertices are already oc-
cupied – when paa and pab are smaller than 1. Therefore, it describes a
jamming effect that will be relevant when the concentration of particles is
large, that is, at short times. However, for sufficiently long times and low
concentration of particles, this jamming effect becomes weak and this term
can be neglected [27]. In this situation, the dynamics becomes equivalent
to the A + A → ∅ one. For this reason, in this section we will mainly focus
in the numerical results obtained for this particular case.

3.1. Density decay in uncorrelated networks

The main prediction of Section 2 is that the total density of particles should
decay with time in uncorrelated networks of infinite size as predicted by
Eq. (42) and as given by Eq. (32) in uncorrelated finite size networks in the
diffusion-limited regime. In Fig. 1, we represent the inverse particle density
220 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

from computer simulations in networks with different degree exponent γ in

UCM networks. At the initial time regime, the growth of this function is
faster for smaller values of the exponent γ, in agreement with the theoretical
prediction Eq. (42). At longer times, on the other hand, finite size effects
take over and one observes the linear regime described by Eq. (32).

Fig. 1. Inverse average particle density ρ(t) as a function of time for the A + A → ∅
process in uncorrelated SF networks with different degree exponents and size N = 106 .
The dashed line corresponds to the finite size behavior 1/ρ(t) ∼ t.

The size dependence of the slope of the linear behavior in the diffusion-
limited regime can also be checked using numerical simulations. In Fig. 2,
one can see that, for a fixed degree exponent γ = 2.5, the curves for
increasing values of N show an increase of the slope in the final linear
region. Linear fits to the final part of each curve give an estimation of the
increase of the slope as a function of N . We show this slope in the inset of
Fig. 2, in very good agreement with the theoretical prediction Eq. (32).

Fig. 2. Inverse average particle density ρ(t) as a function of time for the A + A → ∅
process in uncorrelated networks with degree exponent γ = 2.5 for different network
sizes. Inset: slope in the linear regime as a function of N , according to the prediction in
Eq. (32)
Reaction-diffusion Processes in Scale-free Networks 221

Fig. 3. Local exponent αL as a function of the network size, obtained by fitting a power
law function in the time domain before the diffusion-limited regime. Data form Fig. 2.

As we have discussed in section 2.3, even for the large network sizes
considered in our simulations, the diffusion-limited regime takes over so
quickly that it is very difficult to perform a direct quantitative check of the
predicted infinite size limit regime. Using the simulation results presented
in Fig. 2, we can give a rough estimate of the size of the network needed to
recover the predicted behavior ρ(t) ∼ t−1/(γ−2) . We perform fits of the form
ρ(t) ∼ t−αL in the time window from t ≈ 10 to just before the diffusion
limited regime for the different network sizes considered. Figure 3 shows
the dependence of αL on the size of the network. The growth of αL is
extremely slow and can be well fitted by a logarithmic function. Using this
fitting, we can extrapolate at which size the observed exponent would attain
its theoretical value α = 1/(γ − 2). For γ = 2.5, this method gives as a
lower bound N ≈ 6 × 109 far beyond the computing capabilities of modern
computers.

3.2. Depletion, segregation, and dynamical correlations

One of the most relevant differences between diffusion-annihilation processes

on SF networks as compared to regular lattices or homogeneous networks
is that in the first case the dynamics is remarkably faster, with the particle
density decaying in time as a power law with an exponent larger than 1.
This fact corresponds to the absence of two mechanisms that are sponta-
neously generated in lattices and have a slowing effect on dynamics: namely,
depletion [46, 47] and segregation [48]. The first one appears in A + A → ∅
processes: a depletion zone of empty sites is generated around any occupied
site. The second takes place in A + B → ∅ processes: a particle is typically
surrounded by particles of the same type, generating an overall segrega-
222 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

tion between the two species. Since particles have to get in contact or mix
in order to react, this phenomena result in a slowing down of the dynam-
ics, characterized by a power-law decrease of the density with an exponent
smaller than 1 (see Eq. (6)).
A measure to detect this phenomenon in complex networks was intro-
duced in [25, 49]. In the case of the A + A → ∅ process, it was proposed to
measure the quantity

NAA (t)
(49) QAA (t) = + 
n(t) n(t) − 1

that is, the number NAA (t) of close contacts between particles (a close
contact is defined by the existence of a link between two occupied vertices),
divided by the upper bound of the number of possible contacts between
existing particles. A high QAA (t) score (close to 1) corresponds to a case
when nearly all particles form one cluster, while a decrease of this value
suggests that particles are placed apart from each other. With the same
line of reasoning, for the A + B → ∅ process one can measure [25, 50]

NAB (t)
(50) QAB (t) =
NAA (t) + NBB (t)

that is, the number of close contacts between unlike particles compared to
the number of close contacts between particles of the same type. A high
QAB (t) score corresponds to the case when particles of different type are
mixed, while a decrease of this value suggests that particles are segregated
in homogeneous groups.
These measures, however, present some inconvenients. In the first one,
Eq. (49), the denominator does not take into account the specific topology of
the network, and compares the number of contacts with the upper bound of
the number of possible contacts (i.e. the number of possible contacts if all
occupied sites were connected to each other). A proper measure should
take into account the real possible contacts between particles, which is
determined by the special topology of the subgraphs formed by the occupied
sites. In the second case, Eq. (50), the number of connexions NAB is not
clearly related to NAA + NBB , yielding a quantity that could occasionally
diverge.
An alternative measure to quantify depletion and segregation is the
explicit calculation of the density correlations proposed in Ref. [27]. Let
Reaction-diffusion Processes in Scale-free Networks 223

us consider for simplicity the A + A → ∅ process. One can define a particle

correlation function by measuring at a certain time t the average density of
particles in the nearest neighbors of an occupied vertex,
, -
1   aij nj (t)
(51) ρnn (t) = ni (t) ,
n(t) ki
i j

where the brackets denote a dynamical average and ni (t) is the occupation
number of vertex i at time t. Comparing this quantity with the overall
density ρ(t), we can define the normalized correlation function:

ρnn (t)
(52) χ(t) = .
ρ(t)

When χ(t) is smaller than one, the density in the surroundings of an oc-
cupied site is smaller than the average density, which implies the presence
of depletion, segregation or density anticorrelations. In the opposite case,
when χ(t) is larger than one, in the neighborhood of an occupied site the
particle density is larger that the average, signaling accumulation, mixing of
particles or positive density correlations. The case χ(t) = 1 identifies lack
of density correlations, in which particles are homogeneously distributed
across the network substrate.

Fig. 4. Normalized density correlations for the A + A → ∅ process as a function of time

for different time snapshots. Dynamics run on uncorrelated SF networks of degree
exponent γ = 2.5 and size N = 105 .

Figure 4 shows simulation results of the normalized correlation functions

for the A + A → ∅ process in UCM networks. We can see that function χ(t)
is larger than 1 for γ ≤ 3, signaling the presence of positive correlations,
or, correspondingly, the absence of depletion zones, being the surviving
224 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

particles at any given time accumulated in closely connected clusters, a

fact that accelerates the annihilation dynamics with respect to Euclidean
lattices, in which segregation occurs. The absence of depletion is more
marked for smaller values of γ, and only present at small time scales for
values of γ > 3.
Correlation measures can be resolved in degree, in order to yield informa-
tion on the effects of the topological structure of the network, by restricting
the summation in Eq. (51) to the degree class k:
, -
1   aij nj (t)
(53) ρnn (t; k) = ni (t) ,
nk (t) k
i∈V(k) j

and from which a normalized degree resolved correlation function can be

defined, namely

ρnn (t; k)
(54) χ(t; k) = .
ρ(t)

Fig. 5. Normalized density correlations for the A + A → ∅ process as a function of time

resolved in degree for different time snapshots. Dynamics run on uncorrelated SF
networks of degree exponent γ = 2.5 and size N = 105 . The dashed line marks the MF
prediction for lack of dynamical correlations, Eq. (57).

Figure 5 shows simulation results for this quantity for the A + A → ∅

process. We can see here that the connected clusters of particles evidenced
in Fig. 4 correspond to the neighborhoods of the vertices with largest degree,
which show a largest value of χ(t; k).
The increasing values of the function χ(t) as the degree exponent γ
decreases can be understood within the MF formalism. The MF approx-
imation assumes lack of dynamical correlations between the concentration
Reaction-diffusion Processes in Scale-free Networks 225

of nearby vertices. Under this approach, the density product inside the
brackets can be substituted by the product of densities, yielding

" #
1  " #  aij nj (t)
(55) ρ0nn (t; k) = ni (t)
nk (t) k
i∈V(k) j

ρk (t)    aij 
(56) = ρk (t) = P (k  |k)ρk ,
nk (t)  
k 
k i∈V(k) j∈V(k ) k

where we have used the identity Eq. (22). For uncorrelated networks with
P (k  |k) = k P (k )/k and assuming that the dynamics at large times is
in its asymptotic diffusion-limited regime (in finite networks) ρk ∼ kρ/k,
Eq. (36), the degree resolved correlation function in absence of dynamical
correlations takes the form

k 2 
(57) χ0 (t; k) = χ0 = ,
k2

independent of time and degree, and being only a function of the degree
fluctuations, which increases as γ decreases, in agreement with numerical
simulations. However, the degree resolved correlation function shown in
Fig. 5 is flat only for degrees larger than 10, whereas it decreases as the
degree decreases. Besides, the whole curve is slightly smaller than the MF
prediction (dashed line in Fig. 5). This is a direct consequence of fact that,
even if the network is small-world, there are some dynamical correlations
that try to place particles apart.
Density correlations can be extended to the case of the A + B → ∅
process, – or to the more general process defined in Eq. (8) – in order to
account for the lack of particle segregation [27, 25] in this RD systems.
In this case, coupled correlation functions must be defined, starting from
the quantity ρα,β
nn (t; k), defined as the average density of particles of type
β at the nearest neighbors of vertices of degree k filled with α particles
(α, β = A, B), namely,

, -
1   aij nβj (t)
(58) ρα,β
nn (t; k) = α nαi (t) ,
nk (t) k
i∈V(k) j
226 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

the associated normalized density correlation function being given by

ρα,β
nn (t; k)
(59) χα,β (t; k) = .
ρβ (t)

3.3. Degree effects on annihilation

The evolution of the dynamics in SF networks is better understood by

analyzing its resolution in degree. Indeed, the density of particles at any
time t depends on the degree of the vertex. At the MF level, the expression
of the density of particles at vertices of degree k, Eq. (36), for the A+A → ∅
process (μ = 1) takes the form, within the quasi-static approximation,

kρ(t)/k
(60) ρk (t) = ,
1 + 2kρ(t)/k

which implies that high degree vertices host a constant density while low
degree vertices host a density of particles proportional to their degree [26].
In particular, at any given time t the partial density of vertices with degree
larger than k/2ρ(t) is constant and equal to 1/2 up to time t. As the
dynamics evolves, the fraction of degrees associated to a constant density
shrinks and more and more vertices acquire a density proportional to their
degree. For large t, no vertex is left in the network with degree larger than
k/2ρ(t). Therefore, when the density of particles satisfies ρ(t) ! k/2kc
(being kc the maximum degree of the network), then ρk (t) ∼ k in all degree
range. Simulations confirm this picture, as reported by the plots of the
particle density at vertices of degree k for snapshots of the dynamics at
various times, Fig. 6(top).
We can further check the validity of Eq. (60) by noticing that, if it holds,
then the function
kρk (t)
(61) G(k, t) = + 
ρ(t) 1 − 2ρk (t)

should satisfy G(k, t) = k, independently of t. This is confirmed in

Fig. 6(bottom), where we find a perfect collapse for widely separated time
snapshots.
To understand this behavior, one has to consider that high degree ver-
tices are likely to have some of their numerous nearest neighbors occupied
Reaction-diffusion Processes in Scale-free Networks 227

Fig. 6. Degree dependence of A + A → ∅ dynamics on uncorrelated SF networks with

degree exponent γ = 3 and size N = 105 . Top: Particle density at vertices of degree k at
different time snapshots. Bottom: data collapse of the partial densities according to
Eq. (61). The dashed line represents a linear trend.

by particles, given the correlated nature of the process (see section 4). Thus,
in the high density regime, high degree vertices are always surrounded by
a large number of occupied nearest neighbors. Eventually, one of these
nearby particles will diffuse into a hub, annihilating the particle in it, if
there was one, or filling the hub, if it was empty. This implies that during
this phase of the dynamics, hubs spend half of their time occupied and the
other half empty, explaining why, on average, the concentration is 1/2. In
other words, hubs act as drains through which particles vanish while their
density is steadily maintained constant by a continuous replacing of nearby
particles [26]. The absence of such replacing mechanism implies the de-
crease of the partial density of low degree vertices. However, as the global
density decrease, the replacing mechanism gets more and more restricted
to a shrinking fraction of very high degree vertices. In the end, the mech-
anism disappears and the only factor determining the density becomes the
probability of being visited by a diffusing particle. Thus, the density gets
proportional to the degree, ρk (t) ∼ kρ(t)/k, similarly to what happens in
a pure random walk [32].
This picture is further confirmed by the analysis of the annihilation
rates at vertices of different degrees [51]. Let mt (k) be the probability
that an annihilation event takes place in a single vertex of degree k during
the interval t and t + dt. At the MF level, this rate corresponds to the
annihilation term in Eq. (23)
 kP (k  |k)
(62) mt (k) = 2ρk (t) ρk (t),
k
k
228 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

that, in the uncorrelated case, reads

2kρk (t)ρ(t)
(63) mt (k) = .
k

Therefore, the probability per unit of time of an annihilation event in

any vertex increases with the probability that it is occupied and with the
number of links pointing to it. As a consequence, for large degrees and
moderate times, such that ρk (t) ∼ 1/2, we have mt (k) ∼ k, while for small
degrees, for which ρk (t) ∼ k, we expect the behavior mt (k) ∼ k2 . This
picture in confirmed by the numerical simulations reported in Fig. 7 for
different snapshots of the dynamics. To compute mt (k) numerically, we
first compute the quantity Mt (k), defined as the annihilation rate for the
class of degree k. This new quantity is computed as the ratio between
the number of annihilation events happening at vertices of degree k and
the total number of annihilation events during the interval [t, t + δt], for
δt small. Finally,
 the annihilation
 rate at single vertices is computed as
mt (k) = Mt (k)/ δtN P (k) . We consider the time interval δt = 10.

Fig. 7. Probability per unit of time that an annihilation event takes place in a vertex of
degree k, mt (k), at different time snapshots, for the A + A → ∅ dynamics on
uncorrelated SF networks with degree exponent γ = 3 and size N = 105 .

3.4. Effects of the minimum degree of the network

The MF predictions presented in Section 2 refer to generic uncorrelated

SF networks. In all these results, the role of the minimum degree of the
network, m, turns out to be irrelevant. It has been claimed, however, that
dynamics performed in networks with m = 1 deviate from the theoretically
predicted behavior, which is only recovered for m > 1 [38]. In Ref. [38]
Reaction-diffusion Processes in Scale-free Networks 229

this deviation is attributed to some property of the topology related to the

presence of dead ends in the network.
However, the observed deviations can be simply understood by consider-
ing that the dynamics is performed in this case on an effective topology with
anomalies in the degree distribution and correlation spectrum. Indeed, it
is impossible to generate SF networks both uncorrelated and globally con-
nected when the minimum degree is m = 1 [35, 52]. In this case, a SF
network always breaks up into a set of disconnected components – unless
we introduce some correlations in order to avoid it. Therefore, the dynam-
ics must be performed on the giant connected component (GCC) [34] of
the resulting network, which has a topological structure that deviates from
that of the total graph. Fig. 8 shows this effect. The degree distribution
for low degrees of the GCC does not match the one for the whole network,
Fig. 8(top). Besides, degree-degree correlations emerge in the GCC at the
same range of degrees, Fig. 8(bottom), detected in this case by the average
degree of the nearest neighbors of vertices of degree k, k̄nn (k) [28]. The
consequence of this change in the topology of the graph is a slowing down
of the initial part of the dynamics. This can be seen in Fig. 9, where we
plot the inverse of the particle density as a function of time for SF networks
generated with the UCM algorithm with m = 1 and m = 2.

Fig. 8. Degree distribution (top) and correlations (bottom) for a UCM networks with
γ = 2.7, m = 1, and size N = 105 The difference between the topology of the whole
network and the GCC affects low degree vertices, where the GCC distribution deviates
from the theoretical one and correlations appear.

A deeper insight is gained by looking at the degree resolution of the

density, Fig. 9(inset). While in a connected (m > 1) network, low degree
vertices immediately acquire a density proportional to their degree (see
Sec. 3.3), this process is much slower on the GCC of a network with m = 1,
230 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

where low degrees exhibit deviations from the expected degree distribution
and correlations. Nevertheless, the global dynamics is dominated in the long
term by high degree vertices, whose degree distribution and correlations are
coherent with the expected ones. This explains why the density displays for
long times the expected linear trend. Moreover, one can see that, after a
transient period, the degree resolution of the density behaves as predicted
by the MF theory.

Fig. 9. Density of surviving particles in SF networks with γ = 2.5, N = 105 , generated

with the UCM algorithm for m = 1 (restricted to the GCC) and m = 2. Inset: degree
resolution of the density at times t = 10, 100 and 1000 (from up to down).

3.5. Effects of a tree topology

Substantial deviations from the theoretically predicted trends in RD

processes are observed when the dynamics is performed on networks with
m = 1 and absence of loops, i.e. on trees [53], deviations that also show up
in other dynamical systems [54, 55]. In this case, the process experiences a
slowing-down and never reaches the asymptotic linear trend that dominates
the dynamics in looped networks.
While the UCM model generates looped networks with arbitrary degree
distributions3 , in the case of loopless networks we have only available the
Barabási–Albert model (BA) [6] with m = 1 and γ = 34 Thus, we restrict
our comparison to this value of the degree exponent. In Fig. 10, we compare
results of numerical simulations of the diffusion annihilation process on
3
The length of these loops is, however, of the order log N .
4
Notice that the linear preferential attachment model [56] can generate tree networks
with arbitrary γ, but intrinsically correlated except for γ = 3 [57].
Reaction-diffusion Processes in Scale-free Networks 231

uncorrelated SF networks with and without loops. The Figure shows that
the dynamics on trees is systematically slowed down and never reaches a
linear trend. A direct fit of the power law slope at long times yields the
exponent αT ∼ 0.9. The effective exponent smaller than 1 is further stressed
in the inset of Fig. 10, where we plot the inverse density divided by t.

Fig. 10. Density of surviving particles in a tree structured and looped network: BA
networks with size N = 106 and minimum degree m = 1, 2. The dashed line represents
the linear trend. Inset: plot the same quantity divided by t.

This slowing down is a general feature of dynamical systems with a

diffusive component, which can be related with the observed slowing down
of simple diffusion (random walks) in tree networks [53, 58].
Interestingly, the degree dependence of the annihilation diffusion dy-
namics on trees is different from the looped networks case, as it was first
noticed in Ref. [53]. In Fig. 11(top) one can see that the partial density in
the hubs never reaches a linear trend and rather displays a sub-linear in-
crease. This results in a kind of depletion of the trees’ hubs, that impedes
them to ever gain a density proportional to their degrees. The effects of
high degree depletion are reflected as well in the probability mt (k) of an
annihilation event in a vertex of degree k. The asymptotic quadratic trend
observed in looped networks is never reached in trees, Fig. 11(bottom).
An insight in the origin of this depletion is gained by measuring the
occupation probability of a single site of degree k. We further restrict this
measure by considering the surviving occupation probability of vertices of
degree k at time t

Pts (k)
(64) st (k) = ,
n(k)
232 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

Fig. 11. Degree dependence of A + A → ∅ dynamics on uncorrelated SF trees with

degree exponent γ = 3 and size N = 105 . Top: Particle density at vertices of degree k at
different time snapshots. Bottom: Probability that an annihilation event takes place in
a vertex of degree k at different time snapshots.

where Pts (k) is the occupation probability of degree class k at time t,

computed including not all particles present at any t < t, but only those
that will survive until t, that is, that have performed a random walk of
length t without ever encountering other particles. The results for this
quantity as measured by simulations are reported in Fig. 12 for trees and
looped uncorrelated networks with γ = 3.

Fig. 12. Surviving occupation probability of A + A → ∅ dynamics on uncorrelated SF

with degree exponent γ = 3 and size N = 105 and different time snapshots. Top: results
for looped networks (m = 2). Bottom: results for tree networks (m = 1).

For any time t the ensemble of surviving particles is composed by those

that have performed only diffusive steps until t, without experiencing any
annihilation. However, the outcome of the simulations suggest that their
random walk is biased. Indeed, st (k) is strongly sublinear at high degrees,
Reaction-diffusion Processes in Scale-free Networks 233

suggesting that the surviving random walks are those that have “avoided”
high degree vertices. This is reasonable, since high degree vertices are those
that host the higher density of particles and therefore falling on a high
degree vertex almost surely implies for a walker to annihilate.
In reaction diffusion dynamics on looped networks this bias progressively
disappear when the global density decreases, so that the particles surviving
for large times end up exploring the network as pure random walkers. On
the contrary, the bias is always present in tree structured networks.

4. Outlook

In this chapter we have reviewed the main results reported for reaction-
diffusion processes on complex networks. As we have seen, many unexpected
and interesting effects arise, as compared to regular lattices. More new
results can be expected in the future, as both the analytical framework that
has been developed and the numerical techniques employed can be easily
generalized to deal with different RD processes.
However, all the results and derivations presented so far correspond to
a fermionic version of the RD dynamics, which suffers from two conceptual
drawbacks: first, there is no systematic framework for the description of this
kind of processes, and both numerical models and theoretical approxima-
tions (through heterogeneous MF theory) must be considered on a case by
cases basis; and second, while it is relatively easy to deal with RD processes
with reactions of order 2 (involving at most two particles), it becomes more
problematic to implement reactions among three or more particles. Some
attempts have been made in this direction. For example, in Ref. [59] the
three-species process A + B + C → ∅ was considered. However, to imple-
ment it within a fermionic formalism, the authors of [59] had to consider
the presence of “intermediate” particles, produced by the reaction of two of
the species, that afterwards can react with the third species to produce an
actual annihilation event.
These type of multi-body interactions can be easily accounted for in
a more general bosonic version of the dynamics, in which more than one
particle are allowed to occupy a single vertex simultaneously. In this frame-
work, reactions take place inside the vertices, with no restrictions regarding
their order, and the interaction between vertices is mediated exclusively by
diffusion. While some preliminary interesting work has been done in this
234 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

direction [60, 61, 62], the full possibilities that the bosonic approach offers
to the investigation of RD processes remain to be explored.

Acknowledgments. We thank M. Ángeles Serrano for a careful reading of

the manuscript. M. B. acknowledges financial support from DGES, Grant
FIS2007-66485-C02-02 and Generalitat de Catalunya Grant No. SGR00889.
R.P.-S. acknowledges financial support from the Spanish MEC FEDER
under Project No. FIS2007-66485-C02-01. M. C. acknowledges financial
support from Universitat Politècnica de Catalunya.

References

[1] R. Albert and A. L. Barabási, Rev. Mod. Phys., 74 (2002), 47.

[2] S. N. Dorogovtsev and J. F. F. Mendes, Evolution of networks: From biological nets
to the Internet and WWW (Oxford University Press, Oxford, 2003).
[3] M. E. J. Newman, SIAM Review, 45 (2003), 167.
[4] B. Bollobás, Modern Graph Theory (Springer-Verlag, New York, 1998).
[5] D. J. Watts and S. H. Strogatz, Nature, 393 (1998), 440.
[6] A. L. Barabási and R. Albert, Science, 286 (1999), 509.
[7] S. Dorogovtsev, A. Goltsev and J. Mendes, Critical phenomena in complex networks
(2007), E-print arXiv:0705.0010v2 [cond-mat.stat-mech].
[8] R. Pastor-Satorras and A. Vespignani, Evolution and structure of the Internet:
A statistical physics approach (Cambridge University Press, Cambridge, 2004).
[9] R. M. Anderson and R. M. May, Infectious diseases in humans (Oxford University
Press, Oxford, 1992).
[10] R. Cohen, K. Erez, D. ben Avraham and S. Havlin, Phys. Rev. Lett., 86 (2001),
3682.
[11] D. S. Callaway, M. E. J. Newman, S. H. Strogatz and D. J. Watts, Phys. Rev. Lett.,
85 (2000), 5468.
[12] R. Pastor-Satorras and A. Vespignani, Phys. Rev. Lett., 86 (2001), 3200.
[13] A. L. Lloyd and R. M. May, Science, 292 (2001), 1316.
[14] N. G. van Kampen, Stochastic processes in chemistry and physics (North Holland,
Amsterdam, 1981).
[15] J. Marro and R. Dickman, Nonequilibrium phase transitions in lattice models (Cam-
bridge University Press, Cambridge, 1999).
[16] O. Diekmann and J. Heesterbeek, Mathematical epidemiology of infectious diseases:
model building, analysis and interpretation (John Wiley & Sons, New York, 2000).
Reaction-diffusion Processes in Scale-free Networks 235

[17] J. D. Murray, Mathematical biology, 2nd edn. (Springer Verlag, Berlin, 1993).
[18] M. Doi, J. Phys. A: Math. Gen., 9 (1976), 1465.
[19] M. Doi, J. Phys. A: Math. Gen., 9 (1976), 1479.
[20] L. Peliti, J. Phys. I, 46 (1985), 1469.
[21] D. C. Mattis and M. L. Glasser, Rev. Mod. Phys., 70 (1998), 979.
[22] M. Le Bellac, Quantum and statistical field theory (Clarendon Press, Oxford, 1991).
[23] A. A. Ovchinnikov and Y. B. Zeldovich, Chem. Phys., 28 (1978), 215.
[24] B. P. Lee, J. Phys. A: Math. Gen., 27 (1994), 2633.
[25] L. K. Gallos, P. Argyrakis, Phys. Rev. Lett., 92 (2004), 138301.
[26] M. Catanzaro, M. Boguñá and R. Pastor-Satorras, Phys. Rev. E, 71 (2005), 056104.
[27] S. Weber and M. Porto, Phys. Rev. E, 74(4) (2006), 046108.
[28] M. A. Serrano, M. Boguñá, R. Pastor-Satorras and A. Vespignani, in: Large scale
structure and dynamics of complex networks: From information technology to fi-
nance and natural sciences, ed. by G. Caldarelli, A. Vespignani (World Scientific,
Singapore, 2007), pp. 35–66.
[29] R. Pastor-Satorras and A. Vespignani, Phys. Rev. E, 63 (2001), 066117.
[30] M. Boguñá, R. Pastor-Satorras and A. Vespignani, in: Statistical Mechanics of
Complex Networks, Lecture Notes in Physics, vol. 625, ed. by R. Pastor-Satorras,
J. M. Rubı́, A. Dı́az-Guilera (Springer Verlag, Berlin, 2003).
[31] M. Boguñá and R. Pastor-Satorras, Phys. Rev. E, 66 (2002), 047104.
[32] J. D. Noh and H. Rieger, Phys. Rev. Lett., 92 (2004), 118701.
[33] L. Lovász, in Combinatorics, Paul Erdős is Eighty, vol. 2, ed. by V. T. Sós,
D. Miklós, T. Szőnyi (János Bolyai Mathematical Society, Bupadest, 1996), pp.
353–398.
[34] S. N. Dorogovtsev and J. F. F. Mendes, Adv. Phys., 51 (2002), 1079.
[35] M. Catanzaro, M. Boguñá and R. Pastor-Satorras, Phys. Rev. E, 71 (2005), 027103.
[36] M. Boguñá, R. Pastor-Satorras and A. Vespignani, Euro. Phys. J. B, 38 (2004),
205.
[37] M. E. J. Newman, in: Handbook of Graphs and Networks: From the Genome to
the Internet, ed. by S. Bornholdt, H. G. Schuster (Wiley-VCH, Berlin, 2003), pp.
35–68.
[38] L. K. Gallos and P. Argyrakis, Phys. Rev. E, 72 (2005), 017101.
[39] M. Abramowitz and I. A. Stegun, Handbook of mathematical functions (Dover,
New York, 1972).
[40] A. Bekessy, P. Bekessy and J. Komlos, Stud. Sci. Math. Hungar., 7 (1972), 343.
[41] E. A. Bender and E. R. Canfield, Journal of Combinatorial Theory A, 24 (1978),
296.
[42] B. Bollobás, Eur. J. Comb., 1 (1980), 311.
236 M. Catanzaro, M. Boguñá and R. Pastor-Satorras

[43] M. Molloy and B. Reed, Random Struct. Algorithms, 6 (1995), 161.

[44] J. Park, M. E. J. Newman, Phys. Rev. E, 66 (2003), 026112.
[45] S. Maslov, K. Sneppen and A. Zaliznyak, Physica A, 333 (2004), 529.
[46] D. Torney and H. McConnell, J. Phys. Chem., 87 (1983), 1441.
[47] D. Torney and H. McConnell, Proc. R. Soc. Lond. A, 387 (1983), 147.
[48] D. Toussaint and F. Wilczek, J. Chem. Phys., 78 (1983), 2642.
[49] L. K. Gallos and P. Argyrakis, J. Phys.: Condens. Matter, 19 (2007), 065123.
[50] J. Newhouse and R. Kopelman, J. Phys. Chem., 92 (1988), 1538.
[51] L. K. Gallos, P. Argyrakis, Phys. Rev. E, 74 (2006), 056107.
[52] R. Cohen, K. Erez, D. ben Avraham and S. Havlin, Phys. Rev. Lett., 85 (2000),
4626.
[53] J. D. Noh and S. W. Kim, Journal of the Korean Physical Society, 48 (2006), S202.
[54] C. Castellano et al., Phys. Rev. E, 71 (2005), 066107.
[55] L. Dall’Asta, A. Baronchelli, A. Barrat and V. Loreto, Phys. Rev. E, 74 (2006),
036105.
[56] S. N. Dorogovtsev, J. F. F. Mendes and A. N. Samukhin, Phys. Rev. Lett., 85
(2000), 4633.
[57] A. Barrat and R. Pastor-Satorras, Phys. Rev. E, 71 (2005), 036127.
[58] A. Baronchelli, M. Catanzaro and R. Pastor-Satorras, Random walks on scale-free
trees (2008), E-print arXiv:0801.1278v1 [cond-mat.stat-mech].
[59] K. H. Chang, K. G. Park, K. D. Ahan, S. Y. Kim, D. H. Ha and K. Kim, Journal
of the Physical Society of Japan, 76 (2007), 035001.
[60] V. Colizza, R. Pastor-Satorras and A. Vespignani, Nature Physics, 3 (2007), 276.
[61] V. Colizza and A. Vespignani, Phys. Rev. Lett., 99 (2007), 148701.
[62] A. Baronchelli, M. Catanzaro and R. Pastor-Satorras, Bosonic reaction-diffusion
processes on scale-free networks (2008), E-print arXiv:0802.3347v1 [cond-mat.stat-
mech].

Michele Catanzaro Marián Boguñá

Departament de Fı́sica i Enginyeria Departament de Fı́sica Fonamental
Nuclear Universitat de Barcelona
Universitat Politècnica de Catalunya 08028 Barcelona
08034 Barcelona Spain
Spain e-mail:
e-mail:
Reaction-diffusion Processes in Scale-free Networks 237

Romualdo Pastor-Satorras
Departament de Fı́sica i Enginyeria
Nuclear
Universitat Politècnica de Catalunya
08034 Barcelona
Spain
e-mail:
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 239–275.

Chapter 6
Toward Understanding the Structure and
Function of Cellular Interaction Networks

JUILEE THAKAR, CLAIRE CHRISTENSEN and RÉKA ALBERT

The components of a biological cell (such as proteins and molecules) interact on

several levels, forming a variety of networks. Newly-developed high-throughput
molecular biology methods now allow for the construction of genome-level in-
teraction graphs. In parallel, high-throughput molecular abundance data paired
with computational algorithms can be used to infer graphs of interactions and
causal relationships. Graph-theoretical measures and network models are more
and more frequently used to discern the functional constraints in the organization
of biological networks. Perhaps most importantly, the combination of interaction
and abundance information allows the formulation of predictive dynamic models.
Here we review some of the dominant theoretical and computational methods
used for the inference, graph-theoretical analysis and dynamic modeling of cellu-
lar networks, and we present two specific projects that integrate these three steps.
Throughout this chapter we focus on presenting the significant advances in un-
derstanding biological systems allowed by graph theory. Conversely, we hope that
the new biological data and results may inspire new directions in graph theory.

1. Definition of Cellular Interaction Networks

The cell is the structural and functional unit of almost all known living
organisms (viruses are an exception). It is the smallest unit of an organ-
ism that is classified as living, and is sometimes called the building block
of life [5]. The functions of each cell, such as converting nutrients into en-
240 J. Thakar, C. Christensen and R. Albert

ergy, or reproducing, are determined by the interactions among the macro-

molecules that make up cells. There is a large diversity in these cellular
components and in the interactions possible among them. The units of
hereditary information encoded in each cell’s DNA are called genes. Each
gene contains most of the information for the synthesis of (usually) one pro-
tein. When the gene is expressed, an intermediary molecule called messenger
RNA (mRNA) is first made (transcribed) using the gene as a template, and
then the corresponding protein is synthesized (translated). Proteins make
up most of the biomolecules of the cell and their functions range from pro-
viding structure to cells and tissues, working as molecular motors, sensing
chemicals in the environment (these proteins are called receptor proteins),
driving chemical reactions (these proteins are called enzymes) or regulating
gene expression (these proteins are called transcription factors). Cells also
contain a variety of small molecules. The content of each cell is in a con-
stant flux of synthesis and decay. Protein abundances range from zero to
thousands, and their lifetimes range from minutes to days. Most cells are
also rapidly recycled. Components come and go, and therefore the essence
of a living process is the interconnections between them, in other words, the
graph of cellular components and interactions.
A system of elements that interact or regulate each other can be rep-
resented by a mathematical object called a graph (or network) [13]. At
the simplest level, the system’s elements are reduced to graph nodes (also
called vertices) and interactions are reduced to edges connecting pairs of
nodes. Edges can be either directed, specifying a source (starting point)
and a target (endpoint), or non-directed. Thus, the interactions among
genes and proteins can be depicted graphically as graphs with either di-
rected or nondirected edges. Directed edges are suitable for representing
the flow of material from a reactant to a product in a reaction or the flow
of information from a transcription factor to the gene whose transcription
it regulates. Non-directed edges are used to represent mutual interactions
such as protein-protein binding. Graphs can be augmented by assigning
various attributes to the nodes and edges: multi-partite graphs allow rep-
resentation of different classes of node (such as mRNA and protein), and
edges can be characterized by signs (positive for activation, negative for
inhibition), or weights quantifying confidence levels, strengths, or reaction
speeds.
In this chapter, we will focus on graphs constructed from molecular-
to-cellular level interactions. At the genomic level, transcription factors
activate or inhibit the transcription of genes to give mRNA; since these
Toward Understanding the Structure 241

transcription factors are themselves products of genes, the ultimate effect

is that genes regulate each other’s expression as part of gene regulatory
networks. Similarly, proteins participate in diverse post-translational inter-
actions that lead to modified protein functions or they bind to form protein
complexes; the totality of these processes is called a protein-protein interac-
tion network. The biochemical reactions of cellular metabolism can likewise
be integrated into a metabolic network whose fluxes are regulated by en-
zymes catalyzing the metabolic reactions. In many instances these different
levels of interactions are integrated, as is done, for example, when the pres-
ence of an external signal triggers a cascade of interactions that involves
both biochemical reactions and transcriptional regulation.
Genome-level information concerning cellular networks is often described
using five ‘omes’: genome (DNA sequence information), transcriptome (the
totality of transcribed genes), proteome (all proteins in a cell), metabolome
(all metabolites in a cell), and interactome (the totality of protein inter-
actions). During the last decade, the respective omics have produced an
incredible quantity of molecular expression and interaction data, contribut-
ing to the construction of several types of interaction graphs [70, 14].
The full representation of transcriptional regulatory graphs asso-
ciates two separate node classes with transcription factors and mRNAs,
respectively, and has two types of directed edge, which correspond to tran-
scriptional regulation and translation [52]. The edges describing transcrip-
tional regulation can have two regulatory effects (signs): activation and
inhibition. In protein interaction graphs, the nodes are proteins, and
two proteins are connected by a nondirected edge if there is strong evidence
of their association.
A metabolic network can be represented as a directed and weighted
tri-partite graph, whose three types of node are metabolites, reactions, and
enzymes, and whose two types of edge represent mass flow and catalytic reg-
ulation, respectively. Mass flow edges connect reactants to reactions and re-
actions to products, and are often marked by the stoichiometric coefficients
of the metabolites [53]; enzymes catalyzing the reactions are connected by
regulatory edges to the nodes signifying the reaction [45]. Several simpli-
fied representations include the substrate graph, whose nodes are reactants,
joined by an edge if they occur in the same chemical reaction [90], and the
reaction graph, whose nodes are reactions, connected if they share at least
one metabolite.
Signal transduction networks, connecting extracellular signal inputs
to the control of transcription factors, share a significant overlap with pro-
242 J. Thakar, C. Christensen and R. Albert

tein interaction networks and metabolic networks, as they involve both pro-
tein interactions and biochemical reactions. The nodes of signal transduc-
tion networks can be categorized by the function of the corresponding pro-
tein or molecule, and the edges are mostly directed, indicating the direction
of signal propagation.
This chapter focuses on three major questions: (i) How can one map
or infer the regulatory graph underlying a biological system, and at what
scale should such a graph be constructed for an optimal understanding of
the system? (ii) What are the best measures that capture the most salient
features of a biological interaction graph? and (iii) How does the topology
of biological systems influence their dynamics and function? We present
a representative (while by necessity not comprehensive) picture of recent
progress made in answering these questions. We follow up with two example
models that integrate interaction graph synthesis, analysis and predictive
dynamic modelling.

2. Detecting or Inferring Interactions

Advancements in molecular biology techniques, in tandem with improve-

ments in computational techniques for analysis of complex systems, have
produced high-throughput interaction data. These improvements promise
to change the focus of cell biology from an understanding of local, binary in-
teractions to an understanding of the aggregation of these interactions into
a functional system.
Transcriptome data convey the identity of each expressed gene and
its level of expression (the abundance of its transcribed mRNA) for a de-
fined population of cells. While the transcriptome constitutes node (com-
ponent)-level information without information on interactions, it can be
used to generate networks of functional relationships between genes. Gene
expression data studied across time can be combined with computational
methods to extract interactions [40]. The construction of transcriptional
regulatory graphs requires the large-scale detection of transcription factor –
DNA interactions. Known information about these interactions is available
in databases such as the Transcription Factor Database (TRANSFAC) [94],
Regulon Database (RegulonDB) [43], and the Kyoto Encyclopedia of Genes
and Genomes (KEGG) [49].
Toward Understanding the Structure 243

The interactome represents protein-protein interactions, including the

formation of protein complexes and interactions involved in signal propaga-
tion. One of the most frequently used interaction detection method, called
the yeast two hybrid method, creates two hybrid proteins by fusing two pro-
teins of interest to the DNA binding and activation domain of a transcription
factor. The two hybrid proteins are introduced into yeast, and if the two
proteins bind, the transcription factor is reconstituted, and transcription of
its target genes can be observed. This method yielded the first interaction
maps of metazoans, namely D. melanogaster [30], Caenorhabditis elegans
[54] and Homo sapiens [79]. Many interactome databases are maintained,
including the Database of Interacting Protein (DIP) [96], the Biomolecular
Interaction Network (BIND) [6], the Munich Information Center for Protein
Sequences (MIPS) [66], the Human Protein Reference Database (HPRD)
[74], and the Yeast Proteome Database (YPD) [28].
The metabolome represents the totality of metabolites (molecules)
used or produced by an organism during its interconversion of nutrients into
energy. Experimental techniques of metabolome analysis include methods to
trace and identify metabolites and to characterize the enzymes catalyzing
the metabolic reactions. The function and metabolic pathways of identi-
fied metabolites can be looked up in the Encyclopedia of Metabolic Path-
ways (MetaCyc) [50] and the Kyoto Encyclopedia of Genes and Genomes
(KEGG) [49].
Though all ‘omes’, with the exception of the interactome, essentially
provide vertex information, they can be used to infer networks of indirect
interactions. Computational inference (also referred to as reverse engineer-
ing) aims to extract causal relationships from transcriptome, proteome, and
metabolome data. Inference of cellular networks allows for a clearer com-
prehension of the inner machinery of the cell, and when combined with
modeling, can also be used to make experimentally-verifiable predictions
about cellular networks. A variety of computational methods for network
inference exists; choosing a specific computational method depends on the
nature of the data from which inferences will be made, on the type of net-
work under consideration, on the features of the system one would most
like to illuminate, and on the amount of computational time available to
the researcher.
Probabilistic methods, including clustering analyses [22], data-min-
ing [62], and naı̈ve Bayesian networks [26], are applicable both to the in-
ference of protein-protein interactions, and to the inference of functional
relationships among genes based on similarities in gene expression profiles.
244 J. Thakar, C. Christensen and R. Albert

Clustering of genes on the basis of their expression profiles gives insight into
classes of genes that respond in a similar manner to varying conditions, and
that might therefore be co-regulated [75, 27], and these functional clusters
can be used as a starting point for determining gene-regulatory relation-
ships. Recently-developed methods use topological measures to establish
correlation thresholds for determining whether or not pairs of genes are co-
expressed: transitivity criteria lead to excellent agreement with Bacillus
subtilis operon structure and differential regulation [36], and heterogeneity
criteria lead to experimentally validated gene modules in human glioblas-
toma [42].
Data-mining schemes typically extract information on relationships be-
tween two entities based on the statistical co-occurrence of features of in-
terest associated with the entities, for example, their inclusion in databases
and biomedical journals [84]. Data mining algorithms have been used to
augment the yeast protein-protein interaction network [62] and for the in-
ference of both direct and indirect protein-protein interactions in eukaryotes
and prokaryotes [65].
Bayesian network methods aim to infer a directed, acyclic graph that
summarizes the dependency relationships among variables in the system,
and a set of local joint probability distributions that statistically convey
these relationships [26]. The initial links of the dependency graph are es-
tablished either randomly or based on prior knowledge, and the network is
refined by an iterative search-and-score algorithm in which multiple can-
didate networks are scored against experimental observations and against
one another [98]. Bayesian networks have been employed to sort yeast pro-
teins into functional groupings [23] and to infer protein interactions in the
S. cerevisiae cell cycle [26].
Deterministic methods employed for the inference of gene-regulatory
networks from time-course gene expression (microarray) data seek to cor-
relate the rate of change in the expression level (mRNA concentration) of
a given gene with the levels of other genes in the network. These methods
describe interdependence in gene expression in one of two ways: continuous
deterministic methods postulate a system of differential equations [17], while
Boolean and other logical methods predict a discrete relationship [82]. De-
terministic methods based on systems of linear differential equations have,
for example, been used to infer gene-regulatory networks in B. subtilis [37]
as well as regulatory networks specific to the central nervous system of the
rat [17]. Experimental data on gene expression levels is substituted into the
Toward Understanding the Structure 245

relational equations, and the ensuing system of equations is then solved for
the regulatory relationships between two or more components [37].
Deterministic Boolean methods assume that the expression level of
each gene can be approximated with a binary variable: expressed/ON or
not expressed/OFF and describe gene regulation by Boolean logical func-
tions [82]. Each node’s logical function is found by determining the min-
imum set of nodes whose (changing) expression levels can explain the ob-
served changes in state of the given node in all experimental trials. Proba-
bilistic Boolean methods incorporate uncertainty and fluctuations in expres-
sion levels by assigning several alternative logical functions to each gene [82];
a machine learning algorithm then selects the most probable logical func-
tion at each time point. This method was used to infer the regulatory
networks involved in embryonic segmentation and muscle development in
D. melanogaster [100].
The majority of network inference methods presented in this section use
node-level (expression) information to infer causal relationships. There also
exists a complementary problem: inferring interactions from indirect causal
relationships. Indeed, experimental information about the involvement of a
protein in a process is often restricted to evidence of differential responses
to a stimulus in wild-type organisms versus an organism where the respec-
tive protein’s expression or activity is disrupted. These observations can be
incorporated by two intersecting paths in an (incompletely mapped) inter-
action network; the inference algorithm must integrate indirect and direct
evidence to find a network consistent with all experimental observations [55].
This inference problem is less studied [55, 92, 4], but we expect it will play
an increasing role when integrating information from disparate data sources.

3. Network Measures

With a set of experimentally-determined or inferred interactions in hand, the

relevant cellular interaction graph can be constructed. A graph-theoretical
analysis can then provide powerful biological insights into the structural
organization and function(s) of the system. Here we present a handful of
graph-theoretical measures that can serve as the basis of such an analysis.
246 J. Thakar, C. Christensen and R. Albert

Fig. 1. Illustration of frequently used graph measures. The node degree (k) quantifies
the number of edges that start at or end in a given node, for example node K has both
an in-degree and an out-degree of two. The clustering coefficient (C) characterizes the
cohesiveness of the first neighborhood of a node, for example the clustering coefficient of
node F is zero because its two first neighbors are not connected, and the clustering
coefficient of node C is 1, indicating that it is a part of the BCD clique. The graph
distance (d) between two nodes is defined as the number of edges in the shortest path
between them. For example, the distance between nodes L and K is one, the distance
between nodes K and L is two (along the KM L path), and the distance between nodes
J and I is infinite because no path starting from J and ending in I exists. The
betweenness centrality (b) of a node quantifies the number of shortest paths in which the
node is an intermediary (not beginning or end) node. For example, the betweenness
centrality of node C is zero because it is not contained in any shortest paths that do not
start or end in C, and the betweenness centrality of node G is one and a half because it
is an intermediary in the F GH shortest path and in one of two alternative shortest
paths between E and F (EGF and EDF ). The degree distribution, P (k) (P (kin ) and
P (kout ) in directed networks) quantifies the fraction of nodes with degree k. The
clustering-degree function C(k) gives the average clustering coefficient of nodes with
degree k. For example, in the top graph the sole node with degree one (A) has an
undefined clustering coefficient because it only has a single first neighbor, and all three
nodes of degree 3 (B, D, G) have a clustering coefficient of one third, giving C(3) = 1/3.
The distance distribution P (d) denotes the fraction of node pairs having the distance d.
The betweenness centrality distribution P (b) quantifies the fraction of nodes with
betweenness centrality b.
The top, undirected graph is connected, has a range of degrees from one to four,
clustering coefficients between zero and one, a range of pairwise distances from one to
four, and node betweenness centralities between zero and twelve and a half. The BCD
and FGH subgraphs are cliques (completely connected subgraphs) of three nodes.
The bottom, directed graph contains two source nodes (I and O, both with kin = 0),
one sink node (J, with kout = 0), one feed-forward loop (OLN ; both O and L feed
into N ) and two feedback loops (M LN and M LK). The nodes K, L, M and N form
the graph’s strongly connected component. The in-component contains the source nodes
I and O, while the out-component consists of the sink node J.
Toward Understanding the Structure 247

Degree and degree distribution

The degree of a node is the number of edges adjacent to that node. If the
directionality of interaction is important, a node’s total degree can be broken
into an in-degree and an out-degree, quantifying the number of incoming and
outgoing edges adjacent to the node. In a graph whose edges are quantified
by weights one can also define a node strength, the sum of the weights of the
edges adjacent to the node (Figure 1). While the degree of a specific node is
a local topological measure, this local information can be synthesized into a
global description of the network by reporting the degrees of all nodes in the
network in terms of a degree distribution, P (k), which gives the probability
that a randomly-selected node will have degree k (Figure 1). The degree
distribution is obtained by first counting the number of nodes, N (k), with
k = 1, 2, 3, . . . edges, and then dividing this number by the total number of
nodes, N , in the network (the same procedure can be employed to find in-
and out-degree distributions for a given directed network).
The majority of cellular networks has been shown to have (out-) degree
distributions that are scale-free, meaning that the high diversity of node
degrees precludes the existence of a typical node that could be used to
characterize the rest of the nodes in the network (reviewed in [1]). The
degree distribution of these scale-free networks is close to a power-law:
P (k) ∼
= Ak −γ , where A is a normalization constant, and where the degree
exponent γ is typically similar for similar networks. The degree distributions
of protein interaction networks, metabolic networks and the out-degree
distribution of most gene-regulatory networks, for example, are power-laws
with 2 < γ < 3 [52, 45, 34].

Clustering coefficient

The clustering coefficient quantifies the extent to which a node’s first neigh-
borhood is a completely-connected subgraph (clique) [93]. Mathematically,
2Ei
the local clustering coefficient is given by Ci= ki (k i −1)
, where Ei is the number
of edges connecting the immediate neighbors of node i, and ki is the degree
of node i. Protein-protein interaction networks and metabolic networks [90]
exhibit large average clustering coefficients, indicating a high level of redun-
dancy and cohesiveness. Alternatively, the average clustering coefficient of
nodes with degree k can be plotted as a function of node degree, C(k) (Fig-
ure 1). It has been found that for a wide variety of cellular networks, this
248 J. Thakar, C. Christensen and R. Albert

clustering-degree relation has the functional form C(k) = B/kβ , where the
exponent β typically falls between 1 and 2 [97, 77].

Connectivity, paths, distances, efficiency, and graph components

Two nodes of a graph are connected if a sequence of adjacent nodes, a path,

links them [13]. A path can signify a transformation route from a nutrient
to an end product in a metabolic network, or it could represent a chain of
ligand-induced reactions from the source to the sink in a signal transduction
network. The distance (path length) between any two nodes in a network
is defined to be the number of edges in the shortest path connecting those
nodes. If the edges of a network are weighted (e.g. with rate constants), then
the distance between two nodes will be the sum of edge weights along the
path for which this sum is a minimum [21]. Often, the average path length
d = dij , i.e. the average number of edges in the shortest path between any
two nodes in a network (Figure 1), scales with the natural logarithm of the
number of nodes in the graph: d ∼ ln(N ). This small world [93] property
implies that path lengths remain small, even if the networks become very
large, and it is a feature of metabolic and protein interaction networks.
Particularly in the case that a network is directed, it is possible that by
starting at an edge adjacent to a given node and tracing a path along consec-
utive edges, only a fraction of the nodes in the network will be accessible to
the starting node. If a network is either directed or unconnected, it is often
more advantageous to define the graph’s global efficiency eff = 1/dij  [51].
Unconnected nodes’ distance is infinite by definition, and thus these node
pairs do not contribute to the network’s efficiency. If, however, a path
does exist between every pair of nodes in a network, the network is said to
be connected. Directed networks having directed paths in both directions,
between every pair of nodes are said to be strongly connected. Even if a
cellular network is not (strongly) connected it is beneficial to identify con-
nected partitions of the network. For example, a directed network has one or
several strongly-connected components, a subgraph(s) whose node pairs are
connected in both directions. Each strongly-connected component is asso-
ciated with an in-component (nodes that can reach the strongly-connected
component, but that cannot be reached from it) and an out-component (the
converse) (Figure 1). It has recently been suggested that the nodes of each
of these components share a component-specific task within a given network.
In signal transduction networks, for example, the nodes of the in-component
Toward Understanding the Structure 249

tend to be involved in ligand-receptor binding; the nodes of the strongly-

connected component form a central signaling subnetwork; and the nodes
of the out-component are responsible both for the transcription of target
genes as well as for phenotypic changes [59]. By identifying large connec-
tivity classes within a network, one may be able to gain a sense of how the
network is organized functionally.

Betweenness centrality, sources and sinks

The number, (net) directionality, and strength of connections associated

with a given node are measures of that node’s local centrality, and can be
synthesized into distributions over all nodes in a network to give valuable in-
sight both into the heterogeneity of node interactivity levels within a cellular
network, and into the flow of information, mass, or other entities through the
network. In particular, the sources and sinks of the network – those nodes
with only outgoing or incoming edges, respectively – represent the initial
and terminal points of the flow. In signal transduction networks, for exam-
ple, extracellular ligands and/or their receptors are typically sources, acting
as the injection points for chemical signals; these chemical signals then ter-
minate at effectors – the networks’ sinks [58]. If a node is neither a source
nor a sink, its betweenness centrality – the number of (shortest) paths from
node s to node t passing through the node, divided by the total number of
(shortest) st-paths (Figure 1) indicates the importance of that node to the
propagation of flow within the network [25]. One can similarly define the
betweenness centrality of an edge [31]. Even though a node’s betweenness
centrality is not necessarily correlated with its degree, betweenness central-
ities are usually power-law distributed, with a characteristic exponent close
to 2 [33]. The most ubiquitous substrates in biochemical pathways often
have the highest betweenness centralities [41].

4. Properties of Cellular Interaction Graphs

To date, few cellular networks have been reconstructed and analyzed in

full. However, transcriptional regulatory maps exist for E. coli [81] and
S. cerevisiae [52, 34, 56], and protein-protein interaction maps have been
constructed for a variety of organisms, including Helicobacter pylori [76],
250 J. Thakar, C. Christensen and R. Albert

S. cerevisiae [88], C. elegans [54], D. melanogaster [30] and H. sapiens [79].

Graph-theoretical analysis of these and other cellular networks has yielded
a wealth of information regarding structural organization and functioning
at the cellular level.

Both prokaryotic and eukaryotic transcription networks exhibit an

approximately scale-free out-degree distribution, signifying the potential of
transcription factors to regulate a multitude of target genes. The in-degree
distribution is a more restricted exponential function, illustrating that com-
binatorial regulation by several transcription factors is observed less than
regulation of several targets by the same transcription factor. Neither the
E. coli nor the yeast transcription network has strongly-connected compo-
nents, suggesting a unidirectional, feed-forward type regulation mode. Note
that although purely transcriptional feedback loops are under-represented,
feedback is still frequently accomplished through the protein-level regula-
tion of transcription factors, often via auto-regulation. The subgraphs found
by following the paths that start from non-transcriptionally regulated genes
have relatively little overlap [7], reflecting the fact that distinct environmen-
tal signals tend to initiate distinct transcriptional responses. The source –
sink distances are small in both networks, and the longest regulatory chain
has only four (in E. coli ) or five (in S. cerevisiae) edges.

The current versions of protein interaction maps are, by necessity,

incomplete, and also suffer from a high rate of false positives. Despite these
drawbacks, there is an emerging consensus in the topological features of the
maps of different organisms. For example, all protein interaction networks
have a giant connected component and the distances on this component are
close to the small-world limit given by random graphs [30, 97]. This finding
suggests an illustration of pleiotropy, since perturbations of a single gene or
protein can propagate through the network, and have seemingly unrelated
effects. The degree distribution of the yeast protein interaction network is
approximately scale-free (see Figure 2). The Drosophila protein network
exhibits a lower-than-expected fraction of proteins with more than 50 in-
teracting partners; this deviation is suspected to be caused by incomplete
coverage and could change as more interactions are discovered, as was the
case for the yeast protein interaction network [30, 97]. The heterogeneous
clustering-degree function C(k) = B/kβ , where the exponent β is around 2
[97], and the inverse correlation between the degree of two interacting pro-
teins [63] indicate that the neighborhood of highly-connected proteins tends
to be sparser than the neighborhood of less-connected proteins.
Toward Understanding the Structure 251

Fig. 2. Topological properties of the yeast protein interaction network constructed from
four different databases. a) Degree distribution. The solid line corresponds to a power
law with exponent γ = 2.5. b) Clustering coefficient – degree function. The solid line
corresponds to the function C(k) = B/k 2 . c) The size distribution of connected
components. All networks have a giant connected component of more than 1000 nodes
(on the right) and a number of small isolated clusters. Reproduced from [97] Copyright
(2004), with permission from Wiley InterScience.

The largest reconstructed signal transduction network was synthe-

sized from more than 1200 articles in the experimental literature and con-
tains 1259 interactions among 545 cellular components of the hippocampal
CA1 neuron [59]. This network exhibits an impressive interconnectivity: its
strongly-connected component (the central signaling network) includes 60%
of the nodes, and the subgraphs that start from various ligand-occupied re-
ceptors reach most of the network within 15 steps. The average input-output
path-length is near 4, suggesting the possibility of very rapid response to
signaling inputs. Both the in- and out-degree distributions of this network
are consistent with a power-law with an exponent around 2, and the high-
est degree nodes include the four major protein kinases (MAPK, CaMKII,
PKA and PKC).

All metabolic network representations indicate an approximately

scale-free or at least broad-tailed [45, 90, 85] metabolite degree distribu-
tion. The degree distribution of enzymes is strongly peaked, indicating
that enzymes catalyzing several reactions are rare [45]. The variability of
metabolite degrees can be accounted for if they are functionally separated
into high-degree carriers and low-degree metabolites unique to separate re-
action modules (such as catabolism or amino acid biosynthesis) [85]; how-
ever, such a picture does not seem to explain the frequency of intermediate
degrees. The clustering-degree function follows the relationship C(k) ∼
= 1/k.
252 J. Thakar, C. Christensen and R. Albert

The substrate and reaction graphs indicate a remarkably small and

organism-independent average distance among metabolites and reactions
[45, 90]. If the preferred directionality of the reactions is known and is
taken into account, only the largest strongly-connected component (whose
nodes can reach each other in both directions) has well-defined average path
length. While this average path length is still small in all the organisms
studied, the strongly connected component itself contains less than 50% of
the nodes [57].
The general architectural features of molecular interaction networks de-
scribed so far are shared to a large degree by other complex systems, ranging
from technological networks to social networks. While this universality is
intriguing, it is arguably more important to discern whether, and how, the
graph properties of cellular networks reflect their functional constraints.
Hubs: In a scale-free network small-degree nodes are the most abun-
dant, but the frequency of high-degree nodes decreases relatively slowly.
Thus, nodes that have degrees much higher than average, so-called hubs,
exist. Because of the heterogeneity of scale-free networks, random node dis-
ruptions do not lead to a major loss of connectivity, but the loss of the hubs
causes the breakdown of the network into isolated clusters [1]. The validity
of these general conclusions for cellular networks can be verified by corre-
lating the severity of a gene knockout with the number of interactions the
gene’s products participate in. A gene knockout is a mutation that induces
the non-expression of a gene, or the non-functionality of its products, thus
in effect removing the nodes corresponding to that gene’s products from the
interaction graph. Indeed, as much as 73% of the S. cerevisiae genes are
non-essential – i.e. their knockout has no phenotypic effects [29]. This con-
firms cellular networks’ robustness in the face of random disruptions. The
likelihood that a gene is essential (i.e. that its knockout strain does not sur-
vive) or toxicity-modulating (i.e. that its knockout strain is toxin-sensitive)
correlates with the number of interactions its protein product has [46, 80].
This indicates that the cell is vulnerable to the loss of highly interactive
hubs1 . Among the most well-known examples of hub proteins is the tumor
suppressor protein p. 53, which has an abundance of incoming edges – in-
teractions regulating its activity – and which also has many outgoing edges
in the genes whose transcription it activates. The tumor suppressor p53 is
inactivated by a mutation in its gene in 50% of human tumors, corrobo-
rating the fact that cellular networks are vulnerable to loss of their most
1
We note here that different network representations can lead to distinct sets of hubs
and there is no rigid boundary between hub and non-hub genes or proteins.
Toward Understanding the Structure 253

connected hubs [89]. High interactivity is not the only marker of functional
importance, however: low-degree nodes in genome-wide metabolic networks
of various microorganisms are almost as likely to be critical to the overall
network functions as high-degree nodes [60].
Modularity: Cellular networks have long been thought to be modular,
composed of functionally-separable subnetworks corresponding to specific
biological functions [39]. Since genome-wide interaction networks are highly
connected, modules should not be understood as disconnected components
but rather as components that have dense intra-component connectivity
but sparse inter-component connectivity. Several methods have been pro-
posed to identify functional modules on the basis of the physical location or
function of network components [78], or on the topology of the interaction
network [83, 35]. The challenge is that modularity does not always mean
clear-cut subnetworks linked in well-defined ways, since there is a high de-
gree of overlap and cross-talk between modules [38]. A heterogeneous degree
distribution, inverse correlation between degree and clustering coefficient (as
seen in metabolic and protein interaction networks) and modularity, taken
together, suggest hierarchical modularity, in which modules are made up
of smaller and more cohesive modules, which themselves are made up of
smaller and more cohesive modules etc. [77].
Subgraphs and motifs: Cellular networks contain recurring subgraphs
with well-defined topologies, termed interaction motifs (Figure 3). Exam-
ples include loops (corresponding to positive or negative autoregulation),
short cycles (called positive or negative feedback loops in this context), and
a subgraph with three nodes of which one is a source and one is a sink
and three directed edges (called a feed-forward loop in this context). In-
teraction motifs such as autoregulation (usually a negative feedback) and
feed-forward loops have a higher abundance in transcriptional regulatory
networks than expected from randomly-connected graphs with the same
degree distribution [7, 67]. Protein interaction motifs such as short cy-
cles and small, completely-connected subgraphs are both abundant [30] and
evolutionarily-conserved [95], partly because of their enrichment in protein
complexes. Feedforward loops and triangles of scaffolding (protein) interac-
tions are also abundant in signal transduction networks, which additionally
contain a significant number of feedback loops, both positive and negative
[59]. The abundant motifs of integrated mRNA/protein networks are often
signatures of higher-order network structures that correspond to biologi-
cal phenomena such as interacting transcription factors regulating the same
target gene or co-regulation of members of a protein complex [99].
254 J. Thakar, C. Christensen and R. Albert

Fig. 3. Examples of network motifs in the yeast transcriptional regulatory network.

Regulators are represented by circles; target genes are represented by rectangles.
Transcriptional regulation is indicated by a solid edge. Genes that encode regulators are
linked to them by dashed edges. For example, in the autoregulation motif, the Ste12
regulatory protein activates the transcription of the STE12 gene, which is transcribed
and translated into Ste12 protein. Reprinted from [52] Copyright (2002), with
permission from the American Association for the Advancement of Science.

The functional relevance of recurring network motifs has been investi-

gated both theoretically and experimentally. Coherent feed-forward loops
have been shown to cause sign-sensitive delay: coherent feed-forward loops
filter out brief signal fluctuations [61], while incoherent feed-forward loops
accelerate signaling through an overshoot dynamics [48]. The comparative
abundance of negative feedback loops in the early steps of signal transduc-
tion networks and of positive feedback loops at later steps suggests that
weak or short-lived signals are filtered by early barriers posed by negative
feedback loops, while strong and persistent signals are amplified and are
able to evoke a biological response [59].
Toward Understanding the Structure 255

Path redundancy: Any response to a perturbation requires that in-

formation about the perturbation spreads within the network. Thus the
short path lengths of metabolic, protein interaction and signal transduction
networks (their small world property) [45, 59, 46] is a very important fea-
ture that ensures fast and efficient reaction to perturbations. Another very
important global property related to paths is path redundancy, or the avail-
ability of multiple paths between a pair of nodes [71]. Either in the case
of multiple flows from input to output, or contingencies in the case of per-
turbations in the preferred pathway, path redundancy enables the robust
functioning of cellular networks by relying less on individual pathways and
mediators.

5. Graph Models

An often-used means of testing how organizational features such as those de-

scribed in the previous section reflect the networks’ functions is to construct
model graphs based on features or assembly principles deemed most salient.
The predictive power of the model is determined by a graph-theoretical
comparison of the model network to the original. In the following we exam-
ine three general families of network models as well as a family of network
models formulated specifically for intracellular networks.

Erdős–Rényi (ER) random graphs

Mid-twentieth-century work on uniformly random graphs [24] pioneered

many of the basic measures and analytic techniques mentioned in the pre-
vious sections. An Erdős–Rényi (ER) random graph is formed by randomly
connecting N nodes with E edges. For large N , the degree distribution
of such a graph is close to a binomial distribution, implying that most
nodes have degree k, close to the average degree in the graph, k = 2E/N .
Therefore, unlike most cellular networks, ER random graphs are largely
homogeneous in degree. In addition, the average clustering coefficient of
ER random graphs scales inversely with the size of the network, such that
C = k/N , and the clustering coefficient distribution of an ER random
graph (unlike that of most cellular networks) is independent of node degree,
peaking at a value equal to the connection probability p. Finally, the aver-
256 J. Thakar, C. Christensen and R. Albert

age path length of ER random graphs, d ≈ ln (N )/ ln k, remains small

even for large networks [13], and is consistent with the average path length
of several real networks [1].

Scale-free random graphs

Scale-free random graphs adhere to a prescribed degree distribution, though

individual links in the graphs are established randomly [69]. Scale-free
random graphs have smaller average path lengths [19] than comparably-
sized ER random graphs, and they exhibit clustering coefficients similar
to their ER counterparts [68]. Thus, while the clustering coefficient of
biological networks is not captured by scale-free random graphs, scale-free
random graphs are approximately comparable to scrambled, but degree
distribution-preserving versions of real networks. For this reason, they serve
as a better null model of biological networks than do ER random graphs, a
point corroborated by the fact that scale-free random graphs are often used
as baselines from which to establish statistical significance thresholds for
features and properties of biological networks [67]. While scale-free random
graphs are, by definition, able to capture the scale-free degree distribution
found in most real networks, they make no attempt to explain why such
heterogeneity in connectivity exists in real networks. Accounting for this
feature necessitates a shift from modeling network topology to modeling
network assembly.

Evolving network models

A large (and growing) class of network models addresses the question of

how scale-free topologies arise in real networks by describing network as-
sembly and evolution. The simplest of these evolving network models is
the Barabási–Albert (BA) model [8], which introduced two core assump-
tions: growth (i.e. an increase in the number of nodes and edges over time)
and preferential attachment (i.e. a greater likelihood that nodes of high de-
gree will acquire new edges). The BA model assumes linear growth and
proportional preferential attachment, and leads to a power-law degree dis-
tribution P (k) = Ak −3 that captures the upper end of the range of observed
degree-distribution exponents in biological networks. Networks generated
with the BA algorithm have small average clustering coefficients, a con-
Toward Understanding the Structure 257

stant clustering-degree function C(k) [77] and slightly smaller average path
lengths than what is found in comparable random graphs [12], features that
prevent them from capturing all topological characteristics of real networks.
Numerous models, augmenting linear growth and proportional preferential
attachment with features such as nonlinear attachment, initial attractive-
ness of isolated nodes, accelerated growth, aging, fitness and node loss (re-
viewed in [1]) have offered successful solutions for the shortcomings of the
basic BA model. As described in a companion article in this volume, Prof.
Béla Bollobás has developed a general graph theory of evolving networks.
A recent model proposed by Ravasz et al. is based on a self-similar
growth pattern and not on preferential attachment [77]: here, the network
grows by iterative network duplication and subsequent integration of the
duplicated elements to the network’s original core. The net result of the
replication model is a degree-distribution exponent, γ = 1 + loglog(n−1)
(n)
, where
n is the size of the seed graph. For small n, the replication model produces
a degree-distribution exponent very close to 2, comparable to what is seen
in cellular networks. In addition, in contrast to all previous models, the
replication model produces a clustering coefficient that is independent of the
size of the network and that scales inversely with node degree, properties
that also seem to characterize protein interaction and metabolic networks.

Models of cellular network evolution

The topology of cellular networks is shaped by dynamical processes occur-

ring on evolutionary time scales, such as gene duplication (two copies of
the same gene in the DNA) or point mutations (changes in the nucleotide
sequence of the gene). Interestingly, both gene duplications and point mu-
tations – unique biological processes- lead to a preferential increase of the
degree of highly connected proteins, i.e. to preferential attachment [91]. Sev-
eral growing network models based on random gene duplication or mutation
and subsequent functional divergence have displayed good agreement with
the topology of protein interaction networks [73, 11, 10]. Note that these
network models aim to identify the main evolutionary mechanisms shaping
the topology of cellular interaction networks across organisms and not to
predict individual gene duplication events.
258 J. Thakar, C. Christensen and R. Albert

6. Modeling the Dynamics of Cellular Networks

The graphs comprising cellular networks are static representations of dy-

namic processes. Moreover, the nodes of cellular networks often represent
entire populations of proteins or molecules without reflecting the abundance
or activity of these populations. To capture the changes in gene expression
levels or in protein/molecule abundances in response to external or internal
stimuli, the interaction networks must be augmented with node-level infor-
mation characterizing the expression, abundance, and activity – in short,
the state – of each node. A dynamical model of an interacting system
is based on the premise of locality in network space and consists of a set
of equations indicating how the state of each node changes in response to
changes in the state of its regulators (including itself), where the identity
of the regulators is given by the interaction network.
Dynamic models have as input information (i) the interactions and
regulatory relationships among components (i.e. the interaction network),
(ii) the manner in which the strength of the interactions depends on the state
of the interacting components (i.e. the transfer functions, including kinetic
parameters for each interaction) and (iii) the initial state of each component
in the system. Given these inputs, the model will output the time evolution
of the state of the system, for example the system’s response to the presence
or absence of a given signal. Due to the demanding prerequisites of dynamic
modeling, most dynamic networks so far constructed have been quite small
(see, for example, [87, 9, 20]). In this section we briefly outline some of
the most promising techniques so far developed for modeling the dynamics
of cellular networks, and we review the successes and implications of these
models’ results.
Continuous and deterministic models are formulated as differen-
tial equations based on mass-action chemical kinetics for the production
and decay of all components [44]. With sufficiently thorough knowledge of
the biochemical interactions comprising a system – i.e. a compilation of all
pairwise interactions among the system’s components, measurements or es-
timates of kinetic parameters such as dissociation constants and half-lives,
and a known initial state for the system – it is possible to quite accurately
reproduce the dynamic behavior of a complex biological system by describ-
ing the constituent interactions as coupled (usually, nonlinear) differential
equations. For example, using a continuous deterministic model, von Das-
sow et al. reproduced the expression patterns of segment polarity genes
Toward Understanding the Structure 259

in D. melanogaster, and demonstrated that these patterns are remarkably

robust to changes in the kinetic parameters governing the biochemical re-
actions that result in gene expression [20].
Boolean models assume binary states for network nodes and are for-
mulated as logical rules describing the change in state of each regulated
node as a function of its regulators. The utility of Boolean dynamic mod-
els lies in their ability to predict dynamic trends in the absence of detailed
kinetic parameters. For example, Boolean models successfully described
the wild type or mutant expression patterns of segment polarity genes in
D. melanogaster and predicted a significant error correcting ability for the
gene regulatory network [2, 16]. Boolean modeling also reproduced ob-
served gene expression patterns and mutant behavior of the floral organs of
Arabidopsis thaliana [64]. The fact that simple Boolean methods can cap-
ture the same (broad) dynamic trends and predict the same asymptotic,
biologically-viable behaviors as their far more detailed differential equation-
based counterparts often makes them a more feasible first-order approach
for studying a system’s dynamics.
Hybrid dynamic models meld Boolean logical functions with contin-
uous synthesis and decay [15, 32]. In this approach, the rates of change in
concentration of effector molecules are expressed as differential equations,
and combinatorial regulation is described by Boolean functions. This type of
hybrid approach is appealing in that its continuous features can incorporate
a great deal of quantitative detail while its discrete features allow potential
uncertainty in interactions [15]. Being less computationally-intensive than
strict deterministic models, and requiring fewer initial estimates and less
initial input, hybrid approaches can be applied to larger networks than is
possible with pure continuous models.

7. Examples of Network-Based Dynamics Models

Modeling signal transduction in plants

The surface of plant leaves contains numerous microscopic openings (pores)

called stomata. Plants take up carbon dioxide for photosynthesis through
these pores, and also lose water, by evaporation, through them. Plants must
regulate the size of the pores to maximize carbon intake and minimize water
loss, and do so with two “muscle cells” called guard cells. When the guard
260 J. Thakar, C. Christensen and R. Albert

cells are full of water, the pores are larger, and when the guard cells lose
water, the pores decrease in size. Plants sense the environment, for example
the light level, the atmospheric carbon dioxide and humidity, and signal to
the guard cells to open or close the pores. The pores usually open in the light
to participate in photosynthesis. During drought, when the roots do not
have enough moisture, they synthesize a chemical (hormone) called abscisic
acid (ABA). This chemical travels to the leaves and signals to the guard
cells to close the pores regardless of light conditions. This ABA-induced
closure of stomata is a universal process and is a very important component
of drought survival. In a collaborative project with plant biologist Sarah
Assmann and graduate student Song Li, we aimed to model this process
and understand how its sensitivity and reliability is maintained [55].
More than 20 proteins and molecules are implicated in the process of
ABA- induced closure. However, only a few direct interactions among these
proteins and molecules are known. To confirm a component’s role in the
process, biologists knock the component out and compare the response of
these perturbed plants with the response of normal plants. If the perturbed
plants have a weaker response, it is inferred that the knocked-out component
has a positive role in the process. Thus, it was found that the flow of positive
ions out of the guard cells, or the increase of calcium concentration, has a
positive role in the process, while the protein ABI1 has a negative role in
the process. To understand how ABA-induced closure is orchestrated we
need to gather all the clues and determine how they fit together into an
interaction graph.
Constructing the graph: First, we included all proteins and mole-
cules known to be involved in the process as network nodes. We introduced
two kinds of edges for positive and negative regulatory effects. We gath-
ered all the information from the model plant Arabidopsis thaliana, in total
more than 140 interactions from more than 100 published articles. To find
the most parsimonious representation of indirect regulatory effects inferred
from knock-out experiments, we assumed that each indirect causal relation-
ship between two nodes corresponds to a path in the interaction graph. We
also assumed that an indirect effect on a process corresponds to two inter-
secting paths. For example, the inference that ABI1 inhibits ABA-induced
closure corresponds to having a positive path from ABA to closure, and a
negative path from ABI1 to closure, and these two paths intersect at an
unknown intermediary node. Then we aimed to identify these intermediary
nodes by known nodes if additional information was available. For example,
if we knew that the A → B and C → B paths intersect at an intermediary
Toward Understanding the Structure 261

node, but A is also directly promoting C, we identified the intermediary

node with C itself. In summary, we expanded the network with interme-
diary nodes and then we collapsed it again. We did this manually, but a
later collaboration with Bhaskar Dasgupta yielded the rigorous formulation
of the graph algorithms involved, termed binary transitive reduction and
pseudo-vertex collapse, and its approximate solution using combinatorial
optimization [4]. The graph synthesis procedure is now implemented in the
software NET-SYNTHESIS [47].
Graph analysis: The resulting graph (see Figure 4) has ABA as a
source node, and closure as a sink node. The 54 named nodes span enzymes
that catalyze chemical reactions (denoted by red nodes in Figure 4) signal
transduction proteins that regulate other proteins (green nodes), chemicals
such as calcium and pH level (orange nodes), and transport related nodes
such as ion pumps and channels (blue nodes). The 43 unnamed black dots
represent unidentified intermediary nodes. The node in-degrees range be-
tween 0 (for ABA and four other source nodes) and 6 (for Closure). The
out-degrees range between 0 (for Closure and ROP10) and 8 (for Ca2+ c ).
The largest strongly connected component includes 18 nodes that are mostly
related to (the regulation of) ion flows. There is a considerable path redun-
dancy in the strongly connected component. For example, calcium increase
can be due to release of calcium from cellular stores, or to flow of calcium
from the outside of the cell. The in-component of the strongly connected
component contains most of the remaining nodes, while its out-component
contains three nodes: Closure, Actin and Malate. Indeed, ion flow causes
the flow of water out of the cell, the reorganization of the actin cytoskeleton
and stomatal closure.
The graph contains edges with negative sign. These edges tend to be
distributed in such a way that most of the paths they participate in are pos-
itive. Biologically, these paths correspond to inhibiting an inhibitor. There
are at least four node-independent paths from ABA to closure, and indeed
at least four nodes, corresponding to actin reorganization, cytosolic pH in-
crease, malate breakdown and membrane depolarization, respectively, need
to be simultaneously disrupted to block all ABA- closure paths. A relatively
small fraction of node disruptions (e.g. 10% of four-node disruptions) cause
an increase in the ABA- Closure path length.
Dynamic modeling of the process: The graph analysis of the ABA
signal transduction network indicates a significant degree of path redun-
dancy. But these paths could be dynamically dependent on each other, for
example by a conditional dependence between regulatory processes converg-
262 J. Thakar, C. Christensen and R. Albert

Fig. 4. The network for induction of stomatal closure by the plant hormone ABA, based
on data available as of 2006.
The nodes in this graph include: the hormone ABA as the input (source) node; enzymes
(red) such as Guanyl cyclase (GC) and Sphingosine kinase (SphK); signal transduction
proteins (green) such as the heterotrimeric G protein α. (GPA1) and β. component
(AGB1) and Protein kinase (OST1); membrane transport related nodes (blue) such as
anion efflux at the plasma membrane (AnionEM), Ca2+ influx to the cytosol from
intracellular stores (CIS), K+ efflux through slowly activating outwardly-rectifying K+
channels at the plasma membrane (KOUT); secondary messengers and small molecules
(orange) such as cytosolic Ca2+ increase (Ca2+i ), phosphatidic acid (PA) and Malate;
Closure as the output (sink) node. Small black filled circles represent putative
intermediary nodes mediating indirect regulatory interactions. Arrowheads represent
activation, short perpendicular bars indicate inhibition. Purple lines denote interactions
derived from species other than Arabidopsis or inferences made during the network
synthesis process. Nodes involved in the same metabolic pathway or protein complex are
bordered by a grey box; only those arrows that point into or out of the box signify
information flow. Figure reproduced from [55].

ing on the same node. Thus we set out to model the dynamics of the process.
Since there is very little kinetic information available on the regulatory in-
teractions among the elements of the ABA signal transduction network,
we could not describe the process in full quantitative detail. Instead we
Toward Understanding the Structure 263

chose a Boolean framework, because it can incorporate conditional activa-

tion (the required synergy between two nodes) through the AND operator,
and the effects of inhibitors through the NOT operator. We constructed the
Boolean transfer functions for each node based on our best interpretation
of the available literature. To reflect the lack of knowledge regarding the
duration of the interactions we implemented an asynchronous state update
procedure where the nodes’ states are updated one by one in a randomly
selected order which changes in each round of update. To reflect the lack
of information regarding the resting levels of the elements of the network
before receiving the ABA signal, we randomly sampled initial conditions
among those biologically feasible. In effect, we constructed a population of
simulated stomata and followed what fraction of the population is closed
at any given time (represented by the node closure being in the state 1).
We did these simulations for normal guard cells and for perturbed (knock-
out) guard cells where up to three nodes are maintained in the off state.
We performed a detailed analysis of the resulting behaviors and classified
perturbations into five classes with respect to their severity.

The most effective perturbations cause complete insensitivity to ABA

(none of the simulated stomata close). Three single perturbations, namely
the knockout of the nodes corresponding to anion flow, membrane depo-
larization, and actin reorganization, are in this category, and 18% (25%)
of double(triple) perturbations. A less severe effect is reduced sensitivity,
meaning that the fraction of closed stomata settles at a value that is less
than 100% but more than zero. Five nodes (17.5%) are in this category, and
27% (31%) of double(triple) node combinations. For the remaining three
classes all simulated stomata ultimately reach closure, but with identifiably
different timescales. 65% of single node disruptions have a trajectory indis-
tinguishably close to the normal (wild type) response. The frequency of this
behavior decreases for multiple disruptions, to 38% (23%) for double(triple)
knockouts. Two disruptions, namely the knockout of the inhibitors ABI1
and Ca2+ ATPase, lead to a hypersensitive response, or an ABA-induced
closure that is faster than wild type. The percentage of hypersensitive re-
sponses stays at similar levels (6%) for multiple perturbations. Two disrup-
tions, namely the blocking of the increase of the cytosolic Ca2+ levels and
of reactive oxygen species production display hyposensitivity, or a response
slower than wild type. The frequency of this response increases to 12%
(13%) in double(triple) perturbations. In conclusion, the existence of con-
ditional dependence between processes does make the network less robust
than the simplest graph representation and analysis suggests.
264 J. Thakar, C. Christensen and R. Albert

One of the important predictions of our model is that calcium disruption

has less severe effects than pH disruption. Specifically, according to the
model when Ca2+ signaling is perturbed the simulated stomata still close
but more slowly than normal. However when pH increase is perturbed only
a fraction of the simulated stomata is closed at any given time. This may
have been a controversial result because it is has previously been thought
that Ca2+ is a required master regulator, and there has been relatively little
experimental work focused on pH mediated processes. Thus we decided to
test this prediction experimentally. Song Li treated plant leaf peels with the
compound BAPTA to disable the increase in cytosolic Ca2+ or with the acid
butyrate to clamp the cytosolic pH levels, and compared the treated peels’
response to ABA with normal peels’ response. He found that the BAPTA
treated leaves had an ABA response closer to normal than the butyrate
treated leaves (for details see [55]). Thus the experiments confirmed that
Ca2+ disruption has a less severe effect than pH disruption.

Modeling the systems-level regulation of immune responses to

Bordetella

The immune response to pathogens such as viruses or bacteria is based on

a few very important cell types. Cells such as neutrophils and macrophages
phagocytose (ingest) bacteria. Other cells work to produce molecules that
alert neutrophils to the presence of bacteria and prime them to be more
effective. For example, B cells produce antibodies that bind to bacteria
and make them easily recognizable by phagocytic cells. Several cell types
produce chemicals such as cytokines that attract neutrophils to the site of
infection. There are two main branches of the immune system: innate im-
munity is an immediate and general response to invaders, while adaptive
immunity is a pathogen-specific response which also includes recognition of
previously encountered pathogens. Even from this brief description it is
clear that the immune response is a network whose nodes represent diverse
entities from molecules to cells and whose edges represent interactions, sig-
nal transduction pathways and regulatory processes. In addition, pathogens
can often manipulate cells or signaling processes by an arsenal of so-called
virulence factors, aiming to divert the immune responses targeted against
them. In a collaborative project with Prof. Eric Harvill’s group, we con-
structed a dynamic network model of a respiratory infection by two related
bacteria in the Bordetella family [86].
Toward Understanding the Structure 265

Constructing the graph: First we constructed a consensus graph of

mammalian immune responses to generic respiratory pathogens by curat-
ing the general body of knowledge on mammalian immune responses. The
vertices of the graph represent immune cells or chemicals. As the biological
outcome of the infection is clearance of bacteria after sustained phagocy-
tosis, or conversely the persistence of bacteria if phagocytosis is weak, we
also included a separate vertex for phagocytosis. The directed edges are
processes such as signal transduction pathways, coating of bacteria by anti-
gen, or B cell stimulation, and they have positive (activating) or negative
(inhibitory) signs. We incorporated regulatory relationships that modulate
a process (or an unspecified mediator of a process) as edges directed toward
another edge.

Then we augmented this consensus graph with bacterial virulence factors

and effector mechanisms of two specific bacteria, to obtain two bacterium-
specific graphs. The first pathogen, Bordetella bronchiseptica, has an O
antigen on its surface that inhibits the binding of immune materials to
it. It also has a syringe-like protrusion called the type three secretion
system through which it injects chemicals into immune cells, interfering with
their signaling and ultimately killing many of them. The second pathogen,
Bordetella pertussis, produces the pertussis toxin that inhbits neutrophils’
recruitment to the site of infection. We constructed the two bacterium-
specific graphs based on the Bordetella literature and on the research of the
Harvill group.

Graph analysis: The signal to the immune response is the presence of

bacteria, however this signal does not correspond to a source node. Instead,
the node Bacteria is in an indirect negative feedback loop with the node
Phagocytosis. The consensus immune response network (Figure 5) contains
18 nodes and is strongly connected. The biological reasons of this strong
connectivity are two-fold. First, the previously mentioned negative feedback
between bacteria and phagocytosis. Second, there is a positive feedback
loop between innate and adaptive immune responses in which the adaptive
immune responses offer additional ways to activate phagocytic cells and lead
to ultimate clearance of bacteria. Another interesting graph feature is the
presence of a double negative feedback loop regulating the differentiation
of T cells, in such a way that only one type of T cell is predominant
in a given condition. Each of the two bacterium-specific graphs contains
two additional source nodes corresponding to bacterial virulence factors.
The Bordetella bronchiseptica graph also has a 19th node of the strongly
266 J. Thakar, C. Christensen and R. Albert

Fig. 5. The consensus network of immunological steps and processes activated upon
invasion by Bordetellae species. Network nodes denote components of the immune
system and edges represent interactions and processes. Edge labels give a brief biological
description of the underlying process. The edges are classified into two regulatory
effects, activation and inhibition, and are represented by incoming black arrows and
incoming grey blunt segments, respectively. Figure reproduced from [86].

connected component, corresponding to the phagocytic cells killed by the

bacterial type three secretion system.
Dynamic model: The experimental information on the Bordetellae
(and for infections in general) is generally restricted to measurements of
bacterial numbers at several instances (e.g. at 1, 3, 7, 14, 21, 28 days af-
ter infection). As there is no kinetic information regarding single edges,
we again chose a Boolean representation with two discrete states for each
node and an asynchronous update of the node states. This model augments
the classical parameter-less Boolean framework by two quantitative mea-
sures: finite decay times for cytokines and bacterial virulence factors, and a
Toward Understanding the Structure 267

threshold duration for effective phagocytosis. The decay times incorporate

the observation that the effect of cytokines and of bacterial virulence fac-
tors decreases after an initial peak, even when the conditions that caused
their activation persist. The threshold phagocytosis duration incorporates
the fact that the phagocytic activity during the first (innate) stages of the
infection is not sufficient for clearing the bacteria. We expressed these pa-
rameters as multiples of the Boolean model’s timesteps, which respectively
corresponds to one to two days. We verified that that a relatively large
range of each parameter gives identical results.
Our model reflects the characteristic timecourse of a bacterial infection
as activation patterns for the network nodes. Each pattern is in good agree-
ment with known infection features, such as the sequential activation of the
two T cell types. It also captures experimental observations regarding the
effect of giving antibodies at the time of infection. Indeed, antibody treat-
ment greatly reduces the duration of a Bordetella bronchiseptica infection,
both in our model, and in experiments. However, antibodies are unable to
shorten a Bordetella pertussis infection, neither in experiments nor in our
model. The model predicts that a second infection of hosts that are recov-
ering from an infection will be cleared much more easily than the primary
disease, and this conclusion holds for both pathogens. Our experimental col-
laborators tested this prediction and they confirmed the shorter duration of
the second infection.

8. Discussion

Our understanding of cellular processes at the systems level grows as the

result of an ongoing dialogue and feedback among experimental, compu-
tational and theoretical approaches (Figure 6). High-throughput experi-
ments reveal, or allow the inference of, the edges of genome-wide cellular
interaction networks. Graph-theoretical analysis of these networks enables
general insight into the topological and functional organization of cellular
regulation. Comparative network analysis feeds back to network inference
[36, 42, 3, 18], also expanding the tools of graph theory to incorporate the
diversity of molecular interactions.
Most inference methods operate on the assumption that a system can be
reduced to interactions taking place on a single temporal scale. However, as
experimental data continues to amass, it is becoming increasingly evident
268 J. Thakar, C. Christensen and R. Albert

Fig. 6. An overview of applications of graph theory in studying cellular networks.

Inference methods used in conjunction with experimental data allow an optimal
construction of the network. Graph theoretical analysis of interaction networks allows
insight into their functionality, and also feeds back to fine-tune network inference
methods. Integration of network topologies with temporal information can be used to
develop predictive dynamic models which can then be tested by experimentation.

both that biochemical reactions occurring within or between cells take place
on timescales that can differ from one another by many orders of magnitude
[72], and that these timescales can be dictated by any number of factors,
including the spatial scale on which interactions take place, the types of
biomolecules or complexes that are interacting, and the environmental con-
ditions to which the system is subjected [7, 38]. Dynamic models integrate
interaction (topological), state (expression or activity) and temporal infor-
mation regarding a system into a description of its predicted dynamical tra-
jectory in normal and perturbed conditions. Experimental testing of these
predictions validates the models’ initial hypotheses or generates alternate
hypotheses, in either case leading to new biological discoveries.

While genome-level interaction maps help us in understanding regula-

tory design features and evolutionary rules, dynamic modeling of systems
with a less than genome-wide scope and specified inputs and outputs allows
the identification of key regulatory components or parameters (as opposed
to general trends). While the use of extensive dynamic modeling is limited
Toward Understanding the Structure 269

by computational complexity and the availability of detailed transfer func-

tions and kinetic parameters, the examples we have shown illustrate how
including the full complexity of the interaction network allows us simplicity
in describing the individual interactions. The success of discrete dynamic
models suggests that the interaction/regulatory graph underlying a biologi-
cal system has a more important role in determining the dynamic behavior of
the system than the kinetic details of the individual interactions. It has also
been suggested that it is possible to glean a pseudodynamical view of cellu-
lar processes from their graph representations by mapping the propagation
of a chemical signal through the graph from an extracellular signal source
to sinks within the cell [59]. Similarly augmenting the currently available
directionless interactome networks with information regarding the sources
(signals) and outputs of the network and the cause-and-effect (directional)
relationships along the edges will significantly enhance their functional in-
formation content.
As experimental, theoretical and computational techniques continue to
improve, adding to the scope and detail captured by network models, fu-
ture generations of (pseudo)dynamic network models promise to greatly
augment our current understanding of biological systems, perhaps causing
us to modify our initial assumptions about cellular network structure and
function. As the applicability and effectiveness of network-based modeling
is increasingly recognized, graph theory will play a leading role in an emerg-
ing framework for understanding complex biological systems. We hope that
this chapter will contribute to establishing a dialog between graph theorists
and systems biologists.

References

[1] R. Albert and A. L. Barabási, Statistical mechanics of complex networks, Reviews

of Modern Physics., 74 (2002), 47–97.
[2] R. Albert and H. G. Othmer, The topology of the regulatory interactions predicts
the expression pattern of the segment polarity genes in Drosophila melanogaster,
Journal of Theoretical Biology, 223 (2003), 1–18.
[3] I. Albert and R. Albert, Conserved network motifs allow protein-protein interaction
prediction, Bioinformatics, 20 (2004), 3346–3352.
[4] R. Albert, B. DasGupta, R. Dondi, et al., A novel method for signal transduction
network inference from indirect experimental evidence, J. Comput. Biol., 14 (2007),
927–949.
270 J. Thakar, C. Christensen and R. Albert

[5] B. Alberts, Molecular biology of the cell, Garland Pub (New York, 1994).
[6] G. D. Bader, I. Donaldson, C. Wolting, et al., Bind–the biomolecular interaction
network database, Nucleic Acids Res., 29 (2001), 242–245.
[7] G. Balázsi, A. L. Barabási and Z. N. Oltvai, Topological units of environmental
signal processing in the transcriptional regulatory network of escherichia coli, Proc.
Natl. Acad. Sci. U.S.A., 102 (2005), 7841–7846.
[8] A. L. Barabási and R. Albert, Emergence of scaling in random networks, Science,
286 (1999), 509–512.
[9] N. Barkai and S. Leibler Robustness in simple biochemical networks, Nature, 387
(1997), 913–917.
[10] J. Berg, M. Lassig and A. Wagner, Structure and evolution of protein interaction
networks: A statistical model for link dynamics and gene duplications, BMC Evol.
Biol., 4 (2004), 51.
[11] H. Bersini, T. Lenaerts and F. C. Santos, Growing biological networks: Beyond the
gene-duplication model, J. Theor. Biol., 241 (2006), 488–505.
[12] B. Bollobás and O. M. Riordan, Mathematical results on scale-free random graphs,
in: S. Borhholdt, H. G. Schuster, editors, Handbook of graphs and networks, Wiley
(Weinheim, 2003), pp. 1–32.
[13] B. Bollobás, Graph theory: An introductory course, Springer-Verlag (New York,
1979).
[14] H. Caron, B. van Schaik, M. van der Mee, et al., The human transcriptome map:
Clustering of highly expressed genes in chromosomal domains, Science, 291 (2001),
1289–1292.
[15] M. Chaves, E. D. Sontag and R. Albert, Methods of robustness analysis for Boolean
models of gene control networks, Syst. Biol. (Stevenage), 153 (2006), 154–167.
[16] M. Chaves, R. Albert and E. D. Sontag, Robustness and fragility of Boolean models
for genetic regulatory networks, J. Theor. Biol., 235 (2005), 431–449.
[17] T. Chen, H. L. He and G. M. Church, Modeling gene expression with differential
equations, Pac. Symp. Biocomput. (1999), 29–40.
[18] C. Christensen, A. Gupta, C. D. Maranas, et al., Inference and graph-theoretical
analysis of bacillus subtilis gene regulatory networks, Physica A, 373 (2007), 796–
810.
[19] R. Cohen, S. Havlin and D. ben-Avraham, Structural properties of scale free net-
works, in: S. Borhholdt, H. G. Schuster, editors, Handbook of graphs and networks,
Wiley (Weinheim, 2003), pp. 85–110.
[20] G. von Dassow, E. Meir, E. M. Munro, et al., The segment polarity network is a
robust developmental module, Nature, 406 (2000), 188–192.
[21] E. W. Dijkstra, A note on two problems in connection with graphs, Numerische
Math., 1 (1959), 269–271.
[22] E. R. Dougherty, J. Barrera, M. Brun, et al., Inference from clustering with appli-
cation to gene-expression microarrays, J. Comput. Biol., 9 (2002), 105–126.
Toward Understanding the Structure 271

[23] A. Drawid and M. Gerstein, A bayesian system integrating expression data with
sequence patterns for localizing proteins: Comprehensive application to the yeast
genome, J. Mol. Biol., 301 (2000), 1059–1075.
[24] P. Erdős and A. Rényi, On the evolution of random graphs, Bull. Inst. Int. Stat.,
38 (1961), 343.
[25] C. L. Freeman, A set of measures of centrality based on betweenness, Sociometry,
40 (1977), 35.
[26] N. Friedman, M. Linial, I. Nachman, et al., Using bayesian networks to analyze
expression data, J. Comput. Biol., 7 (2000), 601–620.
[27] P. S. Gargalovic, M. Imura, B. Zhang, et al., Identification of inflammatory gene
modules based on variations of human endothelial cell responses to oxidized lipids,
Proc. Natl. Acad. Sci. U.S.A., 103 (2006), 12741–12746.
[28] J. I. Garrels, Ypd-a database for the proteins of Saccharomyces cerevisiae, Nucleic
Acids Res., 24 (1996), 46–49.
[29] G. Giaever, A. M. Chu, L. Ni, et al., Functional profiling of the Saccharomyces
cerevisiae genome, Nature, 418 (2002), 387–391.
[30] L. Giot, J. S. Bader, C. Brouwer, et al., A protein interaction map of Drosophila
melanogaster, Science, 302 (2003), 1727–1736.
[31] M. Girvan and M. E. J. Newman, Community structure in social and biological
networks, Proc. Natl. Acad. Sci. (USA), 99 (2002), 7821.
[32] L. Glass and S. A. Kauffman, The logical analysis of continuous, non-linear bio-
chemical control networks, J. Theor. Biol., 39 (1973), 103–129.
[33] K. I. Goh, B. Kahng and D. Kim, Universal behavior of load distribution in scale-
free networks, Phys. Rev. Lett., 87 (2001), 278701.
[34] N. Guelzim, S. Bottani, P. Bourgine, et al., Topological and causal structure of the
yeast transcriptional regulatory network, Nature Genetics, 31 (2002), 60–63.
[35] R. Guimera and L. A. Nunes Amaral, Functional cartography of complex metabolic
networks, Nature, 433 (2005), 895–900.
[36] A. Gupta, C. D. Maranas and R. Albert, Elucidation of directionality for co-
expressed genes: Predicting intra-operon termination sites, Bioinformatics, 22
(2006), 209–214.
[37] A. Gupta, J. D. Varner and C. D. Maranas, Large-sale inference of the transcrip-
tional regulation of bacillus subtilis, Computers and Chemical Engineering, 29
(2005), 565–576.
[38] J. D. Han, N. Bertin, T. Hao, et al., Evidence for dynamically organized modularity
in the yeast protein-protein interaction network, Nature, 430 (2004), 88–93.
[39] L. H. Hartwell, J. J. Hopfield, S. Leibler, et al., From molecular to modular cell
biology, Nature, 402 (1999), C47–52.
[40] J. Hasty, D. McMillen, F. Isaacs, et al., Computational studies of gene regulatory
networks: In numero molecular biology, Nat. Rev. Genet., 2 (2001), 268–279.
272 J. Thakar, C. Christensen and R. Albert

[41] P. Holme, M. Huss and H. Jeong, Subnetwork hierarchies of biochemical pathways,

Bioinformatics, 19 (2003), 532–538.
[42] S. Horvath, B. Zhang, M. Carlson, et al., Analysis of oncogenic signaling networks
in glioblastoma identifies aspm as a molecular target, Proc. Natl. Acad. Sci. U.S.A.,
103 (2006): 17402–17407.
[43] A. M. Huerta, H. Salgado, D. Thieffry, et al., Regulondb: A database on transcrip-
tional regulation in escherichia coli, Nucleic Acids Res., 26 (1998), 55–59.
[44] D. H. Irvine and M. A. Savageau, Efficient solution of nonlinear ode’s expressed in
S-system canonical form, SIAM Journal of Numerical Analysis, 27 (1990), 704–
735.
[45] H. Jeong, B. Tombor, R. Albert, et al., The large-scale organization of metabolic
networks, Nature, 407 (2000), 651–654.
[46] H. Jeong, S. P. Mason, A. L. Barabási, et al., Lethality and centrality in protein
networks, Nature, 411 (2001), 41–42.
[47] S. Kachalo, R. Zhang, E. Sontag, et al., Net-synthesis: A software for synthesis,
inference and simplification of signal transduction networks, Bioinformatics, 24
(2008), 293–295.
[48] S. Kalir, S. Mangan and U. Alon, A coherent feed-forward loop with a sum input
function prolongs flagella expression in escherichia coli, Mol. Syst. Biol., 1 (2005),
2005 0006.
[49] M. Kanehisa and S. Goto, Kegg: Kyoto encyclopedia of genes and genomes, Nucleic
Acids Res., 28 (2000), 27–30.
[50] P. D. Karp, M. Riley, M. Saier, et al., The ecocyc and metacyc databases, Nucleic
Acids Res., 28 (2000), 56–59.
[51] V. Latora and M. Marchiori, Efficient behavior of small-world networks, Physical
Review Letters, 87 (2001), 198701–198704.
[52] T. I. Lee, N. J. Rinaldi, F. Robert, et al., Transcriptional regulatory networks in
Saccharomyces cerevisiae, Science, 298 (2002), 799–804.
[53] N. Lemke, F. Heredia, C. K. Barcellos, et al., Essentiality and damage in metabolic
networks, Bioinformatics, 20 (2004), 115–119.
[54] S. Li, C. M. Armstrong, N. Bertin, et al., A map of the interactome network of the
metazoan C. elegans, Science, 303 (2004), 540–543.
[55] S. Li, S. M. Assmann and R. Albert, Predicting essential components of signal
transduction networks: A dynamic model of guard cell abscisic acid signaling, PLoS
Biol., 4 (2006), e312.
[56] N. M. Luscombe, M. M. Babu, H. Y. Yu, et al., Genomic analysis of regulatory
network dynamics reveals large topological changes, Nature, 431 (2004), 308–312.
[57] H. W. Ma and A. P. Zeng, The connectivity structure, giant strong component and
centrality of metabolic networks, Bioinformatics, 19 (2003), 1423–1430.
[58] A. Ma’ayan, R. D. Blitzer and R. Iyengar, Toward predictive models of mammalian
cells, Annu. Rev. Biophys. Biomol. Struct. (2004), 319–349.
Toward Understanding the Structure 273

[59] A. Ma’ayan, S. L. Jenkins, S. Neves, et al., Formation of regulatory patterns during

signal propagation in a mammalian cellular network, Science, 309 (2005), 1078–
1083.
[60] R. Mahadevan and B. O. Palsson, Properties of metabolic networks: Structure
versus function, Biophys. J., 88 (2005), L07–09.
[61] S. Mangan, A. Zaslaver and U. Alon, The coherent feedforward loop serves as a
sign-sensitive delay element in transcription networks, J. Mol. Biol., 334 (2003),
197–204.
[62] E. M. Marcotte, I. Xenarios and D. Eisenberg, Mining literature for protein-protein
interactions, Bioinformatics, 17 (2001), 359–363.
[63] S. Maslov and K. Sneppen, Specificity and stability in topology of protein networks,
Science, 296 (2002), 910–913.
[64] L. Mendoza and E. R. Alvarez-Buylla, Dynamics of the genetic regulatory network
for Arabidopsis thaliana flower morphogenesis, J. Theor. Biol., 193 (1998), 307–
319.
[65] C. von Mering, L. J. Jensen, M. Kuhn, et al., String 7-recent developments in the
integration and prediction of protein interactions, Nucleic Acids. Res. (2006).
[66] H. W. Mewes, C. Amid, R. Arnold, et al., Mips: Analysis and annotation of proteins
from whole genomes, Nucleic Acids Res., 32 (2004), D41–44.
[67] R. Milo, S. Shen-Orr, S. Itzkovitz, et al., Network motifs: Simple building blocks
of complex networks, Science, 298 (2002), 824–827.
[68] M. E. J. Newman, Random graphs as models of networks, in: S. Bornholdt, H. G.
Schuster, editor, Handbook of graphs and networks, Wiley (Weinheim, 2003), pp.
35–65.
[69] M. E. J. Newman, S. H. Strogatz and D. J. Watts, Random graphs with arbitrary
degree distributions and their applications, Physical Review E., 64 (2001), 026118.
[70] A. Pandey and M. Mann, Proteomics to study genes and genomes, Nature, 405
(2000), 837–846.
[71] J. A. Papin and B. O. Palsson, Topological analysis of mass-balanced signaling
networks: A framework to obtain network properties including crosstalk, J. Theor.
Biol., 227 (2004), 283–297.
[72] J. A. Papin, T. Hunter, B. O. Palsson, et al., Reconstruction of cellular signalling
networks and analysis of their properties, Nat. Rev. Mol. Cell. Biol., 6 (2005),
99–111.
[73] R. Pastor-Satorras, E. Smith and R. V. Sole, Evolving protein interaction networks
through gene duplication, Journal of Theoretical Biology, 222 (2003), 199–210.
[74] S. Peri, J. D. Navarro, T. Z. Kristiansen, et al., Human protein reference database
as a discovery resource for proteomics, Nucleic Acids Res., 32 (2004), D497–501.
[75] J. Qian, M. Dolled-Filhart, J. Lin, et al., Beyond synexpression relationships:
Local clustering of time-shifted and inverted gene expression profiles identifies new,
biologically relevant interactions, J. Mol Biol., 314 (2001), 1053–1066.
274 J. Thakar, C. Christensen and R. Albert

[76] J. C. Rain, L. Selig, H. De Reuse, et al., The protein-protein interaction map of

helicobacter pylori, Nature, 409 (2001), 211–215.
[77] E. Ravasz, A. L. Somera, D. A. Mongru, et al., Hierarchical organization of mod-
ularity in metabolic networks, Science, 297 (2002), 1551–1555.
[78] A. W. Rives and T. Galitski, Modular organization of cellular networks, Proceedings
of the National Academy of Sciences of the United States of America, 100 (2003),
1128–1133.
[79] J. F. Rual, K. Venkatesan, T. Hao, et al., Towards a proteome-scale map of the
human protein-protein interaction network, Nature, 437 (2005), 1173–1178.
[80] M. R. Said, T. J. Begley, A. V. Oppenheim, et al., Global network analysis of
phenotypic effects: Protein networks and toxicity modulation in Saccharomyces
cerevisiae, Proc. Natl. Acad. Sci. U.S.A., 101 (2004), 18006–18011.
[81] S. S. Shen-Orr, R. Milo, S. Mangan, et al., Network motifs in the transcriptional
regulation network of escherichia coli, Nat. Genet., 31 (2002), 64–68.
[82] I. Shmulevich, E. R. Dougherty, S. Kim, et al., Probabilistic Boolean networks:
A rule-based uncertainty model for gene regulatory networks, Bioinformatics, 18
(2002), 261–274.
[83] V. Spirin and L. A. Mirny, Protein complexes and functional modules in molecular
networks, Proc. Natl. Acad. Sci. U.S.A., 100 (2003), 12123–12128.
[84] B. J. Stapley and G. Benoit, Biobibliometrics: Information retrieval and visualiza-
tion from co-occurrences of gene names in medline abstracts, Pac. Symp. Biocomput
(2000), 529–540.
[85] R. Tanaka, Scale-rich metabolic networks, Phys. Rev. Lett., 94 (2005), 168101.
[86] J. Thakar, M. Pilione, G. Kirimanjeswara, et al., Modeling systems-level regulation
of host immune responses, PLoS. Comput. Biol., 3 (2007), e109.
[87] J. J. Tyson, K. C. Chen, B. Novak, Sniffers, buzzers, toggles and blinkers: Dynamics
of regulatory and signaling pathways in the cell, Curr. Opin. Cell. Biol., 15 (2003),
221–231.
[88] P. Uetz, L. Giot, G. Cagney, et al., A comprehensive analysis of protein-protein
interactions in Saccharomyces cerevisiae, Nature, 403 (2000), 623–627.
[89] B. Vogelstein, D. Lane and A. J. Levine, Surfing the p53 network, Nature, 408
(2000), 307–310.
[90] A. Wagner and D. A. Fell, The small world inside large metabolic networks, Pro-
ceedings of the Royal Society of London Series B-Biological Sciences, 268 (2001),
1803–1810.
[91] A. Wagner, How the global structure of protein interaction networks evolves, Proc.
Biol. Sci., 270 (2003), 457–466.
[92] A. Wagner, Reconstructing pathways in large genetic networks from genetic per-
turbations, J. Comput. Biol., 11 (2004), 53–60.
[93] D. J. Watts and S. H. Strogatz, Collective dynamics of ‘small-world’ networks,
Nature, 393 (1998), 440–442.
Toward Understanding the Structure 275

[94] E. Wingender, P. Dietze, H. Karas, et al., Transfac: A database on transcription

factors and their DNA binding sites, Nucleic Acids Res, 24 (1996), 238–241.
[95] S. Wuchty, Z. N. Oltvai and A. L. Barabási, Evolutionary conservation of motif
constituents in the yeast protein interaction network, Nat. Genet., 35 (2003), 176–
179.
[96] I. Xenarios, L. Salwinski, X. J. Duan, et al., Dip, the database of interacting
proteins: A research tool for studying cellular networks of protein interactions,
Nucleic Acids Res., 30 (2002), 303–305.
[97] S. H. Yook, Z. N. Oltvai and A. L. Barabási, Functional and topological character-
ization of protein interaction networks, Proteomics, 4 (2004), 928–942.
[98] J. Yu, V. A. Smith, P. P. Wang, et al., Advances to bayesian network inference for
generating causal networks from observational biological data, Bioinformatics, 20
(2004), 3594–3603.
[99] L. V. Zhang, O. D. King, S. L. Wong, et al., Motifs, themes and thematic maps of
an integrated Saccharomyces cerevisiae interaction network, J. Biol., 4 (2005), 6.
[100] W. Zhao, E. Serpedin and E. R. Dougherty, Inferring gene regulatory networks from
time series data using the minimum description length principle, Bioinformatics
(2006).

Juilee Thakar Claire Christensen

Department of Physics School of Information,
The Pennsylvania State University University of Michigan,
University Park Ann Arbor,
PA 16802 MI 48109

Réka Albert
Department of Physics
Huck Institutes of the Life Sciences
The Pennsylvania State University
University Park
PA 16802
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 277–324.

Chapter 7
Scale-Free Cortical Planar Networks

WALTER J. FREEMAN and ROBERT KOZMA

with an appendix by
BÉLA BOLLOBÁS∗ and OLIVER RIORDAN

Modeling brain dynamics requires us to define the behavioral context in which

brains interact with the world; to choose appropriate mathematics, here ordinary
differential equations (ODE) and random graph theory (RGT); to choose the
levels of description and scales in the hierarchy of neurodynamics; to define
an appropriate module for each level; and to address questions of boundary
conditions, linearity, time-variance, autonomy, and criticality.
ODE applied to the olfactory system serves to model perception by a phase
transition that reorganizes the background activity. Feedback control theory
is used to model the dynamics of self-organized criticality and simulate the
background activity and its reorganization, by which microscopic input triggers
the construction of an order parameter that retrieves a mesoscopic spatiotemporal
pattern expressing the class of input. Perception is shown to depend on the
coincidence of three conditions: intentional prediction of a sensory input by an
attractor landscape; emergence of a null spike in the background activity; and
the presence in the sensory input of the expected stimulus.
RGT serves to model criticality and the phase transition and the basic oper-
ations of perception in three-layered allocortex. Modeling six-layered neocortex
faces the major problem of reconciling the global formation of very large-scale
activity patterns in cognition that appear to be scale-free with the high degree
of specificity in fine structure of neural activity relating to sensation and action.
The developmental anatomy of neocortex provides essential guidelines on how
to construct random graphs that can model both the large-scale and small-scale
operations by which high-level cognition engages with the world, and their in-
teractions across hierarchical levels. It is likely that use of RGT will succeed
in describing these complex neural mechanisms, which cannot be expected for
ODE alone, provided that close collaboration is maintained by mathematicians,
neuroanatomists, and neurophysiologists.
∗
Research supported in part by NSF grants DMS-0906634, CNS-0721983 and CCF-
0728928, and ARO grant W911NF-06-1-0076.
278 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

1. Introduction

1.1. The context of quantitative brain dynamics from psychology,

biology and philosophy

Animals and humans use their finite brains to comprehend and adapt to
infinitely complex environments. The dynamics employed by brains con-
forms to the scientific method of hypothesis formation and experimental
testing. As mathematicians we seek to devise formal systems by which to
describe the neural processes of construction and evaluation of hypotheses
in brain dynamics by reformulating the existing hypotheses that have been
developed by psychologists, neurobiologists, and philosophers to describe
the dynamics underlying behavior.
Experimental psychologists describe the process in terms of trial-and-
error behaviors guided by reinforcement [41] and by temporal difference
learning [106]. Trial behaviors constitute hypotheses that, if successful, are
rewarded by release of chemicals such as dopamine and CCK into the brain.
The power of chemical reinforcement is readily apparent in the world-wide
epidemic of cocaine usage, which triggers the reward system. Intermittent
reinforcement establishes persistent behaviors that can be extremely resis-
tant to extinction, when the rules change and reward is no longer forthcom-
ing. Scientists and rats that have been conditioned to expect failure may
persevere indefinitely in unsuccessful behaviors. This criterion shows the
greater dependence of choice of actions on pre-existing brain states of ex-
pectancy than on environmental conditions. In this review we use random
graph theory (RGT) to describe the formation and exercise of such states
of focused attention in brain networks.
Neurobiologists describe the process in terms of the action-perception
cycle [48]. Brains construct predictions of future desired goal states. These
expectancies emerge in the form of spatiotemporal patterns of brain activity
having two aspects. One aspect is motor: the formation of strategic plans
of action in the motor systems [58], [63], in which are embedded a variety
of directions for tactical moves that may be imposed by environmental and
bodily constraints and contingencies. The other aspect is preafference [70]:
the evocation in each of the sensory cortices of attractor landscapes, which
embody the predictions of the several forms that sensory input is expected
to take as the result of successful action. For example, an animal searching
for food can expect to find one of two or more possible types of food, or
Scale-Free Cortical Planar Networks 279

none, or a novel odor that must be explored, or an odor of predator, each

with its attractor in the landscape, only one of which will be selected by
the stimulus. Brains then adapt to the results of the action-based test by
modification of their synaptic networks through learning by association and
habituation. RGT is exceptionally well adapted to modeling these changes
in synaptic connectivity.
Philosophers describe this process as the exercise of intentionality [81].
Goal- directed action is preceded by intent to act that thrusts the body
forth into the world. The senses report the consequences to the brain,
which, from the distortion of the self resulting from action, reorganizes and
reshapes itself and the body so as to accommodate to the world, thereby
coming to know the world by assimilation. This view is predicated on two
forms of unity. The local unity is the integrity of the individual, literally
undividable, forming a closed system with respect to meaning but open
to flows of energy and information. The universal unity is the essential
embedding of every individual in the material, social and spiritual worlds
by the process of embodied cognition through brain dynamics. A valid
mathematical description of the dynamics must conform to the outlines of
the cycle of prediction, action, perception and assimilation.
In the past half century science has opened several new windows for di-
rect experimental observation of normal brain function after 3000 years of
logic and speculation. Our preferred window of observation is on the elec-
tric fields that are generated by brains, because our techniques enable us
to record and measure neural activity patterns from brains of animals and
humans that are actively engaged in intentional behaviors. The behaviors
and the electric activity patterns range in size from single neurons to an
entire cerebral hemisphere, and in duration from fractions of a millisecond
to minutes, days and even years. Indirect measures of neural activity by
measuring blood flow in order to estimate metabolic energy dissipation pro-
vide supportive data. Brains are profligate in energy dissipation as heat at
rates an order of magnitude greater than other organs of comparable mass.
Additional insights come from measurements of magnetic and chemical en-
ergy gradients. The space-time patterns of these observable signs of neural
activity are well suited to precise measurement. Until now our preferred
tool for modeling the dynamics underlying the measured activity has been
ordinary differential equations (ODE) in space-time, because the neurody-
namics of large-scale populations, as manifested by the electric currents of
axonal action potentials and dendritic synaptic potentials, can be modeled
in a mesoscopic continuum. Now we propose to adapt another tool, neurop-
280 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

ercolation, which has been opened to brain science by recent developments

in RGT.

1.2. The mathematical foundations shared by ODE and RGT

Brain function is nonlinear, nonstationary, non-Gaussian and drenched in

noise. It is therefore counterintuitive to find that the dynamics of most
areas of cortex maintain themselves in small-signal domains that are linear,
stationary and Gaussian, as shown by testing for superposition with impulse
inputs. These frames last on the order of 0.1 s and occupy domains at
the scale of cm2. There are temporal nonlinear discontinuities between
these frames, which require very few ms for spatial dissemination, so the
relatively long time spans (∼ 1s) that are required for application of the
standard tools of linear analysis (Fourier transform, Principal Component
Analysis, Independent Component Analysis, regression models, wavelets,
etc.) provide space-time averages over the data that yield reproducible
decompositions, such as the clinical frequency bands (theta, alpha, beta,
gamma, etc.) and that relegate the brief nonlinearities to the residuals.
Neural activity can readily be elicited in response to perturbation by sensory
or direct magnetic and electrical stimulation of the brain. If that evoked
activity remains well within the amplitude and frequency ranges of the
background brain activity, it conforms to superposition and can be measured
by curve-fitting with sums of linear basis functions that are the solutions
to linear ODE. The preponderance of clinical and experimental descriptions
of large-scale neural activity are expressed in terms of frequencies on the
assumption of linearity and stationarity, and the fact that these methods
work as well as they have for over half a century is testimony to the power of
Fourier and Laplacian methods, as well as to the basic simplicity of cortical
dynamics, if it is seen from an optimal point of view.
The coefficients of the fitted basis functions evaluate the poles and zeroes
(eigenvalues) of the optimized ODE. The ODE are structured to represent
the topology of the neural networks, in which the time constants are fixed,
and the synaptic interaction strengths are represented by feedback gains.
Sets of solutions give root loci in the complex plane along the phase curves
at either 0 rad for positive feedback loops or π rad for negative feedback
loops. The feedback gains are given by the intersections of the amplitude
contours with the root loci. Solution sets and root loci have been derived for
changes in cortical function with variations in stimulus intensity, location,
Scale-Free Cortical Planar Networks 281

and timing [44], from which the changes in interaction strengths are inferred.
Changes in functional connectivity during learning have been modeled by
using the driving impulse as a conditioned stimulus. Likewise the root lo-
cus method has been used to model the neural connectivity changes that
underlie changes in the levels of arousal and motivation; type and level of
anesthesia, and the pharmacological effects of common neurotransmitters,
neuromodulators, and their agonists and antagonists. The ODE that have
been evaluated by fitting sums of linear basis functions to cortical impulse
responses have enabled the construction of a hierarchical set of dynamical
networks entitled Katchalsky sets (K-sets) [51]. These models provide the
formal structure, from which we undertake representation of large-scale neu-
rodynamics, first in the ancient, primitive, six- layered allocortex common
to all vertebrates (’reptilian’) and then in the six-layered neocortex found
only in mammals. In this report we summarize the dynamics as seen by use
of ODE, so that RGT can replicate the dynamics as currently known and
then proceed into the unknown.

1.3. The applicability of ODE and RGT depends on the following

properties of cortex

• Hierarchy: Brain activity exists simultaneously at all levels of a hier-

archy from atomic and molecular to psychosocial, whereas observation
and measurement is restricted by the experimental tools to bounded
ranges in space and time. In conformance to our main sources of
data we restrict our concerns to three levels imposed by the tools of
observation: microscopic measurements in microns and ms of synap-
tic potentials and action potentials from microelectrodes; mesoscopic
measurements in mm and tens of ms from electrode arrays of elec-
trocorticograms (ECoG); and macroscopic measurements in cm and
s from scalp electroencephalograms (EEG), magnetoencephalograms
(MEG), and functional magnetic resonance imaging (fMRI) reflecting
levels of cerebral metabolism.
• Scaling: Most descriptions of cortical connectivity emphasize micro-
scopic, local nets. Recent studies on neocortical connectivity have
emphasized conformance of large- scale connectivity to power-law dis-
tributions [115], [74], [28], [57], which opens the likelihood of self simi-
larity of connectivity across scales, and which imposes the necessity of
choosing a scale of description that matches the scale of experimental
282 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

observation. In spatial scale we focus on the mesoscopic population of

neurons, because our reliance on electric fields enables us to interrelate
our observations of ECoG downwardly to the microscopic activities of
neurons that form the population, and upwardly to the macroscopic
activity patterns that are formed by the mesoscopic ensembles. The
most interesting and important brain dynamics concerns the relations
across levels in the hierarchy: millions of microscopic neurons (action
potentials) create mesoscopic order parameters (ECoG) that regulate
the neurons. Likewise in temporal scale we focus on mesoscopic states
having durations on the order of 0.1-1.0 s. Microscopic events such as
action potentials lasting 0.5-2.0 ms are described in statistical aver-
ages such as probability density distributions. The changes in brain
dynamics that result from learning occur on still slower time scales,
which we treat as state changes that we describe as a temporal process
over hours and days. The energy used by brains for electrodynamic
events is drawn from an immense store of electrochemical energy in
the ionic gradients across membranes. Energy deficits incurred during
events of activity are small fractions of the total available; restoration
is by slow metabolic processes that are measured in s, which is the
scale of fMRI and related hemodynamic techniques. This separation
of time scales permits use of ODE as Hamiltonians applied to a con-
servative system, and separately to deal with the dissipation of free
energy in brain dynamics [111], [53].
• Dimensionality: The large-scale dynamics of cortex is optimally de-
scribed in planar networks of ODE and random graphs, owing largely
to its laminar organization.
• Modularity: Description of brain dynamics at every level is based on
choosing a module to represent a node and then constructing a net-
work of modules. In microscopic neural networks the module is a
model neuron having a summing junction and a nonlinear threshold
device. In mesoscopic neural networks the module is the KO set,
which represents an ensemble of microscopic neurons. Its properties
are averages of well-known time and length constants that parameter-
ize the processes of spatiotemporal integration by dendrites and the
transmission of output by action potentials of dendrites. Nonlinear
interactions among the neurons in a KO ensemble support the emer-
gence of mesoscopic order parameters that are observable in sums of
axonal and dendritic fields of potentials, which collectively are called
Scale-Free Cortical Planar Networks 283

brain waves: axonal multiple unit activity (MUA) and the electro-
corticogram (ECoG) and electroencephalogram (EEG) of dendrites.
The optimal module for RGT is the KO set, because our focus is on
large-scale neural activity.
• Boundary conditions: Brains are anatomically bounded structures,
but efforts to identify absorptive or reflective boundaries at the edges
of anatomically defined collections of neurons [89] have not succeeded.
Use of partial differential equations has been limited to describing the
dynamics of core conductors in axons and dendrites. ODEs are ex-
tended into the mesoscopic spatial domain by use of integro-difference
equations on digital computers to model distributed arrays described
by K-sets with periodic boundary conditions. This is the main point
of entry for RGT.
• Linearity and time-invariance: An effective tool for testing super-
position in mesoscopic populations is single-shock and paired-shock
electrical stimulation of axonal afferent pathways, which yields cor-
tical impulse responses. The technique of paired shock testing for
superposition reveals a small-signal, near-linear range of dynamics for
input domains that give impulse responses not exceeding the ampli-
tude of the on-going, self-stabilized background activity. Both addi-
tivity and proportionality hold for time ensemble averages of sums
of superimposed evoked potentials, within limits that are imposed by
the refractory periods of the stimulated axons. However, linearity is
temporally bounded into brief epochs of brain activity that are in-
terrupted by the discontinuities recurring several times each second.
Repeated samples of impulse responses reveal time-variance in the
wave forms, showing that state changes occur at the discontinuities.
Linearity is progressively violated as the intensity of input pulses is
increased and the ensembles are driven outside the small signal range.
Time variance and amplitude-dependent nonlinearity are made com-
patible with description with linear ODE by the introduction of state-
and amplitude-dependent gain coefficients in the ODE and by dis-
play of the changes in the characteristic frequencies with root locus
techniques [44].
• Stationarity: The ongoing background spontaneous ECoG and EEG
from populations reveal endogenous discontinuities, because the syn-
chronized oscillatory signals from multiple electrodes in arrays
abruptly disorganize and then resynchronize quickly at a different fre-
284 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

quency that varies little until the next discontinuity. By this criterion
the populations can be treated on average as stationary over extended
time periods and as having a single mean frequency with a decay enve-
lope that represents a distribution of frequencies. Though in fact the
populations undergo repeated state transitions and exhibit stationar-
ity only briefly, the ODE and RGT suffice to represent the dynamics
averaged over the duration and spatial domain of averaging required
for observation and measurement.
• Autonomy: The brain is deeply engaged with body and the environ-
ment at all times, not as a passive recipient of energy and information
but as a proactive, predictive agent. All normal brain areas exhibit
background spontaneous activity, which is the source of exploratory
drive. It can be completely suppressed under deep anesthesia, but
it returns robustly with recovery of normal brain function. This au-
tonomous background activity is notable for its repetitive discontinu-
ities in multiple frequency ranges, which recur independently of the
engagement of the subject by intentional behavior. The temporal dis-
continuities appear as essential steps in the action-perception cycle by
which cognition proceeds, including the emergence of a goal state, the
successive construction and execution of plans for actions of observa-
tion, the perception of stimuli upon generalization and abstraction,
and assimilation by learning. We propose that the spontaneous dis-
continuities provide the opportunity for phase transitions, by which
cognition proceeds frame by frame. The ODE suffice to represent
the compartmentalized dynamics in functional domains preceding and
following a phase transition; we now seek to describe with RGT the
transition that links the preceding and following frames.
• Criticality: Piece-wise linear analysis using root locus techniques with
ODE have brought to light the existence of a nonzero point attractor
in cortical dynamics, which is manifested as a zero eigenvalue (a pole
at the origin of the complex plane), and at least one limit cycle at-
tractor (a pair of complex conjugate poles on the imaginary axis of
the complex plane). The point attractor is absolutely stable and pro-
vides the background excitatory bias that all brains require for mainte-
nance of their intentionality. The limit cycle attractor is conditionally
stable. Modeling with K-sets and ODE indicates that during a dis-
continuity in temporal dynamics the neural population may closely
approach the complex pole pair [[49], [53], at which the cortical dy-
Scale-Free Cortical Planar Networks 285

namical system becomes undefined and goes critical. At or near that

point all frequencies and wave lengths coexist. This is the gateway to
phase transitions by which order parameters impose synchronized os-
cillations [95] at high frequencies on very large neural populations, in
humans covering large fractions and at times the whole of each hemi-
sphere [47], [13], [100], [107]. The critical state is self-stabilized and
therefore self-organized. As the main route to the adaptiveness and
conditional stability of cognitive functions of brains, criticality is the
main target for description using RGT.

2. Allocortex: The Simplest Model for the Phase

Transition in Criticality

2.1. Microscopic sensation and mesoscopic perception

Olfaction is the least complex of the sensory systems, phylogenetically the

oldest, the algorithmic predecessor of all others, and for most nonhuman
animals by far the most important cognitive organ. In the simplest descrip-
tion the olfactory system consists of the receptor layer in the nose treated
as a KO set, the olfactory bulb, and the olfactory cortex, both having inter-
active excitatory (KIe set) and inhibitory (KIi set) populations in Fig. 1A
that in negative feedback comprise KII sets. These structures along with
the hippocampal formation are examples of three-layered cortex, referring
to the outer layer of input axons and dendrites, the middle layer of cell
bodies, and the inner layer of output axons and interneurons.
The receptor cells numbering 108 send topographically ordered axons
to the olfactory bulb with a convergence ratio of roughly 2000 : 1. The
∼ 103 types of receptor indicate 105 of each type, with convergence to
5 × 105 excitatory bulbar neurons that are coupled in negative feedback
with a roughly equal number of clusters of inhibitory interneurons (granule
cells grouped by gap junctions). The mitral cells send axons to the cortex
not by topographic mapping but by a divergent-convergent pathway. Each
bulbar transmitting neuron excites a broad distribution of cortical neurons,
and each cortical neuron receives from many bulbar mitral cells. This
transmission pathway performs a spatial integral transformation of bulbar
output that selectively amplifies that component of mesoscopic activity that
286 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

Fig. 1. A. Topology of olfactory system. B. Operation of a Hebbian cell assembly

has everywhere at each instant a common frequency of oscillation. That

condition is met only by the endogenously generated activity of the bulb
and not by the activity that is driven by the receptor input.
The benefit from the large number of receptors is the capture of faint
odorants at exceedingly low concentrations, such that only a very small
fraction, perhaps 102 among the 105 available receptor cells is activated on
any one sniff. Owing to turbulence in the nose, the selection and number
differs on every sniff. The problem of generalization from the sample on
each sniff to the class of the odorant is solved in the bulb by Hebbian
learning (Fig. 1B). A reward is given on each trial in which the expected
odorant is present, but not on the trials without the odorant. Chemical
neuromodulators such as dopamine and chole-cystokinin (CCK) increase the
strength of connection between just those pairs of neurons that are coexcited
on reinforced trials. On unreinforced trials the connection strengths among
other pairs of coexcited neurons are automatically reduced by habituation.
Over multiple sniffs a Hebbian nerve cell assembly forms by strengthening
of the excitatory synapses between pairs of excitatory neurons, represented
in the ODE by a gain coefficient kee, and weakening of other excitatory
connections represented by kei. The output connection strengths of the
inhibitory neuron populations, represented by kie and kii, are unchanged.
The effect of the increase in kee is modelled in Fig. 2B. An increase in kee
of 20% increases the output 50-fold; a decrease in kee of −25% decreases
output 1000-fold. The entire assembly is vigorously excited into oscillation
no matter which neurons in the Hebbian assembly are excited by sensitive
receptors, or how many are excited; this demonstrates generalization to the
relevant class of stimulus.
Scale-Free Cortical Planar Networks 287

Fig. 2. A. Increase in the form of induced oscillations in the gamma range (20–80 Hz) by
increase in mutually excitatory feedback gain. B. Amplitude. C. Frequency. The link
between excitatory neurons is the most sensitive synaptic site for increasing and
decreasing the sensitivity of the olfactory system in learning both by association
and by habituation; from [45]

However, the activation of a Hebbian nerve cell assembly constitutes

∼ 0.1% of the bulbar output neurons. The further step that is required is
a phase transition by which the entire olfactory bulb enters into a spatial
pattern of oscillation that is governed by an attractor. Figure 2A shows
examples of oscillatory wave forms in the gamma range that appear on
every channel but are spatially modulated in amplitude, called AM patterns.
These AM patterns manifest an attractor that corresponds to the class to
which the conditioned stimulus (CS) belongs. In each olfactory system a
landscape of attractors embodies the repertoire of odorants that the subject
is capable of classifying at each stage of its lifespan. Each attractor is
based in the synaptic matrix of the bulb in conjunction with other parts of
the brain, which constitutes the memory of the subject for that odorant
and its significance for intentional behavior. The lack of invariance of
the AM patterns for control and odor stimulation shows that they are
288 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

not representations of CS formed by filtering; they are creations through

nonlinear dynamics.

2.2. The origin of the background activity in mutual excitation:

positive feedback

Three types of KO ensemble are active in brains: excitatory, inhibitory,

and modulatory. The dynamics of each module is described by a trans-
fer function that consists of a component for linear pulse to wave density
conversion with gain ka , a distance-dependent part, H(x, y), and a time-
invariant, amplitude-dependent nonlinearity, G(v). The necessary and suf-
ficient minimal representation of ka is by measuring the impulse response of
a KO population in which interactions have been blocked by deep anesthe-
sia. The response to an excitatory impulse is the average synaptic potential
with a rapid depolarization and exponential return to the baseline (Fig. 3).
Paired shock testing reveals proportionality and additivity. The minimal
fitted curve is the sum of two exponential terms, for which the equation is
a linear 2nd order ODE.

d2 pi (t) dpi (t)

(1) (ab) 2
+ (a + b) + pi (t) = Fi (t).
dt dt

A 1st order ODE does not suffice, because the single exponential term
only captures the passive membrane capacitive delay and fails to incorporate
the synaptic and dendritic cable delays, which are crucial in the population
dynamics. Here a and b are biologically determined time constants. pi (t)
denotes the pulse density at the i-th node as a function of time; i ∈ N , where
N is the index set of all nodes. Fi (t) includes the effects of the neighboring
populations influencing node i.
H(x, y) is represented by linear matrix algebra. Most calculations to
date have been with a fully connected squared array of nodes with peri-
odic boundary conditions. The employed models incorporate conduction
velocities for distance-dependent

delays. The source term in Eq. (1) can
be expressed as Fi (t) = G(ka j∈N H(xi , xj )pj (t)), where the summation
runs over the index set N of all populations.
G(v) is determined by calculating the action potential probability con-
ditional on ECoG amplitude. The fitted curve, which was derived for the
population by a generalization from the Hodgkin–Huxley equations for sin-
Scale-Free Cortical Planar Networks 289

Fig. 3. The open loop impulse response (averaged evoked potential under deep
anesthesia) of the olfactory system is fitted with the sum of four exponential terms, two
approximating the synaptic and dendritic cable delays, the passive membrane decay, and
the slow rate (∼ 1/s) of the metabolic recovery process

gle neurons [45], is a sigmoidal function that expresses the axonal pulse
density output of an ensemble as a function of dendritic wave density:
)  *
1
(2) G(v) = q 1 − exp − v
q(e − 1)

In Eq. 2, q is the parameter specifying the slope and maximal asymptote

of the curve. This sigmoid function is modeled from experiment on biological
neural activation [44].
The nonlinear gain can be replaced by a gain constant equal to the
tangent to the sigmoidal curve in piece-wise linear approximation of the
dynamics in the vicinity of a steady-state operating point. The linking of
KO modules into distributed networks of K-sets yields descriptions of brain
dynamics at the macroscopic level, which describe the neural mechanisms
of intentional behavior, including assimilation by learning [51], [66].
290 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

2.3. Source of the repeated discontinuities facilitating phase tran-

sitions: Rayleigh noise

A neural mechanism for the variability of the decay rate of the impulse
response and the aperiodic approach to criticality has been revealed by
detailed analysis of the background ECoG [11]. The characteristic form of
the power spectral density (PSD) of subjects at rest is power-law in log-log
coordinates with a slope close to −2 (brown noise) and decreasing in sleep
near −3 (black noise) [101]. In animals intentionally engaged in perceptual
tasks the PSD deviate from 1/f with peaks reflecting excess power in the
gamma and theta ranges [33], [35], [39]. These peaks result from inhibition
by negative feedback that acts as a band pass filter. The application to the
raw ECoG of a band-pass filter in the gamma range followed by the Hilbert
transform yields the analytic signal, which is decomposed into the analytic
power, and the analytic phase. Calculation of successive differences, divided
by the digitizing step, yields the analytic frequency. The spatial mean, and
spatial standard deviation, SDX(t), co-vary with time, revealing nearly
constant values of analytic frequency and low values of SDX(t) for ECoG
segments lasting 50–120 ms, but bracketed by brief epochs of high SDX(t)
and either low or high values of frequency. These are the times of temporal
discontinuity, across which the values for the frequency change on average
by ∼ 10 Hz (∼ 16% of mean frequency) together with sudden large changes
in spatial patterns of amplitude and phase of the ECoG.
The discontinuities coincide with the decrease in mean analytic power,
A2 (t), to values as low as 10−4 below the modal values. The high spatial
variance in phase revealed by the maxima of SDX(t) at null spikes reflects
the indeterminacy of phase at the minima. This pattern of repeated near-
zero values of analytic power accompanied by maxima in analytic phase
variance in very brief time periods was simulated by band pass filtering
simulated brown noise, applying the Hilbert transform, calculating the an-
alytic power and phase difference at each point in time and space, and
plotting mean power, A2 (t), and phase SDX(t). We infer that these coor-
dinated analytic phase differences are a type of Rayleigh noise, in which the
mesoscopic power nearly vanishes, owing to summation and cancellation of
excitatory and inhibitory currents, but without decrease in the microscopic
power. The observable mesoscopic potential differences vanish, but the mi-
croscopic ionic currents persist undiminished. At this minimum the order
parameter vanishes, and the complete disorganization of the microscopic
activity enables the cortex to transit to a new phase. Being mesoscopic,
Scale-Free Cortical Planar Networks 291

the preceding and succeeding patterns and the transient abeyance in the
null spike are completely invisible in microelectrode recordings at the mi-
croscopic level. What is seen is the sustained firing of neurons in Hebbian
nerve cell assemblies, which is identified as that of feature detector neurons
[99], also pejoratively known as grandmother cells [71].
We postulate that the source of the variation in wave form with fixed in-
put is the variation in the mesoscopic amplitude imposed by the null spikes.
If so, the impulse driving of the KIIei population reveals by the decreased
decay rate the amplification of the response to the evoking stimulus, when-
ever the mesoscopic background activity diminishes. We propose that the
necessary condition for the transposition from the microscopic pattern given
by a Hebbian nerve cell assembly to the mesoscopic order parameter, which
is realized in the AM spatial pattern governed by an attractor, is the occur-
rence of a null spike during the sustained discharge of an activated Hebbian
assembly. Then the initiation of a phase transition that leads to perception
requires the conjunction of three recurring events of independent origin.
One is the macroscopic initiation by the limbic system of a behavioral act
of observation, such as a sniff, saccade or whisk, that brings a sensory volley
to the cortex primed by preafference. The second is the onset of a meso-
scopic null spike, which opens the cortical dynamics to the singularity by
which a state of criticality is closely approached. The third is the pres-
ence in the sensory volley of microscopic action potentials from receptors
that belong to the expected class of input and active a Hebbian assembly.
Therefore the formation of a percept is contingent on the presence in the
environment of that which is sought, on the intentional observation of the
environment, and the local opening of a gate in sensory cortex, ready to
reorganize its background activity.
Empirically the recurrence rates of null spikes are proportional to the
value of the center frequency and to the width of the pass band. The
mathematics remains to be worked out. Null spikes occur independently in
multiple frequency bands spanning the beta and gamma ranges. The ap-
pearance of neocortical ECoG in array windows fixed on the brain surfaces
of active animals resembles that of a pan of boiling water, in which the bub-
bles resemble the epochs of transient stationarity. Most null spikes do not
lead to phase transitions; most incipient phase transitions are quenched, ei-
ther because there is no concomitantly active Hebbian assembly, or because
the reduction in background power is insufficient.
292 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

3. Application of RGT to Neocortex

3.1. Topology of large-scale neocortical connectivity: Generality

versus specificity

Allocortex is relatively simple; 2 quick steps takes it from a selection of

receptor cells to a Hebbian assembly and then to a chaotic attractor [46],
[109], [65]. Neocortex is more complex. The greatest challenge in mod-
eling the dynamics of neocortex is posed by the requirement to meet two
seemingly irreconcilable requirements. One is to model the specificity of
neural action even to the level that a single neuron can be shown to have
the possibility of capturing brain output. The other is to model the gener-
ality by which neural activity is synchronized and coordinated throughout
the brain during intentional behavior. The topology of allocortex summa-
rized in Fig. 1 displays these two essential features of cortex: specificity
in the topographic mapping of receptors into the bulb, and generality in
the divergent-convergent integration of bulb to cortex. Neocortex provides
these two features as well but on a far greater range of spatial scale. The
temporal dynamics of visual, auditory and somatic cortices is very similar to
that of olfaction, including the provision of background spontaneous activ-
ity, power-law distributions of connection lengths and spatial and temporal
power spectral densities, repeated formation of AM spatial patterns with
carrier frequencies in the beta and gamma ranges, and frame recurrence
rates in the theta range. ODE serve to describe the mesoscopic dynamics
of neocortical populations, but falter in attempts to model the entire range
including the transitions between levels, which appear to take place very
near to criticality. RGT offers a new approach, by simulating the results
from mesoscopic recording and ODE modeling.

3.1.1. Neocortical area and neuronal density. Anatomists [104], [14],

[102] agree on the relative invariance across areas of neocortex and species
of mammal, the number of neurons/mm2 being close to 105 within a factor
of 2, compared with wide variations in number of neurons/mm3 . The
logarithm of the area of both allocortex and neocortex varies across species
in accordance with the logarithm of body size, so inversely does the thickness
of cortex and therefore the volume density, respectively in human neocortex
2.6 mm average thickness and 4 × 104 /mm3 . Apart from the scaling factor
of body size that holds for allocortex, the way by which neocortex increases
Scale-Free Cortical Planar Networks 293

in size with evolution is by increase in surface area, adding neurons at

nearly fixed density per unit surface area. By the extra degree of freedom
neocortical area increases more than volume relating to body mass, leading
to gyrification by wrinkling of the surface. The area of allocortex varies over
3 orders of magnitude in proportion to body size. Neocortical area ranges
an additional 2 orders of magnitude. The largest brain in the sperm whale
is 105 greater in area than the smallest in the tree shrew, yet the temporal
dynamics revealed by the ECoG is similar from mouse [43], to whale [77]
giving dramatic evidence for power-law distributions of connectivity and
self- similarity in scale-free dynamics.

The challenge posed by neurobiologists to physicists and mathemati-

cians is to explain the phase transitions that appear in scalp EEG and MEG
recordings from normal humans [10]. The fine structure of neocortex reveals
each hemisphere to be a unified dynamical organ with no interruptions in
connectivity over its entire surface. Given that fact alone it is not surpris-
ing to find synchronized oscillations over distances that often approach the
length and breadth of the hemisphere. The similarity in patterns of phase
synchrony extended across a difference in scale of 34 : 1 between human
and rabbit. What is surprising is that the beta and gamma oscillations at
each spike jump to a new value of frequency that differs from the previous
frequency on average by 10 Hz over the range in animals of 20–80 Hz. The
repetitive resynchronization over relatively immense correlation distances
is surprising, because the coordination of the activity is based in the main
in communication by action potentials with conduction velocities for the
most part well under 10 m/s. [52] have reviewed the variety of proposed
alternatives in field theory to explain the rapid resynchronization and con-
cluded that it is an emergent, macroscopic property. [13] came to the same
conclusion in their study of MEG modelled using small-world theory.

3.1.2. The large-scale embryological growth and development of

neocortex. Anatomists have for over a century recognized the coexistence
of relatively numerous local connections within putative modules and rel-
atively sparse long connections among networks of modules, with the like-
lihood of dynamic modification in formation of Pavlov’s temporary con-
nections by learning. Recently several research groups have attempted to
model cortical structure and function with small world networks [114] by
constructing a cellular neural network and replacing a small proportion of
local connections with long connections. The network is grown by adding
294 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

new cells one at a time and forming new connections for each node as it is
added [e.g., [8]).
This conception lacks important features of how neocortex forms [1], [88],
[62], [80]. At the microscopic level the neurons initially replicate as small
spheres in large numbers, migrate toward the surface of the brain, and then
sprout axons that reach out and form synaptic connections on dendrites.
The dendrites also extend branches that provide the surface area needed
to accommodate synapses, each with a current path to the axonal trigger
zone near the cell body. Neurons continue to branch, extend, and form new
connections, long after replication ceases and excess neurons are pruned
by apoptosis (programmed cell death). Lifelong growth continues in some
areas by cell division and maturation of new neurons, but predominantly
the existing neurons continue to reach out and form new connections that
provide the matrix of connectivity that is shaped by learning. Lengthening
and branching of axons and dendrites continues well into the sixth decade
of human life and perhaps beyond.
Forms of the distributions of connection lengths used in random graphs
are schematized in Fig. 4A. In the initial formulation by Erdős and Rényi
[38] the probability of connections is uniform with distance (lower line). In
cellular neural networks the connections are restricted to nearest or next
nearest neighbor nodes, e.g., [30]. In small-world graphs/networks a low
percentage of the local connections is replaced with uniformly distributed
long connections, step function [114], which dramatically reduces the depth
by bridging across many nodes. Anatomical studies of cortical connectiv-
ity report exponential distributions of the lengths of axon collaterals and
projections (dashed curve) [92], [24], which display predominantly local con-
nections and increasingly sparse connections with increasing distance from
the cell bodies.
Revising the data display from semi-log plots to log-log plots gives
power-law distributions of distances in the middle range with deficits at
both ends. The deficits can be explained by the limitations in the experi-
mental methods. For short distances the observations using light microscopy
omit gap junctions and most unmyelinated axons, which in electron micro-
graphs substantially outnumber the myelinated axons. For long distances
the observations of axon lengths in Golgi preparations are made in tissue
sections of limited thickness (e.g., 300 microns, Fig. 4B) in small mammals
(here mouse). Continuity is very difficult to follow in serial sections, lead-
ing to a deficit in the numbers of long connections. The self-similarity of
many axonal and dendrite trees is well-documented [110], [60], and the dis-
Scale-Free Cortical Planar Networks 295

Fig. 4. A. Distributions of the numbers of connections at varying distances are

schematized in log-log coordinates. The straight line indicates the power-law relation,
1/f α . B. Distributions of measurements of lengths of axons that were made in
histological sections of Golgi preparations [24]. The data were re-plotted in log-log
coordinates

tributions of cortical structural connectivity are power-law [74], [59], [57],

[83]. The most appropriate model for generating an embedding generalist
random graph is a large fixed number of nodes having power-law distrib-
utions of connection distances at fixed slope and with steadily increasing
asymptotes as new stochastically generated links are added. Specializations
within this matrix need then to be undertaken to model the effects of ge-
netic factors on selective targeting biases and apoptosis, and the effects of
learning from experience.
Special significance is attached to the graph with power-law distribu-
tion of connection distances, owing to the introduction of self-similarity at
different scales of observation and measurement. The significance lies in
the dynamical properties endowed by scale-free connectivity [25], [29], [5],
particularly the capacity for repetitive phase transitions to occur almost si-
multaneously over very large areas of neocortex, owing to the approach to
criticality giving the long correlation lengths.

3.2. Evolution of heterogeneous cortical structures

Spatially heterogeneous networks are important in various systems, includ-

ing the cortex. To model the development of such inhomogeneous brain
structures, we describe the evolution of large-scale graphs starting from
296 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

well-defined initial configurations. To model brain development from the

inception of the individual to the formation of detailed brain structures,
we evoke the concepts of pioneer neurons [34], [85] and the transitory sub-
plate [103], [6]. We start from an initial set of 103 pioneer neurons. Pioneer
neurons are modeled as small balls densely packed in a small area of char-
acteristic size of 10μ. During the developmental stage, this tiny cortical
tissue multiplies about 108 -fold, to have the fully developed brain with 1011
neurons. This means a 104 -times increase in linear size, by considering the
essentially 2-dimensional structure of the cortex. For illustration, the stages
of the development of the brain are given in Table 1.

Linear size Number of Neurons Description

10 μ 103 Pioneer neurons in subplate
100 μ 105 Evolving populations
1 mm 107 Cortical columns
10 mm 109 Cortices
100 mm 1011 Full brain
Table 1. Illustration of brain growth

The evolution of the brain is described as a inflation of the initial

graph G0 formed by the interconnected pioneer neurons. Assume an initial
population of about 103 pioneer neurons packed on a small lattice; packing
distance is ∼ 0.33μ. As the cortex develops, the space spanned by the
neurons expands so that the distance between the original neurons increases.
The space created by this expansion is filled in with newly created neurons.
These new neurons grow connection to the existing ones. A schematic view
of the initial stage of development is shown in Fig. 5.

3.2.1. Directedness in cortical networks. An important constraint

in RGT is directedness [8]. Most neural connections by synapses (unlike
gap junctions) are unidirectional; with the exception of dendrodendritic
synapses reciprocal interconnections are by different axons. Topologically
the result is parcellation of a large network into a giant component, one or
more input components, one or more output components, and miscellaneous
islands having insufficient densities of links to participate in the other three
groups. The utility of these distinctions in applications to cortex is obvious.
The concept of the giant component provides an interesting candidate for
a neural correlate of consciousness by virtue of its immense size, unity,
and capacity for rapid state transitions. The islands, evolving in relative
isolation by continuing growth of connections, might be fruitful for using
Scale-Free Cortical Planar Networks 297

Fig. 5. Schematic illustration of the sequential evolution of the cortical graph G; black
circles – pioneer neurons G0 , grey and white circles – consecutive generations of
neurons. The first two steps of the evolution are shown. The illustrated regular
structure is used only for simplicity, while the actual brain tissue evolves random
topological and connectivity structures

scalp EEG to study the process of insight, the ‘aha’ phenomenon described
in [56] and as the collision of matrices of thought by Arthur Koestler [64],
which has been documented in the auditory cortical EEG of gerbils engaged
in category learning [90].

3.2.2. Sparseness in cortical networks. In graphs the number of

synapses between neurons (mono-, di-, poly-synaptic) determine connec-
tion distances. In networks the radial lengths of axons/dendrites from cell
bodies determine distances. Most dendritic trees are small enough (< 1 mm
radius) to sum their currents by passive cable spread; most axons are long
enough (∼ 1 mm) to require communication by pulses that use local energy
so as to transmit with delay but without attenuation. Every synapse re-
quires a dendritic loop current path to the axon like a leaf on a tree, hence
the large surface area of the dendritic tree and the high packing density in
humans of neurons (4 × 104 /mm3 ) and synapses (4 × 108 /mm3 ) to accom-
modate the 104 synapses on each neuron sent by 104 other neurons [104],
[14], [24], [108]. So great is the density of neurons and their connecting
fibers that on average a neuron demarcated by its cell body and nucleus
connects sparsely with < 1% of the neurons within its dendritic arbor, and
298 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

the likelihood of reciprocal connections between pairs is < 10−6 . Given that
each neuron transmits to ∼ 104 others, the number of its targets in 3 steps
would be 1012 , which approaches the 13.7 ± 1.6 109 neurons in each cerebral
hemisphere [23]. This leads to an estimated depth of 3 for a cortical net-
work. As most of the connections are local, real world networks have depths
much larger than this value; see also Appendix. Here we propose a 3-level
hierarchy of cortical organization: microscopic networks of neurons within
cortical columns; mesoscopic populations in dedicated areas of sensory, mo-
tor and associational cortex, and the entire cerebral hemisphere serving as
a macroscopic system that organizes the intentional activity of the brain
and body. A node at each level consists of a network of nodes at the next
level below. To some as yet unknown extent this hierarchical organization is
made possible by self-similarity of connectivity and dynamics across levels.

3.3. Topology of small-scale neocortical connectivity: Specificity

3.3.1. Structural connectivity: Emergence of neocortex from al-

locortex. The laminar neuropil in the phylogenetically older parts of the
cerebrum is called allocortex [86]. Examples include the prepyriform cortex
(paleocortex), the hippocampus (archicortex) and parts of the perirhinal
and entorhinal cortex (mesocortex) [78]; the olfactory bulb is here included
as allocortex owing to its similarity to the others in topology and phylo-
genetic derivation, though not all anatomists accept this taxonomy [24].
Generically allocortex has three layers with differing subdivisions specific to
each area; see Fig. 6. Layer 1 is also called marginal layer, which lies under
the bounding pial membrane and has input axons and the dendritic trees
on which they synapse. Layer 2 has the cell bodies, often with triangular
shapes giving the name pyramidal cells (mitral cells in the bulb). Layer 3
has output axons with recurrent side branches called collateral branches that
synapse on other pyramidal cells and on local interneurons called stellate
cells (internal granule cells in the bulb).
In contrast to allocortex with three layers, neocortex (also called isocor-
tex) is described as having six layers. The layers are defined on the basis of
the spatial patterns of the cell bodies revealed by selective staining of the
DNA in the cell nuclei [72], not on connectivity [26], [104]. In allocortex
(also called reptilian) the layers form by cells migrating from inside-out and
spreading under those already arrived. Comparative neuroanatomists have
shown how neocortex (found only in mammals) has evolved from reptilian
Scale-Free Cortical Planar Networks 299

Fig. 6. Schematic summary of the transition from allocortex to neocortex by intrusion of

cells into Layer 2 of allocortex through transitional mesocortex around the landmark
separating neocortex from allocortex, the rhinal fissure. Adapted from [84], p. 237.

allocortex by intrusion into allocortical (Layer II in Fig. 6) of neurons mi-

grating from the basolateral forebrain [62], [1]. The first neocortical cells
called pioneer cells [34], [85] are guided by a transitory subplate [103], [6]
that enables subsequent cells to leapfrog over pioneer cells by inside-out
progression and form the fine-grain radial bundles of parallel cells that are
oriented perpendicular to the pia in neocortical columns and hypercolumns.
The cytoarchitectures and connectivity patterns in Layers II to IV vary
greatly between local areas. These palisades of axons preferentially fol-
lowing the pioneer cells provide topographic mapping of sensory input and
motor output to and from neocortex. The local areas by Brodmann [26].
These specializations support the specialized functions that are performed
by local networks in the many sensory, associational, and motor areas of
neocortex. The individuation is enhanced by close dependence on topo-
graphically organized specific thalamic inputs by axons forming synapses in
Layer IV. The neocortical marginal Layer I continues to receive input ax-
ons from nonspecific thalamic nuclei and from allocortex. In neocortex the
input to Layer IV percolates outwardly to smaller, local neurons; in allocor-
tex the input to Layer 1 percolates inwardly to larger, long-range projection
neurons in Layer 2.
Braitenberg and Schuz wrote [24]: “A recent hypothesis by Miller [82],
based on differences in both connectivity patterns and spontaneous activity
300 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

between upper and lower layers, assigns to the upper layers the role of a neu-
ronal library, storing most of the information encoded by assemblies, while
the lower layers are assumed to catalyze the process of assembly formation”
(p. 150). Kaas wrote [61]: “Generalizing from cats and monkeys it appears
that the evolutionary advance in brain organization is marked by increases
in the numbers of unimodal sensory fields, not by increases in multimodal
association cortex as traditionally thought” (p. 147). Therefore Layers II–
IV can be conceived to provide highly textured mesoscopic modules that
are most active in learning and adaptation, while the embedding neuropil
in Layers I, V and VI provides the connectivity needed for rapid integra-
tion of modular activity into macroscopic ensembles. We conclude that the
local networks within Layers II to IV are embedded in the continuous sheet
of neuropil in each cerebral hemisphere, which is unified by the global net-
work of Layers I, V and VI. The deep pyramidal cells in Layers V and VI
have long basal dendrites that receive connections from neurons broadly sit-
uated in Layers II to IV, and their apical dendrites extend and branch into
Layer I. The deep pyramidal cells send the longest output axons to other
areas of cortex, to the basal ganglia and brainstem, and by widely radiating
collaterals into the overlying layers. Long cortico-cortical connections are
not fully random; they are often distributed in patches with high local con-
nection density, with intervening regions having few connections [79]. Most
analyses of anatomical data have emphasized point-to-point networks, for
example, the meta-study by Fellemin and Van Essen [40] of connectional ar-
chitecture in visual cortex, and the cortico-subcortical modules of Houk [58].
These linkages are essential for dissemination of activity among the nodes
in the networks. However, they do not address the connectivity required
for the long-range coherence and rapid phase transitions revealed by car-
rier waves. Therefore the best target for modeling neocortical activity by
random graphs appears to be the embedding layers.

3.3.2. Considerations of differences between allocortical and neo-

cortical activity. Three major differences allocortical and neocortical ac-
tivity were revealed between by the functional properties of the beta and
gamma patterns. First, there was only one bulbar phase cone at a time. In
neocortex there were multiple overlapping phase cones at all times. Second,
each bulbar oscillatory pattern occupied the entire bulb with fixed diameter.
The neocortical patterns of shared frequency varied in diameter (correlation
distance), averaging 2-3 times larger than those in the bulb and often ap-
pearing large enough in cats and rabbits to occupy the entire hemisphere.
Scale-Free Cortical Planar Networks 301

Third, the variation in bulbar phase cone duration was limited to half to
two thirds of the interval between frames recurring at the rate of breathing
or sniffing. The distributions of durations from frames with AM patterns
that were classified with respect to CS also conformed to the same power
law. Hence within a range spanning nearly three orders of magnitude the
durations of phase cones showed self-similarity and lack of characteristic
mean and standard deviation, characteristic of scale-free distributions as
seen also for the PSD of the EEG. Each of the myriad small phase cones
without classifiable AM patterns was interpreted as a local state transi-
tion that appeared to abort before accessing a learned basin of attraction.
The superimposed local events resembled the subthreshold fluctuations de-
scribed by Prigogine [96] in a system held very close to criticality and a
transition at the boundary of a basin of attraction.

3.3.3. Preferentiality in the assignment of probability of connec-

tions. A feature of random graphs is preferentiality [8], [4], in which the
probability of a node receiving new connections is proportional to the num-
ber of connections it already has. Constraining a scale-free network with
preferentiality in the assignment of the probabilities of input and output
connections may lead to the emergence of a hub, which is a node of excep-
tionally high density of connections, an outlier in a power-law distribution
that includes long-distance links. Widely studied networks with hubs in-
clude the route maps of major airlines and the patterns of connectivity on
the Internet. Such networks are relatively unaffected by random loss of
nodes but fail with loss of hubs [112], [113]. Likewise the performance of
cerebral cortex is relatively resistant to degradation by local lesions in ar-
eas not affecting input and output pathways. The medical literature is rich
with descriptions of transient loss of circumscribed cognitive functions that
accompany focal damage to limited areas of neocortex [32] and are repaired
by learning, but cortex can fail catastrophically with certain critical lesions,
such as damage to the midbrain reticular formation leading to coma, the
substantia nigra leading to Parkinson’s disease, and the perforant path in
Alzheimer’s disease. Functional evidence for hubs might be seen in the
patches of cortical activity seen in fMRI that are associated with a broad
range of cognitive functions [27].
Significant preferentiality is seen in the mechanism by which pioneer
neurons provide the initial guidance that is required for the construction of
topographic maps, which are displayed in textbooks as motor and sensory
homunculi. This mechanism addresses the local organization that is required
302 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

for sensorimotor integration by neurons in Layers II to IV. The question

is whether another form of preferentiality might be modelled in Layers I,
V and VI, which might help to explain hubs, if indeed they exist. The
connectivity of single neurons does not obey power-law, but have carrying
numbers of input and output synapses numbering in the thousands. The
requirements for macroscopic connectivity supporting hubs in neocortex are
unknown. A mathematical model for large-scale preferentiality is needed to
guide anatomists and physiologists in the search for the requisite structural
and functional properties of cortex.
For example, scale-free connectivity at the population level may give
a fresh start to understanding the enigmas of speech localization, and the
paradox that the higher is the verbal IQ, the smaller is the speech area [91].
Broca’s area is in close conjunction with the motor areas for the lips and
tongue, yet linked by the uncinate fasciculus to the temporal lobe for au-
dition and by U-fibers to the parietal lobe for proprioception. Wernicke’s
area is at the conjunction of somatic, visual and auditory sensory cortices.
In both instances the tendencies for preferential connectivity may support
the growth of hubs respectively for motor and perceptual speech functions.
Focal electrical or magnetic pulse stimulation by neurosurgeons temporarily
inactivates speech, but that does not imply that speech is localized to the
site of stimulation. Instead the elucidation of speech mechanisms requires
investigation of how neurons in a hub participate in macroscopic coordina-
tion of oscillatory neural activity. Definitive exploration of hubs may be
optimized by recording EEG/MEG from a dense array of channels overly-
ing a large percentage of the surface each hemisphere, and with sampling at
1 cm intervals corresponding to the width of gyri so as to capture adequately
the texture of AM patterns of spatially coherent activity [47]. A foreseeable
device with 512 channels for EEG would cover an area of scalp 23 × 23 cm,
comfortably fitting over the calvarium.

4. Structure and Dynamics in Cortex Models

The cortical tissue has a complex anatomical structure of interconnected

neural population producing well-studied behaviors manifested through
spatio-temporal oscillations. These oscillations have been studied inten-
sively using various brain monitoring tools. The major challenge is to de-
termine the relationship between structure and function. The mathematical
Scale-Free Cortical Planar Networks 303

theory of random graphs and percolations [15], [22], [21] have been very use-
ful for th characterization of structural properties of large-scale networks.
There are some promising approaches to establish the structure-function
relationship in complex networks, including brains [8], [105], [67], [7]. How-
ever, a systematic mathematical analysis of this field is still missing. Here
two major aspects of the structure-function relationship will be emphasized:
• Assuming a given structure, in the absence of significant growth and
learning effects, one is interested in studying the dynamic behavior of
the networks. How a desired oscillatory behavior can be achieved by
changing the structural parameters of the network?
• At the next level, structural changes are introduced, which are related
to learning and adaptation. One may assume certain biologically mo-
tivated evolution rules acting through the addition/removal of nodes
and connections. The aim is to study how the behavior of the network
varies under such dynamically changing conditions?
Oscillations have been studied in 2-dimensional lattice using probabilis-
tic cellular automata (PCA) [7]. The state of a node at a given time is
defined by the sum of the active states of its neighbors in the previous
steps, using a probabilistic threshold function. The PCA has been orig-
inally defined with a completely regular and homogeneous neighborhood
structure with a binary function defined over the nodes of the graph. In
addition, mean-field models have been studied as well over random graphs.
The existence of critical behavior and phase transitions have been rigor-
ously shown in the case of very low level of noise and mean field model
[7]. Simulations indicate that mixed models with local and long-range con-
nections exhibit critical behavior of weak Ising type [67]; such models are
also called neuropercolation. An extended neuropercolation model with ex-
citatory and inhibitory nodes can exhibit periodic oscillations, where the
frequency of the oscillations is controlled by the coupling coefficients be-
tween the nodes [98]. These are narrow-band oscillations with well-defined
frequencies in the gamma band.
We described critical oscillations over various PCA. It is of special in-
terest to determine statistical properties of average activation (magnetiza-
tion) as an order parameter. Mean field probabilistic cellular automata
over finite tori can exhibit bi-stable and multi-stable oscillations with expo-
nentially long waiting time between transitions, while each transition takes
polynomial time [7]. Simulations show that the power spectra density (PSD)
of magnetization oscillations exhibit 1/f a behavior near the critical state,
304 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

where f is the temporal frequency. The following major scenarios are dis-
tinguished:

• Far from criticality, the dynamics stays for very long time at a resting
state, i.e., oscillating around a given magnetization level, when clus-
ters grow and dissolve randomly, without the emergence of a giant
component representing large-scale ordering.
• Near critical state, the waiting period is reduced and large clusters
are frequently formed. Simulations show that such large clusters
eventually cover most of the network, and ultimately the system flips
to another basic state.
• Cluster formation and growth, and statistical distribution of the size of
the clusters is a problem of great importance, and should be attempted
using tools of rigorous mathematical analysis.

Functional brain networks have been studied by EEG, MEG and fMRI
tools. In the experimental studies it is typically postulated that there is
a functional link between nodes if their activations show correlation above
a threshold level. Such studies indicate the dominance of relatively simple
connectivity motifs with 3 and 4 nodes [105].
Preferential attachment is a popular model of building networks. Due
to the inherent shortcomings of certain versions of preferential attachment
models of growth, more rigorous approaches are sought [21]. Cortical struc-
tures evolve along the path of pioneer neurons, which can be considered a
form of preferentiality. It is highly desirable to introduce a mathematically
tractable growth model to describe the observed neural processes. Such
models are outlined in the next section.

5. Requirements the Mathematical Models

Several simplifications will be introduced when formulating a mathemati-

cally tractable theory of the structure-dynamics relationship in the cortex.
The cortex is approximated as a 2 dimensional
 sheet
 of interconnected neu-
rons and modelled as a planar graph G V (x),  E over
 the vertex set V
and directed edges E, where x ∈ S ⊂ R2 . G V (x), E is a directed graph
defined by the connected neurons, where axons and dendrites signifying out-
Scale-Free Cortical Planar Networks 305

going and incoming connections, respectively. The following are some key
parameters of the model:

• ωx (r): Distribution function of the lengths of the connections orig-

inating from a node at location x. Biological arguments point to a
1/rk distribution, where r is the connection length, k is the exponent
of the scale-free length distribution, k ≈ 2.
• nx : Spatial distribution (density function) of vertices V in 2D. It can
be uniform or may evolve following a some patterns specified by a
given model V0 , V1 , V2 , . . . , Vt , . . . . Often we consider a sequence of
evolving subsets of vertices in V0 ⊂ V1 ⊂ V2 ⊂ . . . Vt ⊂ R2 .
• d+ −
x /dx : Degree distribution of the outgoing/incoming edges of the
nodes. Neurobiology indicates that the average in-degree of a node
is 104 in the cortex. It is not clear if scale-free degree distribution is
valid in the cortical tissue. The apparent presence of some distributed
hub structure may point towards scale-free degree distribution.

The parameters d+ x , ωx (r), nx describe certain important features of the

planar graph network structure. Depending on the specifics of the model,
various relationships can be derived between these quantities. For a variety
of suitable models, see the Appendix.

6. Conclusions

The challenge of describing and modeling the dynamics of immense numbers

of neurons and synapses in cortex has been met by invoking a continuum
to represent the mesoscopic pulse and wave densities with state variables
of neuronal populations in networks of ODE [44], [12]. This strategy works
well in stationary, small-signal, near-linear ranges and can be extended into
time-varying, amplitude-dependent nonlinear dynamics with piece-wise lin-
earization giving analytic solutions to the equations for simple approxima-
tions. Large-scale modeling has provided a platform in K-sets. The KO set
models the mesoscopic node in a space-time continuum. The KIe set models
the synaptic interaction supporting the non-zero point attractor that sus-
tains cortical background activity. The KIi set models mutual inhibition
and spatial contrast enhancement that is required for spatial AM pattern
formation. The KIIei set predicts one or more limit cycle attractors that
306 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

support sustained oscillations in the gamma and beta ranges. The KIIei set
suffices to model the essential dynamics of allocortex.
ODE in K-sets have been extended in several directions. The KIII set
consisting of a KIe set and three KIIei sets chaos with multiple long feedback
connections model the landscapes of chaotic attractors that form by learning
and provide the memory system for conditioned stimuli [48], [46], [65], [73].
The KIV set consists of three KIII sets devised to model the limbic system
in goal formation, learning, and navigation [66]. KIV has been successfully
demonstrated intentional behavior in embedded robotics [68], [69]. The KV
set is proposed to model neocortex. Advances using ODE are in progress
in three directions: nonequilibrium thermodynamics [49], [50]; dissipative
quantum field theory [53]; and renormalization group theory.
However, further use of ODE is severely limited. Networks of popula-
tions with multiple types of nonlinear feedback cannot be solved analyti-
cally and must be solved by numerical integration with digital computers, in
which case the differential equations are discretized in time and space into
coupled integrodifference equations. Owing to attractor crowding and nu-
merical instabilities from the extremely large number of attractors in high-
dimensional networks, the digital models are unstable and must be tem-
pered with additive noise, which requires stochastic difference equations.
The digital computations for solving the ODE are prohibitively slow. The
alternative route of analog computation using VLSI to solve the ODE has
met with early success [97] but has encountered problems with bias control.
The equations suffice empirically to replicate the pulse and wave densities
and their spatial patterns governed by nonconvergent chaotic attractors [65],
but none of the equations can be used to model phase transitions, only the
states before and after the transitions.
Random graph theory offers a fresh beginning, in which the discreteness
of neurons and synapses can be approximated with numerical representa-
tions in percolation theory. It is readily adapted to describing neural nets at
the microscopic level, the nodes at mesoscopic and macroscopic levels, and
the relations among spatial and temporal variables between levels in phase
transitions. Modeling of structural and functional connectivity in allocortex
by neuropercolation theory is already well advanced, particularly in model-
ing the interplay of long connections, inhibitory feedback, and additive noise
in the genesis of self-regulated spontaneous activity of large nets of nodes
at the mesoscopic level. Results to date suffice to simulate white noise with
a flat PSD (white noise) [101] at low levels of interaction strength, brown
noise with 1/f 2 PSD when close to a phase boundary, and narrow band
Scale-Free Cortical Planar Networks 307

gamma oscillation just beyond a phase transition. The extension to neo-

cortex requires specification of the input and output connection densities of
nodes in a form that is compatible with the known distributions of neuron
sizes, connections, and correlation lengths.

7. Appendix: Mathematical Models

In this Appendix we shall first describe some existing models of random

graphs with properties that may make them useful in the analysis of the
neocortex, and then go on to describe new such models.
We should emphasize that, as always when modelling complex systems,
we have to try to reconcile two contradictory aims: a model should produce
random graphs whose structure resembles that of the neocortex as much
as possible, but it should not be so complicated that it is not susceptible
to mathematical analysis. The hope is that eventually the mathematical
theory of random graphs will be advanced enough that both requirements
can be satisfied: a sophisticated model can be constructed that incorporates
most of the requirements, and this very complicated model can be analyzed
precisely, meaning that one can obtain mathematical descriptions both of
its structure, and of the associated dynamics. At the moment we are very
far from this Holy Grail.
In this appendix we shall approach the task just described from the
mathematical side: we shall examine models of random graphs that can be
analyzed, and that can be viewed as primitive models of the neocortex.
Let us start with the most studied random graph models, namely the
‘classical’ models G(n, M ), introduced by Erdős and Rényi [37], and G(n, p),
introduced by Gilbert [54]. In G(n, M ), the numbers n of vertices and M
of edges are given, and the random graph is chosen uniformly from among
all graphs with vertex set {1, 2, . . . , n} and M edges. In G(n, p), there
are again n vertices, and each possible edge is present with probability p,
independently of the other edges. In many contexts, such as the present
one, the two models (with appropriately related parameters) are essentially
equivalent, and both are now known as Erdős–Rényi random graphs, since
it was they who pioneered the use of probabilistic methods to study random
graphs.
The models G(n, M ) and G(n, p) are homogeneous, in the sense that
the vertex set has no structure; in particular, all vertices are equivalent.
308 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

Most networks in the real world are inhomogeneous, either in the sense that
different vertices have very different properties (for example, in scale-free
networks, some have much higher degree than others), or in the sense that,
while the roles of individual vertices are similar, the vertex set has (usu-
ally geometric) structure which influences the network structure: nearby
vertices are more likely to be connected. In the last few years, many new
inhomogeneous random graph models have been introduced; we shall briefly
describe a few of these in the next subsection.

7.1. A general inhomogeneous model

Bollobás, Janson and Riordan [18] introduced a very general model of sparse
inhomogeneous random graphs, designed for representing graphs with geo-
metric or similar structure. Here ‘sparse’ means that the number of edges
grows linearly with the number n of vertices, so the average degree d¯ is
constant, but the definitions adapt immediately to other density ranges,
for example graphs in which the average degree is n2/5 . In the fully dense
range, with order n2 edges, a closely related but much simpler model was
introduced by Lovász and Szegedy [76]. Note that the mathematical defin-
itions of ‘sparse’ and ‘dense’ involve limits as n → ∞; the distinction may
not be so clear in practice when modelling real-world graphs.
The full definition of the BJR model is rather complicated, so let us
present a simplified version, adapted to the directed graph setting. Given a
probability space (S, μ), we shall call a measurable function κ from S × S
to the non-negative reals a kernel on S. (In [18], where undirected graphs
are considered, the kernel is assumed to be symmetric.) Let p = p(n)
be a normalizing function, for example, p = n−3/5 , corresponding, some-
what crudely, to the probability p = p(n) in G(n, p). The random graph
Gp (n, κ) with vertex set {1, 2, . . . , n} is defined as follows: first choose
x1 , . . . , xn independently from S according to the distribution μ. Then,
given these  ‘vertex types’,
 add each possible directed edge ij with probabil-
ity min pκ(xi , xj ), 1 , independently of the other possible edges. Here the
minimum has no significance: it is needed only to make sure that we work
with probabilities that are at most 1. Much of the time pκ is less than 1,
so the minimum is simply pκ.
One of the simplest instances of this model is defined in a geometric
setting. Let R be a region in the plane, and let μ be Lebesgue measure,
scaled so that μ(R) = 1, say. Then the ‘types’ xi are simply points in
Scale-Free Cortical Planar Networks 309

the plane, chosen independently and uniformly from R. The kernel κ is a

function of two points x, y ∈ R, for example

c
(3) κ(x, y) =  α ,
0 + d(x, y)

where d(x, y) denotes the Euclidean distance between x and y, the constant
0 is a characteristic length scale, α > 0 is the distance exponent, and c is
a normalizing constant. In this case Gp (n, κ) is a random graph on points
distributed randomly in R, with nearby points more likely to be joined than
distant ones, but edges of all lengths possible.
Note that if κ is chosen as above, then it is explicitly built into the
model that the probability that two vertices are joined is a function of their
distance, and that this function follows a power law.
In general, κ can be any function of two points in R; it need not depend
only on their distance, but might also vary in other ways, representing
possible variations in the average number of connections in different parts
of the cortex. Also, the number n of vertices need not be exactly fixed. It is
often convenient to take n to have a Poisson distribution: the vertex types
(in the geometric setting, simply their locations in R) are then distributed
according to a Poisson process. This is the most natural mathematical
model for objects distributed with roughly uniform density but not arranged
in an ordered lattice.
It is very natural to consider graphs whose vertices are points in the
plane, or in a higher (or sometimes lower!) dimensional space, in which
the probability that two vertices are joined is a function of their distance.
The very first such model, dating back almost to the beginning of the
related subjects of random graph theory and percolation theory, is the
disc percolation model of Gilbert [55]. Here the vertex set is a Poisson
process in the plane, and two vertices are joined if they are within a certain
distance 0 . A more recent example is the spread-out percolation model
of Penrose [93], which is essentially the same as the special case of the
(later) BJR model considered above: simplifying his model somewhat, the
connection probability for points at distance d scales as r −2 ϕ(d/r), where
ϕ is some function chosen in advance, and r is a parameter of the model.
In spread-out percolation one studies the limiting behaviour as r → ∞ with
ϕ fixed. In general, random graphs with geometric structure are known as
random geometric graphs; see the book by Penrose [94].
310 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

An advantage of the more general formulation of the model Gp (n, κ) is

that one can incorporate additional features within the same mathematical
framework. For example, it may make sense to model the neocortex by
a random graph in the plane, reflecting the predominantly 2-dimensional
structure, while representing neurons in different layers of the neocortex by
vertices of different classes. To reflect the different patterns of connections
within and between the different layers, the edge probabilities can be taken
to depend

on the vertex classes. For example, given constants p1 , p2 , . . . , pk
with pi = 1 representing the relative numbers of neurons in different
layers, one can take (roughly) pi n vertices of class i distributed uniformly
at random in R, and then join a vertex of class i at point x with one of
class j at point y with a probability pij (x, y) depending on i, j, x and y.
Although this appears to be a generalization of the BJR model, it is in fact
included in the original definitions, by taking the type space S to consist of
k copies of R, one corresponding to each class of vertex.
Furthermore, one can easily include spatial variations in the density of
neurons, as well as in the connection probabilities, by taking the measure
μ to be non-uniform. When there are several vertex classes, their densities
need not be the same.
There is always a danger when introducing a random graph model,
or any mathematical model, that the model becomes too general; while
flexibility is desirable, it is important that the model remains simple and
specific enough for mathematical analysis. Although the BJR model is very
general, as shown in [18], it is still simple enough for analysis. There the
main focus is on the percolation phase transition, but other properties are
also studied. For example, it is shown that the degree distribution is mixed
Poisson: roughly
 speaking, the degree of a vertex of type x is Poisson with
mean λ(x) = κ(x, y) dμ(y).
Translating this to the present setting, suppose we have k vertex classes
(for example, perhaps we should take k = 6, for the six layers of the
neocortex). Suppose that the distribution of vertices of class i is given
by a density function fi (x), x ∈ R, and that there are ni vertices of class i.
Suppose also that the probability that a vertex of class i at point x is joined
to one of type j at point y is given by pij (x, y) as above. Then the expected
out-degree of a vertex of type i at a point x is


k 
d+
i,x = nj pij (x, y)fj (y) dy,
j=1 y∈R
Scale-Free Cortical Planar Networks 311

and the expected in-degree by


k 
d−
i,x = nj pji (y, x)fj (y) dy.
j=1 y∈R

As in [18], the degrees are concentrated around their means, so the in-
and out-degree distributions of vertices of class i are given by sampling
the functions above with respect to the distribution fi (x). In the simple
special case where the connection probabilities depend only on which classes
the vertices are in and the distance between them, then, except near the
boundary, the degrees do not depend on x, and we find that while vertices
of different classes may have very different degrees, and within a class out-
and in-degrees may be different, all vertices in a class have roughly the same
out-degree as each other, and roughly the same in-degree.

7.2. The exponentially expanding graph model

In the study of graphs with power-law degree distributions, there are two
very different basic approaches. One is to build the power law into the
model; this approach was taken by Aiello, Chung and Lu [2, 3], for example.
The other is to seek to explain the power law by incorporating into the
model some simple features also present in the real-world networks that one
hopes lead to power laws. For degree distributions, this was the approach
of Barabási and Albert [9], who showed that growth with ‘preferential
attachment’ leads to scale-free degree distributions. Their mathematically
imprecise model was made precise in [20], and its degree distribution first
analyzed rigorously by Bollobás, Riordan, Spencer and Tusnády in [22].
Many other models soon followed, for example the much more general
(but harder to analyze) model of Cooper and Frieze [31], and the directed
scale-free model of Bollobás, Borgs, Chayes and Riordan [16]. For general
background on the rapidly expanding field of scale-free networks see, for
example, [4, 19, 36, 87].
In the previous subsection we took the analogue of the first approach,
building power-law distribution of edge-lengths into the model. In this sec-
tion we propose a new model, the exponentially expanding graph model,
which incorporates one of the basic features of cortical development, de-
scribed in Subsection 3.2; as we shall see, this automatically leads to a
power-law distribution of edge lengths.
312 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

As described in Subsection 3.2, as the neocortex develops, new neurons

force their way in between existing ones. To a first approximation, existing
neurons retain their connections and relative positions, except that the
distances between them increase with time, and new neurons connect to
nearby neurons; see Fig. 5. In our mathematical model, it is convenient to
rescale time so that the distances grow exponentially with time: we do not
claim that this is the actual growth rate in the brain. At any given time,
the density of neurons is roughly constant, but they do not form a regular
lattice. The appropriate mathematical model for their locations is thus a
Poisson process. Although a realistic model would start with a fairly small
number of neurons (around 103 ), mathematically it happens to make no
difference, and is somewhat cleaner, if we start from nothing.
To define the model formally, let R0 be a region in the plane, with
area A0 . (One can also take R0 to be the whole plane, obtaining
 an infinite
random graph.) For all real numbers t, let R = e t R = (et x, et y) : (x, y) ∈
 t 0
t
R0 be the region
 obtained by scaling R0 by a factor  e in each direction,
and let C = (t, x, y) : (x, y) ∈ Rt , −∞ < t < ∞ be the corresponding
‘exponential cone.’
Let P be a Poisson process on C with intensity 2. In other words, P is
a random set of points of C such that the probability that a small volume
V of C contains a point is 2V , and the numbers of points in disjoint regions
are independent. We take a point (s, x, y) ∈ P to represent a vertex v that
arrives at time s at position (x, y). At a later time t, the vertex is then at
position (et−s x, et−s y). There are many possibilities for the edge-formation
rule. The simplest is to join an arriving vertex v to all older vertices w
that are within a fixed distance 0 of v at the time that v arrives. Let Gt
be the graph formed in this way by time t. Thus the vertex set of Gt is
the set of all points v = (s, x, y) of P with s ≤ t. There is a directed edge
from v = (s, x, y) to w = (s , x , y  ) if and only if s < s and the Euclidean
 
distance between (x, y) and (es −s x , es −s y  ) is at most 0 .
 
Let Ct = (s, x, y) ∈ C : s ≤ t be the part of C corresponding to times
up to t. Then the volume of Ct is
 t  t
vol(Ct ) = area(Rs ) ds = A0 e2s ds = A0 e2t /2.
s=−∞ s=−∞

Let nt denote the number of vertices of Gt . Then the expectation Ent of nt is

exactly 2vol(Ct ) = A0 e2t , and nt has a Poisson distribution with this mean.
By considering the part of Ct corresponding to a small region of Rt , we see
Scale-Free Cortical Planar Networks 313

that the positions at time t of the vertices of Gt form a Poisson process on Rt

with intensity 1. This is the reason for considering exponential growth: the
Poisson process P corresponds to new vertices arriving at a constant rate
(2 per unit time per unit volume); exponential expansion is then exactly
what is needed to maintain a constant density of vertices.
Having defined (the simplest form of) the model, let us turn to its basic
properties. Although the graph Gt grows as t varies, the motivation is to
model the adult cortex, so we are primarily interested in the properties of the
graph Gt with t fixed, where in the application the parameters will be chosen
so that Gt has around 1010 vertices. As noted above, the actual number
nt of vertices has a Poisson distribution, but this is sharply concentrated
around its mean A0 e2t .
The first property we consider is the degree distribution. When a vertex
v arrives at time s, it sends edges to all vertices in a disk of radius 0
around it. Ignoring boundary effects (which are small once the linear size of
Rs is much larger than 0 ), i.e., assuming the whole disk lies inside Rs , the
number of vertices in this disk is Poisson with mean π20 . Hence the limiting
out-degree distribution in Gt is Poisson with mean π20 . If 0 is fairly large
(as in a model of the brain, where the average degree is in the thousands),
then the out-degrees are concentrated around their average value of π20 .
The average in-degree is of course equal to the average out-degree.
Turning to the distribution, let v = (s, x, y) ∈ P be a vertex born at time
s, and set
  2   2
Iv = {(s , x , y  ) : s > s, x − es −s x + y  − es −s y ≤ 20 }.

Then a vertex w sends an edge to v if and only if w = (s , x , y  ) ∈ Iv . Thus

the expected in-degree of v at time t is μv = vol(Iv ∩Ct ), and the distribution
is Poisson with this mean. Ignoring boundary effects, Iv ∩ Ct is made up
of a disk of radius 0 at each time s , s < s < t, so μv = (t − s)π20 . Once
again, to a good approximation, the degrees are close to their expectations,
but now there is some variation due to the ages of the vertices: the expected
in-degree of a vertex v at time t is proportional to its age a = t − s. Since
the number ns of vertices born by time s has mean A0 e2s , the fraction of
vertices with age at least a > 0 is e−2a , so the in-degree distribution follows
(approximately) an exponential distribution:  the number
 of vertices with
in-degree at least d is proportional to exp − 2d/(π20 ) .
Let us turn to the edge-length distribution. Since the length of an edge
increases in proportion to et once it is formed, very roughly speaking, edges
314 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

of length at least r ≥ 0 arise from vertices with age at least a = log(r/0 ),

so the fraction fr of edges with length at least r is roughly the fraction of
vertices with age at least a, i.e., fr ∼ e−2a = 20 /r2 . A little more precisely,
at any time s, for 0 < r < 0 the rate of formation of edges with length (at
the time they are born) between r  and r + dr is 2A0 e2s 2πr dr : the first
factor represents the rate of formation of new vertices, and the second factor
the area at the right distance from a new vertex, and hence the expected
number of old vertices at the right distance.
For r > 0 , an edge at time t with length between r and r +dr must have
been born at some time s = t − a with a > log(r/0 ), and its length when
it was born must have been between re−a and (r + dr)e−a . It follows that
the expected number of edges at time t with length between r and r + dr is
 ∞  ∞
2A0 e2(t−a) 2πre−a dr e−a da = 4A0 πr dr e2t−4a da
a=log(r/0 ) a=log(r/0 )

= 4A0 πre2t (0 /r)4 /4 dr = A0 e2t π40 /r3 dr.

Recalling that the expected number of edges at time t is π20 Ent = π20 A0 e2t ,
the fraction fr of edges with length at least r is
 ∞
 2 
2t −1
A0 e2t π40 /(r ) dr = 20 /(2r 2 ).
3
π0 A0 e
r =r

In other words, the edge-length distribution follows a power law; this power
law is valid over all scales from r = 0 up to r comparable with the linear
size of Rt . The exponent 2 here fits well with the experimental exponent
2.2 seen in Fig. 4B.
The exponent 2 above is equal to the dimension of the space in which the
graph lives: this is no coincidence. Regardless of the real rate of growth, we
may reparameterize time so that linear distances grow proportional to et .
Assuming the spatial density of vertices remains roughly constant, this
means that nt grows with edt , where d is the dimension. The derivation
above is easily seen to lead to a power law with exponent d in this case.
We next turn to the typical graph distance d(v, w) between vertices of
Gt , often known as the diameter. (Note, however, that in graph theory the
diameter of a graph G normally refers to the maximum value of d(v, w),
v, w ∈ V (G), or, if the graph is not connected, the maximum over pairs
(v, w) with v and w in the same component of G.) For the moment we
view Gt as an undirected graph; this is not a reasonable simplification in
Scale-Free Cortical Planar Networks 315

this context, but the argument we give for this case will apply to a related
model treated as a directed graph; we return to this later.
In any reasonable model of the cortex, although the number of ver-
tices is very large (around 1010 ), the typical degrees are also fairly large
(around 104 ), so it is potentially the case that most pairs of vertices are
joined by a path of length 3. On the other hand, the scale-free distribu-
tion of axon lengths (see Fig. 4B) shows that almost all connections have
lengths much shorter than the scale of the whole cortex; if all connections
were short, then clearly the diameter would be very large. As is well known
in graph theory (see, for example, Bollobás and Chung [17]), a few long-
range connections are often enough to bring the diameter down. As we shall
see, that is the case for the model Gt , even if the average degree is not that
large.
The key observation is that among the roughly π20 earlier vertices that
a new vertex v sends edges to, there is a distribution of ages. At any given
time, the proportion of vertices with age at least a is e−2a , so typically
v will send an edge to some vertex w born around Δ = log(π20 )/2 time
units earlier than v. (Edges to vertices this old may or may not appear,
but edges to somewhat less old vertices are very likely.) Starting at a
vertex v, there will typically be a path v = v0 v1 v2 . . . vk where each vi
is older than the previous vi by a time of around Δ. The sequence stops
at around the time t0 where the graph first starts developing, or, more
precisely, when area(Rt0 ) ∼ π20 : at this point, all vertices are joined to
each other. Starting at two vertices and following two such paths, we find
that typically vertices are at graph distance
 at most 2(t − t0 )/Δ. Now
2(t − t0 ) ∼ log(nt /nt0 ) = log nt /(π20 ) , so we find that the diameter is
approximately
 
log nt /(π20 ) log n
diam(Gt ) ∼ 2
=2 − 2.
log(π0 )/2 log(π20 )

This shows that, in the regime where the average degree π20 is constant and
the number n = nt of vertices tends to infinity, the diameter is logarithmic
in n, as is to be expected. Note that the diameter is roughly twice the value
log n/ log d¯ that one would expect in a homogeneous random graph.
The calculation presented above is a simple heuristic, but it is possible to
make it precise, using branching process methods to show that the desired
paths really do exist for a large fraction of the vertices. In the other
direction, one can estimate the expected numbers of paths of various lengths
316 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

to show that the diameter is not significantly smaller than this estimate;
this should not be taken for granted: in the scale-free LCD graph of [20]
(a precise version of the Barabási–Albert model [9]), although the average
degree is constant the diameter is (slightly) sublogarithmic in n.
The exponentially expanding graph model Gt defined above incorporates
one new feature that we believe is essential in any accurate model of the
neocortex, namely a development process involving expansion in space that
guides the formation of connections. So far, we concentrated only on this
one feature, but to obtain a more realistic model, some modifications are
required.
Firstly, assuming that a new vertex joins to all older vertices within
distance 0 leads to unrealistically high clustering: a vertex born just before
time t will be joined to most of the other vertices within distance 0 ; many of
these will also be recent, and joined to each other. The simplest modification
to correct this is to add an additional parameter p, and consider the model
Gt (0 , p) defined as above, except that a new vertex sends an edge to each
older vertex w within distance 0 with probability p, independently of the
other edges. The analysis above carries over essentially unchanged to this
extension, as long as the average degree π20 p is reasonably large. In this
variant a typical (and therefore recent) vertex is joined to a fraction O(p)
of the nearby vertices, so values of p around 10−2 would be realistic.
There is a rather more natural generalization of Gt (0 , p), where we
replace the sharp cutoff at distance 0 by a smooth one. Most generally, we
simply take a function ϕ : R+ → [0, 1], and when a new vertex v arrives,
the probability that it sends an edge to w is simply ϕ d(v, w) , where
d(v, w) is the distance from v to w at the time that v arrives. Note that
the role of the function ϕ here is rather different from that of κ(x, y) in (3);
taking ϕ dropping off sharply beyond some characteristic length, for example
ϕ(r) = ϕ0 e−r /0 , the analysis of Gt above carries over, and we still see
2 2

power-law distribution of the edge lengths in Gt . Unlike in the previous

subsection, where this was built into the model through the choice of κ,
here it comes from the expansion.
The alert reader may have noticed that there is a much more serious
problem with Gt as a directed graph model of the neocortex: if vertices
only send edges to older vertices, then there are no directed cycles, so no
feedback is possible. One possible modification is to assume that when a
new vertex v arrives, not only does v send out edges to nearby vertices w,
but some nearby w send new edges to v.
Scale-Free Cortical Planar Networks 317

An alternative and much more satisfactory modification of the model

is as follows: suppose that we have a function ψ(t) defined on the positive
reals, representing the rate at which a new vertex will send out edges at time
t after it is born; the most natural example is perhaps ψ(t) = ae−bt for t > 0,
with a and b positive constants. Given a function ϕ(r) as above, we may
define a new model Gt (ϕ, ψ) as follows: the vertex set evolves exactly as in
the original model Gt . Given the vertex set, in any timeinterval[s,s+ds] of
infinitesimal length ds, each vertex v has probability ϕ d(v, w) ψ a(v) ds
of sending an edge to each vertex w, where d(v, w) is the distance from v to
w at time s, and a(v) is the age of v at time s. Of course, more generally one
can consider a single 2-variable function ϕ d(v, w), a(v) , or even a function
also depending on the age of w. In all these variants a significant fraction
of edges will go from older vertices to (slightly) newer vertices, so there will
be many directed cycles in the graph.
Once again, the simple heuristic analysis above carries over in this
greater generality, and shows that the edge length distribution remains a
power law with exponent 2, and the diameter remains logarithmic in n
even if the average degree is constant as n → ∞. Of course, turning this
into rigorous mathematics in this generality is likely to require considerable
effort.
Of course, the exponential expansion in Gt can be combined with fea-
tures present in other models, such as the original preferential attachment
mechanism of Barabási and Albert [9]. One natural formulation is to take
the vertex set of Gt , but to join a new
 vertex
 v to a vertex w with prob-
ability proportional to ϕ d(v, w) ψ deg(w) for some functions ϕ and ψ,
where d(v, w) is defined as above and deg(w) is the degree (or perhaps the
in-degree) of vertex w at the time v is born. This model is much harder
to analyze than the LCD model mentioned above, but we believe that if
the ‘preference function’ ψ is linear, then it will also have a power-law de-
gree distribution, with the exponent depending on the exact details of the
model. A closely related (but not exponentially expanding) geometric ran-
dom graph model with preferential attachment has been introduced and
analyzed by Flaxman, Frieze and Vera [42]. Of course it is not clear that
power-law degree distribution is appropriate in modelling the cortex.
For future work, we believe that there are two important directions.
One is the rigorous mathematical analysis of models of the type we have
described here, including the particular model Gt (ϕ, ψ). Another, even more
important, topic is the development of further models that more accurately
represent the structure of the brain, but are still simple enough to analyze.
318 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

References

[1] Aboitiz, F., Montiel, J., Morales, D., Concha, M. [2002], Evolutionary divergence
of the reptilian and the mammalian brains: considerations on connectivity and
development. Brain Res Rev 39: 141–153.
[2] Aiello, W., Chung, F., Lu, L. [2000], A random graph model for massive graphs.
Proceedings of the Thirty-Second Annual ACM Symposium on Theory of Com-
puting (2000), ACM, New York, pp. 171–180.
[3] Aiello, W., Chung, F., Lu, L. [2001], A random graph model for power law graphs.
Experiment. Math. 10: 53–66.
[4] Albert, R., Barabási, A.-L. [2002], Statistical mechanics of complex networks. Rev
Mod Physics 74: 47–97.
[5] Aldana, M. and Larralde, H., Phase transitions in scale-free neural networks:
Departure from the standard mean-field universality class. Phys Rev E 70: 066130,
2004.
[6] Allendoerfer, K. L., Shatz, C. J. [1994], The subplate, a transient neocortical
structure: Its role in the development of connections between thalamus and cortex.
Ann Rev Neurosci 17: 185–218.
[7] Balister, P., Bollobás, B., Kozma, R. [2006], Large deviations for mean field models
of probabilistic cellular automata. Random Structures and Algorithms 29: 399–415.
[8] Barabási, A.-L. [2002], Linked. The New Science of Networks. Cambridge MA:
Perseus.
[9] Barabási, A.-L., Albert, R. [1999], Emergence of scaling in random networks.
Science 286: 509–512.
[10] Barlow, J. S. [1993], The Electroencephalogram: Its Patterns and Origins. Cam-
bridge MA: MIT Press.
[11] Barrie, J. M., Freeman, W. J., Lenhart, M. [1996], Modulation by discriminative
training of spatial patterns of gamma EEG amplitude and phase in neocortex of
rabbits. J. Neurophysiol 76: 520–539.
[12] Basar, E. [2005], Memory as the whole brain work? A large-scale model based on
oscillations in super-synergy. Int. J. Psychophysiol 58: 199–226.
[13] Bassett, D. S., Meyer-Lindenberg, A., Achard, S., Duke, T., Bullmore, E. [2006],
Adaptive reconfiguration of fractal small-world human brain functional networks.
Proc. Nat. Acad. Sci. USA 103: 19518–19523.
[14] Bok, S. T. [1959], Histonomy of the Cerebral Cortex. Amsterdam: Elsevier.
[15] Bollobás, B. [1985/2001], Random Graphs. Cambridge Studies in Advanced Math-
ematics 2nd Ed. Cambridge University Press.
[16] Bollobás, B., Borgs, C., Chayes, T., and Riordan, O. [2003], Directed scale-free
graphs. Proc. 14th ACM-SIAM Symposium on Discrete Algorithms, 132–139.
[17] Bollobás, B., Chung, F. R. K. [1988], The diameter of a cycle plus a random
matching. SIAM J. Discrete Math. 1: 328–333.
Scale-Free Cortical Planar Networks 319

[18] Bollobás, B., Janson, S., Riordan, O. [2007], The phase transition in inhomogeneous
random graphs. Random Structures and Algorithms 31: 3–122.
[19] Bollobás, B., Riordan, O. [2003], Mathematical results on scale-free random graphs.
Handbook of graphs and networks, 1–34, Weinheim, Wiley-VCH.
[20] Bollobás, B., Riordan, O. [2004], The diameter of a scale-free random graph.
Combinatorica 24: 5–34.
[21] Bollobás, B., Riordan, O. [2006], Percolation. Cambridge University Press, 2006.
[22] Bollobás, B., Riordan, O., Spencer, J., Tusnády, G. [2001], The degree sequence of
a scale-free random graph process. Random Structures Algorithms 18: 279–290.
[23] Braendgaard, H., Evans, S. M., Howard, C. V., Gunderson, H. J. G. (2000) The
total number of neurons in the human neocortex unbiasedly estimated using optical
dissectors. J. Microscopy (Oxford) 157(3): 285–304.
[24] Braitenberg, V., Schuz, A. (1998) Cortex: Statistics and Geometry of Neuronal
Connectivity, 2nd ed. Berlin: Springer-Verlag.
[25] Breakspear, M. [2004], Dynamic connectivity in Neural systems: Theoretical and
empirical considerations. Neuroinformatics 2(2): 205–225.
[26] Brodmann, K. [1909], Vergleichende Lokalizationslehre der Grosshirnrinde. Leipzig:
Barth.
[27] Buxton, R. B. [2001], Introduction to Functional Magnetic Resonance Imaging:
Principles and Techniques. Cambridge UK: Cambridge UP.
[28] Changizi, M. A. [2004], Principles underlying mammalian neocortical scaling. Biol
Cybern 84(3): 207–215.
[29] Chen, Q., Shi, D. [2004], The modeling of scale-free networks. Physica A 333: 240–
248.
[30] Chua, L. O. [1998], CNN. A Paradigm for Complexity. Singapore: World Scientific.
[31] Cooper, C., Frieze, A. [2003], A general model of web graphs. Random Structures
and Algorithms 22: 311–335.
[32] Critchley, M. [1979], The Divine Banquet of the Brain. New York: Raven Press.
[33] Dan, Y., Poo, M. (2006) Spike timing-dependent plasticity: From synapse to
percecption. Physiol. Rev. 86: 1033–1048.
[34] Deng, J., Elberger, A. J. [2001], The role of pioneer neurons in the development of
mouse visual cortex and corpus callosum. Anat. Embryol 204 (6): 437–453.
[35] Demiralp, T., Bayraktaroglu, Z., Lenz, D., Junge, S., Busch, N. A., Maess, B.,
Ergen, M., Herrmann, C. S. [2006], Gamma amplitudes are coupled to theta phase
in human EEG during visual perception. Int. J. Psychophysiol 2006 Sep. 4.
[36] Dorogovtsev, S. N., Mendes, J. F. F. [2003], Evolution of networks: from biological
nets to the Internet and WWW. Oxford University Press, Oxford, 2003.
[37] Erdős, P., Rényi, A. [1959], On random graphs I. Publicationes Mathematicae
Debrecen 5: 290–297.
320 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

[38] Erdős, P., Rényi, A. [1960], On the evolution of random graphs. Publ. Math. Inst.
Hung. Acad. Sci. 5: 17–61.
[39] Fell, J., Klaver, P., Elfadil, H., Schaller, C., Elger, C. E., Fernandez, G. [2003],
Rhinal-hippocampal theta coherence during declarative memory formation: inter-
action with gamma synchronization. European J. Neurosci 17: 1082–1088.
[40] Fellemin, D. J., Van Essen, D. C. [1991], Distributed hierarchical processing in the
primate cerebral cortex. Cerebral Cortex 1: 1–47.
[41] Ferster, C. B., Skinner, B. F. [1957], Schedules of Reinforcement. Englewood Cliffs
NJ: Prentice-Hall.
[42] Flaxman, A. D., Frieze, A. M., Vera, J. [2006], A geometric preferential attachment
model of networks. Internet Math. 3: 187–205.
[43] Franken, P., Malafosse, A., Tafti, M. [1998], Genetic variation in EEG activity
during sleep in inbred mice. Amer. J. Physiol 275 RICP 44: R1127–1137.
[44] Freeman, W. J. (1975/2004), Mass Action in the Nervous System. New York:
Academic. Electronic version 2004 – http://sulcus.berkeley.edu/MANSWWW/
MANSWWW.html
[45] Freeman, W. J. [1979], Nonliear gain mediating cortical stimulus-response relation-
ship. Biol Cybern 33:237–247.
[46] Freeman, W. J., Chang, H.-J., Burke, B. C., Rose, P. A., Badler, J. [1997], Taming
chaos: Stabilization of aperiodic attractors by noise. IEEE Trans Circuits and
Systems 44: 989–996.
[47] Freeman, W. J., Burke, B. C., Holmes, M. D. [2003b], Aperiodic phase re-setting
in scalp EEG of beta-gamma oscillations by state transitions at alpha-theta rates.
Human Brain Mapping 19(4): 248–272.
[48] Freeman, W. J. [2000/2006], Neurodynamics. An Exploration of Mesoscopic Brain
Dynamics. London: Springer. Electronic version: http://sulcus.berkeley.edu/.
[49] Freeman, W. J. [2007], Proposed cortical “shutter” mechanism in cinematographic
perception. In: Neurodynamics of Cognition and Consciousness, Perlovsky, L.,
Kozma, R. (eds.) Heidelberg: Springer Verlag, pp. 11–38.
[50] Freeman, W. J. [2008], A pseudo-equilibrium thermodynamic model of information
processing in nonlinear brain dynamics. Neural Networks, 21: 257–265.
[51] Freeman, W. J., Erwin, H. (2008) Freeman K-set. Scholarpedia, 3(2): 3238.
[52] Freeman, W. J., Vitiello, G. [2006], Nonlinear brain dynamics as macroscopic
manifestation of underlying many-body field dynamics. Physics of Life Reviews
3: 93–118. http://repositories.cdlib.org/postprints/1515.
[53] Freeman, W. J., Vitiello, G. (2007), The dissipative quantum model of brain and
laboratory observations. Electronic J. Theoretical Physics 4, 1–18.
http://dx.doi.org/10.1016/j.plrev.2006.02.001.
[54] Gilbert, E. N. [1959], Random graphs. Annals of Mathematical Statistics 30: 1141–
1144.
[55] Gilbert, E. N. [1961], Random plane networks. J. Soc. Indust. Appl. Math. 9: 533–
543.
Scale-Free Cortical Planar Networks 321

[56] Hadamard, J. [1945], The Psychology of Invention in the Mathematical Field.

Princeton NJ: Princeton UP.
[57] Harrison, K. H., Hof, P. R., Wang, S. S. H. [2002], Scaling laws in the mammalian
neocortex: does form provide clues to function? J. Neurocytol 31: 289–298.
[58] Houk, J. C. [2005], Agents of the mind. Biol. Cybern 92 (6): 427–437.
[59] Hwa, R. C. and Ferree, T., Scaling properties of fluctuations in the human elec-
troencephalogram. Physical Rev. 2002, E 66: 021901.
[60] Jelinek, H. J., Elston, G. N. [2003], Dendritic branching of pyramidal cells in the
visual cortex of the nocturnal owl monkey: a fractal analysis. Fractals 11(4): 1–5.
[61] Kaas, J. H. [1987], The organization of neocortex in mammals: implications for
theories of brain function. Ann. Rev. Psychol 38: 129–152.
[62] Karten, H. J. [1997], Evolutionary developmental biology meets the brain: the
origins of mammalian cortex. Proc. Natl. Acad. Sci. 94, 2800–2804.
[63] Kelso, J. A. S. [1995], Dynamic Patterns: The Self Organization of Brain and
Behavior. Cambridge MA: MIT Press.
[64] Koestler, A. [1964], The Act of Creation. New York: Macmillan.
[65] Kozma, R., Freeman, W. J. [2001], Chaotic resonance: Methods and applications
for robust classification of noisy and variable patterns. Int. J. Bifurc Chaos 10:
2307–2322.
[66] Kozma, R., Freeman, W. J., Erdi, P. [2003], The KIV model – Nonlinear spatio-
temporal dynamics of the primordial vertebrate forebrain. Neurocomputing 52–54:
819–826.
[67] Kozma, R., Puljic, M., Balister, P., Bollobás, B., Freeman, W. J. [2005], Phase
transitions in the neuropercolation model of neural populations with mixed local
and non-local interactions. Biol. Cybern 92: 367–379.
[68] Kozma, R., Huntsberger, T., Aghazarian, H., Tunstel, E., Freeman, W. J. [2007],
Computational aspects of cognition and consciousness in intelligent devices. IEEE
Comp Intell. Mag., 2(3): 53–64.
[69] Kozma, R. [2007], Intentional systems: Review of neurodynamics, modeling, and
robotics implementations. Phys. of Life Rev. 5(1): 1–21.
[70] Kay, L. M., Freeman, W. J. (1998) Bidirectional processing in the olfactory-limbic
axis during olfactory behavior. Behavioral Neuroscience 112: 541–553.
[71] Lettvin, J. Y. [1995], J. Y. Lettvin on grandmother cells. pp. 434–435 in The
Cognitive Neurosciences, Gazzaniga MS (ed.). Cambridge MA: MIT Press.
[72] Lewis, B. [1878], On the comparative structure of the cortex cerebri. Brain 1: 79–86.
[73] Li, X., Li, G., Wang, L., Freeman, W. J. [2006], A study on bionic pattern classifier
based on olfactory neural system. Int. J. Bifurc Chaos, 16: 2425–2434.
[74] Linkenkaer-Hansen, K., Nikouline, V. M., Palva, J. M., Iimoniemi, R. J. Long-range
temporal correlations and scaling behavior in human brain oscillations. J. Neurosci.
2001, 15: 1370–1377.
322 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

[75] Lohmann, H., Roerig, B. [1994], Long-range horizontal connections between supra-
granular pyramidal cells in the extrastriate visual cortex of the rat. J. Comp.
Neurol. 344: 543–558.
[76] Lovász, L., Szegedy, B. [2006], Limits of dense graph sequences. J. Combin. The-
ory B 96: 933–957.
[77] Lyamin, O. I., Mukhametov, L. M., Siegel, J. M., Nazarenko, E. A., Polyakova, I. G.,
Shpak, O. V. [2002], Unihemispheric slow wave sleep and the state of the eyes in a
white whale. Behav Brain Res, 129: 125–129.
[78] Maclean, P. D. [1969], The Triune Brain. New York: Plenum.
[79] Malach, R. [1994], Cortical columns as devices for maximizing neuronal diversity.
TINS 17: 101–104.
[80] Mantagnini, A., Treves, A., The evolution of mammalian cortex, from lamination
to arealization. Brain Res. Bull. 2003, 60: 387–393.
[81] Merleau-Ponty, M. [1942/1963], The Structure of Behavior. Fischer AL [trans],
Boston: Beacon.
[82] Miller, R. [1996], Neural assemblies and laminar interactions in the cerebral cortex.
Biol Cybern 75: 253–261.
[83] Miller, R. [2002], Wheels within wheels: Circuits for integration of neural assemblies
on small and large scales. Ch. 18 in: Schuz, A., Miller, R. (eds.) Cortical Areas:
Unity and Diversity. New York: Taylor and Francis, pp. 423–458.
[84] Miller, R., Maitra, R. [2002], Laminar continuity between neo- and meso-cortex:
The hypothesis of the added laminae in neocortex. Ch. 11 in: Schuz, A., Miller, R.
(eds.) Cortical Areas: Unity and Diversity. New York: Taylor and Francis, pp. 219–
242.
[85] Morante-Oria, J., Carleton, A., Ortino, B., Kremer, E. J., Fairen, A., Lledo, P.-
M. [2003], Subpallial origin of a population of projecting pioneer neurons during
corticogenesis. PNAS 100(21): 12468–12473.
[86] Mountcastle, V. B. (ed.) [1974], Medical Physiology, 13th ed. St Louis MO: C. V.
Mosby, p. 232.
[87] Newman, M., Barabási, A.-L., Watts, D. J., eds. [2006], The structure and dy-
namics of networks. Princeton Studies in Complexity, Princeton University Press,
Princeton, NJ. x+582 pp.
[88] Northcutt, R. G., Kaas, J. H. The emergence and evolution of mammalian neocor-
tex. Trends Neurosci 1995, 18(9): 373–379.
[89] Nunez, P. L. and Srinivasan, R. (2005) Electric Fields of the Brain: The Neuro-
physics of EEG. Oxford UK: Oxford UP.
[90] Ohl, F. W., Scheich, H., Freeman, W. J. [2001], Change in pattern of ongoing
cortical activity with auditory category learning. Nature 412: 733–736.
[91] Ojemann, G. A. (2003) The neurobiology of language and verbal memory: Obser-
vations from awake neurosurgery. Intern. J. Psychophysiol 48(2): 141–146.
Scale-Free Cortical Planar Networks 323

[92] Paldino, A., Harth, E. [1977], A computerized study of Golgi-impregnated axons in

rat visual cortex. In Lindsay RD (ed.) Computer Analysis of Neuronal Structures.
New York: Plenum, pp. 189–207.
[93] Penrose, M. D. [1993], On the spread-out limit for bond and continuum percolation.
Annals of Applied Probability 3: 253–276.
[94] Penrose, M. D. [2003], Random Geometric Graphs. Oxford University Press,
xiv+330 pp.
[95] Pikovsky, A., Rosenblum, M., Kurths, J. [2001], Synchronization. A Universal
Concept in Non-linear Sciences. Cambridge UK: Cambridge UP.
[96] Prigogine, I. [1980], From Being to Becoming: Time and Complexity in the Physical
Sciences. San Francisco: WH Freeman.
[97] Principe, J., Tavares, V., Harris, J., Freeman, W. [2001], Design and implementa-
tion of a biologically realistic olfactory cortex in analog VLSI. IEEE J. Proc 89:
569-571.
[98] Puljic, M., Kozma, R. [2008], Narrow-band Oscillations in Probabilistic Cellular
Automata. Phys. Rev. E (in press).
[99] Quiroga, Q. R., Reddy, L., Kreiman, G., Koch, C., Fried, I. [2005], Invariant visual
representation by single-neurons in the human brain. Nature 435: 1102–1107.
[100] Rodriguez, E., George, N., Lachaux, J.-P., Martinerie, J., Renault, B., Varela, F.
[1999], Perception’s shadow: long-distance synchronization of human brain activity.
Nature 397: 430–433.
[101] Schroeder, M. [1991], Fractals, Chaos, Power Laws: Minutes from an Infinite Par-
adise. San Francisco: WH Freeman.
[102] Schuz, A., Miller, R. (eds.) [2002], Cortical Areas: Unity and Diversity. New York:
Taylor and Francis.
[103] Shatz, C. J., Chun, J. J. M., Luskin, M. B. [1988], The role of the subplate in
the development of the telencephalon. Pp. 35–58 in: Jones, E. G., Peters, A. (eds.)
The Cerebral Cortex. Vol. III. The Development of the Cerebral cortex. New York:
Plenum.
[104] Sholl, D. W. [1956], The Organization of the Cerebral Cortex. New York: Wiley.
[105] Sporns, O., Chialvo, D. R., Kaiser, M., Hilgetag, C. C. [2004], Organization, de-
velopment and function of complex brain networks. Trends in Cogn. Sci. 8(9):
418–425.
[106] Sutton, R. S. (1988) Learning to predict by the methods of temporal differences.
Machine Learning 3(1): 9–44.
[107] Traub, R. D., Whittington, M. A., Stanford, I. M., Jefferys, J. G. R., A mechanism
for generation of long-range synchronous fast oscillations in the cortex. Nature
1996, 383: 421–424.
[108] Thompson, H. [1899], The total number of functional nerve cells in the cerebral
cortex of man. J. Comp. Neurol 9: 113–140 [given as 9,282,826,403 which Warren
McCulloch [1967], rounded off to 1010 ].
324 W. J. Freeman and R. Kozma with an appendix by B. Bollobás and O. Riordan

[109] Tsuda, I. (2001) Toward an interpretation of dynamics neural activity in terms of

chaotic dynamical systems. Behav Brain Sci. 24: 793–847.
[110] Uylings, H. B. M., Van Pelt, J. [2002], Measures for quantifying dendritic arboriza-
tions. Network: Computation in Neural Systems 13: 397–414.
[111] Vitiello, G. [2001], My Double Unveiled. Amsterdam: John Benjamins.
[112] Wang, X. F., Chen, G. [2002], Synchronization in scale-free dynamical networks:
Robustness and fragility. IEEE Trans. Circuits Syst. 1 Fund Theory and Appl. 49:
54–62.
[113] Wang, X. F., Chen, G. R. [2003], Complex networks: small-world, scale-free and
beyond. IEEE Trans. Circuits Syst. 31: 6–20.
[114] Watts, D. J., Strogatz, S. H. [1998], Collective dynamics of ‘small-world’ networks.
Nature 393: 440–442.
[115] Zhang, K., Sejnowski, T. J., A universal scaling law between gray matter and white
matter of cerebral cortex. PNAS 2000, 97(10): 5621–5626.

Walter J. Freeman Robert Kozma

University of California at Berkeley, University of Memphis,
Berkeley, Memphis TN 38152,
CA 94720, U.S.A.
U.S.A.

Béla Bollobás Oliver Riordan

University of Memphis, Mathematical Institute,
Memphis TN 38152, University of Oxford,
U.S.A. 24–29 St Giles’,
and Oxford OX1 3LB,
United Kingdom
Department of Pure Mathematics and
Mathematical Statistics,
University of Cambridge,
Cambridge CB3 0WB,
U.K.
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 325–368.

Chapter 8
Reconstructing Cortical Networks:
Case of Directed Graphs with High Level of
Reciprocity

TAMÁS NEPUSZ, LÁSZLÓ NÉGYESSY, GÁBOR TUSNÁDY

and FÜLÖP BAZSÓ

The problem of prediction of yet uncharted connections in the large scale network
of the cerebral cortex is addressed. Our approach was determined by the fact
that the cortical network is highly reciprocal although directed, i.e. the input
and output connection patterns of vertices are slightly different. In order to
solve the problem of predicting missing connections in the cerebral cortex, we
propose a probabilistic method, where vertices are grouped into two clusters
based on their outgoing and incoming edges, and the probability of a connection is
determined by the cluster affiliations of the vertices involved. Our approach allows
accounting for differences in the incoming and outgoing connections, and is free
from assumptions about graph properties. The method is general and applicable
to any network for which the connectional structure is mapped to a sufficient
extent. Our method allows the reconstruction of the original visual cortical
network with high accuracy, which was confirmed after comparisons with previous
results. For the first time, the effect of extension of the visual cortex was also
examined on graph reconstruction after complementing it with the subnetwork
of the sensorimotor cortex. This additional connectional information further
improved the graph reconstruction. One of our major findings is that knowledge
of definitely nonexistent connections may significantly improve the quality of
predictions regarding previously uncharted edges as well as the understanding of
the large scale cortical organization.
326 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

1. Introduction

The cerebral cortex is probably the most prominent example of a natural

information processing network. It is therefore of major importance to learn
how this network is organized. At the lowest level, the cortical network is
composed of physically (i.e. via chemical and electrical synapses) connected
nerve cells. (The chemical synapse is the dominant type and it is rectifying
in contrast to the electrical synapse, which allows bi-directional interactions
between the neurons). The cortex, in general (ignoring species and areal
density differences [17]) consists of approximately 1010 nerve cells, each re-
ceiving numerous connections of order 103 (up to about 104 ) [7]. However,
these data do not necessarily imply a homogenous degree distribution as the
diverse types of neurons could form specific sub-networks. Based on func-
tional constraints and axonal wiring economy, Buzsáki et al. [10] proposed
a scale-free-like, small world architecture for the network of inhibitory local
circuit neurons consisting approximately 20% of the whole neuron popu-
lation. In fact, the significance of such a diversity of neurons is presently
unknown, especially in the case of the pyramidal cells, which is the prin-
cipal cell type of the cerebral cortex (making up the remaining 80%) [17].
Pyramidal cells, considered excitatory in nature, form long distance con-
nections both within the cortex and with subcortical structures. However,
most of the synaptic contacts are formed locally within a short distance, and
it is unclear how the cortical network is organized at the neuron level [18].
Considering anatomical and physiological data, Tononi et al. [65] outlined
a network architecture, which suitably performs segregation and integra-
tion, the fundamental functions of the central nervous system. Integration
is achieved by connections between clusters of neurons, representing func-
tionally specialized units, which are formed by dense local connectivity [65].
Using mutual information as a measure of integration, it was shown that
the proposed network exhibited high complexity, significantly differing from
random and regular lattice networks characterized by low measure of com-
plexity [65]. Differences between cortical and random networks were also
pointed out by Négyessy et al. [45], although on the level of cortical ar-
eas instead of single neurons. On the other hand, based on estimates of
the spreading function, Bienenstock [6] showed that the graph of cortical
neurons has a high dimensionality close to that of an Erdős–Rényi random
graph of similar size. This assumption is consonant with Szentágothai’s
notion of quasi-randomness in neuronal connectivity [63].
Reconstructing Cortical Networks 327

The neuron doctrine (stating that nerve cells are the developmental,
structural and functional units of the central nervous system) has been
challenged by arguing that populations of neurons function as units [62,
19, 43, 65, 22, 5, 8]. This perspective is in close agreement with the so-
called columnar or modular organization of cortical structure and function
as proposed by Mountcastle [43] and Szentágothai [62]. Accordingly, the
cortex is usually viewed as a two dimensional sheet composed of functional
modules with a diameter of 250–500 μm, arranged perpendicularly to the
surface and spanning the layers to the depth of the cortex. Although it is
hard to estimate due to the different types (and size) of columns [11, 43],
a network of such modules would form a graph with millions of vertices
in case of humans. Unfortunately, such a network would be hard to draw
because apart from some cortical regions and specific columns (e.g., [56,
34, 60, 4, 55, 9]) the interconnections among these modules or cell clusters
are obscure. In addition, functional modules may not be fixed structures;
they could dynamically change their extension via neuronal plasticity (e.g.,
[11, 30]). It is noteworthy that the minicolumnar organization apparently
can not resolve the problem of defining structural and functional cortical
units, as momentarily the minicolumn seems to be a similarly vague concept
as the column [54, 18].

At a higher organizational level, the cortex is composed of a set (about a

hundred, roughly four orders of magnitude less than the number of columnar
modules in the human) of structurally and functionally specialized regions or
areas with highly variable shapes and sizes [67]. This level of organization
is of great interest because the available anatomical and imaging (fMRI,
PET, EEG, MEG) techniques made it possible to investigate the network
of cortical areas (hence neuro-cognitive functions) [49, 8, 40, 1, 28]. Most
of our knowledge about this large-scale cortical network comes from studies
charting the neuronal connections between cortical areas. Since the use of
sensitive and powerful tract tracing techniques is not feasible in humans, the
neural connections among the areas (“anatomical connectivity”, [58]) have
been studied intensely in non-human primates, especially in the macaque,
which serves as a model of the human cortex [67]. Large collections of
such data are available at the CoCoMac database [36] and for updates, one
may search PubMed [51]. Although the areas are connected to each other
via varying density of bundles of axonal processes [26] in a complicated
laminar and topographical pattern [20, 53], the network of areas is usually
represented in binary form considering only the knowledge of the existence
of a connection between the areas [57]. Even such a simplification allowed
328 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

the description of the fundamental properties of the network of cortical

areas, e.g., its small-world like characteristics and hierarchical organization
[27, 58, 31]. However, Kötter and Stephan [37] have pointed out that
the lack of information about connectivity can hinder the understanding
of important features of the cortical network.

The cortical network is directed, as long-range connections between

the areas end up in chemical synapses, but strongly reciprocal (reciprocity
reaches 80%) [20]. From the graph theoretic point of view, the high level of
reciprocity presents an obstacle by obscuring directedness in the network.
The global edge density is roughly 0.2–0.4 [57]. A granular organization
representing functional segregation and integration is prevalent also at this
large scale [68, 29, 25, 44, 46], resulting in the density of connections going up
to 0.6 or more within the clusters [15]. The other major characteristic that
makes the cortex a small-world network is that its clustered organization
is accompanied by a short average path length, roughly between 2 and
3 [58, 31]. Because it is reasonable to assume that a considerable part of
the large-scale cortical network is still unknown, the identification of the key
topological features that characterize this network, i.e. understanding its
organizational principles, remained an open issue [29, 50, 37, 32, 33, 15, 59].
A practical way of approaching this problem is to check how exactly the
network can be reconstructed by using a given index or network measure [29,
15]. This approach also has the interesting consequence of predicting missing
data, which can be verified experimentally. The two studies published up
to now present data on such predictions of yet unknown connections in the
cortex [29, 15]. The results of these studies (especially those by Costa et al.
[15], who investigated a broad set of measures) suggest that connectional
similarity of the areas is a good predictor in reconstructing the original
cortical network. However, they also report a relatively large number of
violations, where known existent connections were predicted as nonexistent
in the reconstructed graphs and vice versa [29, 15]. This suggests that
using other approaches could result in better reconstruction of the cortical
network. The aim of the present study was therefore to find a reconstruction
algorithm that predicts the large-scale cortical network more accurately,
i.e. with fewer violations. By considering the similarity of the connections
of the individual areas, our approach is reminiscent to that used by the
previous analyses [29, 15]. However, there are substantial differences as well,
especially the fact that we use a stochastic method which is able to take
into account the amount of uncertainty present in the data being analyzed.
Furthermore, contrary to the previous studies [29, 15], where the in- and
Reconstructing Cortical Networks 329

outputs are either taken into account separately [29] or the network was
symmetrized prior to analysis [15], our approach is principally dependent on
the combination of the areas’ in- and output pattern. Notably, considering
the similarity of the in- and output pattern as the result of the high number
of reciprocated links, a stochastic approach seems advantageous. Finally, in
contrast to Jouve et al. [29], who assumed that a large number of indirect
connections of path length 2 is suggestive of the existence of a direct link
between the areas, our method is free of such assumptions.

2. Methods

In this section, we introduce a simple stochastic graph model based on vertex

types and connection probabilities depending on them. From now on, we
call this the preference model. We will discuss a greedy and a Markov chain
Monte Carlo (MCMC) method for fitting the parameters of the preference
model to a network being studied. The MCMC method shows the rapid
mixing property. In Section 3, we will employ these methods to reconstruct
the cortical network and to predict previously unknown connections.

2.1. General remarks on the method

The problem we would like to solve and the proposed method are not cortex
specific, though the data on which we operate is. As data collection and
mapping are necessarily partial due to unavoidable observational errors, our
method offers the possibility to map interactions, connections, influences
based on the previous knowledge. Given the rough, but in principle cor-
rect summary of such information in the form of appropriate graph model,
one may refine the knowledge of underlying graph representation to some
extent. Applications and extensions of the solution we propose are straight-
forward to apply to any other network, with appropriate caution. The main
assumptions underlying our approach are as follows: the number of nodes is
known in advance, we only wish to predict previously uncharted edges, the
majority of the edges are known, yet a large number of undetected edges are
possible, at least in principle. As with most problems involving prediction,
it is relatively simple to create a model performing slightly better than ran-
dom tossing, but increasing prediction accuracy is a difficult problem. Our
330 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

approach to edge prediction is inspired by one of the most influential re-

sults of graph theory by Endre Szemerédi [61, 35, 64], which became known
as Szemerédi’s regularity lemma. Loosely speaking, the regularity lemma
states that the structures of very large graphs can be described by assigning
vertices to a small number of almost equally sized groups and specifying the
connection probabilities between the groups. The regularity lemma is for-
mulated as an asymptotical and existential statement. The graph we work
with is definitely small, not comparable in size to those graphs to which
one would in principle apply the regularity lemma. Thus our model can
be viewed as a form of fitting, with allowance for error. We do not try to
pretend that the assumptions of the regularity lemma apply to the case of
large-scale cortical networks, but the idea underlying the regularity lemma,
i.e. the probabilistic description of connections between and inside vertex
groups, is exercised in order to find a good reconstruction. The proposed
solution of the graph partitioning problem and its usage in edge prediction
is achieved by using probabilistic methods, which allow finding solutions
close to the optimum and satisfy the precision dictated by the practical
applications.

2.2. The preference model

This graph model starts from an empty graph with N vertices and it assigns
every vertex to one of K distinct groups. The groups are denoted by integer
numbers from 1 up to K. The generation process considers all pairs of
vertices once, and it adds an edge between node v1 and v2 with probability
pij if v1 belongs to group i and v2 belongs to group j. That is, the expected
density of edges between group i and group j is exactly pij , and the existence
of an edge between two vertices depends solely on the group affiliation of the
vertices involved. Fig. 1 shows two possible graphs generated by this model.
The one shown on Fig. 1(a) is a graph with clustered organization: vertices
in similar groups tend to link together while rarely linking to vertices of
the other group. The graph on Fig. 1(b) is a bipartite graph. The model
allows the simultaneous appearance of these two basic patterns in a graph:
pii ≈ 1 results in the one seen on Fig. 1(a) and pij ≈ 1, i = j induces the
one on Fig. 1(b). A method using similar ideas but designed for different
applications was also described in a recent paper of Newman and Leicht [48].
The generalization of the model to directed graphs is straightforward:
vertices of a directed graph will be assigned to an incoming and an outgoing
Reconstructing Cortical Networks 331

Fig. 1. Two graphs generated by the preference model. Black and white colours denote
the groups the nodes belong to. Panel (a) shows a clustered graph where the connection
probability between nodes within the same group is 0.2, while the connection probability
between nodes in different groups is only 0.02. Panel (b) shows a bipartite graph where
only nodes in different groups are allowed to connect with a probability of 0.2.

group (in-group and out-group in short), and the probability of the existence
of an edge between a vertex from out-group i and another vertex from
in-group j is given by pij . The number of parameters in this model is
K 2 + 2N + 1, since there are 2N parameters for the group affiliations of
the vertices, K 2 parameters represent the elements of the preference matrix
and the last parameter is K itself. The probabilities are usually arranged
in a probability matrix P for the sake of convenience. To emphasize the
role of directionality, elements of the preference matrix in the directed
case are sometimes denoted by pi→j instead of pij . We also introduce the
membership vectors u = [u1 , u2 , . . . , uN ] and v = [v1 , v2 , . . . , vN ], where
ui is the out-group and vi is the in-group of vertex i. From now on,
parameterizations of the model will be denoted by M = (K, u, v, P).

This model naturally gives rise to densely connected subnetworks with

sparse connections between them by appropriately specifying the connection
probabilities within and between groups. This is a characteristic property of
cortical networks, and it is assumed that a good reconstruction of the net-
work can be achieved by specifying vertex groups and connection probabil-
ities appropriately. More precisely, given a graph G(V, E) without multiple
or loop edges, the reconstruction task is equivalent to specifying the num-
332 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

ber of groups, finding an appropriate assignment of vertices to groups and

determining the elements of the probability matrix P. The reconstructed
graph then can be generated by the preference model, and new (previously
unknown) connections can also be predicted by checking the probabilities
of the uncertain edges in the fitted model. E.g., a crude reconstruction of
the visuo-tactile network of the macaque monkey (see Section 3 for details
about this dataset) would be a model with two groups (group 1 corresond-
ing to the visual and group 2 to the tactile vertices in the network) and
connection probabilities p1→1 = 0.385, p1→2 = 0.059, p2→1 = 0.035 and
p2→2 = 0.377, based on the density of connections between the groups in
the original network. The introduction of more vertex types results in a bet-
ter reconstruction, and obviously the reconstruction is perfect when N = K
and P is A, the adjacency matrix of the graph. However, such a recon-
struction is not able to predict unknown connections. We will discuss the
problem of overfitting in Section 2.5.

2.3. Measuring the accuracy of reconstruction

Since the preference model is a probabilistic model, every possible graph

with N vertices can theoretically be generated by almost any parameteriza-
tion of the model, but of course some graphs are more likely to be generated
by a specific parameterization than by others. Therefore, we measure the
fitness of a particular parameterization M = (K, u, v, P) with respect to a
given graph G(V, E) by its likelihood, i.e. the probability of the event that
the probabilistic model with parameters M generates G(V, E):

(1) L(M|G) = pui →vj (1 − pui →vj )
(i,j)∈E (i,j)∈E
/
i=j

The restriction i = j in the second product term corresponds to the

nonexistence of loop edges (even if they exist, they are ignored). To avoid
numerical errors when working with small probabilities, one can use the log-
likelihood instead, for the log-likelihood attains its maximum at the same
M where the likelihood does:
   
(2) log L(M|G) = log pui →vj + log 1 − pui →vj
(i,j)∈E (i,j)∈E
/
i=j
Reconstructing Cortical Networks 333

2.4. Fitting the preference model

Fitting a model to a given graph G(V, E) is equivalent to the maximum

likelihood estimation (MLE) of the parameters of the model with respect
to the graph, i.e. choosing M in a way that maximizes log L(M|G). Since
the number of possible group assignments is K N (where K is the number of
groups and N = |V | is the number of vertices), which is exponential in N ,
direct maximization of log L(M|G) by an exhaustive search is not feasible.
An alternative, greedy approach is therefore suggested to maximize the
likelihood.

2.4.1. Greedy optimization. Starting from an initial configuration

M(0) = (K, u(0) , v(0) , P(0) ), the greedy optimization will produce a fi-
nite sequence of model parameterizations M(0) , M(1) , M(2) , . . . satisfying
L(M(k) |G) ≥ L(M(k−1) |G) for k ≥ 1. First we note that the log-likelihood
of an arbitrary configuration M is composed of N local likelihood functions
corresponding to the vertices:


N 
N
 
(3) log L(M|G) = log (Aij pui →vj + (1 − Aij ) 1 − pui →vj )
i=1 j=1
j=i


N
= log Li (G|M)
i=1

where Aij is 1 if there exists an edge from i to j and 0 otherwise. Let

us assume first that K is given in advance. Starting from random initial
membership vectors u(0) and v(0) of M(0) , we can estimate an arbitrary
element pi→j of the real underlying probability matrix P by counting the
number of edges that originate from out-group i and terminate in in-group
j and divide it by the number of possible edges between out-group i and
in-group j. The estimated probabilities are stored in P(0) . After that, we
examine the local likelihoods Li (G|M(0) ) for all vertices and choose the in-
and out-groups of the vertices in a way that greedily maximizes their local
likelihood, assuming that the group affiliations of all other vertices and the
(0)
estimated probabilities remain unchanged. Formally, let ui=k denote the
vector obtained from u(0) by replacing the ith element with k and similarly
(0)
let vi=l denote the vector obtained from v(0) by replacing the ith element
334 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

with l. Let Mi,k,l = (K, ui=k , vi=l , P(0) ), and for every vertex i, every out-
(0) (0) (0)

group k and every in-group l, calculate log Li (G|Mi,k,l ). After that, put
(0)

vertex i in out-group k and in-group l if that maximizes log Li (G|Mi,k,l ).

(0)

Now calculate the next estimation of the probability matrix, P(1) , maximize
the local log-likelihoods based on the new probability matrix and repeat
these two alternating steps until u(k) = u(k−1) and v(k) = v(k−1) .

2.4.2. Markov chain Monte Carlo sampling. The group assignments

obtained by the greedy algorithm suffer from a minor flaw: they correspond
only to a local maximum of the parameter space and not the global one.
The local maximum means that no further improvement could be made by
putting any single vertex in a different group while keeping the group af-
filiations of all other vertices intact. However, there is the possibility of
improving the partition further by moving more than one vertex simulta-
neously. Another shortcoming of the algorithm is the danger of overfitting:
partitions with high likelihood might perform poorly when one tries to pre-
dict connections, because they are too much fine-tuned to the graph being
analyzed. Therefore we also consider employing Markov chain Monte Carlo
(MCMC) sampling methods [3] on the parameter space. (An alternative
MCMC-based data mining method on networks is presented in [13], but
while that method infers hierarchical structures in networks, our algorithm
is concerned with the discovery of densely connected subgraphs and bipar-
tite structures; see Fig. 1(a) and Fig. 1(b), respectively).
Generally, MCMC methods are a class of algorithms for sampling from
a probability distribution that is hard to be sampled from directly. These
methods generate a Markov chain whose equilibrium distribution is equiv-
alent to the distribution we are trying to sample from. In our case, the
samples are parameterizations of the preference model, and the distribution
we are sampling from is the following:

L(M0 |G)
(4) P(M = M0 ) = 
L(M |G) dM
SK

where SK is the space of all possible parameterizations of the probability

model for a given K. Informally, the probability of drawing M as a sample
should be proportional to its likelihood of generating G(V, E), for instance,
Reconstructing Cortical Networks 335

if M1 generates our network with a probability of 0.5 and M2 generates it

with a probability of 0.25, M1 should be drawn twice as frequently as M2 .
The generic framework of the MCMC method we use is laid down
in the Metropolis–Hastings algorithm [24]. The only requirement of the
algorithm is that a function proportional to the density function (that
is, P(M = M0 ) in (4)) can be calculated. Note that P(M = M0 ) ∝
L(M0 |G), since the denominator in (4) is constant. Starting from an
arbitrary random parameterization M(0) , MCMC methods propose a new
parameterization M based on the previous parameterization M(t) using
a proposal density function Q(M |M(t) ). If the proposal density function
is symmetric (Q(M |M(t) ) = Q(Mt |M )), the probability of accepting
the proposed parameterization is min (1, L(M |G)/L(M(t) ) | G). When
the proposal is accepted, it becomes the next state in the Markov chain
(M(t+1) = M ), otherwise the current state is retained (M(t+1) = M(t) ).
MCMC sampling can only approximate the target distribution, since
there is a residual effect depending on the starting position of the Markov
chain. Therefore, the sampling consists of two phases. In the first phase
(called burn-in), the algorithm is run for many iterations until the residual
effect diminishes. The second phase is the actual sampling. The burn-in
phase must be run long enough so that the residual effects of the starting
position become negligible.
A desirable property of a Markov chain in a MCMC method is rapid
mixing. A Markov chain is said to mix rapidly if its mixing time grows at
most polynomially fast in the logarithm of the number of possible states in
the chain. Mixing time refers to a given formalization of the following idea:
how many steps do we have to take in the Markov chain to be sure that
the distribution of states after these steps is close enough to the stationary
distribution of the chain? Given a guaranteed short mixing time, one can
safely decide to stop the burn-in phase and start the actual sampling after
the number of steps taken exceeded the mixing time of the chain.
Several definitions exist for the mixing time of a Markov chain (for an
overview, see [42]). To illustrate the concept, we refer to a particular variant
called total variation distance mixing time, which is defined as follows:

Definition 2 (Total variation distance mixing time). Let S denote the set
of states of a Markov chain C, let A ⊆ S be an arbitrary nonempty subset of
the state set, let π(A) be the probability of A in the stationary distribution
of C, and πt (A) be the probability of A in the distribution observed after
336 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

step t.
 The total variation
 distance mixing time of C is the smallest t such
that  πt (A) − π(A) ≤ 1/4 for all A ⊆ S and all initial states.

However, many practical problems have resisted rigorous theoretical

analysis. This applies also to the method presented here, mostly due to the
fact that the state transition matrix of the Markov chain (and therefore its
stationary distribution) is a complicated function of the adjacency matrix of
the network and the number of vertex groups, and no closed form description
exists for either. In these cases, a common approach to decide on the length
of the burn-in phase is based on the acceptance rate, which is the fraction
of state proposals accepted during the last m steps. Sampling is started
when the acceptance rate drops below a given threshold (a typical choice
is 20% or 0.2). Local maxima are avoided by accepting parameterization
proposals with a certain probability even when they have a lower likelihood
than the last one, but being biased at the same time towards partitions with
high likelihoods. In the case of multiple local maxima with approximately
the same likelihood, MCMC sampling tends to oscillate between those local
maxima. By taking a large sample from the equilibrium distribution, one
can approximate the probability of vertex i being in out-group k and in-
group l and extract the common features of all local maxima (vertices that
tend to stay in the same groups despite randomly walking around in the
parameter space).
The only thing left to clarify before employing MCMC sampling on
fitting the preference model is the definition of an appropriate symmetric
proposal density function. We note that the number of groups K is constant
and the probability matrix P can be approximated by the edge densities for
a given out- and in-group assignment, leaving us with only 2N parameters
that have to be determined. We take advantage of the fact that the condi-
tional distribution of each parameter (assuming the others are known) can
be calculated exactly as follows:

Li (G|Mi,k,∗ )
(5a) P(ui = k) =
K
l=1 Li (G|Mi,l,∗ )

Li (G|Mi,∗,k )
(5b) P(vi = k) =
K
l=1 Li (G|Mi,∗,k )

where Mi,k,∗ = (K, ui=k , v, P) and Mi,∗,k = (K, u, vi=k , P). Since the
conditional distribution of each parameter is known, Gibbs sampling [23]
can be used. The Gibbs sampling alters a single variable of the parameter
Reconstructing Cortical Networks 337

vector in each step according to its conditional distribution, given all other
parameters. It can be shown that the proposal distribution defined this way
is symmetric if the variable being modified is picked randomly according to
a uniform distribution. In practice, it is sufficient to cycle through the
variables in a predefined order as long as the Markov chain can access all
states under this ordering. To speed up the burn-in process, one can apply
the greedy optimization method described in Section 2.4.1 and revert to the
MCMC sampling when the algorithm reached the first local maximum.

2.5. Choosing the number of groups

As mentioned earlier in Section 2.2, the key parameter that controls the
balance between accurate reconstruction and meaningful prediction is the
number of vertex groups used in the preference model. A very small number
of groups yields an inaccurate reconstruction and most likely meaningless
predictions. Increasing the number of groups gradually improves the accu-
racy of reconstruction, attaining perfection when the number of groups is
equal to the number of vertices, but in this case no new edges are predicted.
This is the classical problem of overfitting: by increasing the number of
groups, the ability of the model to generalize beyond the original data di-
minishes. Therefore, the goal is to select the number of groups in a way that
achieves good reconstruction while still allowing the model to predict con-
nections by assigning a high probability to vertex pairs where an uncertain
connection is suspected.
We tried multiple approaches to infer the appropriate number of groups
in the networks we studied. The exact results will be discussed in Sec-
tion 2.6.2 and Section 3; here we only outline the basic ideas. We will
make use of the eigenvalues of the Laplacian matrix of the graph, the sin-
gular value decomposition (SVD) of the adjacency matrix and the Akaike
information criterion [2].
Given an undirected graph G(V, E) without loops and multiple edges,
its Laplacian matrix is defined as L = D − A, where A is the adjacency
matrix and D is a diagonal matrix composed of the degrees of the vertices.
A basic property of the Laplacian matrix is that its smallest eigenvalue is
zero, and its multiplicity is equal to the number of connected components
of the graph. The number of eigenvalues close to zero is frequently used for
determining the number of dense subgraphs (communities, clusters) in the
graph and, based on similar reasoning, this could be a good estimate of the
338 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

number of groups that have to be used in the preference model; however, we

cannot use D − A directly, since this form of the Laplacian is defined only
for undirected graphs. An extension of the Laplacian to directed graphs
was introduced in [12]. This involves calculating the Perron vector φ of
the transition probability matrix P of the graph. The transition probability
matrix P is derived from the adjacency matrix by normalizing the row sums
to be 1. The Perron vector φ is a unique (up to scaling) left eigenvector
of P satisfying φP = φ. The existence of this vector is guaranteed by the
Perron–Frobenius theorem. There is no closed-form solution for φ, but it is
easy to calculate in polynomial time numerically. The directed Laplacian is
then defined as:

Φ1/2 PΦ−1/2 + Φ−1/2 P∗ Φ1/2

(6) L=I−
2
where P∗ is the conjugate transpose of P and
Φ is a diagonal matrix com-
n
posed of the elements of φ, assuming that i=1 φi = 1. The properties
emphasized above for the undirected Laplacian hold for the directed Lapla-
cian as well.
The singular value decomposition of an m×n matrix M is a factorization
process that produces an m × m and an n × n unitary matrix (U and V,
respectively) and an m × n matrix Σ with non-negative numbers on the
diagonal and zeros off the diagonal in a way that M = UΣV∗ . The diagonal
of Σ contains the singular values, while the columns of U and V are the left
and right singular vectors, respectively. Plotting the singular values on a
scree plot (sorted from large to small) is a good visual cue to determining the
number of groups in the model: the number of groups can simply be assigned
according to the number of large singular values. It is noteworthy that one
can approximate the original matrix M by setting all singular values other
than the l largest to zero and disregarding the appropriate rows of U and
V that correspond to the zeroed singular values. The remaining parts of
U and V can serve as an input for a k-means clustering algorithm in an
l-dimensional space, and the results of the clustering yield a good candidate
of an initial position of the greedy optimization process of the preference
matrix. In practice, however, performing a complete SVD is less efficient
than optimization from a random initial position.
The Akaike information criterion (AIC) [2] is a measure of the good-
ness of fit of a statistical model (the preference model in our case). It is
an unbiased estimator of the Kullback–Leibler divergence [38], and it is an
operational way of determining the appropriate trade-off between the com-
Reconstructing Cortical Networks 339

plexity of a model and its predictive power. AIC is calculated as 2k−2 log L,
where k is the number of parameters in the model and L is the likelihood.
In the preference model, k = K 2 + 2N + 1. The suggested number of groups
can be determined by fitting the model with various numbers of groups and
choosing the one that minimizes the Akaike information criterion.

The AIC can also be used to detect situations when the network be-
ing studied is in fact completely random, and therefore its appropriate
description is simply an Erdős–Rényi random graph model instead of the
preference model. This is done by estimating the probability parameter
p of the Erdős–Rényi model from the edge density of the network and
then calculating the log-likelihood of the network according to the Erdős–
Rényi model. Given a directed network with n vertices and m edges, the
m
maximum likelihood estimator of p is n(n−1) , resulting in a log-likelihood
of m log p + (n2 − n − m) log (1 − p) (assuming that there are no loop
edges).
 The baseline AIC corresponding to the Erdős–Rényi model is then
2 − 2 m log p + (n − n − m) log (1 − p) , since the model has only a sin-
2

gle parameter. If the network being studied is completely random, the AIC
corresponding to the case of two groups will be larger than the baseline AIC
of the Erdős–Rényi model, for we introduced more parameters without ac-
tually improving the likelihood. On the other hand, networks possessing a
structure that can be described by the preference model will show signifi-
cant improvement in the log-likelihood compared to the pure random case,
resulting in a lower AIC.

2.6. Performance measurements

To demonstrate the validity of the fitting algorithms presented above, we

conducted several benchmarks on computer-generated test graphs. First,
we generated graphs according to the preference model, ran the fitting
algorithm on the graphs by supplying the appropriate number of groups
beforehand and then compared the known and the estimated parameters of
the model. These benchmarks were performed in order to test the validity of
the fitting algorithm and to assess the quality of the results obtained. Next,
we ran the fitting algorithms without specifying the number of groups to
show that the Akaike information criterion is suitable for determining the
right value of k.
340 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

2.6.1. Fitting the model with given number of groups. This bench-
mark proceeded as follows: graphs with 128 vertices were generated accord-
ing to the preference model using 4 in- and out-types. The type distribution
was uniform, so there were 32 vertices of each type on average. The pref-
erence matrix was chosen as follows: each element pij was set to one of
two predefined values p1 and p2 with probability 0.5. p1 and p2 was var-
ied between 0 and 1 with a step size 0.05. For each (p1 , p2 ) combination,
we generated 50 graph instances using the preference model. Values of the
quality functions (described below) were averaged over these instances and
the results were plotted as a function of p1 and p2 . We used only two prob-
abilities because the results can then be visualized on a heat map or a 2.5D
plot.
To assess the fitness of the fitted model, we had to define some quality
functions that compare the fitted parameters to the original (expected) ones.
First we note that the number of groups and the probability matrix do not
have to be compared, since the former is fixed and the latter one is calculated
from the group assignments, so errors in the elements of the probability
matrices are simply due to errors in the group assignments. Therefore, only
the group assignments matter. The following quality functions were defined:
Normalized mutual information of the confusion matrix.
This measure was suggested by Fred and Jain [21] and later applied to
community detection in graphs by Danon et al. [16]. The measure is
based on the confusion matrix C = [cij ] of the expected and observed
group assignments. cij is the number of vertices that are in group i in
the original and group j in the fitted model. The confusion matrix can
be calculated separately for in- and out-groups, but they can safely be
added together to obtain a single confusion matrix and then a single
quality measure, which is the normalized mutual information of the
confusion matrix:

k
k c c ∗∗
i=1 j=1 cij log cij
i∗ c∗j
(7) I(C) = −2
 
k
i=1 ci∗ log cc∗∗
i∗
+ c∗i log cc∗∗
∗i

where ci∗ is the sum of the i-th row, c∗j is the sum of the j-th column of
the confusion matrix. c∗∗ is the sum of cij for all i, j. It is assumed that
0 log 0 = 0. When the fitted group assignment is completely identical
to the expected one (apart from rearrangement of group indices), I(C)
attains its maximum at 1. I(C) = 0 if the two group assignments
Reconstructing Cortical Networks 341

are independent. Danon et al. [16] argue that this measure is in

general stricter than most other quality measures proposed so far.
For instance, a completely random assignment of groups still has an
expected success ratio of 0.25 for 4 groups (since each pair is consistent
with probability 1/4). In this case, the normalized mutual information
is close to zero, which is a more intuitive description of what happened
than a success ratio of 0.25. See the paper of Danon et al. [16] for a
list of other measures they considered.
Likelihood ratio. This measure is simply the ratio of the likelihoods of the
original and the fitted parameterizations, given the generated graph.

The likelihood ratios and the mutual information indices are plotted on
Figure 2. As expected, the mutual information index is low when p1 ≈ p2 .
This is no surprise, since p1 ≈ p2 implies that the actual difference between
different vertex types diminish: they all behave similarly, and the random
fluctuations at this network size render them practically indistinguishable.
The overall performance of the algorithm is satisfactory in the case of
p1 ! p2 and p1 $ p2 , with success ratios and mutual information indices
larger than 0.9 in all cases. In cases when p1 ≈ p2 , the likelihood ratio is
greater than 1, which indicates that the fitted model parametrization is more
likely than the original one. This phenomenon is an exemplar of overfitting:
apparent structure is detected by the algorithm where no structure exists
at all if we use too many groups.

Fig. 2. Mean likelihood ratios of the fitted parameterizations to the expected ones (left)
and mean normalized mutual information conveyed by the confusion matrices (right) as
a function of p1 and p2 .
342 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

2.6.2. Fitting the model without a predefined number of groups. In

Section 2.5, we described three different methods for estimating the number
of groups one should use for a given network when fitting the preference
model. Two of these methods requires some human intervention, since one
had to choose a threshold manually for the eigenvalues of the Laplacian
matrix or for the singular values of the adjacency matrix.
We investigated the eigenvalues of the directed Laplacian matrix first.
After some experiments on graphs generated according to the preference
model, it became obvious that the number of eigenvalues of the Laplacian
close to zero correlate to the number of groups only if the vertex groups
coincide with densely connected subgraphs. In other words, pii must be
large and pij for i = j must be small. This is illustrated on Figure 3.
The left panel shows the case when pij = 0.2 + 0.6 δ(i, j) (the graph is
clustered) and the right panel shows the case when pij is 0.2 or 0.8 with
1/2 probability. There is indeed a relatively large jump after the eighth
eigenvalue for the former case, but the transition is smooth for the latter.
Therefore, the eigenvalues of the directed Laplacian matrix were excluded
from further investigations.

Fig. 3. The 15 smallest nonzero eigenvalues of the Laplacian matrix for graphs
generated by the preference model with 8 groups, either with or without a strong
clustered structure (left and right panel, respectively)

In the case of SVD analysis, one has to count the large singular values.
“Large” is definitely a subjective term, therefore a scree plot of the singular
values is often used as a visual aid. The scree plot is simply a bar graph of
the singular values sorted in their decreasing order. The plot usually looks
like the side of a mountain with some debris at the bottom: the singular
values decrease rapidly at first, but there is an elbow where the steepness
of the slope decreases abruptly, and the plot is almost linear from there on
(see Figure 4 for an illustration). The number of singular values to the left
Reconstructing Cortical Networks 343

of the elbow is the number of groups we will choose. To allow for automated
testing, we implemented a simple method to decide on the place of the elbow.
The approach we used is practically equivalent to the method of Zhu and
Ghodsi [69]. It is based on the assumption that the values to the left and
right of the elbow behave as independent samples drawn from a distribution
family with different parameters. The algorithm first chooses a distribution
family (this will be the Gaussian distribution in our case), then considers all
possible elbow positions and calculates the maximum likelihood estimation
of the distribution parameters based on the samples to the left and right
side of the elbow. Finally, the algorithm chooses the position where the
likelihood was maximal. Assuming Gaussian distributions on both sides,
the estimates of the mean and variance are as follows:

q
n
i=1 xi i=q+1 xi
(8) μ̃1 = μ̃2 =
q n−q

q 2
n 2
2 i=1 (xi − μ1 ) + i=q+1 (xi − μ2 )
σ̃ =
n−2

where xi is the i-th element in the scree plot (sorted in decreasing order),
n is the number of elements (which coincides with the number of vertices)
and q is the number of elements standing to the left of the elbow. Note
that the means of the Gaussian distributions are estimated separately, but
the variance is common. Zhu and Ghodsi [69] argue that allowing different
variances makes the model too flexible. The common variance is calculated
by taking into account that the first q elements are compared to μ1 and the
remaining ones are compared to μ2 . See the paper of Zhu and Ghodsi [69]
for a more detailed description of the method.

Fig. 4. The largest 20 singular values of the adjacency matrix of a graph generated by
the preference model with 8 groups
344 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

In this benchmark, 100 networks were generated with 128 vertices each.
Elements of the preference matrix were chosen to be p1 or p2 with equal
probability, as in Section 2.6.1 before, but the case of p1 ≈ p2 was avoided
by constraining p1 to be above 0.6 and p2 to be below 0.4. The number
of groups was varied between 2 and 8 according to a uniform distribution.
The number of groups in the fitted model was estimated by the SVD and
the AIC methods, the best AIC was chosen by trying all possible group
counts between 2 and 10. The AIC method proved to be superior to the
SVD method: the estimation was perfect in 79% of the cases. The number
of groups was underestimated by 1 group in 14, 2 groups in 3 and 3 groups
in 2 cases. There were 2 overestimations by 1 group as well, resulting in a
mean squared error of 0.46 groups. On the other hand, the SVD method
made severe mistakes at times; in fact, only 7% of its estimations matched
the prior expectations, all other cases were overestimations, sometimes by 7
or 8 groups. This is due to the unsupervised choice of the elbow in the scree
plot. It is assumed that better results can be achieved by making the choice
manually, therefore the conclusion is that the SVD-based estimation should
be handled with care and the AIC method is preferred when one would like
to choose the number of groups automatically.

2.7. Handling uncertain connections

Despite being concerned about predicting unknown connections in a network

where some parts are uncertain, we only discussed fitting the preference
model to a graph where all connections were known and all uncertain
connections were assumed to be nonexistent. As a refinement of the model,
we can include our a priori assumption about the probability of the event
that a particular, presently uncharted connection exists in the network. Let
us denote by bi→j our degree of belief in the existence of an edge going from
vertex i to j. We write bi→j = 1 if we are fully convinced that the edge
actually exists and bi→j = 0 for edges that are known to be nonexistent.
Intermediary values of bi→j can be thought about as probabilities, e.g.,
bi→j = 0.3 means that the probability of an edge from vertex i to j is 0.3.
(Note that bi→j acts as a generalization of Aij : bi→j is 1 if and only if we
are convinced that Aij = 1, bi→j is 0 if and only if we are convinced that
Aij = 0. Uncertain connections result in 0 < bi→j < 1). In this sense,
not only our model but the graph being fitted is also probabilistic, and we
are trying to find the model whose expected likelihood with respect to the
Reconstructing Cortical Networks 345

whole ensemble of possible graphs parameterized by the degrees of belief

is maximal. All the optimization methods described earlier also work in
this case, only the likelihood and the log-likelihood functions have to be
adjusted:

N
N
  
(9a) L(M|G) = (bi→j pui →vj + 1 − bi→j 1 − pui →vj )
i=1 j=1, j=i


N 
N
  
(9b) log L(M|G) = log (bi→j pui →vj + 1 − bi→j 1 − pui →vj ).
i=1 j=1, j=i

The elements of the optimal probability matrix P can then be thought

about as the posterior probabilities of the edges in the network. An edge
whose prior probability is significantly lower than its posterior probability
is then likely to exist, while connection candidates with significantly higher
prior than posterior probabilities are likely to be nonexistent.

3. Results and Discussion

In this section, we will present our results on application of the preference

model to the prediction of unknown connections in the visual and sensori-
motor cortex of the primate (macaque monkey) brain.
The dataset we are concerned with in this section is a graph model of the
visuo-tactile cortex of the macaque monkey brain. Connectivity data was
retrieved from the CoCoMac database [36] and it is identical to the dataset
previously published in [44]. The whole network contains 45 vertices and
463 directed links among them. The existence of connections included in
the network were confirmed experimentally, while connections missing from
the network were either explicitly checked for and found to be nonexistent,
or never checked experimentally. To illustrate the uncertainty in the dataset
being analyzed, we note that 1157 out of the 1980 possible connections were
uncertain (never checked experimentally) and only 360 were known to be
absent.
The network consists of two dense subnetworks corresponding to the vi-
sual and the sensorimotor cortex (30 and 15 vertices, respectively). The
visual cortex can also be subdivided to the so-called dorsal and ventral
346 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

parts using a community detection algorithm based on random walks [39].

Most of the uncertain connection candidates are heteromodal (originating
in the visual and terminating in the sensorimotor cluster, or the opposite),
and it is assumed that the vast majority of possible heteromodal connec-
tions are indeed nonexistent. The basic properties of these networks are
shown in Table 1, while the adjacency matrix of the visuo-tactile network
is depicted on Fig. 5. Note that since the visual and sensorimotor cor-
tices are subnetworks of the visuo-tactile networks, their adjacency ma-
trices are the upper-left 30 × 30 and lower-right 15 × 15 submatrices of
the adjacency matrix of the visuo-tactile cortex. In order to compare our
results with previous reconstruction attempts that were only concerned
with the visual cortex [29, 15], we will present results based on the vi-
sual subnetwork as well as the whole visuo-tactile cortex. Data analysis
was performed using the open source igraph network analysis library [14]
(http://cneurocvs.rmki.kfki.hu/igraph).

Visual Sensorimotor Visuo-tactile

Vertices 30 15 45
Known connections (edges) 335 85 463
Known nonexistent connections 310 0 360
Unknown connections 225 125 1157
Density 0.385 0.404 0.233
Density (excl. unknowns) 0.519 1.000 0.548
Diameter 3 3 5
Average path length 1.6632 1.767 2.149
Reciprocity 0.850 0.888 0.815
Table 1. Basic properties of the original networks

3.1. Rapid mixing of the MCMC process

First, we illustrate the rapid convergence of the MCMC process to the

equilibrium distribution. This property is crucial, since a chain with short
mixing time tends to get close to its equilibrium distribution quickly, thus
ensuring that a short burn-in period is sufficient. Since a rigorous proof
of the mixing time of the Markov chain designed for fitting the preference
model is well beyond the scope of this chapter, and we are mostly concerned
with its applicability to the visual and visuo-tactile networks, we check the
fast convergence of the method by plotting the log-likelihood of the states of
the Markov chain and the acceptance rate as the fitting progresses from an
Reconstructing Cortical Networks 347

Fig. 5. Adjacency matrix of the visuo-tactile cortex dataset. Black cells denote known
existing connections, white cells denote known nonexistent connections. Gray cells are
connections not confirmed or confuted experimentally. The upper left 30 × 30 submatrix
is the adjacency matrix of the visual cortex, the lower right 15 × 15 submatrix describes
the sensorimotor cortex.

arbitrary random starting position. Fig. 6 illustrates that the chain mixes
rapidly, reaching its equilibrium distribution in roughly n2 steps, where n is
the number of vertices in the network. This satisfies the criterion of rapid
mixing, since the number of possible states in the Markov chain is k2n (k
is the number of groups), therefore the mixing time is polynomial in the
logarithm of the number of states. The number of groups was fixed at 7 for
the visual and 10 for the visuo-tactile cortex, these choices will be explained
348 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

later. Unknown edges were treated as nonexistent. Fig. 6 suggests that one
can start sampling from the Markov chain after roughly n2 steps or after
the acceptance rate drops below 0.2.

Fig. 6. Log-likelihood of the states of the Markov chain (left) and acceptance rates in a
window of 100 samples (right) for the visual and visuo-tactile cortices, normalised by n2 ,
on a logarithmic time scale

3.2. Methodological comparison with other prediction approaches

Our method allows prediction of nonreciprocal connections, and the network

data is not symmetrised for the sake of computational and methodological
tractability, in contrast to [15]. Furthermore, we only use the connectional
data for prediction,, no other anatomical facts were taken into account. An
approach where additional neuroanatomical facts were used as predictional
input is described in [15]. Jouve et al. [29] use a specific property of the
visual cortex: the existence of indirect connections of length 2 between
areas are presumed to support the existence of a direct connection. This
property need not hold for other large cortical structures, especially when
investigating the interplay of different cortices (e.g., the visual and the
sensorimotor cortices). In fact, this assumption is difficult to prove or
disprove due to the poor knowledge of connection structure in other parts of
the cortex. The problem is even more pronounced in the case of heteromodal
connections, thus other guiding principles had to be sought.
Reconstructing Cortical Networks 349

3.3. Visual cortex

Since the visual cortex is a part of the visuo-tactile cortex, the adjacency
graph of the visual cortex can be found in Fig. 5 as the upper left 30 × 30
submatrix. It is noteworthy that most of the unknown connections are
adjacent to the areas VOT and V4t, and the subgraph consisting of the
vertices PITd, PITv, CITd, CITv, AITd, AITv, STPp and STPa (all
belonging to the ventral class) is also mostly unknown. Based on the
connection density of the visual cortex (assuming unknown connections to
be nonexistent), the probability of the existence of a connection classified as
unknown was set to 0.385. These degrees of belief were taken into account in
the likelihood function as described in Sect. 2.7. The search for the optimal
configuration started from a random initial position, first improved by a
greedy initial phase, then followed by MCMC sampling after reaching the
first local maximum. The sampling process was terminated when at least
106 samples were taken from the chain. The sample with the best likelihood
became the final result.

The optimal number of groups in the preference model was determined

by studying the eigenvalues of the Laplacian and the singular values of
the adjacency matrix (unknown connections were treated as nonexistent)
as well as the Akaike information criterion of the obtained partitions at
various group numbers from 2 to 15. Partitions having more than 15 groups
do not seem feasible, since in these cases, at least one of the groups will
contain only one vertex. The eigenvalues and the singular values are shown
on Fig. 7. A visual inspection suggests using only two groups (which is
congruent with the anatomical fact that the visual cortex is composed of
two major pathways, namely the dorsal and the ventral stream), but the
minimal AIC value was achieved using 7 groups (see Table 2). Since two
groups are intuitively insufficient for an accurate reconstruction, we decided
to use 7 groups in the rest of the analysis. This is further supported by the
mediocre success rate of the model with only two groups. Success rates were
calculated as follows: for every possible threshold τ between 0 and 1 (with
a granularity of 0.01), the percentage of known edges that had a predicted
probability greater than τ was calculated. The final threshold used for
calculating the success rate was chosen to be the one that produced the
highest ratio of correctly predicted known edges. τ fluctuated around 0.5 in
all cases. As it was expected based on our reasoning outlined in Sect. 2.5,
the success rate increased steadily as we increased the number of groups,
350 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

but the divergence of τ from 0.5 after having more than 7 groups is likely
to be a precursor of overfitting.

Fig. 7. Eigenvalues of the Laplacian and singular values of the adjacency matrix of the
visual cortical graph

K Log-likelihood AIC τ Success rate

2 −481.608 1091.216 0.50 80.7%
3 −440.280 1018.560 0.50 82.4%
4 −413.027 978.055 0.50 84.2%
5 −394.664 959.328 0.50 85.1%
6 −378.271 948.543 0.50 87.4%
7 −363.146 944.292 0.50 87.6%
8 −353.071 954.143 0.50 88.5%
9 −340.886 963.773 0.47 89.6%
10 −331.626 983.253 0.43 90.2%
11 −319.771 1001.543 0.49 90.7%
12 −307.766 1023.532 0.48 91.5%
13 −300.657 1059.315 0.48 91.9%
14 −297.540 1107.081 0.46 92.0%
15 −288.615 1147.231 0.49 92.4%
Table 2. Likelihoods, AIC values and success rates in the visual cortex

The fitted model with 7 groups provided probabilities for the 225 un-
known connections, 137 of them were above the optimal threshold τ = 0.5.
The ratio of predicted edges approximately matches the density of the vi-
sual cortex when we exclude the unknown connections from the density
calculation (see Table 1). However, if we wanted the ratio of predicted con-
nections to match the density of known connections in the visual cortex, we
would have to increase τ to 0.654, predicting only 81 connections. This ra-
tio matches the one reported in [15], although the connection matrix in [15]
included an additional area in the analysis. The predicted adjacency matrix
Reconstructing Cortical Networks 351

with τ =0.654 is shown on Fig. 8 and its basic descriptive graph measures
are to be found in Table 3.

Fig. 8. The predicted adjacency matrix of the visual cortex with 7 vertex groups. White
cells denote confirmed existing and absent connections. Dark gray cells denote
mismatches between the known and the predicted connectivity. Light gray cells denote
predictions for unknown connections.

Besides the overall success rate, we also calculated the ratio of correctly
predicted 1’s and 0’s (R1 and R0 , respectively) for the case of K = 7. With
τ = 0.5, 84.8% of known 0’s and 93.1% of known 1’s √ were predicted correctly
(R0 = 0.848, R1 = 0.931). The geometric mean ( R0 R1 ) was 0.888, which
dropped to 0.8245 when raising τ to 0.654 (R0 = 0.925, R1 = 0.734), thus
a higher τ seems to be better at reconstructing non-existing connections.
352 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

Predicted visual Predicted visuo-tactile

Vertices 30 45
Edges 358 757
Density 0.412 0.382
Diameter 3 4
Average path length 1.478 1.833
Reciprocity 0.517 0.645
Reciprocity of predicted connections 0.674 0.824

Table 3. Basic properties of the predicted networks

We compared our results to earlier studies [29, 15]. Comparisons were

based on the percentage of matching predictions. Since both studies took
a slightly different sets of areas into consideration, we did not take into
account those areas that were not present in any of the matrices.
The predictions of Jouve et al. [29] are based solely on topological
features of the network model of the visual cortex, similarly to the method
presented here. The agreement between the two predicted matrices is
moderate: 61.6% of the predictions match for τ = 0.5 and only 47% for
τ = 0.654. Most of the disagreements involved areas V4t (28), VOT (22),
FEF (17) and DP (16). Area MSTd was joined together with MSTp in
our study (resulting in the vertex denoted by MSTd/p), therefore neither
MSTd nor MSTd/p was taken into account. We note that the matrix used
in the present paper incorporated the results of anatomical experiments that
could not have been included in the matrix in [29], therefore the moderate
match between the two matrices can be explained by the differences in the
initial dataset, see Table 4 for the number of mismatches involving each
area. Since the prediction method of Jouve et al. was not concerned with
reconstructing the entire network (predictions were made only on unknown
connections), no comparison could be made based on the success rates of
the two methods.
The predictions published by Costa et al. [15] are based on several topo-
logical (e.g., node degree, clustering coefficient) and spatial features (e.g.,
area sizes, local density of the areas in the 3D space, based on their known
positions in the cortex). In this sense, the reconstruction method based on
the preference model is simpler, for it depends solely on the connection ma-
trix. We also note that Costa et al. inferred the topological features from a
symmetrized connectivity matrix, thus their predicted matrix is also com-
pletely symmetric, while our method produced a matrix where only 67.4% of
the connections (67.4% of the predicted, previously unknown connections)
Reconstructing Cortical Networks 353
Known connections Jouve et al. [29] Costa et al. [15]
V1 5 0 0
V2 5 1 0
V3 6 0 0
V3A 1 4 1
V4 21 0 0
V4t 2 28 15
VOT 7 22 18
VP 10 0 –
MT 3 2 0
MSTd/p 5 – –
MSTl 7 7 0
PO 10 3 1
LIP 8 3 3
PIP 5 13 4
VIP 4 6 3
DP 12 16 8
7a 9 10 7
FST 11 8 1
PITd 6 6 2
PITv 4 7 6
CITd 2 9 3
CITv 3 8 5
AITd 7 2 1
AITv 4 8 3
STPp 11 9 8
STPa 6 12 3
TF 13 8 11
TH 6 8 2
FEF 8 17 3
46 15 7 9
Table 4. Number of mismatches in the predicted matrix, grouped by areas.
In the second column, the known connections of the original matrix are compared to our
predictions. In the last two columns, only the unknown (predicted) connections are
compared to the unknown connections of our dataset. The 4 largest number of
mismatches in each column are underlined.

were reciprocal. The ratios of correctly predicted 1’s and 0’s in the visual
cortex reported by Costa√et al. were slightly worse (R0 = 244/350 = 0.697,
R1 = 207/295 = 0.701, R0 R1 = 0.699, loop connections excluded). Note
that the comparison can not be fully accurate because of the slightly dif-
ferent set of areas used in the analysis (MIP and MDP were present only
in [15], whereas MSTd/p and VP were present only in the matrix used in
354 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

this study). 69.8% of the predictions presented here matched the predic-
tions of [15], and all predicted edges with a probability larger than 0.8 were
predicted in [15] as well.
One may note that in spite of the improvement of the reconstruction
as compared to the previous studies [29, 15], there is still a relatively high
number of mismatches on Fig. 8. The distribution of mismatches in the ad-
jacency matrix can be suggestive of the methodological shortcomings and
the state of knowledge in the investigated network. It appears that most
of the mismatches are to be found within the two major visual clusters,
the dorsal and ventral visual subsystems, where connectional densities are
higher than in the lower left and upper right quadrants of the matrix repre-
senting the intercluster connections. Interestingly, most of the mismatches
affected either the input or output patterns of areas V4 and 46, and to a
lesser degree of TF and FEF in the intercluster regions. These areas are
central nodes in the visual cortical network, establishing connections be-
tween different clusters. In fact, the inclusion of the sensorimotor cortex
improved the reconstruction (see Section 3.4). It is also noteworthy that
relatively few mismatches/violations occurred in case of the lower order ar-
eas (listed in the upper left corner of the matrix). This is an important point
as low-level areas establish connections mostly within their cluster and the
connections of these areas are relatively well explored. These observations
indicate the dependence of reconstruction quality on the actual knowledge
of the network.
To summarize without going into the details, we conclude that our
reconstruction is biologically realistic and reflects our understanding of the
organization of the visual cortical connectivity.

3.4. Visuo-tactile cortex

The network model of the visuo-tactile cortex is an extension of the visual

cortex, obtained by adding the 15 areas of the sensorimotor cortex and
their respective connections. Connections going between a visual and a
sensorimotor area are called heteromodal connections. The density of the
sensorimotor cortex is slightly higher than that of the visual cortex. Based
on the connection densities, the probability of the existence of an unknown
connection was assumed to be 0.385 inside the visual cortex and 0.404 inside
the sensorimotor cortex. Unknown heteromodal connections were assumed
to exist with probability 0.1. Note that the vast majority of heteromodal
Reconstructing Cortical Networks 355

connections is unknown. There was no confirmed nonexisting sensorimotor

connection indicated in the data set. The adjacency matrix is shown on
Fig. 5. The optimal configuration was found by combining the greedy
optimization with the MCMC method, similarly as above.
The number of groups in the preference model was determined again by
the Akaike information criterion. The eigenvalues of the Laplacian and the
singular values of the adjacency matrix suggested 5 groups, which is again in
concordance with anatomical considerations, but as shown above, 5 groups
were insufficient to reproduce only the visual cortex (part of the visuo-tactile
cortex). Log-likelihoods, AIC values and success rates are shown in Table 5,
from 5 to 15 groups. The optimal number of groups with the lowest AIC
was 10.
K Log-likelihood AIC τ Success rate
5 −814.956 1859.913 0.42 83.6%
6 −783.935 1819.871 0.23 84.4%
7 −756.352 1790.705 0.46 84.8%
8 −736.163 1780.327 0.37 86.1%
9 −718.422 1778.844 0.43 86.4%
10 −697.078 1774.156 0.49 87.3%
11 −683.335 1788.671 0.46 89.3%
12 −684.105 1836.210 0.46 89.3%
13 −665.337 1848.674 0.47 89.4%
14 −653.755 1879.510 0.48 89.4%
15 −652.173 1934.347 0.40 90.1%
Table 5. Likelihoods, AIC values and success rates in the visuo-tactile cortex

The fitted model with 10 groups predicted 225 connections √ with τ =

0.47 out of the 1157 unknown ones (R0 = 0.883, R1 = 0.892, R0 R1 =
0.887). This is 19.4% of the unknown connections and it roughly matches
the overall density of the visuo-tactile cortex (23.3%). However, only 5
heteromodal connections (all originating from LIP) were predicted apart
from the known existing ones. This is due to the fact that very little
is known about the heteromodal connections, and the algorithm cannot
generalize beyond them with higher confidence. We also note that the
posterior probability of many heteromodal connections in this case stayed
at 0.1, the same as their prior probability. Taking into account that even
a significant difference between the prior and the posterior probabilities
of the heteromodal connections may not reach the threshold of 0.49, we
decided to use different thresholds for non-heteromodal and heteromodal
connections (τ1 and τ2 , respectively). τ1 was left at 0.49, while τ2 was
356 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

lowered to 0.137, the average a posterori probabilities of the unknown

heteromodal √ connections. This new configuration yielded R0 = 0.831,
R1 = 0.927, R0 R1 = 0.877 and 132 predicted heteromodal connections,
related mainly to areas LIP, VIP, DP, 7a, FST, TF, FEF and 46 in the visual
cortex. It is noteworthy that four of these areas (46, 7a, LIP and VIP) were
classified as structural overlaps between the two subnetworks in the fuzzy
community analysis of Nepusz et al. [46]. Anatomical considerations also
support the bridge-like role of these areas between the cortices. It was
previously suggested in the literature that area VIP should be split into two
areas (VIPm and VIPp), establishing stronger connections with visual or
sensorimotor areas, respectively [41]. VIP and LIP are involved with hand
and eye coordination, respectively, requiring a combined input of visual and
tactile signals. Area 46 is a part of the dorsolateral prefrontal cortex, and it
does not have functions related to low-level sensory information processing.
Being a higher level (supramodal) area, it integrates visual, tactile and
other information. Area 7a integrates visual, tactile and proprioceptive
signals. Finally, areas TF and FEF are also high level structures integrating
widespread cortical information (e.g., [20]).
The predicted connectivity matrix is shown on Fig. 9, the basic graph
measures are depicted in Table 3. To show the subtle differences between
predicted connections, the exact probabilities are shown on Fig. 10, encoded
in the background colour of the matrix cells (white indicating zero probabil-
ity and black indicating 1). The latter figure shows the prediction in its full
detail, especially in the sensorimotor cortex where the predicted clique-like
subgraph reveals its internal structure more precisely. The group affiliations
of the individual vertices are shown in Table 6.
We also examined the ratios of correctly predicted known 0’s and 1’s
with respect to pure visual and pure sensorimotor connections. As ex-
pected, the algorithm performed better in the visual cortex, which is more
thorougly charted than the√ sensorimotor cortex. The calculated ratios were
R0 = 0.865, R1 = 0.902, R0 R1 = 0.882 for the visual cortex. Since the sen-
sorimotor cortex contained no known non-existing connections, R1 could not
be calculated for it. All known connections in the sensorimotor cortex were
predicted correctly (R0 = 1), however, this is due to the lack of information
on nonexisting connections in the sensorimotor cortex. The ratios of the
visual cortex were similar to the ones obtained when analyzing the visual
cortex alone.
Out of the 225 unknown connections in the visual cortex, 46 were pre-
dicted differently when we took into account the sensorimotor cortex. The
Reconstructing Cortical Networks 357

Fig. 9. The predicted adjacency matrix of the visuo-tactile cortex with 10 vertex groups,
τ1 = 0.49 and τ2 = 0.137. White cells denote confirmed existing and absent connections.
Black cells denote mismatches between the known and the predicted connectivity. Light
gray cells denote predictions for unknown connections.

most discrepancies involved the outgoing edges of VOT (10 mismatches),

PIP (6 mismatches) and TF (6 mismatches). These can be caused by the
additional information present in the system in the form of heteromodal
connections. At the same time, prediction errors related to the known vi-
sual connections of visual areas having heteromodal connections decreased
(e.g., area TF: 13 to 6, area 46: 15 to 4), due to the same additional in-
formation. Other notable improvements were at V4 (21 to 13) and DP (12
to 7).
358 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

Fig. 10. Probability of the connections in the visuo-tactile cortex with 10 vertex groups,
τ1 = 0.49 and τ2 = 0.137. Probabilities are denoted by colours, with white
corresponding to 0 and black corresponding to 1. The predicted adjacency matrix is
shown in the matrix cells.

The reconstruction quality of the visual cortex was improved by adding

the information about the heteromodal connections and the sensorimotor
cortex (see Fig. 8 and the upper left part of Fig. 9). This was not a simple
consequence of increasing the number of clusters from 7 to 10, but the corol-
lary of the additional information about the connections that visual areas
form with the sensorimotor cortex. This contextual information may also
give guidelines for understanding the mechanisms of heteromodal interac-
tions.
Reconstructing Cortical Networks 359
Out-group 1 V1, VOT, MSTl
Out-group 2 V2, V3, V3A, V4t, VP, MT, PO, PIP
Out-group 3 V4
Out-group 4 MSTd/p, FST, FEF
Out-group 5 LIP, VIP
Out-group 6 DP, 7a
Out-group 7 PITd, PITv, CITd, CITv, AITd, AITv, STPp, STPa, TH
Out-group 8 TF, 46
Out-group 9 3a, 1, 2, 5, SII, 7b, 4, 6, SMA
Out-group 10 3b, Ri, Ig, Id, 35, 36
In-group 1 V1, PIP
In-group 2 V2, V3, V3A, V4t, VP, MT, PO
In-group 3 35, 36
In-group 4 V4, FST, FEF
In-group 5 VIP
In-group 6 MSTd/p, MSTl, LIP, DP, 7a
In-group 7 PITd, PITv, CITd, CITv, AITd, AITv, STPp, STPa, TH
In-group 8 VOT, TF, 46
In-group 9 3a, 1, 2, 5, SII, 7b, 4, 6, SMA
In-group 10 3b, Ri, Ig, Id
Table 6. Group affiliations of the areas in the visuo-tactile cortex

3.5. Major structural changes after reconstruction

An interesting feature of the reconstructed network is the complete clique

induced by the following vertices: PITd, PITv, CITd, CITv, AITd, AITv,
STPp, STPa, TF and TH. We note that this region was mostly uncharted
in the original matrix, and there were only three confirmed nonexisting
connections (STPp → AITd and STPa ↔ AITv). This clique also appeared
in the earlier analysis of Jouve et al. [29] and a similar but smaller clique
also emerged in our predictions of the visual cortex. Notably, a similar
tendency could also be observed in the analysis of Costa et al. [15], as only
a few connections were missing among the areas mentioned above to form
a clique. Our assumption is that our prediction based on solely the visual
cortex is a more accurate approximation of the true connectional pattern of
these areas.
Finally, we compared the community structure of the original and the
predicted connectivity matrix in order to obtain further support for the
validity of our predictions. We argue that a valid predicted connectivity
matrix should not only obtain a high ratio of correctly predicted 1’s and
360 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

0’s, but also keep the community structure of the network mostly intact. In
order to take the directionality of the edges into account, we employed the
community detection method of Latapy & Pons [39]. This method is based
on the idea that given a significant community structure in the network,
short random walks tend to stay within the same community. We tried
random walks with length ranging from 3 to 7. The quality of the obtained
partitions was assessed by the modularity function Q [47]. Best results
were achieved with random walks of length 7 for the original and 4 for the
predicted cortex (Q = 0.339 and Q = 0.301, respectively). Table 7 shows
the detected communities of the original and the predicted visuo-tactile
cortex.
Original visuo-tactile cortex (Q = 0.339)
Community 1 V1, V2, V3, V3A, V4t, VP, MT, MSTd/p, MSTl, PO, LIP,
PIP, VIP, DP, 7a, FST, FEF, 46
Community 2 V4, VOT, PITd, PITv, CITd, CITv, AITd, AITv, STPp,
STPa, TF, TH
Community 3 3a, 3b, 1, 2, 5, Ri, SII, 7b, 4, 6, SMA
Community 4 Ig, Id, 35, 36
Predicted visuo-tactile cortex (Q = 0.379)
Community 1 V1, V2, V3, V3A, V4t, VP, MT, MSTd/p, MSTl, PO, LIP,
PIP, VIP, DP, 7a, FST, FEF, 46
Community 2 V4, VOT, PITd, PITv, CITd, CITv, AITd, AITv, STPp,
STPa, TF, TH
Community 3 3a, 3b, 1, 2, 5, Ri, SII, 7b, 4, 6, SMA, Ig, Id, 35, 36
Table 7. Community structure of the original and the predicted visuo-tactile cortex

Community 1 corresponds approximately to the dorsal stream of the

visual cortex while community 2 contains the areas of the ventral stream
in both cases. Community 3 and community 4 form the sensorimotor
cortex. These groups were joined together to form a single community
of the sensorimotor cortex in the case of the predicted matrix. Since there
is no known anatomical meaning of community 4, the predicted community
structure represents our understanding of the sensorimotor cortex more
accurately than the original one. Apart from this mismatch, we can conclude
that the major structure of the original matrix was preserved during the
reconstruction process, but the reconstructed network seems less modular
than the original, as indicated by the decrease in the modularity of the
partitions.
Other notable differences were the smaller diameter and shorter average
path length (Table 1, 3). Considering segregation and integration as major
Reconstructing Cortical Networks 361

cortical functions [65], these findings together predict a more efficient corti-
cal information processing as it can be deduced on the basis of our present
state of knowledge about this network.

4. Conclusion

From a biological point of view, the stochastic representation of the cortical

network described in the present study demonstrates the high level of uncer-
tainty of our knowledge about the connectivity of the ventral visual cluster
(community 2) as well as the sensorimotor cortex (communities 3 and 4).
Similarly to our findings, both previous studies [29, 15] predicted numer-
ous connections within the ventral visual cluster, making it almost a large,
fully connected clique. A similar observation was made here in the sensori-
motor cortex. We should note that in contrast to the dorsal visual cluster
(including the majority of areas in community 1) there is a massive lack of
information regarding the non-existence of the connections in the ventral
and sensorimotor clusters (see also [20]). Considering the deterministically
imperfect knowledge of the cortical, and probably most of the real-life com-
plex networks, these findings point out the significance of information about
known nonexistent connections. Interestingly, similar conclusion was drawn
by Kötter and Stephan [37], who investigated the role of components in cor-
tical information processing by defining their network participation indices.
This is an important point when considering that neurobiologists focus on
the existence of connections between the cortical areas and often ignore the
absence of them while mapping cortical connections in their experiments.
In fact, this point is in agreement with our expectations considering that
the cortical network architecture is shaped by evolutionary and functional
constraints.
Considering individual cortical areas , our predictions resulted in a rela-
tively large number of mismatches of the connections of some mid- (e.g., V4
and DP) and high level areas (e.g., TF and 46) when compared to the orig-
inal data. The connections of high level areas, which form the connections
between the different cortical clusters, are hard to predict in individual sub-
networks where the connections between these areas and the neglected part
of the network is necessarily missing from the data. The largest number of
mismatches occurred in the case of V4. This area is a functionally heteroge-
neous, complex structure, which maybe divided to sub-regions, and forms a
362 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

bridge between the dorsal and ventral clusters of the visual cortex [46]. This
could at least partly explain the large number of mismatches occurred in the
reconstruction process. The assumption outlined here is further supported
by the fact that the number of mismatches involving areas V4, DP, TF
and 46 decreased after taking into account the sensorimotor cluster and the
heteromodal connections. On the other hand, V4t and VOT are examples
of structures with largely unexplored connectivity. The existing data sug-
gest that these areas exhibit roughly similar pattern of connectivity to their
topographical neighbor area V4, which has much better explored connec-
tivity. The large differences observed in the predicted connectivity of V4t
and VOT in the three studies (including ours as well as Costa et al. [15] and
Jouve et al. [29]) is most probably due to the large uncertainty in the con-
nectional data of these two areas. Interestingly, a more careful examination
of Fig. 8 suggest that the prediction of the connections of these three areas,
especially that of V4 and VOT tended to be somewhat complementary in
the visual cortex. This was less evident in the larger visuo-tactile network
(Fig. 9). Based on these observations it is suggested that the mismatches of
V4-connections occurred during the reconstruction as a consequence of the
optimization process.

From the methodological point of view it is important to note that

our approach is not a community structure detection. There are many
sophisticated methods which reliably detect communities within networks
[66, 52, 39, 47], yet the information about the community structure is not
sufficient for an accurate network reconstruction. One of the main ad-
vantages of our method is a higher accuracy in reconstruction of charted
connections (known existent or nonexistent), compared to earlier methods
[29, 15]. It is important to emphasize that our method allows the predic-
tion of nonreciprocal connections. The presented approach is general and
applicable to other networks not necessarily related to cortical structure.
Its generality is reflected in the fact that we do not use information about
the underlying spatial structure, as these data may be unavailable or unin-
telligible for some networks, nor we assume that there is a given property
of paths related to the clustering coefficient in the network. Our method
is based on an exact goal function that we optimize. The existence of a
goal function allows for a comparison of different solutions which may be
difficult to carry out in the absence of a quantitative measure. The draw-
backs of our approach are related to the necessity of knowledge regarding
existing and confirmed nonexisting connections. This necessity is reflected
in the prediction of heteromodal connections, as the number of confirmed
Reconstructing Cortical Networks 363

or refuted heteromodal connections is surprisingly small, see Fig. 5. One

has to note that the method is computationally expensive, which may be
a severe limiting factor in case of large networks. Finally, inspired by the
Szemerédi regularity lemma, the present study exemplifies the usefulness of
a theoretical approach in analyzing real world data.

Acknowledgments. László Négyessy acknowledges the support of the

János Bolyai Research Scholarship of the Hungarian Academy of Sciences.

References

[1] S. Achard and E. Bullmore, Efficiency and cost of economical brain functional
networks, PLoS Comp. Biol., 3(2) (2007), e17.
[2] H. Akaike, A new look at the statistical model identification, IEEE Trans. Automat.
Contr., 19(6) (1974), 716–723.
[3] C. Andrieu, N. de Freitas, A. Doucet and M. I. Jordan, An introduction to MCMC
for machine learning, Mach Learn, 50 (2003), 5–43.
[4] A. Angelucci and J. Bullier, Reaching beyond the classical receptive field of V1
neurons: horizontal or feedback axons? J. Physiol., 97(2–3) (2003), 141–54.
[5] M. A. Arbib and P. Érdi, Précis of Neural organization: structure, function and
dynamics, Behav. Brain Sci., 23(4) (2000), 513–33, Discussion: pp. 533–71.
[6] E. Bienenstock, On the dimensionality of cortical graphs, J. Physiol., 90(3–4)
(1996), 251–6.
[7] V. Braitenberg and A. Schüz, Cortex: Statistics and Geometry of Neuronal Con-
nectivity, Springer (1998).
[8] S. L. Bressler and E. Tognoli, Operational principles of neurocognitive networks,
Int. J. Psychophysiol., 60(2) (2006), 139–48.
[9] P. Buzás, K. Kovács, A. S. Ferecskó, J. M. Budd, U. T. Eysel and Z. F. Kisvárday,
Model-based analysis of excitatory lateral connections in the visual cortex,
J. Comp. Neurol., 499(6) (2006), 861–81.
[10] G. Buzsáki, C. Geisler, D. A. Henze and X. J. Wang, Interneuron diversity se-
ries: Circuit complexity and axon wiring economy of cortical interneurons, Trends
Neurosci, 27(4) (2004), 186–93.
[11] W. H. Calvin, Cortical columns, modules and Hebbian cell assemblies, in: M. A.
Arbib, editor, Handbook of Brain Theory and Neural Networks, Bradford Books/
MIT Press (1995), pages 269–272.
[12] F. R. K. Chung, Laplacians and the Cheeger inequality for directed graphs, Ann.
Combinator, 9 (2005), 1–19.
364 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

[13] A. Clauset, C. Moore and M. E. J. Newman, Structural inference of hierarchies

in networks, in: Proceedings of the 23rd International Conference on Machine
Learning (ICML), June 2006.
[14] G. Csárdi and T. Nepusz, The igraph software package for complex network re-
search, InterJournal Complex Systems, (2006), page 1695.
[15] L. da Fontoura Costa, M. Kaiser and C. C. Hilgetag, Predicting the connectivity
of primate cortical networks from topological and spatial node properties, BMC
Syst. Biol., 1(16) (2007).
[16] L. Danon, J. Duch, A. Diaz-Guilera and A. Arenas, Comparing community struc-
ture identification, J. Stat. Mech. (Oct. 2005), page P09008.
[17] J. DeFelipe, L. Alonso-Nanclares and J. I. Arellano, Microstructure of the neocor-
tex: comparative aspects, J. Neurocytol, 31(3–5) (2002), 299–316.
[18] R. J. Douglas and K. A. Martin, Mapping the matrix: the ways of neocortex,
Neuron, 56(2) (2007), 226–38.
[19] G. M. Edelman, Neural Darwinism: selection and reentrant signaling in higher
brain function, Neuron, 10(2) (1993), 115–25.
[20] D. J. Felleman and D. C. Van Essen, Distributed hierarchical processing in the
primate cerebral cortex, Cerebr. Cortex, 1(1) (1991), 1–47.
[21] A. L. N. Fred and A. K. Jain, Robust data clustering, in: Proceedings of the
IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
volume 2, pages 128–133. IEEE (2003). DOI 10.1109/CVPR.2003.1211462.
[22] W. J. Freeman, Mesoscopic neurodynamics: from neuron to brain, J. Physiol.,
94(5–6) (2000), 303–22.
[23] S. Geman and D. Geman, Stochastic relaxation, Gibbs distributions and the
Bayesian restoration of images, IEEE Trans Pattern Anal Mach Intell, 6 (1984),
721–741.
[24] W. K. Hastings, Monte Carlo sampling methods using Markov chains and their
applications, Biometrika, 57(1) (1970), 97–109. DOI 10.2307/2334940.
[25] C. C. Hilgetag, G. A. Burns, M. A. O’Neill, J. W. Scannell and M. P. Young,
Anatomical connectivity defines the organization of clusters of cortical areas in
the macaque monkey and the cat, Philos. Trans. R. Soc. Lond. B. Biol. Sci.,
355(1393) (2000), 91–110.
[26] C. C. Hilgetag and S. Grant, Uniformity, specificity and variability of corticocortical
connectivity, Philos. Trans. R. Soc. Lond. B. Biol. Sci., 355(1393) (2000), 7–20.
[27] C. C. Hilgetag, M. A. O’Neill and M. P. Young, Hierarchical organization of
macaque and cat cortical sensory systems explored with a novel network processor,
Philos. Trans. R. Soc. Lond. B. Biol. Sci., 355(1393) (2000), 71–89.
[28] A. A. Ioannides, Dynamic functional connectivity, Curr. Opin. Neurobiol, 17
(2007), 161–70.
[29] B. Jouve, P. Rosenstiehl and M. Imbert, A mathematical approach to the connec-
tivity between the cortical visual areas of the macaque monkey, Cerebr. Cortex,
8(1) (1998), 28–39.
Reconstructing Cortical Networks 365

[30] J. H. Kaas, Organizing principles of sensory representations, Novartis Foundation

Symposium, 228 (2000), 188–198.
[31] M. Kaiser, Brain architecture: a design for natural computation, Philos. Trans.
R. Soc. Lond. B. Biol. Sci., 365(1861) (2007), 3033–45.
[32] M. Kaiser and C. C. Hilgetag, Modelling the development of cortical systems
networks, Neurocomputing, 58–60 (2004), 297–302.
[33] M. Kaiser and C. C. Hilgetag, Nonoptimal component placement, but short process-
ing paths, due to long-distance projections in neural systems, PLoS. Comp. Biol.,
2(7) (2006), e95.
[34] Z. F. Kisvárday, A. S. Ferecskó, K. Kovács, P. Buzás, B. J. M. and U. T. Eysel,
One axon-multiple functions: specificity of lateral inhibitory connections by large
basket cells, J. Neurocytol., 31(3–5) (2002), 255–64.
[35] J. Komlós and M. Simonovits, Szemerédi’s regularity lemma and its applications
in graph theory, in: D. Miklós, V. T. Sós and T. Szőnyi, editors, Combinatorics,
Paul Erdős is Eighty, Vol. 2, János Bolyai Mathematical Society, Budapest (1996),
pages 295–352.
[36] R. Kötter, Online retrieval, processing and visualization of primate connectivity
data from the cocomac database, Neuroinformatics, 2 (2004), 127–144.
[37] R. Kötter and K. E. Stephan, Network participation indices: characterizing com-
ponent roles for information processing in neural networks, Neural Network, 16(9)
(2003), 1261–75.
[38] S. Kullback and R. A. Leibler, On information and sufficiency, Ann. Math. Stat.,
22 (1951), 79–86.
[39] M. Latapy and P. Pons, Computing communities in large networks using random
walks, J. Graph Algorithm Appl., 10(2) (2006), 191–218.
[40] L. Lee, K. Friston and B. Horwitz, Large-scale neural models and dynamic causal
modelling, NeuroImage, 30 (2006), 1243–54.
[41] J. W. Lewis and D. C. Van Essen, Corticocortical connections of visual, sensorimo-
tor and multimodal processing areas in the parietal lobe of the macaque monkey,
J. Comp. Neurol., 428 (2000), 112–137.
[42] R. Montenegro and P. Tetali, Mathematical aspects of mixing times in Markov
chains, Foundations and Trends in Theoretical Computer Science, 1(3) (2006),
237–354. DOI 10.1561/0400000003.
[43] V. B. Mountcastle, The columnar organization of the neocortex, Brain, 120(4)
(1997), 701–22.
[44] L. Négyessy, T. Nepusz, L. Kocsis and F. Bazsó, Prediction of the main cortical
areas and connections involved in the tactile fuction of the visual cortex by network
analysis, Eur. J. Neurosci, 23(2) (2006), 1919–1930.
[45] L. Négyessy, T. Nepusz, L. Zalányi and F. Bazsó, Convergence and divergence are
mostly reciprocated properties of the connections in the network of cortical areas,
Proc. Roy. Soc. B., 275(1649) (2008), pp. 2403–2410.
DOI 10.1098/rspb.2008.0629.
366 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

[46] T. Nepusz, A. Petróczi, L. Négyessy and F. Bazsó, Fuzzy communities and the
concept of bridgeness in complex networks, Phys. Rev. E., 77 (2008), 016107.
[47] M. E. J. Newman, Modularity and community structure in networks, Proc. Natl.
Acad. Sci. Unit. States. Am., 103(23) (June 2006), 8577–8582.
[48] M. E. J. Newman and E. A. Leicht, Mixture models and exploratory analysis in
networks, Proc. Natl. Acad. Sci. Unit. States. Am., 104(23) (2007), 9564–9569.
DOI 10.1073/pnas.0610537104.
[49] R. E. Passingham, K. E. Stephan and R. Kötter, The anatomical basis of functional
localization in the cortex, Nat. Rev. Neurosci, 3(8) (2002), 606–16.
[50] F. Petroni, S. Panzeri, C. C. Hilgetag, R. Kötter and M. P. Young, Simultaneity of
responses in a hierarchical visual network, NeuroReport, 12(12) (2001), 2753–9.
[51] http://www.ncbi.nlm.nih.gov/PubMed/.
[52] J. Reichardt and S. Bornholdt, Detecting fuzzy community structures in complex
networks with a Potts model, Phys. Rev. Lett., 93 (2004), 218701.
[53] K. S. Rockland, Non-uniformity of extrinsic connections and columnar organiza-
tion, J. Neurocytol, 31(3–5) (2002), 247–53.
[54] K. S. Rockland and N. I., Some thoughts on cortical minicolumns, Exp. Brain.
Res., 158(3) (2004), 256–77.
[55] A. W. Roe, Modular complexity of area V2 in the macaque monkey, in: C. Collins
and J. Kaas, editors, The Primate Visual System, CRC Press (2003), pages 109–
138.
[56] A. W. Roe and D. Y. Ts’o, Specificity of color connectivity between primate V1
and V2, J. Neurophysiol, 82(5) (1999), 2719–30.
[57] O. Sporns, Graph theory methods for the analysis of neural connectivity patterns,
in: R. Kötter, editor, Neuroscience Databases. A Practical Guide, Klüwer (2002).
[58] O. Sporns, D. R. Chialvo, M. Kaiser and C. C. Hilgetag, Organization, development
and function of complex brain networks, Trends Cognit. Sci., 8(9) (2004), 418–25.
[59] O. Sporns, C. J. Honey and R. Kötter, Identification and classification of hubs in
brain networks, PLoS ONE, 2(10) (2007), e1049.
[60] D. D. Stettler, A. Das, J. Bennett and C. D. Gilbert, Lateral connectivity and
contextual interactions in macaque primary visual cortex, Neuron, 36(4) (2002),
739–50.
[61] E. Szemerédi, Regular partitions of graphs, in: E. du Centre National de la
Recherche Scientifique (CNRS), editor, Problèmes combinatoires et théorie des
graphes (Colloq. Internat. CNRS, Univ. Orsay, Orsay, 1976, Paris (1978), pages
399–401.
[62] J. Szentágothai, The “module-concept” in cerebral cortex architecture, Brain Res.,
95(2–3) (1975), 475–96.
[63] J. Szentágothai, Specificity versus (quasi-) randomness revisited, Acta Morpholog-
ica Hungarica, 38 (1990), 159–167.
Reconstructing Cortical Networks 367

[64] T. Tao, Szemerédi’s regularity lemma revisited, Contributions to Discrete Mathe-

matics, 1 (2006), 8–28.
[65] G. Tononi, G. M. Edelman and O. Sporns, Complexity and coherency: integrating
information in the brain, Trends Cognit. Sci., 2 (1998), 474–84.
[66] S. van Dongen, A stochastic uncoupling process for graphs, Technical Report INS-
R0011, National Research Institute for Mathematics and Computer Science in the
Netherlands (2000).
[67] D. C. Van Essen and D. L. Dierker, Surface-based and probabilistic atlases of
primate cerebral cortex, Neuron, 56 (2007), 209–25.
[68] M. P. Young, Objective analysis of the topological organization of the primate
cortical visual system, Nature, 358(6382) (1992), 152–5.
[69] M. Zhu and A. Ghodsi, Automatic dimensionality selection from the scree plot via
the use of profile likelihood, Computational Statistics and Data Analysis, 51(2)
(2006), 918–930. DOI 10.1016/j.csda.2005.09.010.

Tamás Nepusz László Négyessy

KFKI Research Institute for Particle Neurobionics Research Group
and Nuclear Physics of the Hungarian Academy of Sciences –
Hungarian Academy of Sciences Péter Pázmány Catholic University –
H-1525 Budapest Semmelweis University
P.O. Box 49. H-1094 Budapest
Hungary Tűzoltó u. 58.
and Hungary
Budapest University of Technology e-mail: [email protected]
and Economics
Department of Measurement and
Information Systems
H-1521 Budapest
P.O. Box 91.
Hungary
e-mail: [email protected]
368 T. Nepusz, L. Négyessy, G. Tusnády and F. Bazsó

Gábor Tusnády Fülöp Bazsó

Alfréd Rényi Institute of KFKI Research Institute for Particle
Mathematics and Nuclear Physics of the
Hungarian Academy of Sciences Hungarian Academy of Sciences
Reáltanoda u. 13–15. H-1525 Budapest
1053 Budapest P.O. Box 49.
Hungary Hungary
e-mail: [email protected] and
Polytechnical Engineering College
Subotica
Marka Oreškovića 16
24000 Subotica
Serbia
e-mail: [email protected]
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 369–408.

Chapter 9
k-clique Percolation and Clustering

GERGELY PALLA, DÁNIEL ÁBEL, ILLÉS J. FARKAS,

PÉTER POLLNER, IMRE DERÉNYI and TAMÁS VICSEK∗

We summarise recent results connected to the concept of k-clique percolation.

This approach can be considered as a generalisation of edge percolation with
a great potential as a community finding method in real-world graphs. We
present a detailed study of the critical point for the appearance of a giant k-
clique percolation cluster in the Erdős–Rényi-graph. The observed transition is
continuous and at the transition point the scaling of the giant component with
the number of vertices is highly non-trivial. The concept is extended to weighted
and directed graphs as well. Finally, we demonstrate the effectiveness of k-clique
percolation as a community finding method via a series of real-world applications.

1. Introduction

In recent years there has been a growing interest in the dense, highly
interconnected parts of real-world graphs, often referred to as communities,
modules, clusters of cohesive groups [74, 76, 26, 46, 32, 54, 73, 64]. These
structural subunits can correspond to multi-protein functional units in
molecular biology [70, 78], a set of tightly coupled stocks or industrial
sectors in economy [59, 41], groups of people [74, 85, 63], cooperative players
[80, 84, 79], etc. The location of such building blocks can be crucial to the
∗
Corresponding author.
370 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

understanding of the structural and functional properties of the systems

under investigation.
Cliques (maximal complete subgraphs, in which every vertex (also re-
ferred to as node) is linked to every other vertex) correspond to the most
dense parts of a network [14, 25, 12, 11], therefore, they serve as an ideal
starting point to search for communities. However, limiting the community
definition to cliques only would be too restrictive in most cases. k-clique
percolation offers a similar, but more flexible alternative for network cluster-
ing [21, 64, 1]. This approach has been proven successful in many applica-
tions ranging from the study of cancer-related proteins in protein interaction
networks [44, 45], through the analysis of stock correlations [41] to the ex-
amination of various social networks [34, 63]. This success has inspired the
extension of the approach to weighted and directed networks as well [27, 67].
In this chapter we overview the recent results connected to this approach.
At this point we note that k-cores are also closely related to the concepts
described above and turned out to be of fundamental importance in the de-
composition of large complex networks as well [13, 75, 22, 33, 23, 17]. In this
approach a graph can be treated as a set of successively enclosed k-cores,
similar to a Russian nested doll (getting denser and denser inside).
As k-clique percolation can be considered to be a generalisation of edge
percolation, it provides a set of very interesting problems in random graph
theory by itself. (The first rigorous mathematical results on this subject
were given recently by Bollobás and Riordan in [15]). One of the most
conspicuous early results in random graph theory was related to the (edge)
percolation transition of the Erdős–Rényi (E-R) uncorrelated random graph
[24, 14]. The various aspects of this classical model remain still of great
interest since such a graph can serve both as a test-bed for checking all sorts
of new ideas concerning complex networks in general, and as a prototype
of random graphs to which all other random graphs can be compared. The
mentioned percolation transition of the E-R graph takes place at p = pc ≡
1/N , where p is the probability that two nodes are connected by an edge
and N is the total number of nodes in the graph. The appearance of a
giant component in a network, which is also referred to as the percolating
component, results in a dramatic change in the overall topological features
of the graph and has been in the centre of interest for other networks as
well.
A similar critical linking probability can be derived for the emergence
of a giant cluster composed of adjacent k-cliques (complete subgraphs of
k nodes) [21, 66]. Naturally, this critical probability grows with increasing
k-clique Percolation and Clustering 371

with k, as the conditions for the formations of a k-clique percolation cluster

become more restrictive. Although k-clique percolation is a generalisation
of edge percolation, it has a couple of unique features, e.g. a node can be
part of several k-clique percolation clusters at the same time, and multiple
order parameters can be defined to describe the percolation transitions,
which show different behaviour at the critical point. These aspects are
detailed in Sect. 2, which is aimed at summarising the most important
results concerning k-clique percolation in the E-R graph. The concept is
extended to weighted and directed networks in Sects. 3–4. Finally, in Sect. 5
we show how k-clique percolation can be applied to community finding and
network clustering.

2. k-clique Percolation in the E-R-graph

We begin with a few definitions laying down the fundamentals of k-clique

percolation [21, 66]:
• k-clique: a complete (fully connected) subgraph of k vertices [14].
• k-clique adjacency: two k-cliques are adjacent if they share k − 1
vertices, i.e., if they differ only in a single vertex.
• k-clique chain: a subgraph, which is the union of a sequence of adja-
cent k-cliques.
• k-clique connectedness: two k-cliques are k-clique-connected, if there
exists at least one k-clique chain containing the two k-cliques.
• k-clique percolation cluster (or component): a maximal k-clique-
connected subgraph, i.e., it is the union of all k-cliques that are k-
clique-connected to a particular k-clique.
The above concept of k-clique percolation can be illustrated by “k-clique
template rolling” (see Fig. 1). A k-clique template can be thought of as an
object that is isomorphic to a complete graph of k nodes. Such a template
can be placed onto any k-clique of the network, and rolled to an adjacent k-
clique by relocating one of its nodes and keeping its other k − 1 nodes fixed.
Thus, the k-clique-communities of a graph are all those subgraphs that can
be fully explored by rolling a k-clique template in them but cannot be left by
this template. We note that a k-clique percolation cluster is very much like
a regular edge percolation cluster in the k-clique adjacency graph, where
372 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

Fig. 1. Illustration of k-clique template rolling at k = 4. Initially the template is placed

on A–B–C–D (left panel) and it is “rolled” onto the subgraph A–C–D–E (middle
panel). The position of the k-clique template is marked with thick black lines and black
nodes, whereas the already visited edges are represented by thick gray lines and gray
nodes. Observe that in each step only one of the nodes is moved and the two 4-cliques
(before and after rolling) share k − 1 = 3 nodes. At the final step (right panel) the
template reaches the subgraph C–D–E–F , and the set of nodes visited during the
process (A–B–C–D–E–F ) are considered as a k-clique percolation cluster.

the vertices represent the k-cliques of the original graph, and there is an
edge between two vertices, if the corresponding two k-cliques are adjacent.
Moving a particle from one node of this adjacency graph to another one
along an edge is equivalent to rolling a k-clique template from one k-clique
of the original graph to an adjacent one. Note that a node can be part of
several k-clique percolation clusters at the same time, the simplest example
of this is given by two triangles (k-cliques at k = 3) overlapping in a single
node.
A k-clique percolation cluster fulfilling the above definition is a very good
candidate for a community in real networks. We shall detail this aspect in
Sect. 5. Here we mention that these objects can be considered as interesting
specific cases of the general graph theoretic objects defined in [25, 8] in very
different contexts.

2.1. Derivation of the critical point with heuristic arguments

The threshold probability (critical point) of k-clique percolation in the E-

R random graph can be obtained using the template rolling picture with
the following simple heuristic arguments. At the percolation threshold we
have to require that after rolling a k-clique template from a k-clique to an
adjacent one (by relocating one of its vertices), the expectation value of
the number of adjacent k-cliques, where the template can roll further (by
k-clique Percolation and Clustering 373

relocating another of its vertices), be equal to one. The intuitive argument

behind this criterion is that a smaller expectation value would result in
premature k-clique percolation clusters, because starting from any k-clique
the rolling would quickly come to a halt and, as a consequence, the size
of the clusters would decay exponentially. A larger expectation value, on
the other hand, would allow an infinite series of bifurcations for the rolling,
ensuring that a giant cluster is present in the system. The above expectation
value can be estimated as (k − 1)(N − k − 1)pk−1 , where the first term
(k − 1) counts the number of vertices of the template that can be selected
for the next relocation, the second term (N − k − 1) counts the number
of potential destinations for this relocation, out of which only the fraction
pk−1 is acceptable, because each of the new k − 1 edges (associated with the
relocation) must exist in order to obtain a new k-clique. For large N , our
criterion can thus be written as

(1) (k − 1)N pk−1

c = 1,

from which we get

1
(2) pc (k) = +  1
(k − 1)N k−1

for the threshold probability. The subscript “c” throughout this Chapter
indicates that the system is at the percolation threshold (or critical point).
Obviously, for k = 2 the above result agrees with the known percolation
threshold (pc = 1/N ) for E-R graphs, because 2-clique connectedness is
equivalent to regular (edge) connectedness.

2.2. Generating function formalism

The above results can be made stronger with the help of the generating
function formalism [66] in a fashion similar to that of [58]. (Note that the
following derivation is still heuristic from a rigorous mathematical point of
view). We first summarise the definition and the most important properties
of the generating functions. If a random variable ξ can take non-negative
integer values according to some probability distribution P(ξ = n) ≡ ρ(n),
374 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

then the corresponding generating function is given by

∞

(3) Gρ (x) ≡ x  =
ξ
ρ(n)xn .
n=0

The generating-function of a properly normalised distribution is absolutely

convergent for all |x| ≤ 1 and hence has no singularities in this region. For
x = 1 it is simply
∞

(4) Gρ (1) = ρ(n) = 1.
n=0

The original probability distribution and its moments can be obtained from
the generating-function as

1 dn Gρ (x) 
(5) ρ(n) = ,
n! dxn x=0
∞
&  '
 d l
(6) ξ  =
l l
n ρ(n) = x Gρ (x) .
dx
n=0 x=1

And finally, if η = ξ1 + ξ2 + · · · + ξl , where ξ1 , ξ2 , . . . , ξl are independent

random variables (with non-negative integer values), then the generating
function corresponding to P(η = n) ≡ σ(n) is given by
" # " #" # " #
(7) Gσ (x) = xη  = xξ1 xξ2 · · · xξl = xξ1 xξ2 · · · xξl
= Gρ1 (x)Gρ2 (x) · · · Gρl (x).

Now, we can proceed to the derivation of the critical point in the N → ∞

limit. First, let us consider the probability distribution r(n) of the number
of k-cliques adjacent to a randomly selected k-clique. Finding a k-clique
B adjacent to a selected k-clique A is equivalent to finding a node outside
A linked to at least k − 1 nodes in A. The number of possibilities for this
node is N − k. Edges in the E-R graph are independent of each other,
therefore the probability that a given node is linked to all nodes in A is pk ,
whereas the probability that it is linked to k − 1 nodes in A is k(1 − p)pk−1 .
Therefore, to leading order in N the average number of k-cliques adjacent
to a randomly selected one is
+ 
(8) r = (N − k) k(1 − p)pk−1 + pk N kpk−1 .
k-clique Percolation and Clustering 375

From the independence of the edges it also follows that the probability
distribution r(n) becomes Poissonean, which can be written as
n
k−1 (N kpk−1 )
(9) r(n) = exp (−N kp ) .
n!

Let us suppose that we are below the percolation threshold, therefore,

k-cliques are rare, adjacent k-cliques are even more rare, and loops in the
k-clique adjacency graph are so rare that we can assume it to be tree-like1 .
In this case the size of a connected component in the k-clique adjacency
graph (corresponding to a k-clique percolation cluster) can be evaluated by
counting the number of k-cliques reached in a branching process as follows.
We start at an arbitrary k-clique in the component, and in the first step
we invade all its neighbours in the k-clique adjacency graph. From then on,
whenever a k-clique is reached, we proceed by invading all its neighbours,
except for the one the k-clique has been reached from. In terms of the
original graph, this is equivalent to rolling a k-clique template to all adjacent
k-cliques except for the one we arrived from in the previous step.
In the process described above, we can assign to each k-clique the
subgraph in the k-clique percolation cluster that was invaded from it. (Note
that we assumed the k-clique adjacency graph to be tree-like). Let us denote
by I(n) the probability, that the subgraph reached from an arbitrary starting
k-clique in the branching process contains n number of k-cliques, including
the starting k-clique as well. This subgraph is actually equal to a k-clique
percolation cluster. Similarly, let H(n) denote the probability that the
subgraph reached from a k-clique appearing later in the branching process
(i.e., from a k-clique that is not the starting one) contains n number of k-
cliques. This is equivalent to the probability that by starting at a randomly
selected k-clique and trying to roll a k-clique template via all possible subsets
of size k −1 except for one, then by subsequently rolling the template on and
on, in all possible directions without turning back, a k-clique percolation
“branch” of size n is reached. And finally, let Hm (n) be the probability,
that if we pick m number of k-cliques randomly, then the sum of the sizes
of the k-clique branches that we can reach in this way consists of n number
of k-cliques. Since we are below the percolation threshold, the k-clique
adjacency graph consists of many dispersed components of small size, and
the probability that two (or more) k-cliques out of m belong to the same
k-clique percolation cluster is negligible. Hence, according to Eq. (7), the
1
This assumption is an approximation since the adjacency graph is weakly assortative.
376 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

generating functions corresponding to H(n) and Hm (n), denoted by GH (x)

and GHm (x) respectively are related to each other as:
+ m
(10) GHm (x) = GH (x) .

Let q(n) denote the probability, that for a randomly selected k-clique,
by excluding one of its possible subsets of size k − 1, we can roll a k-
clique template through the remaining subsets to n adjacent k-cliques. This
distribution is very similar to r(n), except that in this case we can use only
k−1 subsets instead of k in the k-clique to roll the k-clique template further,
therefore
 
  N (k − 1)pk−1 n
(11) q(n) = exp − N (k − 1)p k−1
.
n!
By neglecting the loops in the k-clique adjacency graph, Hn can be expressed
as

(12) H(n) = q(0)H0 (n − 1) + q(1)H1 (n − 1) + q(2)H2 (n − 1) + . . . .

By taking the generating function of both sides and using Eqs. (5) and (10),
we obtain
∞ $ ∞ %
(13) GH (x) = q(m)Hm (n − 1) xn
n=0 m=0

∞
& ∞  '
  1 dn−1 +  m
= q(m) GH (x)  xn =
(n − 1)! dxn−1 x=0
n=0 m=0
∞
 + m  
= q(m) GH (x) x = xGq GH (x) ,
m=0

where Gq (x) denotes the generating function of the distribution q(n).

We can write an equation similar to Eq. (12) for I(n) as well, in the
form of

(14) I(n) = r(0)H0 (n − 1) + r(1)H1 (n − 1) + r(2)H2 (n − 1) + . . .

Again, by taking the generating functions of both sides we arrive at

 
(15) GI (x) = xGr GH (x) ,
k-clique Percolation and Clustering 377

where Gr (x) denotes the generating function of r(n). From Eqs. (6) and
(15) we get
   
(16) I = GI (1) = Gr GH (1) + Gr GH (1) GH (1) = 1 + Gr (1)GH (1)

for the average size of the components invaded from a randomly selected
k-clique. Using Eq. (13) we can write
   
(17) GH (1) = Gq GH (1) + Gq GH (1) GH (1) = 1 + Gq (1)GH (1),

from which GH (1) can be expressed as

1
(18) GH (1) = .
1 − Gq (1)

By substituting this back into Eq. (16) we get

Gr (1) r

(19) I = 1 + =1+ .
1 − Gq (1) 1 − q

The above expression for the expected size of the connected components
in the k-clique adjacency graph invaded from a randomly selected k-clique
diverges when

(20) q = N (k − 1)pk−1 = 1.

This point marks the phase transition at which a giant component (corre-
sponding to a giant k-clique percolation cluster) first appears. Therefore,
our final result for the critical linking probability for the appearance of the
giant reassures Eq. (2), found via heuristic arguments.

2.3. Partial differential equation approach to k-clique percolation

The critical point of k-clique percolation was studied in a more general

framework by Ráth and Tóth [69]. (Similarly to Sects. 2.1–2.2., the results
of this Section are based on heuristics and do not provide rigorous proofs in
the mathematical sense). In this approach the E-R graph is constructed in a
stochastic process: from an initially empty graph containing√only nodes but
no edges, the possible edges are introduced at√
a rate of 1/ N . At time t,
the ratio of “occupied” edges equals 1 − e −t/ N , therefore, for N → ∞ the
378 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

resulting graph at√any time t is equivalent to an E-R graph with a linking

probability p = t/ N .
The above model can be naturally extended by replacing the initial state
with a non-empty graph. The method used by Ráth and Tóth to derive pc
for this general case was based on partial differential equations (PDE) and
can be applied to arbitrary initial component size distribution. However,
the initial state must fulfil the following conditions:
• If a small subset of edges is selected, then the distribution of the sizes
of the percolation clusters of these edges is asymptotically indepen-
dent from the probability distribution of the sizes of the rest of the
percolation clusters as N → ∞.
• The k-clique percolation clusters of the initial graph correspond to
trees of the k-clique adjacency graph.
The starting point of this method is the approximation of the change in
the number of k-clique percolation clusters having a given number of edges
m between time t and t + dt. Although this approach can be generalised to
higher k in a straightforward way, Ráth and Tóth focused exclusively on the
k = 3 case. By denoting the number of k-clique percolation clusters with
m edges at time t by Cm (N, t), they introduced the following quantities:

Cm (N, t)
(21) cm (t) ≡ lim 3 ,
N →∞ 1
2N
2

(22) vm (t) ≡ m · cm (t).

The Laplace-transforms of cm (t) and vm (t) are given by


(23) V (t, x) ≡ vm (t) · e−m·x ,
m∈N

(24) C(t, x) ≡ cm (t) · e−m·x .
m∈N

By introducing
 
 E(N, t)
(25) E(t) ≡ lim 3 ,
N →∞ 1
2N
2
k-clique Percolation and Clustering 379

Ráth and Tóth showed that in the mean-field approximation and in the
N → ∞, dt → 0 limits the C(t, x) satisfies the following PDE:

∂ 2 2
(26) C(t, x) = eV (t,x) −E(t) −x − 2V (t, x) · E(t)2 .
∂t

Equation (26) was solved with the method of characteristics [69], result-
ing in the following expression:
  1
(27) t2c · (2ab + 1) + tc · b + a · (2ab + 1) = ,
2

where a = V (0, 0) = m vm (0), b = − ∂x ∂

V (0, 0) and tc denotes the time of
the appearance of the giant k-clique percolation cluster.
√ In the special case
of an empty initial graph a = b = 0, yielding tc = 1/ 2 and

tc 1
(28) pc = √ = √ ,
N 2N

in agreement with the results of Sect. 2.1. From Eq. (26) Ráth and Tóth also
derived an equation for the rescaled size of the giant component (number
of links compared to the total number of links), denoted by v∞ (t). By
introducing
2 2
(29) W (t, x) ≡ eV (t,x) −E(t) −x ,
 
(30) V( (t, w) ≡ V t, X(t,
( w) ,

( w) denotes the inverse function of W (t, x) in the x variable, this

where X(t,
equation was formulated as [69]
 
(31) v∞ (t) = V( (0, 1) + t − V( 0, W (t, 0) − tW (t, 0).

2.4. Numerical simulations

The numerical studies of k-clique percolation in the E-R graph are in full
agreement with the results obtained in Sects. 2.1–2.2. The observed transi-
tion is continuous, characterised by non-universal critical exponents, which
depend on both k and the way the size of the giant component is measured.
380 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

There are two plausible choices for measuring the size of the giant com-
ponent: The most natural one, which we denote by N ∗ , is the number of
vertices belonging to this cluster. We can also define an order parameter
associated with this choice as the relative size of that cluster:

(32) Φ = N ∗ /N.

The other choice is the number N ∗ of k-cliques of the largest k-clique per-
colation cluster (or equivalently, the number of vertices of the largest com-
ponent in the k-clique adjacency graph). The associated order parameter is
again the relative size of this cluster:

(33) Ψ = N ∗ /N ,

where N denotes the total number of k-cliques in the graph (or the total
number of vertices in the adjacency graph). N can be estimated as
 
N k(k−1)/2 N k k(k−1)/2
(34) N ≈ p ≈ p ,
k k!
 
because k different vertices can be selected in Nk different ways, and any
such selection makes a k-clique only if all the k(k − 1)/2 edges between
these k vertices exist, each with probability p. Note that the classical E-R
percolation is equivalent to our k = 2 case, and the E-R order parameter
(relative number of nodes) is identical to Φ. Also note that in general the
size of the largest cluster could be measured as the number of its l-cliques,
N(l)∗ , for 1 ≤ l ≤ k. However, for simplicity we restrict ourselves to the two
limiting cases (N ∗ ≡ N(1)∗ and N ∗ ≡ N ∗ ) defined above.
(k)
Computer simulations indicate that the two order parameters behave
differently near the threshold probability. To illustrate this, in Figs. 2. we
plotted Φ and Ψ, respectively, as a function of p/pc (k) for k = 4 and for
various system sizes (N ), averaged over several runs. The order parameter
Φ for k ≥ 3 converges to a step function as N → ∞. The fact that the step
is located at p/pc (k) = 1 is actually the numerical proof of the validity of
the theoretical prediction (2) for pc (k). The order parameter Ψ for k ≥ 2,
on the other hand, similarly to the classical E-R transition, converges to a
limit function, which is 0 for p/pc (k) < 1 and grows continuously from 0 to
1 if we increase p/pc (k) from 1 to ∞. The limiting shape of this curve (with
proof) is given in [15].
k-clique Percolation and Clustering 381

Fig. 2. (a) Simulation results for the order parameter Φ as a function of p/pc (k) at
k = 4, averaged over several runs, such that the statistical error is smaller than the size
of the symbols. Φ converges to a step function in the N → ∞ limit. (b) The order
parameter Ψ as a function of p/pc (k) for the same simulations as in (a). Ψ converges to
a limit function (which is 0 for p/pc (k) < 1 and grows continuously to 1 above
p/pc (k) = 1) in the N → ∞ limit. Figure from [21]

−α
 ∼αN , with some
The width of the steps in Fig. 2. follows+ a power law,
exponent α. Plotting Φ as a function of p/pc (k) − 1 N , i.e., stretching
out the horizontal scale by N α , the data collapse onto a single curve. This
is shown for k = 3, 4, and 5 in Fig. 3a. The exponent α is around 0.5 for
every k ≥ 3. Although for k = 3 a slight deviation from α = 0.5 has been
obtained, it cannot be distinguished from a possible logarithmic correction.
One of the most fundamental results in random graph theory concerns
the behaviour of the largest component at the percolation threshold, where
it becomes infinitely large in the N → ∞ limit. Erdős and Rényi showed
[24] that for the random graphs they introduced, the size of the largest
component N ∗ (measured as the number of its nodes) at p = pc ≡ 1/N
diverges with the system size as N 2/3 , or equivalently, the order parameter
Φ scales as N −1/3 . A similar scaling behaviour can be observed for k-clique
percolation at the threshold probability pc (k) as well. If we assume, that
the k-clique adjacency graph is like an E-R graph, then at the threshold
the size of its giant component Nc∗ scales as Nc . Plugging p = pc from
2/3

Eq. (2) into Eq. (34) and omitting the N -independent factors we get the
scaling

(35) Nc ∼ N k/2

for the total number of k-cliques. Thus, the size of the giant component
Nc∗ is expected to scale as Nc
2/3
∼ N k/3 and the order parameter Ψc as
/Nc ∼ N −k/6 .
2/3
Nc
382 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

Fig. 3. (a) The width of the steps in Fig. 2a follows a power law, ∼ N −α , as the steps
collapse onto a single curve if we stretch them out by N α horizontally. We have plotted
the results obtained at k = 3 and k = 5 as well. The data for k = 4 and k = 5 are
shifted upward by 0.4 and 0.8, respectively, for clarity. (b) The order parameter at the
threshold, Ψc , scales as some negative power of N , in good agreement with
expression (36). Figure from [21].

This is valid, however, only if k ≤ 3. The reason for the breakdown

of the above scaling is that for k > 3 it predicts that the number of k-
cliques of the giant k-clique percolation cluster, i.e., the number of vertices
2/3
of the giant component in the k-clique adjacency graph, Nc ∼ N k/3 ,
grows faster than N . On the other hand, in analogy with the structure
of the giant component of the classical E-R problem, we expect that the
giant component in the adjacency graph also has a tree-like structure at
the threshold, with very few loops. As a consequence, almost every node
of the adjacency graph corresponds to a node of the original graph. Thus,
in the adjacency graph the giant component should not grow faster than N
at the threshold. Therefore, for k > 3 we expect that Nc∗ ∼ N , and using
Eq. (35), Ψc = Nc∗ /Nc ∼ N 1−k/2 . In summary:

N −k/6 for k ≤ 3
(36) Ψc ∼ .
N 1−k/2 for k ≥ 3

Numerical results showing the scaling of Ψc at pc as a function of N are

depicted in Fig. 3b, and the results are in good agreement with the above
arguments.
k-clique Percolation and Clustering 383

3. Percolation of Weighted k-cliques

As mentioned in the introduction, k-clique percolation is important from

the point of view of community finding, as k-clique percolation clusters
can be considered as dense communities. For the majority of the networks
occurring in nature and society, the edges connecting the nodes have an
associated weight as well, referring to the strength/intensity of the relation
between its endpoints.
Plain k-clique percolation can be used even in these cases, as described
in Sect. 5.1, on a truncated graph that contains only those edges that have a
weight higher than a given threshold. However, a method that incorporates
edge weights is expected to produce better results. Thus, we suggest a
generalisation of clique percolation to weighted networks [27]. For binary
graphs, using percolating k-clique clusters instead of only full cliques is a
less restrictive approach that highlights extended, less dense regions of the
graph. In the same way, our aim when defining weighted clique percolation
is to introduce concepts that can be used to examine those parts of the
graph that are denser than a given lower limit.

3.1. Definitions

In a weighted network, to each edge, (i, j), a weight wij ∈ R is assigned.

The intensity of a k-clique C is defined as the geometric mean of its edge
weights [62]:
 2/[k(k−1)]

I(C) = wij .
i<j
i,j∈C

Compared to unweighted k-clique percolation clusters, we define

weighted k-clique percolation clusters by considering only k-cliques having
an intensity greater than a given threshold I. In analogy with the definition
of k-clique adjacency, a weighted k-clique chain is a k-clique chain where
the intensity of all cliques is above I.
Compared to k-clique percolation on the truncated graph, the above
definition is less restrictive: a k-clique containing weak edges (low weights)
can be part of the percolation cluster if it contains a considerable number
of strong (large weights) edges as well.
384 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

3.2. Percolation transition in the weighted E-R graph

To define a weighted version of the E-R graph, we assign to each edge of this
graph a weight selected independently and randomly from a uniform distri-
bution on the interval (0, 1]. At a fixed I, the critical linking probability,
pc (I), of k-clique percolation is the edge probability where a giant module
(containing k-cliques fulfilling the intensity condition) emerges. The special
case I = 0 is equivalent to the unweighted case.
The results derived in Sect. 2. for unweighted clique percolation provide
an upper limit for pc in a weighted E-R graph. Consider weighted E-R
graph as defined above, remove edges weaker than I and ignore the weights
of the remaining edges. The resulting graph will an E-R graph with link
probability p∗ = p(1 − I). The (unweighted) k-cliques of this truncated
graph are a subset of those weighted k-cliques of the original graph which
pass the intensity threshold, resulting in a higher percolation threshold.
This gives the upper limit pc (I) < pc (0)/(1 − I).
A better approximation can be given by modifying the heuristic argu-
ment considered in Sect. 2.1. by taking into account the intensities of the
cliques. We keep the main idea (a k-clique template is rolled and percola-
tion of the k-cliques is required) and modify only the condition for rolling
the template further. In the process of rolling if a k-clique C1 precedes an-
other k-clique C2 , then we will say that C1 is the parent of C2 , and C2 is a
child of C1 . The k-clique template can be rolled from parent to child, if the
child k-clique passes the intensity threshold I.
We consider two approximations for this. First we assume that the prob-
ability distribution of edge weights in the child k-cliques is the original uni-
form distribution on the interval (0, 1]. (The actual probability distribution
of an edge weight is different from this, as we shall see shortly.) Denoting
by Pk < 1 the probability that the child k-clique has an intensity larger
than I, the expected number of cliques available for the template to roll to
is pk−1 N (k − 1)Pk . Applying the criterion for percolation, i.e. that the ex-
−1/(k−1)
pectation value of this number should be 1, we get pc (I) pc (0) Pk .
The probability Pk is simply the probability that the product of
k(k − 1)/2 independent random variables, with uniform distribution on
(0, 1], reaches I k(k−1)/2 , and can be expressed with straightforward but te-
dious integrals.
As noted above, the probability distribution of the edge weights in the
child k-cliques is not the original uniform distribution: the parent clique has
k-clique Percolation and Clustering 385

Fig. 4. Numerical analysis of the percolation of k-cliques fulfilling the intensity condition
in weighted E-R graphs. The sample numerical results shown in panels (a-c) were
obtained for N = 100 and k = 3 using 1 run for each (p, I) grid point. In panel (d)
points were computed from 3 to 100 runs for each (k, I) parameter pair and error bars
are smaller than the sizes of the symbols. (a) The order parameter, Φ, in the points of a
grid on the (k, I) plane. (b) We computed the transition line, pc = pc (I), as the curve
with Φ = 1/2 on the (k, I) plane. From the values of Φ at nearby grid points we
increased the precision of the transition line with linear interpolation. (c) Numerical
curve for the percolation threshold and the second order analytical approximation. The
area between the two curves, D, measures the difference between the two results.
(d) Difference between the numerical and analytical results for pc (I) at various system
sizes, N , and clique size parameters. Figure from [27].

an intensity above I, and it shares (k − 1)(k − 2)/2 edges with candidate k-

cliques. With an appropriate simplifying approximation for the probability
distribution of edge weights of the parent clique, one can calculate a second-
order approximation for pc . Higher-order approximations can be generated
by taking into account the probability distribution of the edge weights of
the grandparent clique, grand-grandparent clique and so on.
To numerically check our results, we generated weighted E-R networks
of size N = 10 . . . 400, and measured pc (I) in the following way (see Fig. 4):
we calculated the order parameter Φ, as defined in Eq. (32) on a square
386 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

grid, and approximated the points where Φ = 1/2 by linear interpolation

along the grid-lines. The algorithm we used exploited the fact that, for
a given initial random seed, one can reuse the cluster structure at (p1 , I1 )
for calculating the cluster structure at a different (p2 , I2 ), using a Hoshen–
Kopelman (also known as Union-Find) algorithm.

4. Directed k-clique Percolation

In practice many real networks contain directed connections among the ver-
tices, where the direction of a single link signals either the direction of some
kind of flow (e.g. the flow of information, energy), or the asymmetry of the
relation between the vertices (e.g. a superior-inferior relation). This raises
the question of whether a community searching algorithm that inherently
takes into account the directionality of links is more suitable for directed
networks than the usual undirected algorithms. Along this idea, in this sec-
tion we define the notion of directed k-cliques (in which the configuration of
the directed links has to meet certain criteria), and study the percolation
of these objects in the directed equivalent of the E-R graph [67].

4.1. Definitions

In undirected graphs a pair of nodes is either connected or not, whereas

in a directed graph the same pair, (A, B), can be connected in three ways:
either by a “single link” as (i) A → B and (ii) A ← B or by a “double link”
as (iii) A  B. Multiple links (i.e., more than one link between A and B in
the same direction) and self-links (such as A → A) are not allowed.
In a complete subgraph of size k the k(k − 1)/2 links can be directed
in 3k(k−1)/2 ways. Since in the undirected k-clique percolation we treat
these alternatives as identical, introducing link directions allows a large
variety of possible rules for defining directed k-cliques. A natural concept,
however, is to aim for objects preserving some kind of directedness as a
whole, rather than just being a collection of nodes connected by directed
links. Therefore, we define directed k-cliques as complete subgraphs of size
k in which an ordering can be made such that between any pair of nodes
there is a directed link pointing from the node with the higher order towards
the lower one. Since the presence of double links usually leads to multiple
k-clique Percolation and Clustering 387

Fig. 5. Groups of nodes forming a directed k-clique (a, c) and groups (b, d) that do
not. (a) A directed k-clique without double links. The index of each node corresponds
to its order (equivalent to the number of its out-links) within the directed k-clique.
(b) A complete subgraph without double links, but not accepted as a directed k-clique,
because it contains a directed loop. (c) A directed k-clique with a double link. Note
that the order of the nodes depends on which link is deleted from the double link.
(d) Double link in a complete subgraph that is not a directed k clique. It is not possible
to remove a link from the double link in a way that all directed loops disappear.

possibilities to order the nodes in a way fulfilling the above requirement, for
simplicity we first concentrate on directed k-cliques without double links.
In this case, the higher the order of a node, the more out-neighbours it has
in the k-clique (see illustration in Fig. 5a). Thus, the restricted out-degree
of a node in the k-clique (the number of its out-neighbours in the k-clique,
ranging from 0 to k − 1) can be assigned as its order. From this, it can
be seen easily (for details see Appendix of [67]) that the condition for a k-
clique with no double links to qualify as a directed k-clique is equivalent to
the following three conditions:

• Any directed link in the k-clique points from a node with a higher
order (larger restricted out-degree) to a node with a lower order.
• The k-clique contains no directed loops (where a “directed loop” is a
closed directed path).
• The restricted out-degree of each node in the k-clique is different.

The overall directionality of such an object naturally follows the ordering

of the nodes: the node with highest order is the one which has only out-
neighbours, and can be viewed as the “source” or “top”-node of the k-clique,
388 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

whereas the node with lowest order has only incoming links from the others,
and corresponds to a “drain” or “bottom” node.
None of the above three conditions holds in the presence of double links:
directed loops appear in the k-clique, the restricted out-degree of at least two
nodes in the k-clique become the same (see Appendix of [67]), and we can
find directed links pointing in the direction of increasing order. However,
based on the ordering of the nodes, it is always possible to eliminate the
double links (by removing all links that point towards higher order) from
a directed k-clique in such a way that the remaining single links fulfil all
three conditions. See Fig. 5c as an example.
The k-clique adjacency is defined similarly to the undirected case: two
directed k-cliques are adjacent if they share k − 1 nodes. The directed k-
clique percolation clusters arise as the union of directed k-cliques that can
be reached from each other through a series of k-clique adjacency. The k-
clique template rolling picture can be applied to illustrate these clusters in
the same fashion as in the undirected case. The directed k-clique percolation
clusters provide a community definition for real networks in a similar fashion
to the undirected case (see Sect. 5. for details).

4.2. Percolation transition in the directed E-R graph

The directed equivalent of the E-R graph consists of N nodes providing

N (N − 1) possible “places” for the directed links, and these are filled inde-
pendently with uniform probability p, producing on average M N (N −1)p
links. (Note that in the original undirected E-R graph there are only
N (N − 1)/2 possibilities to introduce an edge, therefore, at linking proba-
bility p, there are only M N (N − 1)p/2 connections). In the following
we shall evaluate the critical linking probability, pdir
c , for directed k-clique
percolation using similar heuristic arguments as in Sect. 2.1.
c is decreasing with increasing N , and converges to zero as N → ∞.
pdir
We restrict ourselfs to the large N limit, and evaluate pdir
c to leading order
only. Let us suppose that we approach the critical point from below: the
directed k-cliques do not assemble yet into a giant cluster and we can find
only small, isolated clusters, i.e. the system is dispersed. In our k-clique
template rolling picture this means that when trying to explore the directed
percolation clusters by rolling such a template on them, we must stop the
rolling after a few steps as we run out of unexplored adjacent directed k-
cliques.
k-clique Percolation and Clustering 389

One can estimate pdir

c from the condition that at the critical point the
average number of yet unexplored directed k-cliques adjacent to the k-clique
we have just reached becomes equal to one. (This makes it possible to roll
our template on and on for a long time). Since we are going to evaluate
pdir
c to leading order only, we can neglect the possibility to roll our k-clique
template using double links between the same nodes: When reaching a
directed k-clique, the minimal number of further links that must be present
to enable the continuation of the template rolling is k−1. The probability of
such a case is therefore proportional to pk−1 . Even though it is not forbidden
in the first place to continue using double links as well, each double link in
the new directed k-clique we are going to roll onto multiplies the probability
by p. In other words, the probability to roll further to a k-clique containing
one double link is smaller by a factor of p, the probability to roll further to
a k-clique containing two double links is smaller by a factor of p2 , etc.
Consider the branching process exploring a directed k-clique percolation
cluster at the point when we are about to roll our template further on. We
can choose the next node for relocation in k − 1 different ways, which can
then be relocated to approximately N places. If there were no restrictions
for directing the links inside a directed k-clique, then the k − 1 new links
connecting the new node to this k − 1 shared nodes could be directed
in 2k−1 ways. However, the new directed k-clique has to fulfil the three
conditions detailed in Sect. 4.1. as well, therefore the actual number of
allowed configurations is much smaller. The rank of the new node in the
new directed k-clique can be chosen in k ways. The k − 1 nodes shared
with the previous k-clique are already ordered, and we can “insert” the new
node anywhere into this hierarchy. By fixing the order of the new node we
fix the direction of the new links as well, therefore, we can allow only k
different configurations for the directionality of these links. By combining
these factors together, the condition for reaching the critical point of the
percolation transition can be written as
+  k−1
(37) pdir
c N (k − 1)k = 1,

from which we gain

+  −1/(k−1)
(38) c = N k(k − 1)
pdir = pc /k 1/(k−1)

as a first order approximation for the critical linking probability. Note that
in the limiting case of k = 2 (the directed link percolation), the pdir
c = pc /2
390 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

relation holds, which is consistent with the 2 : 1 ratio for the number of
links in the directed and undirected E-R graph, respectively.
To measure the size of the largest directed k-clique percolation cluster
we can use Φ and Ψ, defined in Eqs. (32–33), in complete analogy with
the undirected case. In Figs. 6a–b we display Φ and Ψ as functions of
p/pdir
c obtained in numerical simulations, where the directed k-clique size
was k = 4, and the system size varied between N = 50 and N = 1600.
Similarly to the undirected k-clique percolation, the order parameter Φ
converges to a step function for increasing system sizes, whereas Ψ converges
to a limit function (which is 0 for p/pc (k) < 1 and grows continuously to 1
above p/pc (k) = 1). We have evaluated the transition point numerically
as well, by computing the second moment of the distribution of Ni values,
excluding the largest one, N1 = N ∗ :

(39) χ= (Ni /N )2 .
i>1

Note that this quantity is analogous to the percolation susceptibility. Both

below and above the transition point the Ni (i > 1) values follow an expo-
nential distribution, and only at pc do they have a power-law distribution.
Thus, χ is maximal at the numerical transition point, pnum c . In Fig. 6c we
show χ calculated for the curves shown in Fig. 6b, as a function of p/pdir c .
In order to check the theoretical prediction for the critical point obtained
in (38) we have carried out a finite-size scaling analysis of the numerical re-
sults. In Fig. 6d we show the ratio pnum
c /pdir
c as a function of 1/N . Indeed,
for large systems, the above ratio converges to 1 roughly as 1 + cN −1/2 .

5. Applications: Community Finding and Clustering

The study of the intermediate-scale substructures in networks, made up

of vertices more densely connected to each other than to the rest of the
network, has become one of the most highlighted topics in complex network
theory. A reliable method to pinpoint such objects has many potential
industrial applications, e.g. it can help service providers (phone, banking,
Internet, etc.) identify meaningful groups of customers (users), or support
biomedical researchers in their search for individual target molecules and
novel protein complex targets [47, 4]. Since communities have no widely
accepted unique definition, the number of available methods to pinpoint
k-clique Percolation and Clustering 391

Fig. 6. Numerical results for directed k-clique percolation in ER-graphs. In each

sub-figure, points show an average over 4 to 100 simulations depending on system size.
a) The order parameter Φ (the number of nodes in the largest percolation cluster
divided by N ) as a function of p/pdir dir
c , where pc was obtained from Eq. (38). b) The
order parameter Ψ (the number of directed k-cliques in the largest percolation cluster
divided by the total number of directed k-cliques) as a function of p/pdir c . c) The
numerically determined value for the critical linking probability, pnum
c , defined as the
average location of the maximum of χ(p), playing the role of the normalised percolation
susceptibility (see Eq. 39). d) Verification of the theoretical prediction for the critical
point. The pnum
c /pdir
c ratio converges to 1 for large N . Figure from [67].

them is vast [74, 76, 26, 46, 32, 54, 73, 64, 27, 67, 71, 72, 37, 36, 38, 52]. The
majority of these algorithms classify the nodes into disjoint communities,
and in most cases a global quantity called modularity [56, 55] is used to
evaluate the quality of the partitioning. However, as pointed out in [29, 49],
the modularity optimisation introduces a resolution limit in √ the clustering,
and communities containing a smaller number of edges than M (where M
is the total number of edges) cannot be resolved. One of the big advantages
of the clique percolation method (CPM) is that it identifies communities as
k-clique percolation clusters, and therefore, the algorithm is local, and does
not suffer from resolution problems of this type [64, 21].
392 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

Along with the rapid development of network clustering techniques, the

ability of revealing overlaps between communities has become very impor-
tant as well [86, 9, 39, 83, 31, 89, 57, 71, 52]. Indeed, communities in real-
world graphs are often inherently overlapping: each person in a social web
belongs usually to several groups (family, colleagues, friends, etc.), proteins
in a protein interaction network may participate in multiple complexes [39]
and a large portion of webpages can be classified under multiple categories.
Prohibiting overlaps during module identification strongly increases the per-
centage of false negative co-classified pairs. As an example, in a social web
a group of colleagues might end up in different modules, each corresponding
to e.g. their families. In this case, the network module corresponding to
their workgroup is bound to become lost. The other big advantage of CPM
beside its local nature is that it allows overlaps between communities in a
natural way: a node can be part of several k-clique percolation clusters at
the same time.

5.1. The CPM in practice

In principle, the CPM detailed in Sect. 2 can be only applied to binary net-
works (i.e. to those with undirected and unweighted connections). However,
an arbitrary network can always be transformed into a binary one by ignor-
ing any directionality in the connections and keeping only those connections
that are stronger than a threshold weight w∗ . Changing the threshold is
similar to changing the resolution (as in a microscope) with which the com-
munity structure is investigated: by increasing w ∗ the communities start
to shrink and fall apart. A very similar effect can be observed by chang-
ing the value of k as well: increasing k makes the communities smaller and
more disintegrated, but at the same time, also more cohesive. When we
are interested in the community structure around a particular node, it is
advisable to scan through a ranges of k and w ∗ values and monitor how the
communities change. Meanwhile, when analysing the modular structure of
the entire network, the criterion used to fix these parameters is based on
finding a modular structure as highly structured as possible [64]. This can
be achieved by tuning the parameters just below the critical point of the
percolation transition. In this way we ensure that we find as many modules
as possible, without the negative effect of having a giant module that would
smear out the details of the modular structure by merging (and making
invisible) many smaller modules.
k-clique Percolation and Clustering 393

The edge weights can be also taken into account in a somewhat refined
way when using weighted k-cliques (fulfilling an edge-weight intensity cri-
terion), as described in Sect. 3. This approach is referred to as the CPMw
method [27], and the optimal k-clique intensity threshold can be adjusted
similarly to the calibration of the optimal edge weight threshold described
above.
The directed k-clique percolation clusters defined in Sect. 4 provide a
suitable community definition for directed networks (this is the CPMd
method [67]). Due to the asymmetry of the directed connections, nodes
with mostly incoming links are expected to play a very different role in a
given community from those having mostly outgoing links or from those
having a similar amount of both kinds of links. A member (node) having
only out-neighbours amongst the others can be viewed as a “source” or a
“top-node”, whereas a node with only incoming links from the rest of the
community is a “drain” or a “bottom-node”. Most nodes, however, fall
usually somewhere between these two extremes. To quantify this property,
we can introduce the relative in-degree and relative out-degree [67] of node
i in a community α as

dαi,in
(40) α
Di,in ≡ ,
dαi,in + dαi,out

dαi,out
(41) α
Di,out ≡ ,
dαi,in + dαi,out

where dαi,in and dαi,out denote the numbers of in- and out-neighbours amongst
the other nodes in the community, respectively. Obviously, the values of
α α
both Di,out and Di,in are in the range between 0 and 1, and the relation
α +D α
Di,in = 1 holds. For weighted networks, Eqs. (40, 41) can be replaced
i,out
by the relative in-strength and relative out-strength defined as
α
wi,in
(42) α
Wi,in ≡ α + wα ,
wi,in i,out
α
wi,out
(43) α
Wi,out ≡ α + wα ,
wi,in i,out

α α
and wi,out denote the aggregated weight of outgoing and incom-
where wi,out
ing connections with other nodes in the community α. As an illustration,
in Fig. 7. we show the directed communities of the word “GOLD” in a
394 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

word association network studied in [67]. The weight of a directed link in

this case indicates the frequency at which people in questionnaires associ-
ated the endpoint of the link with its starting point. The communities are
colour coded with the overlaps emphasised in red. According to its differ-
ent meanings, the word “GOLD” participates in four, strongly internally
connected communities. Beside the node labels we display the relative out-
strength of the nodes in the communities using Eq. (43). Apparently, nodes
with a special/particular meaning (e.g. “SAPPHIRE”) tend to have high
relative out-strength, whereas commonly used words with general meaning
(e.g. “MONEY”) have low relative out-strength. Thus, it seems that the
overall directionality of the communities in this case is from special words
towards more general words.

Fig. 7. The directed modules of the word “GOLD” in a word association network. The
modules are colour coded and the overlaps between the modules are displayed in red.
The size of each node is proportional to the number of modules it participates in (some
of them are not shown in this figure). Beside the name of the nodes we display their
α α α α
Wi,out = wi,out /(wi,in + wi,out ) values as well. Nodes with high W (e.g. “SAPPHIRE”)
usually correspond to special, rarely used words, whereas nodes with low relative
out-degree (e.g. “MONEY”) are very common. Figure from [67].
k-clique Percolation and Clustering 395

5.2. Applying the CPM to real networks

In this section we summarise the most important results obtained so far with
the help of the CPM in the analysis of real networks. These achievements
are related to a wide spectrum of problems, ranging from cancer metastasis
through the formation of social groups to the study of the directed commu-
nities of webpages. Here we focus solely on the results closely related to the
CPM in the cited works.

5.2.1. The graph of communities. As we already pointed out, one of the

big advantages of the CPM is that it allows overlaps between the communi-
ties. These overlaps naturally lead to the definition of the community graph
[64, 68]: a network representing the connections between the communities,
with the nodes referring to communities and edges corresponding to shared
members between the communities of the original graph. The community
graph can be treated as a “coarse-grained” view of the original network, and
can be used to study the organisation of the system at a higher level. As an
illustration, in Fig. 8. we show the community graph of the protein-protein
interaction (PPI) network obtained from the DIP core list of protein-protein
interactions of the yeast, S. cerevisiae [87]. The biological functions or pro-
tein complexes that can be associated with the communities shown in the
left panel were looked up by using the GO-TermFinder package [16] and the
online tools of the Saccharomyces Genome Database [20].
It is well known (see e.g. [5, 3, 51]) that the nodes of large real networks
have a power law degree distribution. Studies of various complex systems
showed that if we consider the network of communities instead of the nodes
themselves, we still observe a degree distribution with a fat tail, but a char-
acteristic scale is introduced, below which the distribution is exponential
[64]. This is in agreement with our understanding of a complex system
having different levels of organisation with units specific to each level. In
addition, in the present case the principle of organisation (scaling) is pre-
served (with some specific modifications) when going to the next level in
the hierarchy.
In a wide range of graph models the basic mechanism behind the emerg-
ing power law degree distribution is that new nodes appearing in the system
attach to the “old” ones with a probability proportional to their degrees
[5, 3, 51]. Furthermore, the occurrence of preferential attachment was di-
rectly demonstrated in several real-world networks with scale free degree
396 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

Fig. 8. The community graph at k = 4 for the PPI network of S. cerevisiae obtained
from the DIP core list. The area of a node and the width of an edges are proportional to
the size of the corresponding community (number of members) and to the size of the
overlap (number of shared nodes), respectively. The coloured communities are cut out
and magnified to reveal their internal structure in the left panel. In this magnified
picture the nodes and edges of the original network have the same colour as their
communities, those that are shared by more than one community are emphasised in red,
and the grey edges are not part of these communities. The area of a node and the width
of an edge are proportional to the total numberof communities they belong to. Figure
from [64].

distribution [6, 43, 53]. In the study of the community structure in a scien-
tific co-authorship network it has been shown that similar processes control
the growth of communities and the development of the community graph
as well [68].

5.2.2. Molecular biological networks. Over the past decade in biology,

especially in Bioinformatics and Systems Biology, the network approach has
become very popular and successful [40, 7, 10, 50, 90, 2]. The CPM (together
with its free software implementation, CFinder [65, 1], capable of detecting
and visualising k-clique percolation clusters) provides a flexible and handy
tool for identifying modules in such graphs. In Sect. 5.2.1. we demonstrated
the concept of community graphs with the help of a biological example, the
PPI network of yeast. The communities detected with the CPM in such
networks can be associated mostly with either protein complexes or certain
functions [64, 1, 88]. For some proteins no function is available yet. Thus,
k-clique Percolation and Clustering 397

the fact that they show up as members of communities can be interpreted

as a prediction for their functions.
Moreover, Jonsson et. al. used the CPM to validate the reliability of
the connections in a PPI network [45]. The available direct experimental
data concerning protein-protein interactions is not equally broad for the
different species. Thus, in some cases the construction of the PPI network
is based on other methods, too, for example, homology (DNA sequence
similarity) between the proteins. The weight of an edge, A-B, in the
PPI network is obtained by integrating (e.g. summing) the weights from
several sources of evidence: experimentally measured physical interactions,
homology and many others. Each experimental and prediction technique
can identify different groups of interactions with high efficiency, therefore,
to find the largest possible portion of the “real” biological list of interactions
it is necessary to integrate data from a large number of sources. The scoring
function used to calculate the edge weights in the integrated network can
be validated in a number of ways; one is based on the assumption that
interactions within densely linked communities are more acceptable than
interactions between communities, i.e. a higher score is expected for intra-
community connections. This assumption was confirmed by Jonsson et. al.
in a study of the rat proteome, where edges inside the CPM communities
were observed to be significantly stronger than edges connecting nodes in
different communities [45].
In the example above one major goal was the automatic identification
of protein communities involved in cancer metastasis. In metastasis (a
cellular state) cancer cells have the ability to break away from the primary
tumour and move to different organs, making the cancer more difficult to
treat. Little is known about the molecular biology of metastasis, but it
is now broadly accepted that these cells have an increased motility and
invasiveness. These novel behaviours involve protein-protein interactions
which have to be identified and characterised if an effective treatment is to
be developed. The main results of [45] showed that the CPM can help to
identify key protein communities involved in cancer metastasis.
A closely related study by Jonsson and Bates was aimed at the inves-
tigation of the topological properties of cancer proteins (proteins closely
related to the development of cancer) as nodes of the human PPI net-
work [44]. The community structure was examined with the help of the
CPM, and the results showed that (among various other topological dif-
ferences between cancer- and non-cancer proteins) cancer proteins appear
in community overlaps more frequently than predicted from their overall
398 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

ratio amongst all proteins. Since communities usually represent different

cellular processes, proteins in the overlaps may be participating in multiple
processes, and can be considered to be at the “interface” of distinct but ad-
jacent cellular processes. Therefore, cancer proteins seem to be mediators
between different pathways. In one of the examples presented by Jonsson
and Bates, four communities were tied together by cancer proteins with
functions ranging from signal transduction to the regulation of cell growth
and cell death. Furthermore, the ratio of cancer proteins in the communi-
ties was increasing with k, and cancer proteins seemed to take part in larger
communities. A plausible explanation of this effect is that cancer proteins
participate in more complex cellular processes. It is also conceivable that
the larger communities correspond to larger or more complicated cellular
machineries, where cancer proteins play a role [44].
A cancer-related investigation using CPM was carried out recently by
Finocchiaro et. al. in [28] as well. In this case the network was constructed
from gene expression data: groups of up to 10 genes with significant co-
expression were fully connected. The communities in the resulting network
were extracted using several methods (including CPM), and according to
the results, the identified communities were enriched with genes responsi-
ble for the regulation of the cell cycle, apoptosis, phosphorylation cascades,
extracellular matrix, immune and interferon response regulation. For the
majority of communities, promoter searches for enriched cis-regulatory mod-
ules support the conclusion that the communities identified here reflect bi-
ologically relevant sets of co-regulated genes whose expression is altered in
human cancer. As such, the identified communities may provide marker
genes useful for clinical applications as well as hitherto unknown regulators
of cancer signalling pathways that may constitute novel entry points for
pharmacological intervention.

5.2.3. Social networks. The CPM was successfully applied to various

networks related to the social contacts between people as well. The study
of social networks has a long history; in its early period sociologist used
questionnaires and personal interviews to reveal the graph of social ties.
The spectrum of social interactions that can be probed in this approach is
very wide, however, the size of the sample that can be examined in this way
is rather limited. Nowadays, due to the rapid developments in computer
technology, new possibilities opened up for the exploration of social ties,
enabling the construction of networks on a much larger scale. A prominent
example of this is given in [60, 61], where a network consisting of more than
k-clique Percolation and Clustering 399

4·106 customers of a mobile phone company is analysed (the edges represent

mutual calls between the people).
The community structure of this system was analysed with the help of
CPM in [63], and according to the results, the majority of the found com-
munities contained individuals living in the same neighbourhood, and with
comparable age, a homogeneity that supports the validity of the uncovered
community structure. Interestingly, the time evolution of the small com-
munities (e.g. a smaller collaborative or friendship circles) and the large
communities (e.g. institutions) showed major differences. At the heart of
small cliques were a few strong relationships, and as long as these persisted,
the community around them was stable. It appeared to be almost impossi-
ble to maintain this strategy for large communities, however. In contrast,
the condition for stability for large communities was continuous changes in
their membership, allowing for the possibility that after some time practi-
cally all members are exchanged. Such loose, rapidly changing communities
are reminiscent of institutions, that can continue to exist even after all
members have been replaced by new ones. For example, in a few years
most members of a school or a company could change, yet the school and
the company will be detectable as a distinct community at any time step
throughout its existence. This effect was observed in the community evolu-
tion of a co-authorship network as well [63]. (The edges between co-authors
in this case corresponded to articles published together).
Another interesting study of social networks was given in [34] by
González et. al., investigating the community structure and ethnic pref-
erences in high schools. The friendship networks between the students for
84 schools were constructed from the Add Health database [82], and the
communities were extracted using the CPM. The communities at k = 3
covered the majority of the students in most of the schools, and the corre-
sponding community graphs showed complex, richly interconnected struc-
tures. In contrast, at k = 4 the community graphs became rather sparse,
and the involved communities covered less than 20% of the students. At
the same time, the number of communities belonging to the different eth-
nic groups became balanced even for cases when the ratio of the sizes of
the ethnic groups was far from unity (and, correspondingly, on the level of
less cohesive groups, e.g. for k = 3, the students who were in majority, had
much larger friendship circles). A plausible explanation of this effect is that
when in minority, the students tend to form stronger ties, thus, the num-
ber of more densely interconnected communities becomes over-represented
compared to what happens in the k = 3 case [34].
400 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

Fig. 9. Different network properties averaged over the community graphs (circles) and
the underlying networks of students (squares) in the studied schools. (a) The cumulative
degree distribution P (d). (b) The average clustering coefficient C (the fraction of edges
between the nearest neighbours of a vertex compared to the number of edges that could
possibly exist between them) in function of the vertex degree d. (c) The average degree
of the nearest neighbours in function of the degree. Note that the base network shows
assortativity (increasing tendency at low degrees), whereas the community graph is
disassortative. Figure from [34].

The other important result in this study connected to the CPM is

that the graph of communities turned out to be disassortative in spite of
the fact that the underlying network of friendship showed assortativity, as
demonstrated in Fig. 9c. (A network is said to be assortative if the average
degree of the nearest neighbours is increasing with the node degree, i.e.
high degree nodes “like” to connect to high degree nodes, and disassortative
in the opposite case). This is another indication of the differences in the
interactions at different levels in the hierarchy of a complex system.

The results in [60, 61, 34] (described partly in this section) also inspired
a couple of new models for the development of social networks [81, 48, 35].
In the works of Toivonen et. al. [81] and Kumpula et. al. [48], the empha-
sis is on the balance between two different type of attachment mechanisms:
cyclic closure and focal closure. The first one corresponds to the formation
of new ties or the enhancement of the strength of existing ties within an
already densely connected neighbourhood (i.e. two people who have many
k-clique Percolation and Clustering 401

common acquaintances will get to know each other as well sooner or later).
The second one refers to the formation of ties independently of the geodesic
distance and is attributed to shared activities (hobbies, etc.). By changing
the relative strength of the two types of attachment mechanisms, the form-
ing network undergoes a transition from a homogeneous state (where the
majority of the edges were formed by focal closure) to an inhomogeneous
state with apparent, dense communities (in which the edges result mostly
from cyclic closure). This transition, and the appearing communities were
studied with the help of the CPM [81, 48]. According to the results, by
adequately balancing the two types of attachment mechanisms, the statisti-
cal properties of the model network match the mobile-phone network in all
studied aspects.
In [35] González et. al. introduced a network model based on colliding
(finite sized) particles travelling in a finite cell with periodic boundary con-
ditions. Each collision results in a new edge between the involved particles,
and the updating of the velocities depend on the degree of the particles.
With suitably chosen collision rules and aging scheme (particles die after a
certain amount of time) the quasi-stationary states of the resulting network
reproduce accurately the main statistical and topological features (e.g. the
community size distribution for communities obtained by CPM) of the high
school friendship networks mentioned earlier.
We close the overview of the CPM related results in social networks by
mentioning the study of the collaboration network among rappers by Smith
in [77]. The edges in this network correspond to co-appearance as artists
in lyrics obtained from several sources. The community structure of the
resulting graph was analysed with several methods including the CPM.

5.2.4. Further results. Finally, we collect a few other results related to the
CPM ranging from the investigation of economical networks to the graph
of certain webpages. In [41] a subset of the New York Stock Exchange
was analysed by Heimo et. al. with both spectral methods and the asset
graph method. In the latter case, the asset graph was constructed from the
correlation matrix of the stocks: the edges represent correlations stronger
than a certain threshold. The emerging graph was studied with the help of
the CPM. The results show that the first few eigenvectors of the correlation
matrix are localised on the communities, however their borders are fuzzy and
do not define clear cluster boundaries. With increasing eigenvector index
(the eigenvectors are ordered according to the corresponding eigenvalue),
the eigenvectors appear to localise increasingly less regularly with respect
402 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

to the asset graph topology. Therefore it appears that identifying the

strongly interacting clusters of stocks solely based on spectral properties of
the correlation matrix is rather difficult; the asset graph method (coupled
with the CPM) seems to provide more coherent results.
Gao and Wong applied the CPM to document clustering in [30]. The
graph of the documents was constructed using document similarity (more
similar documents are connected by a stronger edge). According to the
results, the communities obtained via the CPM can outperform some typical
algorithms on benchmark data sets, and shed light on natural document
clustering.
An interesting application of the CPM is shown by Castelló et. al. in the
study of the dynamics of competing opinions [19]. In the voter model, the
state of the agents can be either A or B [42], whereas in the AB model a
third, intermediate AB state is included as well [18]. The network of voters
is constructed with the help of a variation of the social network models
based on the balance between cyclic closure and focal closure [81, 48],
briefly discussed in Sect. 5.2.3. At each time step, the state (opinion) of
a randomly selected agent is changed with a probability depending on the
states of its neighbours. Starting from a random initial opinion distribution,
in both models the system converges to consensus, where all nodes are in the
same state. However, in the AB model the average time needed to reach
the ordered state is highly dependent on the structure of the underlying
network. According to the results of Castelló et. al., when a rich, apparent
community structure can be detected with the help of the CPM, the lifetime
distribution of the meta-stable (disordered) states becomes a power-law,
so that the mean lifetime is not representative of the dynamics. These
trapped meta-stable states, which can order at all time scales, originate in
the mesoscopic network structure.
Finally, we mention the comparative analysis of the directed communi-
ties in Google’s own webpages (the links follow the direction of the hyper-
links), a word association network (the directions of the links indicate that
people in the survey associated the end point of the link with its start point),
a university e-mail network (the links point from the sender to the recip-
ient), and a transcriptional regulatory network (the links point from the
regulating protein to the protein of the regulated gene) in [67]. The iden-
tified directed modules were inherently overlapping and the investigated
networks could be classified into two major groups in terms of the over-
laps between the modules. In the word association network and Google’s
webpages overlaps were likely to contain in-hubs, whereas the modules in
k-clique Percolation and Clustering 403

the email and transcriptional regulatory networks tended to overlap via out-
hubs. In other words, in these two major classes of directed graphs, directed
modules “point” towards and away from their shared regions.

References

[1] B. Adamcsek, G. Palla, I. J. Farkas, I. Derényi and T. Vicsek, CFinder: Locating

cliques and overlapping modules in biological networks, Bionformatics, 22 (2006),
1021–1023.
[2] T. Aittokallio and B. Schwikowski, Graph-based methods for analysing networks
in cell biology, Briefings in Bioinformatics, 7 (2006), 243–255.
[3] R. Albert and A.-L. Barabási, Statistical mechanics of complex networks, Rev.
Mod. Phys., 74 (2002), 47–97.
[4] A. V. Antonov and H. W. Mewes, Complex functionality of gene groups identified
from high-throughput data, J. Mol. Biol., 363(1) (2006), 289–296.
[5] A.-L. Barabási and R. A. and, Emergence of scaling in random networks, Science,
286 (1999), 509–512.
[6] A.-L. Barabási, H. Jeong, Z. Néda, E. Ravasz, A. Schubert and T. Vicsek, Evolution
of the social network of scientific collaborations, Physica A, 311 (2002), 590–614.
[7] A.-L. Barabási and Z. N. Oltvai, Network Biology: Understanding the Cells’s
Functional Organization, Nature Reviews Genetics, 5 (2004), 101–113.
[8] V. Batagelj and M. Zaveršnik, Short cycle connectivity, Discrete Mathematics, 307
(2007), 310–318.
[9] J. Baumes, M. Goldberg and M. Magdon-Ismail, Efficient Identification of Over-
lapping Communities, Lect. Notes in Computer Science, 3495 (2005), 27–36.
[10] A. Beyer, S. Bandyopadhyay and T. Ideker, Integrating physical and genetic maps:
from genomes to interaction networks, Nature Reviews Genetics, 9 (2007), 699–710.
[11] G. Bianconi and M. Marsili, Emergence of large cliques in random scale-free net-
works, Europhys. Lett., 74 (2006), 740–746.
[12] G. Bianconi and M. Marsili, Number of cliques in random scale-free network
ensembles, Physica D-Nonlinear Phenomena, 224 (2006), 1–6.
[13] B. Bollobás, Graph Theory and Combinatorics: Proceedings of the Cambridge Com-
binatorial Conference in honour of Paul Erdős, Academic, New York, 1984.
[14] B. Bollobás, Random graphs, Cambridge University Press, Cambridge, 2nd edition,
2001.
[15] B. Bollobás and O. Riordan, Clique percolation, arXiv:0804.0867, 2008.
[16] E. I. Boyle, S. A. Weng, J. Gollub, H. Jin, D. Botstein, J. M. Cherry and G. Sher-
lock, GO::TermFinder–open source software for accessing Gene Ontology informa-
tion and finding significantly enriched Gene Ontology terms associated with a list
of genes, Bioinformatics, 20 (2004), 3710–3715.
404 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

[17] S. Carmi, S. Havlin, S. Kirkpatrick, Y. Shavitt and E. Shir, A Model of Internet

Topology Using k-shell Decomposition, Proc. Natl. Acad. Sci. USA, 104 (2007),
11150–11154.
[18] X. Castelló, V. M. Eguı́luz and M. S. Miguel, Ordering dynamics with two non-
excluding options: bilingualism in language competition, New Journal of Physics,
8 (2006), 308.
[19] X. Castelló, R. Toivonen, V. M. Eguı́luz, K. Kaski and M. S. Miguel, Anomalous
lifetime distributions and topological traps in ordering dynamics, Europhys. Lett.,
79 (2007), 66006.
[20] J. M. Cherry, C. Ball, S. Weng, G. Juvik, R. Schmidt, C. Adler, B. Dunn, S. Dwight,
L. Riles, R. K. Mortimer and D. Botstein, Genetic and physical maps of Saccha-
romyces cerevisiae, Nature, 387(6632 Suppl) (1997), 67–73.
[21] I. Derényi, G. Palla and T. Vicsek, Clique percolation in random networks, Phys.
Rev. Lett., 94 (2005), 160202.
[22] S. N. Dorogovtsev, A. V. Goltsev and J. F. F. Mendes, k-core architecture and
k-core percolation on complex networks, Physica D-Nonlinear Phenomena, 224
(2006), 7–19.
[23] S. N. Dorogovtsev, A. V. Goltsev and J. F. F. Mendes, k-core organization of
complex networks, Phys. Rev. Lett., 96 (2006), 040601.
[24] P. Erdős and A. Rényi, On the evolution of random graphs, Publ. Math. Inst. Hung.
Acad. Sci., 5 (1960), 17–61.
[25] M. G. Everett and S. P. Borgatti, Analyzing Clique Overlap, Connections, 21
(1998), 49–61.
[26] B. S. Everitt, Cluster Analysis, Edward Arnold, London, 3th edition, 1993.
[27] I. J. Farkas, D. Ábel, G. Palla and T. Vicsek, Weighted network modules, New
Journal of Physics, 9 (2007), 180.
[28] G. Finocchiaro, F. M. Mancuso, D. Cittaro and H. Muller, Graph-based identifica-
tion of cancer signaling pathways from published gene expression signatures using
PubLiME, Nucl. Ac. Res., 35 (2007), 2343–2355.
[29] S. Fortunato and M. Barthelemy, Resolution limit in community detection, Proc.
Natl. Acad. Sci. USA, 104 (2007), 36–41.
[30] W. Gao and K.-F. Wong, Natural document clustering by clique percolation in
random graphs, Lect. Notes in Comp. Sci., 4182 (2006), 119–131.
[31] D. Gfeler, J.-C. Chappelier and P. D. L. Rios, Finding instabilities in the commu-
nity structure of complex networks, Phys. Rev. E., 72 (2005), 056135.
[32] M. Girvan and M. E. J. Newman, Community structure in social and biological
networks, Proc. Natl. Acad. Sci. USA, 99 (2002), 7821–7826.
[33] A. V. Goltsev, S. N. Dorogovtsev and J. F. F. Mendes, k-core (bootstrap) percola-
tion on complex networks: Critical phenomena and nonlocal effects, Phys. Rev. E.,
73 (2006), 056101.
k-clique Percolation and Clustering 405

[34] M. C. González, H. J. Herrmann, J. Kertész and T. Vicsek, Community structure

and ethnic preferences in school friendship networks, Physica A-Statistical Mechen-
ics and its Applications, 379 (2007), 307–316.
[35] M. C. González, P. G. Lind and H. J. Herrmann, System of mobile agents to model
social networks, Phys. Rev. Lett., 96 (2006), 088702.
[36] R. Guimerà and L. A. N. Amaral, Functional cartography of complex metabolic
networks, Nature, 433 (2005), 895–900.
[37] R. Guimerà, S. Mossa, A. Turtschi and L. A. N. Amaral, The worldwide air
transportation network: Anomalous centrality, community structure and cities’
global roles, Proc. Natl. Acad. Sci. USA, 102 (2005), 7794–7799.
[38] R. Guimerà, M. Sales-Pardo and L. A. N. Amaral, Module identification in bipartite
and directed networks, Phys. Rev. E., 76 (2007), 036102.
[39] U. Guldener, M. Munsterkotter, G. Kastenmuller, N. Strack and J. van Helden,
CYGD: the Comprehensive Yeast Genome Database, Nucl. Ac. Res., 33 (2005),
D364–D368.
[40] L. H. Hartwell, J. J. Hopfield, S. Leibler and A. W. Murray, From molecular to
modular cell Biology, Nature, 402 (1999), 6761, supplement C47–C52.
[41] T. Heimo, J. Saramäki, J.-P. Onnela and K. Kaski, Spectral and network meth-
ods in the analysis of correlation matrices of stock returns, Physica A-Statistical
Mechanics and its Applications, 383 (2007), 147–151.
[42] R. A. Holley and T. M. Liggett, Ergodic theorems for weakly interacting infinite
systems and voter model, Annals of Probability, 3 (1975), 643–663.
[43] H. Jeong, Z. Néda and A.-L. Barabási, Measuring preferential attachment for
evolving networks, Europhysics Letters, 61 (2003), 567–572.
[44] P. F. Jonsson and P. A. Bates, Global topological features of cancer proteins in the
human interactome, Bioinformatics, 22 (2006), 2291–2297.
[45] P. F. Jonsson, T. Cavanna, D. Zicha and P. A. Bates, Cluster analysis of net-
works generated through homology: automatic identification of important protein
communities involved in cancer metastasis, BMC Bioinformatics, 7 (2006), 2.
[46] S. Knudsen, A Guide to Analysis of DNA Microarray Data, Wiley-Liss, 2nd edition,
2004.
[47] N. J. Krogan, G. Cagney, H. Y. Yu, G. Q. Zhong, X. H. Guo, A. Ignatchenko, J. Li,
S. Y. Pu, N. Datta, A. P. Tikuisis, T. Punna, J. M. Peregrin-Alvarez, M. Shales,
X. Zhang, M. Davey, M. D. Robinson, A. Paccanaro, J. E. Bray, A. Sheung,
B. Beattie, D. P. Richards, V. Canadien, A. Lalev, F. Mena, P. Wong, A. Starostine,
M. M. Canete, J. Vlasblom, S. W. C. Orsi, S. R. Collins, S. Chandran, R. Haw,
J. J. Rilstone, K. Gandi, N. J. Thompson, G. Musso, P. S. Onge, S. Ghanny,
M. H. Y. Lam, G. Butland, A. M. Altaf-Ui, S. Kanaya, A. Shilatifard, E. O’Shea,
J. S. Weissman, C. J. Ingles, T. R. Hughes, J. Parkinson, M. Gerstein, S. J. Wodak,
A. Emili and J. F. Greenblatt, Global landscape of protein complexes in the yeast
Saccharomyces cerevisiae, Nature, 440 (2006), 637–643.
[48] J. M. Kumpula, J.-P. Onnela, J. Saramäki, K. Kaski and J. Kertész, Emergence of
communities in weighted networks, Phys. Rev. Lett., 99 (2007), 228701.
406 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

[49] J. M. Kumpula, J. Saramäki, K. Kaski and J. Kertész, Limited resolution in com-

plex network community detection with Potts model approach, European Physical
Journal B., 56 (2007), 41–45.
[50] O. Mason and M. Verwoerd, Graph theory and networks in Biology, IET Systems
Biology, 1 (2007), 89–119.
[51] J. F. F. Mendes and S. N. D. and, Evolution of Networks: From Biological Nets to
the Internet and WWW, Oxford University Press, Oxford, 2003.
[52] T. Nepusz, A. Petróczi, L. Négyessy and F. Bazsó, Fuzzy communities and the
concept of bridgeness in complex networks, Phys. Rev. E., 77 (2008), 016107.
[53] M. E. J. Newman, Clustering and preferential attachment in growing networks,
Phys. Rev. E., 64 (2001), 025102.
[54] M. E. J. Newman, Detecting community structure in networks, Eur. Phys. J. B.,
38 (2004), 321–330.
[55] M. E. J. Newman, Fast algorithm for detecting community structure in networks,
Phys. Rev. E., 69 (2004), 066133.
[56] M. E. J. Newman and M. Girvan, Finding and evaluating community structure in
networks, Phys. Rev. E., 69 (2004), 026113.
[57] M. E. J. Newman and E. A. Leicht, Mixture models and exploratory analysis in
networks, Proc. Natl. Acad. Sci. USA, 104 (2007), 9564–9569.
[58] M. E. J. Newman, S. H. Strogatz and D. J. Watts, Random graphs with arbitrary
degree distribution and their applications, Phys. Rev. E., 64 (2001), 026118.
[59] J.-P. Onnela, A. Chakraborti, K. Kaski, J. Kertész and A. Kanto, Dynamics of
market correlations: Taxonomy and portfolio analysis, Phys. Rev. E., 68 (2003),
056110.
[60] J.-P. Onnela, J. Saramäki, J. Hyvönen, G. Szabó, M. A. de Menezes, K. Kaski, A. L.
Barabási and J. Kertész, Analysis of a large-scale weighted network of one-to-one
human communication, New Journal of Physics, 9 (2007), 179.
[61] J.-P. Onnela, J. Saramäki, J. Hyvönen, G. Szabó, D. Lazer, K. Kaski, J. Kertész
and A.-L. Barabási, Structure and tie strengths in mobile communication networks,
Proc. Natl. Acad. Sci. USA, 104 (2007), 7332–7336.
[62] J.-P. Onnela, J. Saramäki, J. Kertész and K. Kaski, Intensity and coherence of
motifs in weighted complex networks, Phys. Rev. E., 71 (2005), 065103.
[63] G. Palla, A.-L. Barabási and T. Vicsek, Quantifying social group evolution, Nature,
446 (2007), 664–667.
[64] G. Palla, I. Derényi, I. Farkas and T. Vicsek, Uncovering the overlapping com-
munity structure of complex networks in nature and society, Nature, 435 (2005),
814–818.
[65] G. Palla, I. Derenyi, I. Farkas, T. Vicsek, P. Pollner and D. Ábel, Free software
for finding overlapping dense groups of nodes in networks, based on the clique
percolation method.
[66] G. Palla, I. Derényi and T. Vicsek, The critical point of k-clique percolation in the
Erdős–Rényi graph, J. Stat. Phys., 128 (2007), 219–227.
k-clique Percolation and Clustering 407

[67] G. Palla, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek, Directed network

modules, New Journal of Physics, 9 (2007), 186.
[68] P. Pollner, G. Palla and T. Vicsek, Preferential attachment of communities: The
same principle, but a higher level, Europhys. Lett., 73 (2006), 478–484.
[69] B. Ráth and B. Tóth, Triangle percolation in mean field random graphs – with
PDE, arXiv:0712.2646v1, 2007.
[70] E. Ravasz, A. L. Somera, D. A. Mongru, Z. N. Oltvai and A.-L. Barabási, Hier-
archical organization of modularity in metabolic networks, Science, 297 (2002),
1551–1555.
[71] J. Reichardt and S. Bornholdt, Detecting Fuzzy Community Structures in Complex
Networks with a Potts Model, Phys. Rev. Lett., 93 (2004), 218701.
[72] J. Reichardt and S. Bornholdt, Statistical mechanics of community detection, Phys.
Rev. E., 74 (2006), 016110.
[73] A. W. Rives and T. Galitski, Modular organization of cellular networks, Proc. Natl.
Acad. Sci. USA, 100 (2003), 1128–1133.
[74] J. Scott, Social Network Analysis: A Handbook, Sage Publications, London, 2nd
edition, 2000.
[75] S. B. Seidman, Network structure and minimum degree, Social Networks, 5 (1983),
269–287.
[76] R. M. Shiffrin and K. Börner, Mapping knowledge domains, Proc. Natl. Acad. Sci.
USA, 101 (2004), 5183–5185.
[77] R. D. Smith, The network of collaboration among rappers and its community
structure, J. Stat. Mech., page P02006, 2006.
[78] V. Spirin and K. A. Mirny, Protein complexes and functional modules in molecular
networks, Proc. Natl. Acad. Sci. USA, 100 (2003), 12123–12128.
[79] G. Szabó and G. Fáth, Evolutionary games on graphs, Physics Reports-Review
Section of Physics Letters, 446 (2007), 97–216.
[80] G. Szabó, J. Vukov and A. Szolnoki, Phase diagrams for an evolutionary prisoner’s
dilemma game on two-dimensional lattices, Phys. Rev. E., 72 (2005), 047107.
[81] R. Toivonen, J.-P. Onnela, J. Saramäki, J. Hyvönen and K. Kaski, A model for
social networks, Physica A-Statistical Mechanics and its Applications, 370 (2006),
851–860.
[82] J. R. Udry, P. S. Bearman and K. M. Harris, Public-use data set from Add
Health, funded by a grant from National Institute of Child Health and Human
Development.
[83] T. Vicsek, Phase transitions and overlapping modules in complex networks, Physica
A-Statistical Mechanics and its Applications, 378 (2007), 20–32.
[84] J. Vukov, G. Szabó and A. Szolnoki, Cooperation in the noisy case: Prisoner’s
dilemma game on two types of regular random graphs, Phys. Rev. E., 73 (2006),
067103.
408 G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek

[85] D. J. Watts, P. S. Dodds and M. E. J. Newman, Identity and search in social

networks, Science, 296 (2002), 1302–1305.
[86] D. M. Wilkinson and B. A. Huberman, A method for finding communities of related
genes, Proc. Natl. Acad. Sci. USA, 101 (2004), 5241–5248.
[87] I. Xenarios, D. W. Rice, L. Salwinski and M. K. Baron, DIP: the Database of
Interacting Proteins, Nucl. Ac. Res., 28 (2000), 289–291.
[88] S. Zhang, X. Ning and X.-S. Zhang, Identification of functional modules in a PPI
network by clique percolation clustering, Comp. Biology and Chemistry, 30 (2006),
445–451.
[89] S. Zhang, R.-S. Wang and X.-S. Zhang, Uncovering fuzzy community structure in
complex networks, Phys. Rev. E., 76 (2007), 046103.
[90] X. Zhu, M. Gerstein and M. Snyder, Getting connected: analysis and principles of
biological networks, Genes & Development, 21 (2007), 1010–1024.

Gergely Palla, Illés J. Farkas, Dániel Ábel, Illés J. Farkas, Imre

Péter Pollner and Tamás Vicsek Derényi and Tamás Vicsek
Statistical and Biological Physics Department of Biological Physics
Research Group of HAS Eötvös University
H-1117 Budapest, Pázmány Péter H-1117 Budapest, Pázmány Péter
sétány 1/A sétány 1/A
Hungary Hungary
[email protected]
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 409–443.

Chapter 10
The Inverse Problem of Evolving Networks –
with Application to Social Nets

GÁBOR CSÁRDI, KATHERINE J. STRANDBURG, JAN TOBOCHNIK

and PÉTER ÉRDI

1. Introduction

Many complex systems can be modeled by graphs [8]. The vertices of

the graph represent objects of the system, and the edges of the graph the
relationships between these objects. These relationships may be structural
or functional, according to the modeler’s needs [1, 29, 7].
Following scientific convention, in this paper we assume that all impor-
tant information about a given system is encoded in this simple model; more
precisely, in an attributed graph, where the vertices and edges may possess
labels from a given alphabet.
Let us give a few examples. In a citation network of scientific papers [11],
the vertices model the papers, the edges the citations between them; the
vertices can be labeled by the date of the paper, its authors, and probably a
list of keywords. By including these and only these properties in the model,
we implicitly neglect everything else, such as the title of the paper and the
nationality and gender of the authors.
In a network of airline flights [3], the vertices “are” airports and the
edges represent direct flights between them. The vertices might be labeled
410 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

with each airport’s number of terminals and home country, and the edges
with the airlines and possibly the flight time.
In a neural network model of synchronization in the hippocampus, one
usually has to distinguish between pyramidal cells and interneurons [31];
here the edges are the synapses from the presynaptic cells to the postsynap-
tic ones. Furthermore, the synapses can be excitatory or inhibitory.
In this paper we study network evolution. For us, this term means that
the structure of the networks, in other words the modeled complex systems,
change in time. In the citation network, new papers are published and make
new citations to other papers. In the airline network, new airports might be
built or old ones might be shut down, new flights introduced between two
already existing airports, or conversely, underused flights might be removed.
In the neural network example, network evolution means the ontogenesis of
the hippocampal structure.
In other words, the evolution of a network involves the addition and/or
deletion of edges and vertices. We intend to model these kinds of phenom-
ena, by using a discrete time, stochastic framework. We are particularly
interested in the inverse problem: what is the best description of the evolu-
tion of a given network, or set of networks, in a well-defined mathematical
framework? This is a data-driven approach; one can even interpret it as ex-
perimental science: we make experiments on the networks to see how they
behave, although our possibilities are often quite limited. For example we
cannot change the flight time of an “edge” just to see how the network will
react. If, however, such an event has already happened for some particular
reason, we are able to observe one possible reaction.
Put a different way, in our approach the input is the data collected about
the evolution of a network, the output is a set of parameters for a stochastic
model framework. In our case, the parameters are so-called kernel functions.
This chapter is organized as follows. First, we define the stochastic
kernel-based framework in Sect. 2 and formalize the inverse problem. In
Sect. 3 we show two methods for estimating kernel functions, the frequentist
and the maximum likelihood methods. In Sect. 4 we show some possible
applications of the methodology.
The Inverse Problems of Evolving Networks 411

2. The Model Framework

The model framework we introduce in this section is a discrete time, discrete

space, stochastic system. Discrete time means that we are required to create
time steps: snapshots of the network. This is like composing a movie from
still images. Just like there is no motion in each snapshot of such a movie,
all structural changes in our evolving graphs happen between the time steps.
Our model framework is a set of simple equations describing the struc-
tural changes of the network in the past and the possible changes in the
future. The framework is stochastic, however; in other words does not give
changes exactly, only their probabilities. The assumption is that everything
is possible, but that some outcomes are less probable than others.
The evolution equations depend on the properties of the vertices and
edges. These include both structural properties and vertex and edge labels.
The dependence is formalized by kernel functions: these depend on the
vertex/edge properties and map them to non-negative real numbers, which
are in turn used as the odds of the various possible structural changes.

2.1. Events

Four elementary structural changes are possible in a network: addition of

an edge, deletion of an edge, addition of a vertex, deletion of a vertex.
This gives us four kernel functions: the first will predict which edge will be
added, assuming that an edge will be added, the second does the same with
edge deletions and the third and fourth the same for vertices. A fifth kernel
function might be defined to predict which elementary event happens next.
Which of these events a modeler needs to include in the framework
depends on the purpose of the modeling, and also on the modeled system.
In a citation network, edges and vertices are never deleted from the network,
so there is no need for their corresponding kernel functions. In fact, by
modeling citation networks with a single kernel function – which predicts
which edges will be added – reveals a significant amount of information
about the network.
412 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

2.2. The Kernel Functions

The first kernel function we define is for edge additions. Recall, that a
kernel function depends on the properties of the vertices and edges, let us
now assume that only the vertex properties are important, the properties of
a vertex v are denoted by xv . These might be time dependent: xv (t), where
xv (t) is from a set of strings from a finite alphabet. Another way of looking
at the xv values is that they can be vectors composed of arbitrary numbers
or symbols. For example if we think that the important properties of a
vertex in a citation network are 1) the (current) in-degree of the vertex, 2)
the number of time steps passed since the addition of the vertex and 3)
the (current) number of publications by the first author represented by the
vertex, then a property vector might be xv (t0 ) = (1, 123, 10), meaning that
at time step t0 vertex v has in-degree 1, v was added 123 time steps ago
and the first author of v has published 10 papers up to time step t0 . If this
author publishes another paper between time steps t0 and t0 + 1, and this
new paper cites his old paper v, then xv (t0 + 1) = (2, 124, 11). (We assume
that noone else cited v between t0 and t0 + 1, and that this was the only
paper published by the first author of v in this period.)
The kernel function A maps the pairs of all possible labels, or property
vectors (X), to the set of non-negative numbers: A : X × X → R. The higher
a kernel function value, the more probable the realization of the given edge.
More precisely, the probability that a given edge e connects an x-vertex
(a vertex with property vector x) to a y-vertex is given as
 
+  A(x, y)N t(e), x, y
(1) P c(e, x, y) = 1 =
 .
   
(x ,y  )∈X×X A(x , y )N t(e), x , y

Let c(e, x, y) be indicator random variables, one for every edge–property

vector triple; c(e, x, y) = 1 if and only if edge e connects an x vertex to
y-vertex. Moreover, t(e) is the time step before the addition of edge e,
N (t(e), x, y) is the number of possible x-y connections in time step t(e).
For example if there is a single x vertex and a single y-vertex in time step t
then N (t, x, y) = 1. The sum in the denominator goes over all vertex-type
pairs. It is the same as summing over the A(·, ·) values of all missing (but
theoreticallypossible)
 edges. We will often denote the normalization factor
by S(t) or S t(e) .
Eq. (1) gives only the probability that two vertex types will be con-
nected. As two vertices with the same property vector are assumed to be
The Inverse Problems of Evolving Networks 413

completely identical, the probability that two given vertices v and w, with
vertex types xv and xw will be connected is simply
+ 
P c(e, xv , xw ) = 1
(2) P [e connects v and w] =   .
N t(e), xv , xw

Note that an A(·, ·) kernel function is the same in all respects as cA(·, ·),
where c is a positive number. This means that the value of A(·, ·) for one
vertex-type pair can be chosen arbitrarily; we will often use
this fact and
either choose A(x, y) = 1 for some (x, y) vertex-type pair or i A(i) = 1,
where i goes over all vertex-type pairs.

Fig. 1. The direct and inverse problems. In the direct problem we have the A(·, ·) kernel
function (and maybe other kernel functions for more general networks) and we use a
kernel-based network generator algorithm to create an artificial network. The question
is what kind of structures we can obtain from a given class of kernel functions.
In the inverse problem we examine a real network, assembled from observations. The
kernel-based measurement algorithm takes this as the input and gives a possible kernel
function as the output. If the measurement algorithm is good then the output kernel
function is good, in other words it creates the observed network with high probability
when used in a kernel-based generator.

The inverse problem (Fig. 1.) requires the construction of an A(·, ·)

kernel function that describes the evolution of the network well. Certainly,
this makes sense only if we are able to define what well means. A natural
definition would be the probability that the given kernel function generates
exactly the network G under study or, more precisely, the probability that
the kernel function is able to guess the outcome of all edge experiments:

+  A(x∗e , x∗∗
e )
(3) P A(·, ·) generates G =  .
e
S t(e)
414 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

Here, the (x∗e , x∗∗

e ) are the types (i.e. property vectors) of the vertices
connected by edge e (at time step t(e)). If these probabilities are (relatively)
high, then the kernel function fits the network well.
For a large network these probabilities are very small (even if relatively
high); we take the logarithm of Eq. 3 and translate it by subtracting the
logarithmic goodness of the totally random attachment. This way, a kernel
function with goodness value zero gives absolutely no information about the
network. We also divide by the number of edges to obtain a value typically
between zero and one. Thus, the goodness of A(·, ·) is defined as:
& '
1  A(x∗e , x∗∗
e )
 1
(4) log   − log .
|E| e S t(e) e
t(e)

Note that the goodness score is not an absolute measure; it makes no

sense to compare the goodness of kernel functions for different networks.
It does make sense, however, to compare the goodness of different kernel
functions on the same network. This allows alternative descriptions of a
system, and we have a quantitative measure to tell us which one is better.
We shall show example applications for this in Sect. 4.
Assuming we have a procedure to find the kernel function with the
highest goodness on a given set of vertex properties, we can also tell whether
it is worth extending a set of model parameters with an additional one: if
the best kernel function for the new, more difficult model is not substantially
better than the one for the simpler model, then the extension is not needed,
as it adds nothing (or almost nothing) to the goodness.

2.3. The General Framework

Just as edge addition can be described by defining the addition probabilities

based on a simple kernel function, the same can be easily done with edge
deletion, vertex addition and vertex deletion as well. We do not have the
space to give these here, see [12] for a detailed description.
The Inverse Problems of Evolving Networks 415

3. Solving the Inverse Problem

We shall show two ways for solving the inverse problem: the frequentist
and the maximum likelihood methods. Both methods will be formalized for
the edge addition kernel function for citation networks, assuming that the
kernel depends only on the properties of the potentially cited vertex. This
simplification allows for shorter notation, and the results easily generalize
to other kernel functions.

3.1. The Frequentist Method

The Scoring Model. The first method we develop for the inverse problem
follows a simple scoring model. From Eq. 1 we can easily obtain the kernel
function:
+   
P c(e, x) = 1 S t(e)
(5) A(x) =   .
N t(e), x

An observed edge+ e either cites

 an x-vertex or not: if not, then the correct
estimation for P c(e, x) = 1 is zero; thus the approximation, or the score
for A(x), is also zero. In the other case, the approximation is
 
S t(e)
(6) Āe (x) =  ,
N t(e), x

where Āe (x) reads as “the approximation of A(x) based on edge e”. Taking
the average of the scores we get
 
1  S t(e)
(7) Ā(x) =  ,
|Ex | N t(e), x
e∈i(x)

where |Ex | is the number of time steps in which there was at least one
x-vertex present in the network.
To find S(t) and A(·) consistently, we can apply an iterative approach:
1. Assume that A0 (x) (the 0-th approximation of A(·)) is the same for
all x, i.e. A(x) = 1, and determine S0 (t) accordingly: S0 t(e) =
  
|V G t(e) |, the number of vertices in the graph in time step t(e).
416 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi
 
2. Use Sk t(e) to determine
  Ak+1 (x) for all x via the frequentist method
and calculate Sk+1 t(e) from it.
3. Norm the Ak+1 vector using a suitable vector norm, e.g. divide by
the largest A(·) element. This essentially does not change the kernel
function and allows us to handle numerical problems.
4. Repeat the previous two steps until Ak (x) and Ak+1 (x) are closer to
each other than a predefined δ-threshold.
If this procedure converges, then we obtain a consistent solution for A(·).
We shall show that the procedure always converges to a unique solution if
the network fulfills certain minimal requirements.

Convergence and Uniqueness. It is not very difficult to prove that the

procedure discussed in the previous section converges. For this we write
the two substeps updating Ak+1 (x) and Sk+1 (t) into   a single
  equation. We
need some more notation. First, let N t(e) = |V G t(e) | for
simplicity.
We denote the ratio of x-vertices at time step t by px ; obviously, x ptx = 1
t

for all t. Finally, the set of edges citing x-vertices is denoted by i(x). (More
precisely, x-vertices at the time of the citation, as the vertex types may
change in time.) Calculating Sk+1 (t) is given by

  
n
t(e)
t(e)
(8) Sk+1 t(e) = N pj Ak (j).
j=1

After doing one step of the frequentist estimation, we have

1   pj
n t(e)
(9) Ak+1 (x) = A (j).
t(e) k
|Ex |
j=1 e∈i(x) px

The last equation gives a linear transformation of Ak to obtain Ak+1 .

It is very similar to the ‘power method’ that can be used to calculate the
leading eigenvector of a matrix in many cases.
The transformation matrix in (9) is somewhat special, because its ele-
ments are all non-negative. Let us for now assume that all the elements are
positive, because this has important consequences for the power method.
If the elements of the matrix are positive, then according to the Perron–
Frobenius theorem it is true that
The Inverse Problems of Evolving Networks 417

1. there is a unique (real) dominant eigenvalue, and that

2. the eigenvector associated with the dominant eigenvector is all posi-
tive.
3. In addition, having a real dominant eigenvalue ensures that the power
method converges to the leading eigenvector.

Let us take a look at the all-positive matrix assumption. In our case this
requirement means that for every pair of vertex types there must be a case
when the first vertex type is cited and the second vertex type is present in
the network and vice versa. All pairs of vertex types have to be measured
against each other. This ‘restriction’ is sufficient for the convergence, but it
is an open question whether it is required, too.
We have proven that the proposed iteration method always converges
to a positive A(·) function, provided that the representation restriction is
fulfilled. Moreover, we have also proven that the solution of the iteration is
all-positive and unique. To summarize, the following theorem is true:

Theorem 1. Starting from [1, 1, . . . , 1], the frequentist method, as defined

by Eq. (9), converges to a positive vector if the sufficient representation
restriction is fulfilled.

3.2. The Maximum Likelihood Method

The frequentist solution discussed in the previous section is intuitive: we

estimate a quantity by the average of the measurement experiments per-
formed for it. However, it does not generally guarantees that the obtained
kernel function has maximum goodness among all kernels on a set of fixed
vertex properties.
In this section we introduce a maximum likelihood method for extracting
the kernel-function from the network evolution data. Maximum likelihood
methods aim to find the kernel function which is the most probable for a
given network. In our framework this means that we explicitly search for
the kernel function with the highest goodness value.

Citation Networks. The probability that A(·) +is the kernel

 function gen-
erating a particular network G is denoted by P A(·) | G . According to
418 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

Bayes’ rule this is

+  + 
+  P G | A(·) P A(·)
(10) P A(·) | G = ,
P [G]

+  + 
where P G | A(·) is the probability that A(·) generates G, P A(·) is
the probability of the appearance of the kernel function A(·) and P [G] is
the probability of the appearance
+ of the graph G. Sadly, we generally do
not know anything about P A(·) and assume that every A(·) is equally
probable. Here, an “Occam’s razor” assumption could be used to favor
simpler kernel functions, but it is unclear how to do this quantitatively.

+ P [G] is only a constant for our concerns, so we need to maximize

P G | A(·) , which is almost the same as finding the kernel function with
maximum goodness. The goodness of a kernel function is not exactly the
probability that it generates the network, but the probability that it can
guess the outcome of all edge experiments correctly. If we assume that these
experiments are independent then this is the same as the probability that
the kernel generates the network in exactly the same way as it happened
during its evolution.
We need to find

A(xe )
(11) max  .
A(·)
e
S t(e)

If we want to maximize in the space of all possible kernel functions, then

the solution is trivial: it is easy to define vertex properties and a kernel that
generates a given graph with probability 1. It makes more sense to fix the
property vectors and search for the best kernel function based on these fixed
vertex types. From now on, we focus on the problem of finding the best
kernel function values assuming that the property vectors are already fixed.

The Maximization Problem. We first define the function to be maxi-

mized for citation networks. Recall that the S normalization factor is

  
n
t(e)
t(e)
(12) S t(e) = N pi A(i).
i=1
The Inverse Problems of Evolving Networks 419

Let us denote the number of citations to i-vertices by Mi . Then we need

to maximize
A(xe ) n $ n %−1
t(e)
(13)   = A(i)Mi
pi A(i) .
e
S t(e) i=1 e i=1

We have omitted the N t(e) factors here, as they do not effect the position
of the maximum value.

Existence of the Solution

Theorem 2. The target function (13) attains a minimum and a maximum

over the set of non-negative A vectors.

As the kernel function has no scale,

we can restrict ourselves to examine
only kernel functions which satisfy i A(i) = 1, where i goes over all
vertex types. Then the allowed kernel functions form a compact set, as
it is closed and bounded. Any continuous function on a compact set has
a maximum and a minimum according to the basic theorem of calculus by
Bolzano and Weierstrass. The target function is obviously continuous in
all variables (all A(·)), thus the existence of the solution follows from the
Bolzano–Weierstrass theorem.

Uniqueness of the Solution. In practice it is better to work with the

logarithm of the target function; this is


n  $
n %
t(e)
(14) Mi log A(i) − log pi A(i) .
i=1 e i=1

The partial derivatives with respect to all A(k) must be zero in order for
this to have a maximum value:

Mk  pk
t(e)
(15) =
t(e)
, 1 ≤ k ≤ n,
A(k) n
e i=1 pi A(i)

which can be written in fixed-point equation form as

& '−1
 pk
t(e)
(16) A(k) = Mk
n t(e)
, 1 ≤ k ≤ n.
e i=1 pi A(i)
420 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

It gives us an interesting insight if we rewrite these equations in the form

 t(e)
pk A(k)
(17)
n t(e)
= Mk , 1 ≤ k ≤ n.
e i=1 pi A(i)

The left hand side of (17) is the expected number of citations to k-vertices
based on a given A(·) kernel function and the right hand side is the observed
number of citations. Optimality requires that these are exactly the same.
It is true that the target function is minimal if A(k) = 0 for some k
and the minimum value is zero. This means that if A(k) > 0 then all
singular points (the points at which all partial derivatives are zero) are either
maximum points or saddle points. We shall prove that, for any two A1 (·)
and A2 (·) singular points satisfying (17), it is true that A1 (·) = cA2 (·) with
some real c > 0. From this it follows that every such point is a maximum
point, as otherwise no maximum point would exist.
Let us assume that the A1 and A2 kernel functions satisfy (17), from
this we have

 t(e)
pk A1 (k)  t(e)
pk A2 (k)
(18)
n t(e)
=
n t(e)
, 1 ≤ k ≤ n,
e i=1 pi A1 (i) e i=1 pi A2 (i)

and by denoting the two normalization factors by S 1 (e) and S 2 (e) we get

 pt(e) A1 (k)S 2 (e) − pt(e) A2 (k)S 1 (e)

(19) k k
= 0, 1 ≤ k ≤ n.
e
S 1 (e)S 2 (e)

Some simple algebra leads to

n
n t(e) 1
 pt(e) A1 (k) t(e) 2 t(e) 2
k i=1 pi A (i) − pk A (k) i=1 pi A (i)
(20) 1 2
=
e
S (e)S (e)

 pt(e) pt(e) A1 (k)A2 (1) + · · · + pt(e)

n p
t(e) 1
A (k)A2 (n)
1
= k k
−
e
S 1 (e)S 2 (e)

 pt(e) pt(e) A1 (1)A2 (k) + · · · + pt(e)

n p
t(e) 1
A (n)A2 (k)
1
− k k
=
e
S 1 (e)S 2 (e)
The Inverse Problems of Evolving Networks 421

   pt(e)
1 pk
t(e)
= A (k)A (1) − A (k)A (1)
1 2 2 1
+ ··· +
e
S 1 (e)S 2 (e)

   pnt(e) pt(e)
+ A (k)A (n) − A (k)A (n)
1 2 2 k
1 (e)S 2 (e)
1
= 0.
e
S

We assume that

 pit(e) pjt(e)
(21) > 0, 1 ≤ i, j ≤ n,
e
S 1 (e)S 2 (e)

in which case the sum in (20) can be zero in two different ways.
1. All terms are zero, i.e.

(22) A1 (k)A2 (i) = A2 (k)A1 (i), 1 ≤ i ≤ n,

and from here

(23) A1 (i) = cA2 (i), 1 ≤ i ≤ n.

2. There is at least one negative term. There exists 1 ≤ i ≤ n, i = k,

such that A1 (k)A2 (i) − A2 (k)A1 (i) < 0,

A2 (i) A2 (k)
(24) < .
A1 (i) A1 (k)

However, the procedure can be done for all 1 ≤ k ≤ n, so for every

1 ≤ k ≤ n there must be an index 1 ≤ i ≤ n, i = k such that Eq. (24)
holds. This is clearly a contradiction.
The uniqueness of the solution is thus proven. It was also shown that
the existence of any singular point A(i) > 0, 1 ≤ i ≤ n implies that A(·)
is a kernel function with maximum goodness. In other words, the target
function has a single global (and local) maximum if the allowed
solutions
are restricted either by fixing A(x) = 1 for some x or by fixing i A(i) = 1.
This implies that any optimization algorithm which is able to find a local
maximum of a continuous non-linear function in finite time is appropriate
for solving the maximum likelihood problem.
The assumption given in Eq. (21) means that, for every pair of vertex
types, there must be a time step when they are both present in the network.
422 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

This is the sufficient ‘representation restriction’ for the maximum likelihood

method.
We have proved the following theorem:

Theorem 3. The target function (13) has a unique maximum (up to nor-
malization) over the set of positive vectors A if the representation restriction
of the maximum likelihood method holds.

Stationary Vertex Type Distribution. In this section we deal with

a simple subclass of networks for which the frequentist method and the
maximum likelihood method give the same result. In these networks the
distribution of vertex types is stationary in time. We use the following
notation: pi is the ratio of i-vertices in the network (1 ≤ i ≤ n); there are
n different vertex types. Obviously it is true that


n
(25) pi = 1.
i=1

The number of citations to an i-vertex is denoted by Mi and the total

number of edges is M . N t(e) is the number of vertices in the network in
time step t(e) and the total number of vertices is N . In stationary networks
the following holds:

  +  
n
(26) S t(e) = N t(e)
p1 A(1) + p2 A(2) + · · · + pn A(n) = N t(e)
pi A(i).
i=1

For simplicity we will sometimes use the notation


n
∗
(27) p = pi A(i).
i=1

Thus we need to maximize

A(xe ) A(x ) A(x ) A(x )
(28)   =  1   2  ···  M  =
e
S t(e) S t(1) S t(2) S t(M )

n
1 1
(29) = A(i)Mi ··· =
i
N t(1) p∗ N t(M ) p∗
The Inverse Problems of Evolving Networks 423
$
n %$ %$ %M
1 1
(30) = A(i)Mi .
e
N t(e) p∗
i

As N t(e) are fixed parameters, we can omit them, and the quantity to
optimize is

$
n %$ 
n %−M
Mi
(31) A(i) pi A(i) .
i i=1

Let us take the logarithm first:


n 
n
(32) Mi log A(i) − M log pi A(i).
i=1 i=1

At the maximum the partial derivatives with respect to all A(k) (1 ≤ k ≤ n)

are zero, which gives us n equations:

Mk M pk
(33) =
n , 1 ≤ k ≤ n,
A(k) i=1 pi A(i)

which can be written as

M
(34) p1 A(1) + · · · + pk (1 − )A(k) + · · · + pn A(n) = 0, 1 ≤ k ≤ n.
Mk

We have a system of n linear equations:

(35) 
M
p1 1 − A(1) + p2 A(2) + ··· + pn A(n) =0
M1  
M
p1 A(1) + p2 1− A(2) + · · · + pn A(n) =0
M2
..
.  
M
p1 A(1) + p2 A(2) + · · · + pn 1− A(n) = 0.
Mn
424 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

If we subtract all other lines from the first line we get

p1 M2
A(2) = A(1)
p2 M1
(36) ..
.
p1 Mk
A(n) = A(1),
pk M 1

which means, that after fixing A(1) = 1 (or some other arbitrary positive
value), we can calculate all values of the kernel function. The maximum
likelihood problem can be easily solved for stationary networks.
It can be shown that if the vertex type distribution is stationary then
the frequentist method is equivalent to the maximum likelihood method.

Two Vertex Types. For two vertex types the general non-stationary
maximum-likelihood problem can be solved by using a fixed point equation.
Let us recall that the solution of the problem is the kernel function satisfying
the fixed point Eq. (16). For two vertex types, if we fix A(1) = 1, then the
fixed point equation reads
& '−1
 p2
t(e)
(37) A(2) = M2 t(e) t(e)
.
e p1 + p2 A(2)

From now on we assume that A(1) ≥ A(2). If this turns out not to be
true, then we simply switch the 1-vertices and the 2-vertices. It is actually
quite simple to determine whether A(1) ≥ A(2) or vice versa:

Lemma 5. In a network with two vertex types,

 t(e)
 t(e)
(38) A(1) ≥ A(2) if and only if p1 ≤ M1 (and p2 ≥ M2 ).
e e

See [12] for the proof.

Let us examine the function
& '−1
 t(e)
p2
(39) f (x) := M2 t(e) t(e)
.
e p1 + p2 x

Theorem 4. f possesses the following properties:

The Inverse Problems of Evolving Networks 425

1. For the unique fixed point x∗ of f we have x∗ ≤ 1.

2. f is strictly monotone increasing.
3. f is concave.
4. If x < x∗ then f (x) < x∗ , where x∗ = f (x∗ ) is the unique fixed point
of f .
5. If x > x∗ then f (x) > x∗ .
6. If x < x∗ then x∗ − f (x) < x∗ − x.
7. If x > x∗ then f (x) − x∗ > x − x∗ .
 
8.  f (x) − x∗  < |x − x∗ | for all x ∈ [0, 1].
9. The fixed point iteration xn+1 := f (xn ) converges to the unique fixed
point from every starting point x0 ∈ [0, 1].
See [12] for the proof.

Convergence of the Fixed Point Equation. Previously we have shown

that the general maximum likelihood problem can be solved using any
optimization method capable of finding a local maximum of a non-linear
continuous function. We have also shown that for two vertex types the
simple fixed point iteration always converges to the solution. It is a natural
question whether we can solve the general problem with the same simple
fixed point iteration, as defined in Eq. (16).
This is still an open question. We give the following conjectures as open
problems:
Conjecture 1. The fixed point equation (16) is convergent from all non-
zero starting vectors, if after each iteration we normalize A(·) according to
A(·) = 1.

Conjecture 2. The fixed point equation (16) is convergent from all non-zero
starting vectors (even without the normalization in the previous conjecture).

Before the third conjecture we define quasi-contractive maps.

Definition 3. Let T : M → M a mapping of a metric space (M, d) into
itself. A mapping T is called a quasi-contraction if and  only if d(T x, T y) ≤
q × max d(x, y), d(x, T x), d(y, T y), d(x, T y), d(y, T x) for some q < 1 and
all x, y ∈ M .
426 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

Conjecture 3. The function defined in Eq. (16) is a quasi-contraction.

(Without the normalization step.)

It is shown in [10] that conjectures 1 and 2 follow from Conjecture 3.

General Networks. It is easy to generalize the maximum likelihood solu-

tion to non-citation networks. For these, the goodness of the kernel function
is defined as
A(x∗ , x∗∗ ) n n $  n
 
%−1
(40)  e  =
e
A(i, j) Mij
N t(e), i, j A(i, j) ,
e
S t(e) i=1 j=1 e i,j=1

where edge e connects an x∗e -vertex to an x∗∗

e -vertex in the analyzed network
and if the network is undirected, then the sum and the products also have
the condition i ≤ j.
The maximization of (40) is equivalent to the citation network case if
we create new variables from the (i, j) pairs. Denoting (i, j) by k we have

$ %−1
∗ ∗
A∗ (xe ) n n
 
(41)   = A(k) Mk
N t(e), k A(k) ,
e
S t(e) k=1 e k=1

where A∗ (·) is the new kernel function and n∗ is the number

 of new variables.

We have n∗ = n2 − 1 for directed networks and n∗ = n × (n + 1) /2 for
undirected networks.
The existence, uniqueness and all the other properties proven for the
citation network case are also valid here.

4. Applications

4.1. The Dynamics of Scientific Collaboration Networks

In this section we briefly present the application of our methods to a non-

decaying network: the cond-mat collaboration network. In this network a
vertex is a researcher who published at least one paper in the online arXiv
cond-mat archive (see http://www.arxiv.org) between 1970 and 1997 (this
is the date when the paper was submitted to cond-mat, not the actual
The Inverse Problems of Evolving Networks 427

publication date, but most of the time these two are almost the same).
There is an edge between two researchers/vertices if they have published
at least one paper together. The data set contains 23,708 papers, 17,636
authors and 59,894 edges.
For this network we used the frequentist method to estimate the kernel
function. First, we measured the attachment kernel for this network based
on the degrees of the two potentially connected vertices. See Fig. 2 for the
Acond-mat (d∗ , d∗∗ ) function. The attractiveness grows with vertex degree,
approximately linearly, unless the degree is zero. Zero degree vertices have
a very high probability of gaining new edges because many papers have
contributions from authors who do not yet have any other papers in the
cond-mat database. These results show that a second, different kernel
function might be appropriate for predicting the connections of the newly
added authors; [25] used a similar framework.

Fig. 2. The attachment kernel for the cond-mat collaboration network; the surface plot
was smoothed by applying a double exponential kernel to it. The right plot shows that
the kernel function has high values for zero-degree nodes. This might be because a new
researcher will usually write a paper with collaborators and thus will have a high
probability of gaining new edges immediately. The right plot has logarithmic axes.

We have tried to fit various functional forms to the two-dimensional

attachment kernel function to check which is a better description of the
dynamics. See Fig. 3 for the shape of the fitted functions and Table 1 for
the functional forms and the results.
The best fit was obtained by

(42) Acond-mat (d∗ , d∗∗ ) = c1 × (d∗ d∗∗ )c2 + c3 ,

where the ci are constants.

428 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

Fitted form Fitted parameters Fit error Fitting method

∗ ∗∗
B c1 max(d , d ) + c2 c1 = 1.26, c2 = −10.56 107357.6 Nelder–Mead
∗ ∗∗
C c1 d d + c2 c1 = 0.0697, c2 = −2.11 4300.2 Nelder–Mead
∗ ∗∗
D c1 (d + d ) + c2 c1 = 1.08, c2 = −18.98 31348.9 Nelder–Mead
c1 d∗ d∗∗ + c2 (d∗ + d∗∗ )+ c1 = 0.0783, c2 = −0.12, 3532.9 BFGS
∗ ∗∗
+c3 max(d , d ) + c4 c3 = −0.093, c4 = 1.50
∗ ∗∗ c2
c1 (d d ) + c3 c1 = 0.016, c2 = 1.22, 3210.4 SANN
c3 = 0.58

Table 1. Fitting the cond-mat network with a degree model. Four optimization methods
were run for each functional form to minimize the least square difference: BFGS,
Nelder–Mead, CG and SANN; the results of the best fits are included in the table. See
[30, 6] for the details of these methods.

Secondly, we used the number of papers as the vertex property, and

measured the Acond-mat (p∗ , p∗∗ ) kernel function. We tried to fit the same
functional forms to the measured kernel functions and found that the form

(43) Acond-mat (p∗ , p∗∗ ) = c1 × (p∗ p∗∗ )c2 + c3

fits best, although with very different parameters from those used for the
degree based measurement.
The goodness of the degree-based kernel was 2.41, whereas the kernel
based on the number of papers has goodness 6.69. This crearly indicates
that the ‘number of papers’-based model is much better at predicting new
collaborations than the degree-based model.
See [4, 26] for other studies on collaboration networks.

4.2. Comparing Alternative Models

In this section we model the US Patent system. In this network vertices

are issued patents and edges are citations from one patent to another; each
patent is required to cite all prior work related to the patented invention.
We create and compare alternative models based on various properties,
and also check the time dependency of the models. The tested models are
summarized in Table 3.
In-Degree. In the simplest model the attachment probability depends
only on the in-degree of the potentially cited vertices. We found that the
The Inverse Problems of Evolving Networks 429

Fig. 3. A shows the smoothed measured degree-based kernel function for the
collaboration network, B, C and D are the fitted functional forms shown in the first
three lines of Table 1. The best fit is clearly obtained by the multiplicative form.

kernel can be well approcimated by the form A(d) = dα + a. This form

may lead to scale-free networks, i.e. networks with power-law in-degree
distribution, if the degree dependence is linear, i.e. if α = 1. Indeed,
we found that the in-degree dependent kernel function can be very well
approximated by the form A(d) = dα + a and the value of the exponent α
is close to unity.

The fact that the obtained degree-dependent kernel function is a smooth

function does not imply that this is the best model of the system and that
all the citations can be explained simply based on vertex degree. It only
tells that this is the best form if we want to model the network based on
in-degree only. In other words, if we add additional vertex properties to the
430 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

Fig. 4. The smoothed measured kernel function based on the number of papers for the
collaboration network.

Fitted form Fitted parameters Fit error Fitting method

B c1 max(p∗ , p∗∗ ) + c2 c1 = 0.58, c2 = −2.54 45930 Nelder–Mead
∗ ∗∗
C c1 p p + c2 c1 = 0.04, c2 = −0.49 7470 Nelder–Mead
∗ ∗∗
D c1 (p + p ) + c2 c1 = 0.57, c2 = −8.81 20513 Nelder–Mead
∗ ∗∗ c2
c1 (p p ) + c3 c1  1, c2 = 1.59, 5312 SANN
c3 = 3.14

Table 2. Fitting the cond-mat network based on the number of papers written by an
author. Four optimization methods were run for each functional form to minimize the
least square difference: BFGS, Nelder–Mead, CG and SANN; the results of the best fits
are included in the table.

model, we might get a better model and each better model can be averaged
out to the preferential attachment rule.

In-Degree and Age. In addition to the in-degree, we consider the age of

the vertices. Here, by age we do not mean real time, but a quantity based
purely on the number of vertices. If a vertex was added to the network before
time step t0 and the current time step is t then we calculate (t−t0 )/w, where
w is the width of an age-window, typically a couple of thousands of time
steps, and the age is the largest smaller integer. For the patent network we
used w = 7100.
The in-degree and age dependent model is better than the simple degree-
dependent one, which is not a big surprise since an extended model is always
at least as good as the original. (Note that in practice this is not always
 Properties Form Goodness Parameter values Figures
* in-degree free 0.3092 – Fig. 5
* in-degree dα + a 0.3048 α = 0.99, a = 2.62 Fig. 5
in-degree free 0.2431 – –
in-degree dα + a 0.2394 α = 0.92, a = 1.95 –
i-d, age free 0.4541 – Fig. 6
i-d, age double Pareto 0.4373 α = 1.08, a = 1.1, αp = 0.3, Fig. 6
βp = 2.25, tp = 26.36
i-d, citing category free 0.2583 – [12]
i-d, citing category dα + a 0.2555 See [12] [12]
The Inverse Problems of Evolving Networks

i-d, cited category free 0.2656 – [12]

i-d, cited category ccat (dα + a) 0.2597 α = 0.89, a = 1.95, –
c = [0.92, 1.54, 1.43, 1.06, 0.90, 0.93]
i-d, cited, citing free 1.1744 – –
i-d, cited, citing ccat (dα + a) 1.1664 See [12] [12]
i-d, age, cited, citing ccat × double Pareto 1.3454 See Fig. 7 Fig. 7
forest fire model free 2.1034 – –
i-d, forest fire free 2.2539 – –
Table 3. Summary of the various models fitted to the US patent citation network. Starred models were fitted to
network data from 1975 to 2005, the others to data from 1975 to 1999. “i-d” means in-degree; see Eq. 44 for the double
Pareto form.
431
432 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

Fig. 5. The measured in-degree based kernel-function for the US patent network. The
red line is the maximum likelihood fitted form A(d) = dα + a. The axes are logarithmic,
so d = 0 is not included in the plot.

Fig. 6. Sections from the in-degree and age based maximum likelihood fitted kernel
function for the US patent citation network. Both plots have logarithmic axes.

true, because of numerical errors.) In fact it is about 80% better than the
degree-based model, which is significant.
The obtained free-form kernel can be well fitted by a function where the
effects of degree and age are separated, A(d, l) = Ad (d)Al (l). The degree-
dependent part is the usual preferential attachment form: Ad (d) = dα + a,
of course with a different exponent and different parameter a. We fitted
the age-dependent part with a double Pareto function, which is a unimodal
function with an initial power-law increase and a power-law decrease in the
tail:

(l/tp )βp −1 if l ≤ tp ,
(44) Al (l) =
(l/tp )−αp −1 if l > tp .
The Inverse Problems of Evolving Networks 433

About 96% of the goodness of the free-form model is preserved with the
fitted form. Consequently, the preferential attachment times double Pareto
form is a very good model if one wants to model the network based on
in-degree and age.
Finding the best kernel function of this form was done by using the
BFGS optimization method, without derivatives, as the function has no
derivative when l = tp if αp = βp .

Patent Categories. Patents are classified into more than 400 patent
classes, defined by the US Patent and Trademark Office. Patent classes
have further subclasses. Researchers at NBER used these classes to create a
patent classification with six big categories and thirty-six subcategories [21].
The six categories are: Chemical, Computers and Communications, Drugs
and Medical, Electrical and Electronic, Mechanical and Others.
A natural question is whether the inclusion of patent categories adds a
substantial amount to the goodness of the model. We expect a significant
goodness increase, based on the fact that the patent network is highly
assortative [27, 28] with respect to categories, i.e. chemical patents tend
to cite chemical patents.

In-Degree, Age, Cited Category, Citing Category. In this model both the
category of the citing and potentially cited vertices are included, as well as
the in-degree and age of the vertices. A separate kernel function of the form
Ac (d, l, p) = cp × Acd (d)Acl (l) was fitted for each citing vertex type c, where
Acd (d) has a preferential attachment form and Acl (l) is the double Pareto
function, as in Eq. 44. The cp are category-dependent constants.

Change in the Dynamics. So far we have assumed that a single kernel

function is valid during the whole development of the network. We now relax
this assumption. More specifically, we address the question of whether the
dynamics of the patent network changed around 1990 (supposedly) because
of the introduced legal changes. This question is currently highly debated
among law scholars [22]. Using the patent categories as vertex properties
we shall also check if the change can be found in the “behavior” of all patent
categories or if just a few of them are affected.
We can handle the time dependence of the kernel function by using a
sliding time-window of 500,000 patents, i.e. separate kernel functions are
estimated for each time window.
434 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

Fig. 7. Measured parameters based on in-degree, age, category of both the cited and
citing patent. Separate kernel functions were fitted for different citing categories. The
fitted form was the product of a degree-dependent term, an age-dependent double
Pareto term and a (cited) category-dependent constant. See the text for details.

First we used a model based on in-degree only, which is appropriate as we

are most interested in the change of the preferential attachment exponent.
Indeed, this model confirms our previous results [14], and shows that the
initially decreasing exponent started to increase around 1990 – see Fig 8.
Next we repeated the study with a model based on in-degree and age.
We used the maximum likelihood fitting method to fit the parameters
of a predefined shape: A(d, l) = (dα + a)Al (l), where Al (l) is a double
Pareto form. The results, shown in Fig. 8, confirm the previous findings;
furthermore, the maximum likelihood method generates much smoother
functions than the ad hoc fitting of the free-form frequentist method in [14].
Finally, in the most complete model we used the categories of both the
citing and cited patents, plus the in-degree and age of the vertices. Kernel
functions with the predefined shape A(d, l, cat) = ccat × (dα + a)Al (l) were
used, where Al (l) is the double Pareto form from Eq. 44. See the results in
Fig. 9. While the change is clearly present in the exponent α, the exponent
αd of the double Pareto aging function is increasing during the whole period.
An increasing parameter αd means that the patents increasingly prefer to
cite more recent patents.
To summarize, our previous findings about the change of the preferential
attachment exponent α [14] were confirmed in all models we have tried for
the patent network. The reason for this change is unclear yet. Although
The Inverse Problems of Evolving Networks 435

Fig. 8. Change of the exponent α in the US patent network using two models. The left
plot is the maximum likelihood fit of the in-degree dependent model A(d) = dα + a. The
right plot is based on an in-degree and age dependent model; the fitted form has an
in-degree dependent increasing term and an age dependent unimodal double Pareto
term. A sliding time-window is used to handle the time-dependence of the parameters;
the width of the time window was 500,000 patents and the difference between
consecutive windows is 100,000 patents.

Fig. 9. Change of the exponent α in the US patent network using an in-degree, age,
cited category and citing category based model. A separate model is fitted for each
citing category. The model has a degree-dependent term, an age dependent double
Pareto term and a (cited) category dependent constant. The width and step size of the
sliding time-window were 500,000 and 100,000 patents respectively.
436 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

patent law change is one likely cause there might be other explanations,
like the recent development in the searching techniques: the patentees and
patent examiners started to use electronic search engines when making the
citations. If the latter explanation is true, then the change should be seen
in other citation networks as well.
For the all the analysis performed in this section, we used the open
source igraph library [13] and the GNU R environment [33].

4.3. Validating Network Models

The Forest Fire Model. The forest fire model was designed by [24], to
explain three facts about real networks:
1. The in-degree and out-degree distributions are power-laws.
2. The networks are becoming denser in time: the number of edges grows
faster than the number of vertices in the graph.
3. The diameter [9, 2] of the network is shrinking; it gets smaller and
smaller as the network evolves.
Let us briefly explain how the forest fire model works. Let us assume
that a new vertex is added to the citation network and it needs to select
some other vertices to cite. First it cites a single vertex (more than one in
some variations) and then it checks the incoming and outgoing links of this
single vertex and with some probability cites these too. Then the process
is repeated for the newly cited vertices as well.
The model imitates how a researcher (or inventor or patent examiner)
finds relevant papers/patents based on the citation list and the “cited-by”
list of the already found relevant papers/patents.
We wanted to validate the forest-fire model for the US patent citation
network. This is, however, not straightforward, even if the forest fire model
fits into the kernel-based framework. The problem is that the actual order
of the citations made by a given vertex is missing from our database: it is
naturally not recorded which citations are made first, second, etc.
Neighbors. First we created a very simple model with two vertex types:
neighbors and non-neighbors. When a new edge is being added to the net-
work, all vertices which are in the same components as the citing vertex are
considered as neighbors, others non-neighbors. As the number of neighbors
is zero when the very first citation of a vertex is being made, we do not
The Inverse Problems of Evolving Networks 437

model the first (outgoing) citation of the vertices, but assume that these
are made randomly. The kernel function in this model has just two values:
one for the neighbors, one for the non-neighbors. If the value for the neigh-
bors is much higher than the one for the non-neighbors, then that indicates
the validity of the forest-fire model.
When trying to estimate the kernel for the forest fire model, it does
matter in which order the citations of a particular vertex are made. For
example, if we carry out the measurement in increasing order of citations
(i.e. oldest vertex is cited first, etc.), that favors the “neighbors” vertex
type and the kernel function is distorted. If we use the opposite ordering,
that favors the non-neighbor vertices. By generating synthetic networks we
found that, by using a random ordering of the citations the kernel is not
biased and on average the correct result is measured.
Indeed, if we fix the kernel function value of the non-neighbors to 1, then
the value for the neighbors in the patent network is 12,723.82. A neighbor
of an already cited vertex has ten thousand times more probability of being
cited (by the same vertex) than a non-neighbor vertex. The goodness of
this model is 2.1034, even if we make the first citation of each vertex totally
randomly!

Neighbors, In-Degree. If we add the in-degree of the vertices to the

property vector, we get an even better model, one with goodness 2.2539.
In this model the first citation is not done randomly but based on the in-
degree of the vertices, and the subsequent citations depend on in-degree and
on whether the possibly cited vertex is in the same component as the citing
vertex or not. The two kernel functions are plotted in Fig. 10.
The forest fire model provides a very good description of the US patent
network.

Preferential Attachment is Required Many real world networks feature

a power-law degree distribution. These networks are also called ‘scale-free’,
as the power-law distribution is invariant to the (multiplicative) scaling of
the independent variable.
First, [32] showed that the preferential attachment mechanism [5] is
capable of generating scale-free networks. This mechanism corresponds to
an in-degree dependent kernel-function of the form A(d) = d + a, where
the (small) increment a ensures that zero-degree vertices have non-zero
attractiveness. This model always generates a scale-free network (with
438 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

Fig. 10. The US patent kernel function, forest fire model, extended with in-degree. The
left plot shows the in-degree based kernel function, which is in effect for the first
(outgoing) citation of all vertices. The right plot shows the kernel function for the
subsequent citations. It is obvious that neighbor vertices have much higher probability
of being cited. Also, note that the kernel for the non-neighbors is much steeper, since
degree matters more for non-neighbors than for neighbors. Both plots have
logarithmic axes.

respect to the in-degree distribution) with exponent α = 3, independently

of the out-degree distribution:

(45) P [d = k] = (k + b)−α .

Here we shall show that the preferential attachment is not only sufficient
to create scale-free networks but is also required.
We state the following. If a network has a stationary scale-free in-degree
distribution, and we take the limit of the infinite network, then the best in-
degree based kernel function describing the evolution of the network is of
the form

(46) A(d) = d + a,

with some parameter a > 0.

Let us now prove this. If the network is scale-free and stationary, then

(47) pi = (i + a)−α , (i ≥ 0),

where pi is the ratio of vertices with in-degree i, and a > 0. This implies
that
∞
 ∞

Mi (t)
(48) ∝ pj = (j + a)−α , t → ∞,
M (t)
j=i+1 j=i+1
The Inverse Problems of Evolving Networks 439

where Mi (t) is the number of edges citing i-degree vertices up to time step
t and M (t) is the total number of edges up to time step t.
If the network is stationary then the best in-degree based kernel function
can be given as (see Sec. 3.2)

M0 M1 Mi
(49) A(0) = , A(1) = , ... A(i) = , ...,
p0 p1 pi

where
Mi (t)
(50) Mi := lim .
t→∞ M (t)

In our case this means that

(51)
∞

A(i) = (i + a)α (j + a)−α ∝ (i + a)α (i + a)−α+1 = i + a, (i ≥ 0),
j=i+1

where we have used

∞

(52) (i + a)−α+1 ∝ (j + a)−α .
j=i+1

Note that the “proof” above is independent of the value of α. This

means that if the assumptions are valid then the best description is linear
preferential attachment, independently of the exponent. Since it is quite
unlikely that this would be true (actually one can check with a couple
of simulated networks that it is not), the assumptions must not be valid
for every possible exponent α. In other words, the proof hints that the
assumptions can be valid only if α = 3, in which case we know that there
is an in-degree based kernel function which is able to generate the network.
This means that scale-free networks with exponents different from α = 3
cannot be generated with any in-degree based kernel-function. In other
words, an in-degree based kernel never generates a scale-free network, except
if A(d) = d + a, in which case it generates one with exponent α = 3.
These “hints” are in good agreement with the fact that non-linear pref-
erential attachment (A(d) = dβ + a, β = 1) never leads to a scale-free
network [23]. If we assume that a kernel function with high goodness is able
to reproduce the degree distribution, then our proof extends this to other
440 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

kernel-functions: no in-degree based kernel function leads to a scale-free

network, except in the case of linear preferential attachment.
Naturally, we do not state that there are no other mechanisms capable
of generating scale-free degree-distributions [17, 16, 19, 18, 34, 20, 15, 35],
nor that in a scale-free network the choices of the participating actors are
based on the degree of the vertices only. But we do state that, whatever
mechanism generated a scale-free network, when we carry out an in-degree
kernel based measurement on it, we will get linear preferential attachment
as the result.
A similar statement can be proved for graphs with exponential
(in-)degree distribution:
If the network has a stationary exponential in-degree distribution, and
we take the limit of the infinite network, then the best in-degree based kernel
function describing the evolution of the network is of the form

(53) A(d) = 1.

We follow in the same way as in the scale-free case. If the exponential

degree-distribution is stationary, then

(54) pi = e−i , (i ≥ 0),

where pi is the ratio of vertices with in-degree i. This implies that

∞
 ∞

Mi (t)
(55) ∝ pj = e−j , t → ∞,
M (t)
j=i+1 j=i+1

where Mi (t) is the number of citations to i-degree vertices up to time step

t and M (t) is the total number of edges up to time step t.
If the network is stationary then the best in-degree based kernel function
can be given as

M0 M1 Mi
(56) A(0) = , A(1) = , ... A(i) = , ...,
p0 p1 pi

where

Mi (t)
(57) Mi := lim .
t→∞ M (t)
The Inverse Problems of Evolving Networks 441

In our case this means

∞

(58) A(i) = e i
e−j ∝ ei e−i = 1, (i ≥ 0),
j=i+1

where we have used

∞

(59) e−i ∝ e−j .
j=i+1

Acknowledgement. This research was partially supported by the Henry

Luce Foundation.

References

[1] Réka Albert and Albert-László Barabási, Statistical mechanics of complex net-
works, Reviews of Modern Physics, 74 (2002), 47.
[2] Réka Albert, Hawoong Jeong and Albert-László Barabási, The diameter of the
world wide web, Nature, 401 (1999), 130–131.
[3] L. A. N. Amaral, A. Scala, M. Barhélémy and H. E. Stanley, Classes of small-world
networks, Proc. Natl. Acad. Sci. USA, 97(21) (2000), 11149–11152, 10.
[4] A-L Barabási, H Jeong, Z Néda, E Ravasz, A Schubert and T Vicsek, Evolution of
the social network of scientific collaborations, Physica A, 311 (2002), 590–614.
[5] Albert-László Barabási and Réka Albert, Emergence of scaling in random networks,
Science, 286(5439) (1999), 509–512.
[6] C. J. P. Belisle, Convergence theorems for a class of simulated annealing algorithms
on rd, Journal of Applied Probability, 29 (1992), 885–895.
[7] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez and D.-U. Hwang, Complex
networks: Structure and dynamics, Physics Reports, 424 (2006), 175–308.
[8] Béla Bollobás, Modern graph theory, Springer (2004).
[9] Béla Bollobás and Oliver Riordan, The diameter of a scale-free random graph,
Combinatorica, 24(1) (2004), 5–34.
[10] Lj. B. Ciric, A generalization of Banach’s contraction principle, Proceedings of the
American Mathematical Society, 45(2) (1974), 267–273.
[11] Gábor Csárdi, Dynamics of citation networks, in: Proceedings of the Conference on
Artificial Neural Networks (2006).
442 G. Csárdi, K. J. Strandburg, J. Tobochnik and P. Érdi

[12] Gábor Csárdi, Modeling complex systems by evolving networks, PhD thesis, Eötvös
Loránd University, 2007.
[13] Gábor Csárdi and Tamás Nepusz, The igraph software package for complex network
research, InterJournal, Complex Systems: 1695 (2006).
[14] Gábor Csárdi, Katherine J. Strandburg, László Zalányi, Jan Tobochnik and Péter
Érdi, Modeling innovation by a kinetic description of the patent system, Physica A,
374 (2007), 783–793.
[15] S. N. Dorogovtsev and J. F. F. Mendes, Evolution of networks with aging of sites,
Phys. Rev. E, 62(2) (2000), 1842–1845.
[16] S. N. Dorogovtsev and J. F. F. Mendes, Effect of the accelerating growth of
communications networks on their structure, Phys. Rev. E, 63 (2001), 025101.
[17] S. N. Dorogovtsev, J. F. F. Mendes and A. N. Samukhin, Growing network with
heritable connectivity of nodes, cond-mat/0011077, 2000.
[18] S. N. Dorogovtsev, J. F. F. Mendes and A. N. Samukhin, Generic scale of the
“scale-free” growing networks, Phys. Rev. E, 63 (2001), 062101.
[19] S. N. Dorogovtsev, J. F. F. Mendes and A. N. Samukhin, Giant strongly connected
component of directed networks, Phys. Rev. E, 64 (2001), 025101.
[20] G. Ergun and G. J. Rodgers, Growing random networks with fitness, Physica A,
303 (2002), 261–272.
[21] Bronwyn H. Hall, Adam B. Jaffe and Manuel Trajtenberg, The nber patent cita-
tion data file: Lessons, insights and methodological tools, in: Adam B. Jaffe and
Manuel Trajtenberg, editors, Patents, Citations, and Innovations: A Window on
the Knowledge Economy. MIT Press (2003).
[22] Adam B. Jaffe and Josh Lerner, Innovation and Its Discontents: How Our Broken
Patent System is Endangering Innovation and Progress, and What to Do About It,
Princeton University Press (2004).
[23] P. L. Krapivsky, G. J. Rodgers and S. Redner, Degree distributions of growing
random networks, Phys. Rev. Lett., 86 (2001), 5401.
[24] Jure Leskovec, Jon Kleinberg and Christos Faloutsos, Graphs over time: densifi-
cation laws, shrinking diameters and possible explanations, in: Proceeding of the
eleventh ACM SIGKDD international conference on Knowledge discovery in data
mining, (2005), pp. 177–187.
[25] M. E. J. Newman, Clustering and preferential attachment in growing networks,
Physical Review E, 64 (2001), 025102.
[26] M. E. J. Newman, Scientific collaboration networks. I. Network construction and
fundamental results, Physical Review E, 64 (2001), 016131.
[27] M. E. J. Newman, Assortative mixing in networks, Phys. Rev. Lett., 89 (2002),
208701.
[28] M. E. J. Newman, Mixing patterns in networks, Phys. Rev. E, 67 (2003), 026126.
[29] M. E. J. Newman, The structure and function of complex networks, SIAM Review,
45 (2003), 167–256.
The Inverse Problems of Evolving Networks 443

[30] Jorge Nocedal and Stephen J. Wright, Numerical Optimization, Springer, 1999.
[31] G. Orbán, T. Kiss, M. Lengyel and P. Érdi, Hippocampal rhythm generation:
gamma related theta frequency resonance, Biological Cybernetics, 84 (2001), 123–
132.
[32] D. J. de S. Price, A general theory of bibliometric and other cumulative advantage
processes, J. Amer. Soc. Inform. Sci., 27 (1976), 292–306.
[33] R Development Core Team, R: A Language and Environment for Statistical Com-
puting, R Foundation for Statistical Computing, Vienna, Austria (2007), ISBN:
3-900051-07-0.
[34] Bosiljka Tadic, Dynamics of directed graphs: the world-wide web, Physica A, 293
(2001), 273–284.
[35] Alexei Vazquez, Knowing a network by walking on it: emergence of scaling, Euro-
phys. Lett., 54 (2001), 430. cond-mat/0006132.

Gábor Csárdi Katherine J. Strandburg

Department of Biophysics New York University School of Law
KFKI Research Institute for Particle New York, NY
and Nuclear Physics of the U.S.A.
Hungarian Academy of Sciences and
Budapest DePaul University
Hungary College of Law
and Chicago, IL
Center for Complex Systems Studies U.S.A.
Kalamazoo College [email protected]
Kalamazoo, MI
U.S.A.
[email protected]

Jan Tobochnik Péter Érdi

Departments of Physics and Department of Biophysics
Computer Science KFKI Research Institute for Particle
Kalamazoo College and Nuclear Physics of the
Kalamazoo, MI Hungarian Academy of Sciences
U.S.A. Budapest
[email protected] Hungary
and
Center for Complex Systems Studies
Kalamazoo College
Kalamazoo, MI
U.S.A.
[email protected]
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 445–488.

Chapter 11
Learning and Representation:
From Compressive Sampling to the ‘Symbol
Learning Problem’

ANDRÁS LŐRINCZ

In this paper a novel approach to neurocognitive modeling is proposed in which

the central constraints are provided by the theory of reinforcement learning. In
this formulation learning is (1) exploiting the statistical properties of the system’s
environment, (2) constrained by biologically inspired Hebbian interactions and (3)
based only on algorithms which are consistent and stable. In the resulting model
some of the most enigmatic problems of artificial intelligence have to be addressed.
In particular, considerations on combinatorial explosion lead to constraints on
the concepts of state–action pairs: these concepts have the peculiar flavor of
determinism in a partially observed and thus highly uncertain world. We will
argue that these concepts of factored reinforcement learning result in an intriguing
learning task that we call the symbol learning problem. For this task we sketch
an information theoretic framework and point towards a possible resolution.

1. Introduction

Reinforcement learning (RL) has successfully reached human level in dif-

ferent games [119, 15, 12, 22, 103, 26, 112, 113]. Furthermore, RL has at-
tractive, near optimal polynomial time convergence properties under certain
conditions [57, 16]. Because RL has been strongly motivated by psychology
and neuroscience [101, 81, 87, 27, 94, 29, 63, 99], too, we take the Markov
446 A. Lőrincz

decision problem (MDP) model of RL as the central organization concept in

modeling the brain. It means that we have to win through the limitations
arising from the Markovian assumption.
The core problem of RL is that even simple agents in simple environ-
ments need a number of variables to detect, such as objects, their shape or
color, other agents, including relatives, friends, and enemies, or the space
itself, e.g., distance, speed, direction, and so on. The size of the state space
grows exponentially in the number of the variables. The base, which cor-
responds to the discretization of the variables, and the number of possible
actions are less harmful. However, another source of combinatorial explo-
sion comes from partial observation; agents need to maintain a history of
sensory information. This temporal depth comes as a multiplier in the ex-
ponent again.
In the so-called multi-agent systems, the problem becomes even harder
as the internal states of the other learning agents are hidden. This fact
has serious consequences. First, the number of agents also multiplies the
exponent. Second, the hidden nature of the internal states violates the
central assumption of RL about the Markov property of the states. For a
review on multi-agent RL, see e.g., [53, 17].
The effect of hidden variables is striking. For example, we have found
that, for two agents with two hidden states (meaning) and two actions
(signals to the other agent), some cost on communication is enough to
prohibit an agreement on signal meaning association [71]. Nevertheless,
there is a resolution if an agent can build an internal model about the other
agent. In this case they can quickly come to an agreement. This situation is
best described as ‘I know what you are thinking and I proceed accordingly’.
However, if both agents build models about each other, then agreement is
again hard, unless one of the agents assumes that the other agent also builds
a model, like ‘I know that you know what I am thinking’. Such observations
highlight the relevance of model construction and the adjustment of the
estimated reward function by, for example, inverse reinforcement learning
[83] and apprenticeship learning [1, 82].
Within the framework of RL, factored description can decrease the
exponent of the state space [13, 14, 61, 44]. In the multi-agent scenario,
for example, if it turns out that a given agent (let us say, agent B) does not
influence the long-term goals in the actual events observed by another agent
(agent A), then the variables corresponding to agent B can be dropped; only
the relevant factors need to be considered by agent A at that time instant.
Learning and Representation 447

Limitations of RL highlight the learning task that finds the relevant factors
for model construction.
In the approach presented in this paper two different types of constraints
will be applied. There is a set of experimental constraints coming from neu-
roscience and cognition. These should be considered as soft constraints, be-
cause the interpretation of the findings is debated in many cases. However,
these constraints help to reduce the number of potential machine learning
approaches.
The other type of constraints comes from mathematics. These are hard
constraints and are the most important supporting tools. However, care
should be taken as they may have also have their own pitfalls hidden deeply
in the assumptions, like the conditions of the different theorems applied.
Ideally, building a model should start from a few philosophical princi-
ples, should be constrained by (1) the need for fast learning (because of
the evolutionary pressure), by (2) the complexity of the environment and
the tasks, and by (3) the different aspects of determinism and stochastic
processes. The emerging model should then be used to derive the global
structure of the brain as well as the tiniest details of the neural organiza-
tion. This dream is certainly beyond our current knowledge and will stay
so for a long time. Instead, a compromise is put forth here: I will start
from the constraints of reinforcement learning, will consider certain facts of
the environment and will build upon those in the quest for the solution. If
there is some reassuring information or guideline from neuroscience then I
will mention those, but the set of choices is certainly subjective. At each
point we will find new sources of combinatorial explosion and will proceed
in order to decrease their impact. Eventually, we will arrive at an unre-
solved problem that we call the symbol learning problem. We shall argue
that the symbol learning problem may eliminate the enigmatic issues of the
symbol grounding problem of cognition. (For the description of the symbol
grounding problem, see [47] and the references therein.) Furthermore, the
symbol learning problem may gain a rigorous information theoretic formu-
lation with some indications that polynomial time algorithms can solve it
under certain conditions.
In the next subsection I provide an overview of the paper. It is a
linear construction with loops at each step. The loops are about the arising
mathematical problems and their possible resolutions. At each step, there
is a large pool of potential algorithms to resolve the emerged problems and
we shall select from those by looking at the soft constraints provided by
experimental neuroscience.
448 A. Lőrincz

1.1. Overview of the structure of the paper

All building blocks or computational principles have neurobiological moti-

vation that I will refer to. In Section 2 I shortly review recent advances on
compressible sensing, the relevance of sparse representation and L0 norm,
and the related neural architecture. In Section 3 I show how a family of
non-combinatorial independent component analysis can be structured to
achieve exponential gains by breaking the sensory information into inde-
pendent pieces. Reassurance and problems both come from neurobiology:
learning can be put into Hebbian form and the particular form provides
insight into the hippocampal formation of the brain, responsible for the for-
mation of declarative memory. However, the same formation points to the
need of action-based labeling; control and factors are interlaced. Unfortu-
nately, action-based labeling is not feasible, because the size of label space
can be enormously huge. As an example, consider the concerted actions of
the muscles. For 600 independent muscle groups with, for simplicity, only
2 (‘on’ or ‘off’) stages, the label space is as large as 2600 . This issue is ad-
dressed in Section 4, where the control problem is reformulated in terms
of speed-field tracking inverse dynamics that can be made robust against
approximation errors. In doing so we may avoid difficulties arising from
kinematics and inverse kinematics. In turn, learning becomes feasible. We
also review MDPs and show that the proposed robust controller scheme fits
well into the RL formalism if used within the so called event-learning frame-
work [74, 116]. In addition, we review factored RL that is also designed to
help avoid other sources of combinatorial explosions. Learning in this setup
is polynomial in the number of states, which can be seen as a novel result
for RL.
In the context of factored RL, we need to consider partially observed
environments, which in turn leads to an enigmatic problem. Namely, there
is a hidden, but quite influential assumption about both the states and the
actions in the traditional formalism, or about the states and the desired
states in the event-based formalism. The problem is that states and actions
are given, that is they have a deterministic description. Such determinism
apparently contradicts to the ‘great blooming, buzzing confusion’ [54] that
we observe: there are so many factors that may influence a ‘given state’
that the description is either huge and the state may never occur, or it must
have stochastic components. The rescuing thought is missing. We shall
formulate this cornerstone of the learning problem that we call the symbol
learning problem.
Learning and Representation 449

Fig. 1. Construction of the paper. Squares: facts about information from nature and
algorithms. Pentagons: emergent problems that need to be solved to proceed further.
Hexagons: reassuring soft information from neuroscience and cognition

2. Sparse Representation

I review certain features of perception and natural stimuli. Then I introduce

the notion of sparse representation and discuss how it enables factored
representation, which unfortunately still leads to combinatorial explosion.
One of the most relevant problems for computational models is the main-
tenance of distinct interpretations of environmental inputs. The presence
of competing representations is not simply a plausible assumption, but it
has strong psychophysical and neurobiological support in rivalry situations,
when two or more likely interpretations ‘fight’ to reach conscious observa-
tion, but our mind chooses between them and only one ‘meaning’ is present
at a time [67, 65]. However, while it inhibits the conscious concurrent ob-
servation, it also enables switches among the different interpretations in a
robust manner. Switching is so robust that usually we can not prevent it
at will [66, 21].
450 A. Lőrincz

In turn, while the inputs are continuous, the existence of distinct repre-
sentations point to some kind of discrete nature of the representation.
It has been observed that typical signals have specific statistical proper-
ties, distributions are not Gaussian but they are heavy-tailed in many cases.
This means that information is compressible in the sense that there exist
bases where the representation has a few large coefficients and many small
ones. This is the case, for example, for natural images.
The accommodation of biological perceptual systems to such peculiar
statistical properties have been extensively studied (see, e.g., [36, 32, 31]).
It is also known that sparse overcomplete representation can be learned ro-
bustly in artificial neural networks for the case of natural images; properties
of the emerging ‘receptive fields’ have little dependence on the sparsifying
non-linearity applied during learning [42, 85, 86, 96, 64]. The reason for
this property remained elusive until the recent discovery of compressive
sampling. It turns out that computations in L1 and in L0 norms result in
the same representation for such heavy tailed distributions [19, 33, 18]. This
is the so called ‘L1 Magic’ that has received considerable attention recently
in diverse fields1 .

2.1. Neural solution to distinct sparse representations

The basic algorithm presented in [85] can be put into a multi-layer recurrent
neural network architecture. The idea can be traced back to the eighties
[8, 91] and has undergone considerable developments during the last two
decades, see, e.g., [56, 93, 76]. There are many variants that can be called
together as autoencoders [3, 49].
In these neural architectures, however, there is no place for competition
between different representations.2 Rubinstein’s global optimization tech-
nique, the so called cross entropy method (CEM) [28] has recently been used
to resolve this problem. The resolution makes use of a novel online vari-
ant [115, 73] of the original batch learning method. In addition, this online
variant

1
For a comprehensive collection of papers visit the compressive sampling resources of
the DSP group at Rice University: http://www.dsp.ece.rice.edu/cs/
2
In this text, we make no difference between the notions of representation and inter-
pretation.
Learning and Representation 451

• applies directly the L0 norm, which is robust in the neural network

sense: the norm is not strict as any Lq norm (where 0 ≤ q ≤ 1) may
produce similar results, and
• encourages us to take a fresh look on the spike-coding versus rate
coding dichotomy.
The basic optimization problem of sparse representation – in a nutshell –
is the following: let x ∈ Rn denote the input we want to reconstruct with a
sparse representation (h ∈ Rm ) using an overcomplete basis set A ∈ Rn×m ,
where m $ n. The corresponding optimization problem is:

(1) h∗ := arg min ξ · hL0 + x − AhL2 ,

h

where .Ln denotes the Ln norm, ξ is a trade-off parameter. The first term
enforces low number of nonzero components, the second term minimizes the
L2 -norm of the reconstruction error ε(t) = x − Ah(t). We are solving the
above problem using CEM as it exploits sparsity. CEM aims to find the
(approximate) solution for global optimization tasks in the following form

h∗ := arg min f (h).

h

where f is a general objective function. Instead of having a single candidate

solution at each time step, CEM maintains a distribution g(t) of h(t) for
guessing solution candidates. The efficiency of this random guess is then
improved by selecting the best samples – the so called ‘elite’ – and iteratively
modifying g(t) to be more peaked around the elite. It is known that,
under mild regularity conditions, CEM converges with probability 1 and
that, for a sufficiently large population, the global optimum is found with
high probability [78]. It is also known that CEM is strikingly efficient in
combinatorial optimization problems. Formally, CEM works as follows. Let
g belong to a family of parameterized distributions, H . Let h(1) , . . . , h(N )
be independent samples (N is fixed beforehand) from g. For each γ ∈ R,
the set of ‘elite’ samples,

(γ := {h(i) | f (h(i) ) ≤ γ, 1 ≤ i ≤ N },
E
 
provides an approximation to the level set Eγ := h | f (h) ≤ γ .
Let Uγ and U (γ be the distribution over the level set Eγ and E
(γ , re-
∗
spectively. For small γ, Uγ is peaked around h . CEM chooses g(t) that
452 A. Lőrincz

is closest in the cross-entropy (or KL divergence) metric to the empirical

distribution U(γ . The algorithm iteratively modifies γ and g to reach the
optimal performance value γ ∗ = f (h∗ ). CEM uses some technical tricks.
It prohibits depletion of E(γ by maintaining the best ρ · N samples (1 − ρ
quantile), where ρ ∈ [0, 1] is a fixed ratio. Also, it deals with parame-
terized distribution families where parameters can easily be approximated
from simple statistics of the elite samples. Note also that updating the dis-
tribution parameters can be too coarse, so a smoothing factor α is applied.
Informally, CEM is a maximum likelihood method, without immature de-
cisions. The conservative nature of CEM comes from its innovation, the
maintenance of the elite, and that it sharpens the distribution of this elite
step-by-step.
Our idea is that the important component of this step-by-step sharpening
of the distribution, the set of elite samples need not be stored [115, 73]. Fur-
ther, sparse representation resolves the dilemma on discrete versus continu-
ous representation: CEM can search for low-complexity (sparse) representa-
tion and within each sparse representation candidate the continuous values
of the representation can be computed by (iterative) pseudoinverse compu-
tation for the allowed sparse non-zero components of h denoted by hL0

(2) ΔhL0 ∝ PL0 AT ε,

where PL0 projects to the allowed non-zero set of the components of the
overcomplete representation. Equation (2) is the gradient update that
follows from (1) for hL0 .
In sum, sparse overcompleteness has some interesting consequences.
Components or features are either present or absent and thus their presence
can be detected independently. Sparseness may recast the reconstruction
task as a combinatorial optimization problem for which we suggested a
modification of the very efficient and simple CEM. CEM is robust, fast and
essentially parallel, but requires batch access to samples. Our modification
resulted in an online version that may enable one to find mapping onto
real neuronal networks. Another extension is that for analog valued inputs
and/or representations, we embedded the online CEM into a reconstruction
network architecture that can directly calculate the reconstruction error
term of the cost function through feedforward and feedback connections.
(See, e.g., [76] about the potential role of such networks in real neuronal
systems).
Learning and Representation 453

Another important aspect of the online CEM is that it is essentially built

on the concurrent use of digital (spike-based feature selection) and analog
valued coding (like the input and its representation). This particular feature
may contribute to the long standing debate (see e.g. [97] and references
therein) about the different neuronal coding schemes by suggesting a new
computational role for spikes and maintaining rate code as well. The novel
interpretation is that spikes are used to improve the probability of the
compositional representation, whereas rate code is about the magnitude
of the selected components. Thus, beyond the advantages of the algorithm
featuring global optimization and parallel sampling in distributed networks,
it might have biological significance and resolve some of the puzzles of
neuronal networks.
A few notes are added here. Sparse representation is promising for nat-
ural signals, because it can overcome the Nyquist noise limit [19, 33, 18],
enables sparse sampling and may offer insight into the two-faceted nature
of neural coding, i.e., spike-coding and rate coding. Further, the presence
or absence of the elements of sparse representation – in principle – enables
tabulated RL methods. However, sparse coding makes combinatorial ex-
plosion even harder, implying that additional algorithmic components are
needed.
An individual element of the sparse neural representation can be seen
as a vector with a dimension that equals the dimension of the input: for
component i of representation h the synaptic vector is the ith column of
matrix A. In the next section we discuss how individual elements of the
neural representation can be collected into (approximately) independent
groups; thus forming (approximately) independent subspaces or ‘factors’.

3. Factored Representation

In the first part of this section, I review our recent results on how to collect
components so that each group is approximately statistically independent
from other groups. The relevance of this section is the following:
1. For groups of statistically independent components, the influence of
other groups may be neglected. Thus machine learning methods may
be restricted to some of the subspaces of the grouped variables making
considerable savings in the exponent of the state space.
454 A. Lőrincz

2. Dynamical dependencies of different orders, e.g., position, momentum,

and acceleration, can also be factored.
3. Sparse coding, independent component analysis (ICA) and the au-
toregressive family of dynamical models reviewed in this section are
closely related to each other [86, 108].
4. Cardoso’s observation that components found by ICA are entailed by
single independent subspaces has received mathematical support for
a limited set of distributions [108]. The observation enables effective
non-combinatorial search for independent subspaces, without a pri-
ori information about the number of subspaces and their dimensions.
Thus, we have a non-combinatorial algorithm that can provide com-
binatorial gains [89].
Independent Subspace Analysis (ISA) [20] is a generalization of ICA.
ISA assumes that certain sources depend on each other, but the dependent
groups of sources are independent of each other, i.e., the independent groups
are multidimensional. The ISA task has been subject of extensive research
[20, 123, 105, 6, 121, 52, 84, 88]. Generally, it is assumed that hidden sources
are independent and identically distributed (i.i.d.) in time. Temporal inde-
pendence is, however, a gross oversimplification of real sources including
acoustic or biomedical data. One may try to overcome this problem by as-
suming that hidden processes are, e.g., autoregressive (AR) processes. Then
we arrive to the AR Independent Process Analysis (AR-IPA) task [51, 90].
Another method to weaken the i.i.d. assumption is to assume moving av-
eraging (MA). This direction is called Blind Source Deconvolution (BSD)
[23], where observation is a temporal mixture of the i.i.d. components.
The AR and MA models can be generalized and one may assume ARMA
sources instead of i.i.d. ones. As an additional step, these models can be
extended to non-stationary integrated ARMA (ARIMA) processes, which
are important, e.g., for modeling economic processes [80]. In the next section
we review the AR-, MA-, ARMA-, ARIMA-IPA generalizations of the ISA
task, when (i) one allows for multidimensional hidden components and (ii)
the dimensions of the hidden processes are not known. In the undercomplete
case, when the number of ‘sensors’ is larger than the number of ‘sources’,
these tasks can be solved, because they can be reduced to the ISA task.
Learning and Representation 455

3.1. Non-combinatorial independent subspace analysis

The ISA task can be formalized as follows:

+ 
(3) x(t) = Ae(t), where e(t) = e1 (t); . . . ; eM (t) ∈ RDe ,

m
that is e(t) is a concatenated vector of components

M em (t) ∈ Rde . The total
dimension of the components is De = m
m=1 de . We assume that for a
m m
given m, e (t) is i.i.d. in time t, and sources e are jointly independent,
i.e., I(E 1 , . . . , E M ) = 0, where I(.) denotes the mutual information of the
arguments, E m (m = 1, . . . , M ) is a dm e dimensional random variable with
cumulative distribution function F (em ) = Pr(E1m ≤ em m
1 , . . . , Edm
e
≤ em
dm
e
).
The dimension of the observation x is Dx . Assume that Dx > De , and
A ∈ RDx ×De has rank De . Then, one may assume without any loss of
generality that both the observed (x) and the hidden (e) signals are white.
For example, one may apply Principal Component Analysis (PCA) as a
preprocessing stage. Then the ambiguities of the ISA task are as follows
[120]: sources can be determined up to permutation and up to orthogonal
transformations within the subspaces.

3.1.1. ISA separation. We are to uncover the independent subspaces.

Our task is to find a matrix W ∈ RDe ×Dx with orthonormal + columns

(that is WWT = I) such that y(t) = Wx(t), y(t) = y1 (t); . . . ; yM (t) ,
ym = [y1m ; . . . ; ydmm ] ∈ Rde , (m = 1, . . . , M ) with the condition that the
m

e
components of the related dm e dimensional random variables – i.e., Y
m
– are
m th
independent. Here, yi denotes the i coordinate of the m estimated th

subspace. This task can be viewed as the minimization of the mutual

information between the estimated components on the orthogonal group:

.  
(4) JI (W) = I Y 1 , . . . , Y M .

Alternatively, because (i) the entropy of the input H(x) is constant and
(ii) WWT = I, (4) is equivalent to the minimization of the following cost
function:

.   m
M
(5) JH (W) = H Y .
m=1
456 A. Lőrincz

Identities for mutual information and entropy expressions can be used to

derive another equivalent cost function:
e
.   
M dm M
(6) JH,I (W) = H(Yim ) − I (Y1m , . . . , Ydm
m ),
e
m=1 i=1 m=1

and arrive finally to the following minimization problem

. 
M
(7) JI,I (W) = I (Y1 , . . . , YdM ) −
1 M
I (Y1m , . . . , Ydm
m ),
e
e
m=1

where Yim (m = 1, . . . , M and i = 1, . . . dem ) denote the stochastic variable

related to yim . The first term of the r.h.s. of (7) is an ICA cost function;
it aims to minimize mutual information for all coordinates. The other
term is a kind of anti-ICA term; it maximizes mutual information within
the subspaces. One may try to apply a heuristics and to optimize (7) in
the following order: (1) Start by any ‘infomax’-based ICA algorithm and
minimize the first term of the r.h.s. in (7). (2) Apply only permutations on
the coordinates to optimize the second term. Surprisingly, this heuristics
leads to the global minimum of (4) in many cases. In other words, ICA
that minimizes the first term of the r.h.s. of (7) solves the ISA task as
well, apart from grouping the coordinates into subspaces. This feature
was first observed by Cardoso [20]. To what extent this heuristic works is
still an open issue. Nonetheless, we consider it as a ‘Separation Theorem’,
because for elliptically symmetric sources and for some other distribution
types one can prove that it is rigorously true [107]. (See also the results
concerning local minimum points [122]). Although there is no proof for
general sources yet, a number of algorithms successfully apply this heuristics
[20, 105, 122, 7, 106, 2].

3.1.2. ISA with unknown components. Another issue concerns the

computation of the second term of (7), because we have to group the ICA
components by means of this term. For subspaces em of known dimensions
dm
e , multi-dimensional entropy estimations can be applied [88], but these
are computationally expensive. Other methods deal with implicit or explicit
pair-wise dependency estimations [7, 122]. Interestingly, if the observations
are indeed from an ICA generative model, then minimization of pair-wise
dependencies is sufficient to solve the ICA task according to the Darmois-
Skitovich theorem [25]. For the ISA problem, in principle, estimation of
Learning and Representation 457

pair-wise dependencies is insufficient to recover the hidden subspaces [88].

Nonetheless, such algorithms seem to work nicely in many practical cases.
A further complication arises if dimensions dm m
e of subspaces e are not
known. Then the dimension of the entropy estimation becomes uncer-
tain. There exist methods that also try to minimize pair-wise dependencies.
A block-diagonalization method has been suggested in [122], whereas [7]
makes use of kernel estimations of the mutual information.
Assume that the separation theorem is satisfied and apply ICA pre-
processing. This step can be followed by the estimation of the pair-wise
mutual information of the ICA coordinates. These quantities can be rep-
resented as the elements of an information adjacency graph, the vertices of
the graph being the ICA coordinates. One can search for clusters of this
graph and may apply different efficient approximations like Kernel Canoni-
cal Correlation Analysis [6] for the estimation of mutual information. Then
variants of the Ncut algorithm [124] can be used for clustering. As a result,
the mutual information within (between) cluster(s) becomes large (small).
Below, we show that the ISA task can be generalized to more realistic
sources.

3.1.3. ISA Generalizations. We need the following notations: Let z stand

for the time-shift operation, that is (zv)(t) := v(t − 1). The N order
1 ×D2
polynomials of z over the D1 × D2 matrices are denoted as R[z]D :=

N ×D
 r
N
F[z] = n=0 Fn z , Fn ∈ R
n D1 2 . Let ∇ [z] := (I − Iz) denote the rth
r

order difference operator, where I is the identity matrix, r ≥ 0, r ∈ Z.

Now, we are to estimate unknown components em from observed sig-
nals x. We always assume that e takes the form like in (3) and that
A ∈ RDx ×Ds is of full column rank.

1. AR-IPA: The AR generalization of the ISA task is defined by the fol-

lowing equations: x = As, where s is a multivariate

AR(p) process i.e.,
s ×Ds
P[z]s = Qe, Q ∈ RDs ×De , and P[z] := IDs − pi=1 Pi z i ∈ R[z]D .
  p
We assume that P[z] is stable, that is det P[η] = 0, for all η ∈ C,
|η| ≤ 1. For dm m
e = 1 this task was investigated in [51]. Case de > 1 is
treated in [90]. The special case of p = 0 is the ISA task.
2. MA-IPA or Blind Subspace Deconvolution (BSSD) task: The ISA
task is generalized to blind deconvolution task

q(movingj average task,
Dx ×De
MA(q)) as follows: x = Q[z]e, where Q[z] = j=0 Qj z ∈ R[z]q .
458 A. Lőrincz

3. ARMA-IPA task: The two tasks above can be merged into an in-
tegrated model, where the hidden s is a multivariate ARMA(p, q):
s ×Ds s ×De
x = As, P[z]s = Q[z]e. Here P[z] ∈ R[z]D p , Q[z] ∈ R[z]D
q .
We assume that P[z] is stable. Thus the ARMA process is stationary.
4. ARIMA-IPA task: In practice, hidden processes s may be non-
stationary. ARMA processes can be generalized to the non-stationary
case of integrated ARMA or ARIMA(p, r, q). The assumption here
is that the rth difference of the process is an ARMA process. The
corresponding IPA task is then

(8) x = As, where P[z]∇r [z]s = Q[z]e.

It is attractive that these algorithms can be put into Hebbian (that

is neurally plausible) forms [75, 72]. The resulting Hebbian forms pose
constraints that can be used to explain several intriguing properties of the
hippocampal formation. These are the reassuring soft considerations that
– up to some extent – we shall review below.

3.2. Model of the hippocampal formation

Inspired by the ideas of Attneave and Barlow [5, 9] ICA has been suggested
to take place in the hippocampal formation (HF) [68]. The idea has been
improved and extended over the years [69, 24, 76, 39] and the resulting model
seems powerful enough to predict several intriguing features of this brain
region, e.g., the independence of neuronal firing in certain areas [95], long
and tunable delays in other places of the HF [48] and different functional
roles for different pathways of the loop [58].
The model is flexible enough to incorporate novel findings about certain
properties of the HF, including the peculiar hexagonal grid structure found
in the entorhinal cortex, which is a part of HF (for a review about this
grid structure and its supposed role in spatial navigation and memory see,
e.g., [79]).
But why is HF so important?
Learning and Representation 459

3.3. Declarative memory: The relevance of the hippocampal

formation

The HF is responsible for the formation of declarative memory, which is

about the representations of facts, events or rules. HF seems quite sim-
ilar in mammals and it is generally considered as the neural correlate of
declarative memory (that is the brain region that carries out the required
computations). Intriguingly, after lesion, remote memories remain mostly
intact but new memories about events that occurred after the lesion can
hardly be formed [104, 102]. In other words, the HF is required for the ac-
quisition (learning or creating representations), but then it carries over the
information to other regions so the information can be kept even without
the HF. It is also relevant that other forms of learning, such as learning
of procedures, or category learning [59] remain intact after hippocampal
lesion.
For the purpose of reinforcement learning, declarative memory is of high
importance. One may doubt if anything can be understood about goal-
oriented functioning of the brain without understanding the advantages of
and the constraints provided by the hippocampal formation.

3.3.1. Learning in the model hippocampus. Learning to perform ICA

may assume many forms. One variant can be given as
 T
(9) ΔW(t + 1)T ∝ W(t)T (I − (
e(t) f (e(t) )

where f (·) is a component-wise nonlinear function with many suitable forms,

(
e(t) = W(t)x(t) and x(t) are the estimation of sources e(t) and the input,
respectively, at time t, and matrix W is the so called separation matrix
belonging to the hidden mixing process x(t) = A(t)e(t) so that WA ap-
proximates the identity matrix upon tuning.
The intriguing feature of this learning rule is that if we write it as
 T
(10) ΔW(t + 1)T ∝ W(t)T (I − W(t)x(t) f W(t)x(t) )

then its self-organizing nature becomes apparent. In this case, matrix

W(t) learns to separate. By contrast, in (9), the learning rule looks like
supervised learning, because learning is driven by the output. This double
option is exploited in the recent version of the model of the hippocampal
460 A. Lőrincz

formation [72]. This double option is relevant for our model, it tells that
symbol learning and symbol grounding are not incompatible with each other.
More details on the learning rule and about the way it fits the sophis-
ticated hippocampal formation can be found in [75] and [72], respectively.
It is also shown in [72] via numerical simulations that the model is capa-
ble of explaining the peculiar hexagonal grid formation in the hippocampal
formation. These are reassuring soft details for us.
However, the precision of the hexagonal grid provided by the model
is not satisfactory for conjunctive (position, direction, and velocity depen-
dent) representations and it seems hard to achieve the required precision
without some internal motion related gauge for general input types. In ad-
dition, the brain somehow separates the motion related information: pieces
of information, which are invariant to straight motion and pieces of informa-
tion, which show rotation invariance, are encoded in different areas; these
factors are physically separated. This grouping seems impossible without
action-related labeling that in turn leads to another source of combinatorial
explosion: the simple concepts of rotation and forward motion hide com-
plex and highly variable action combinations. It is unclear if labeling in
action space is possible at all. Fortunately, we can proceed, because there
exist a different solution. This solution is based on robust controllers and
it transforms the problem to much smaller dimensional spaces.

4. Robust Control, Event Learning and Factored RL

In this section, our inverse dynamics based robust controller is considered

first. The attractive property of this controller is that it reduces continuous
valued action variables to action indices – the same concept that we em-
phasized in the context of sparse representations – while maintaining the
continuity. Then, we rephrase Markov decision problems in the event-based
formalism that suits our robust controller. We will still face combinatorial
explosion in the number of variables. This explosion can be eased by fac-
tored description. Even then, there is a combinatorially large number of
equations to be solved in the emerging factored RL model. We will take
advantage of sampling methods to get over this type of explosion. The
sketched route aims to diminish combinatorial explosion at each step. It is
relevant from the computational point of view, but are in fact, unavoidable
tasks in our modeling efforts. This is so, because combinatorial explosion
Learning and Representation 461

would eliminate the chance to learn the model of an orderly world with
a large number of distinct objects that we can manipulate. We will also
provide some reassuring information from neuroscience and cognition.

4.1. Robust inverse dynamic controller with global stability

properties

One can plan and execute different actions like standing or walking. Does it
mean that we have conscious access to the real actuators, i.e., the muscles?
We can launch the actions, but may have a hard time explaining what
happened in terms of the muscles. The exact description of the kinematics
would also be hard. Furthermore, standing up, walking or sitting are just
too complex to be seen as simple optimized, but reflex like reactions evoked
by the inputs, although this is the view suggested by RL (but see [30] for
a different view). Because of this, there seems to be a conflict between
the higher order concept of ‘actions’ in RL and low level description of the
actuators.
In this subsection we review a particular controller for continuous dy-
namical systems, the so called static and dynamic state (SDS) feedback
controller [109, 110], which is a promising tool and may resolve this conflict.
The SDS control scheme gives a solution to the control problem called
speed field tracking 3 in continuous dynamical systems [50, 38, 111]. The
problem is the following. Assume that the state space of a plant to be
controlled X ∈ RN , its tangent space Ẋ ∈ RN , and a speed field ẋd : X → Ẋ
are given. At time t, the system is in state xt ∈ X, and the dynamics,
including the control actions change the state:

ẋt = B(xt , at )

where ẋt ∈ Ẋ is the actual velocity of the plant, the change of state over
unit time, and at denotes the control. We have freedom in choosing the
control action and we are looking for the action that modifies the actual
velocity ẋt to the desired speed ẋd (xt ). The obvious solution is to apply an
inverse dynamics, i.e., to apply the control signal in state xt which drives
the system into ẋd (xt ) with maximum probability:
   
(11) at xt , ẋdt = Φ xt , ẋdt ,
3
The term, ‘velocity field tracking’, may represent the underlying objective of speed
field tracking better. We shall use the two terms interchangeably.
462 A. Lőrincz

that is   
ẋdt = B xt , Φ xt , ẋdt ,

where for the sake of convenience, d d

 wed used the shorthand ẋt = ẋ (xt ). Of
course, the inverse dynamics Φ xt , ẋt has to be determined some way, for
example by exploring the state space and the effect of the actions first.
The SDS controller provides an approximate solution such that the
tracking error, i.e., ẋd (xt ) − ẋt  is bounded, and this bound can be made
arbitrarily small. The global stability property is attractive and the small
upper bound on the error will enable us to include the controller to the
framework of event learning.
Studies on SDS showed that it is robust, i.e., capable of solving the
speed-field tracking problem with a bounded, prescribed tracking error
[38, 110]. Moreover, it has been shown to be robust also against perturbation
of the dynamics of the system and discretization of the state space [116].
The SDS controller fits real physical problems well, where the variance of
the velocity field ẋd (x) is moderate.
The SDS controller applies an approximate inverse dynamics Φ, ( which
is then corrected by a feedback term. The output of the SDS controller is
 t
   
(12) ( (
at xt , ẋt = Φ xt , ẋt − Φ(xt , ẋt ) + Λ
d d
δ τ dτ,
τ =0

where
 
(13) δτ = Φ ( τ , ẋτ )
( xτ , ẋd − Φ(x
τ

is the correction term, and Λ > 0 is the gain of the feedback. It was shown
that under appropriate conditions, the eventual tracking error of the con-
troller is bounded by O(1/Λ). The assumptions on the approximate inverse
dynamics are quite mild: only ‘sign-properness’ is required [109, 110]. Sign-
properness imposes conditions on the sign but not on the magnitude of the
components of the output of the approximate inverse dynamics. If we dou-
ble the control variables, so that we separate the different signs into different
actions (like pulling or pushing) then the SDS controller requires only the
labels of the control variables and will execute the action. That is, the
problem of continuity of the action space disappears.
Generally, such an approximate inverse dynamics is easy to construct
either by explicit formulae or by observing the dynamics of system during
learning.
Learning and Representation 463

The above described controller cannot be applied directly to event learn-

ing, because continuous time and state descriptions are used. Therefore we
have to discretize the state space. Furthermore, we assume that the dynam-
ics of the system is such that for sufficiently small time steps all conditions of
the SDS controller are satisfied.4 Note that if time is discrete, then instead
of prescribing the desired speed ẋdt we can prescribe the desired successor
yd −x
state ytd and use the difference ẋdt ≈ tΔt t . Because of sign-properness, we
may neglect the multiplier Δt [110, 116]. Therefore the controller takes the
form
    t
d (
at xt , yt = Φ xt , yt − xt + Λ
d
δ τ · Δt,
τ =0

where
 
( xτ , yd − xτ − Φ(x
δτ = Φ ( τ , yτ − xτ ),
τ

and Δt denotes the size of the time steps. Note that xτ and yτ (therefore δ τ )
change at discretization boundaries only. Therefore, event-learning with the
SDS controller has relaxed conditions on update rates.

4.2. Inverse dynamics for robots

First, we review the properties of the SDS controller. Then we argue about
the ease of learning.
The SDS controller requires a description of the state. This description,
however, depends on the order of the dynamics of the plant under control.
So, for Newtonian dynamics in general, we need both the configuration and
how it is changing as a function of time. This description can be complex
for many-segment systems. On the other hand, in typical situations we need
to move the end effector to a given target point. This could be much less
demanding, because the speed-field description for the end effector is easy,
e.g., under visual guidance. The SDS controller is especially simple in this
respect; it does not need the speed of the end effector for stable control as
we describe it below.
In what follows, the time index is dropped to ease notations. Assume
that the plant is a multi-segment arm. The configuration of this plant is
determined by the angle vector θ ∈ Rn that represents the angles between
4
Justification of this assumption requires techniques of ordinary differential equations
and is omitted here. See also [10].
464 A. Lőrincz

the n joints. Assume further that the external space is a 3 dimensional space
where the end effector is at position z ∈ R3 determined by the angle vectors
of the joints, i.e., z = z(θ). The temporal behavior of the end effector can be
expressed by the time dependence of the angle vector. In case of Newtonian
dynamics, the end effector can be characterized by its position z and by its
momentum ż = ż(θ, θ̇), a function of both θ and θ̇. The acceleration of the
end effector z̈ depends on θ, θ̇, and θ̈, i.e., z̈ = z̈(θ, θ̇, θ̈). This acceleration
is needed to describe the dynamics. As an example, we take the case when
we have a desired position zd in the 3 dimensional space. The state of the
plant can be written at time t as q = [z; ż], the vector concatenated from
z and ż, the speed or velocity of the state is q̇ = [ż; z̈]. The SDS controller
allows one to define the desired speed field in any state q for the desired
angle vector zd and in the absence of further prescriptions it is as follows:
   
q̇d q = [(zd − z); (zd − z) − ż ].

(
Thus, the approximate inverse dynamics Φ(q, q̇d ) can be written as
 
(
Φ(q, ( z, ż, (żd − ż)
q̇d ) = Ψ

Furthermore, for real robots subject to SDS control one replace the approx-
(
imate inverse dynamics Ψ(.) with the ‘simplified inverse-dynamics’:
   
( 0 z, z̈ = Ψ
Ψ ( z, 0, z̈

because – as it can be shown easily for plants with masses [110] – (a) setting
ż = 0 corresponds to an additive term, the transformation of state q
provided that the mass of the robot is independent of ż, which should be
a good approximation and (b) such additive terms can be neglected in the
SDS scheme as they cancel in the correcting term (13). This property
further simplifies the application of the inverse dynamics. Simulations on
controlling the end-effector have shown (i) robustness, (ii) that little if any
knowledge is necessary about the details of the configuration (except the
fact if a certain target point is reachable or not), (iii) that little if any
knowledge is needed about the mass of the robotic arm, provided that the
gain is high enough and the noise level is low [70].

4.2.1. Inverse dynamics is easy to learn. As it has been noted before,

the SDS approach to control requires little if any knowledge about the
configuration. If a particular action is feasible then it will be executed
Learning and Representation 465

robustly via speed-field tracking even if the inverse dynamics is crude. The
corresponding theorem [109, 110] says that ‘sign properness’ is satisfactory:
all actuators should move the end effector towards the target. As mentioned
above, if we double the control variables and separate the different signs
into different actions then SDS controller works by means of the labels of
the actions. It is then a black box, which executes the task prescribed by
speed field tracking. However, the applied control values, the state and the
experienced speed, i.e., the low-level sensory information, should be made
available for the black box.

4.2.2. Inverse dynamics can avoid combinatorial explosion. There

are other advantages of speed field tracking based controller using the in-
verse dynamics. The controller can work by using the parameters of the
external space [70]. External space – for the first sight – is 3 dimensional, a
considerable advantage over configuration space. However, we can do better
if we take a look at neurobiology.
The brain creates low dimensional maps of many kinds. Concerning the
limbs, it also creates crude maps that surround the limb: many neurons in
the premotor cortex respond to visual stimuli. The visual receptive fields
of many of these neurons are not changing if the eye moves, so they do
not depend on the retinotopical position. Instead they are arm or hand
centered; they move together with the arm or hand movements [43]. In our
interpretation (see also e.g., [11]) visual information enters as eye-centered
information and then it is transformed from the high dimensional pixel
description to a two-dimensional representation that surrounds the arm and
this representation can be used directly to move any part of the arm, because
this and the visual information together ‘tell’ both the position and the
direction, the two prerequisites for inverse dynamics [70]: From the point
of view of speed-field tracking, this is a crudely discretized two dimensional
manifold serving a large number of ‘end’-effectors, namely, all portions of
the limb.
In the next section, we insert this robust controller into RL.
466 A. Lőrincz

4.3. Markov Decision Processes

An MDP is characterized by a sixtuple (X, A, R, P, xs , γ), where X is a finite

set of states;5 A is a finite set of possible actions; R : X × A → R is the
reward function of the agent, so that R(x, a) is the reward of the agent after
choosing action a in state x; P : X×A×X → [0, 1] is the transition function
so that P (y | x, a) is the probability that the agent arrives at state y, given
that she started from x and executed action a; xs ∈ X is the starting state
of the agent; and finally, γ ∈ [0, 1) is the discount rate on future rewards.
A policy of the agent is a mapping π : X × A → [0, 1] so that π(a | x)
tells the probability that the agent chooses action a in state x. For any
x0 ∈ X, the policy of the agent and the parameters of the MDP determine
a stochastic process experienced by the agent through the instantiation

x0 , a0 , r0 , x1 , a1 , r1 , . . . , xt , at , rt , . . .

The goal is to find a policy that maximizes the expected value of the
discounted total reward. Let the value function of policy π be

∞  
π t
(14) V (x) := E γ rt  x = x0
t=0

and let the optimal value function be

V ∗ (x) := max V π (x)

π

for each x ∈ X. If V ∗ is known, it is easy to find an optimal policy π ∗ ,

∗
for which V π ≡ V ∗ . Provided that history does not modify transition
probability distribution P (y|x, a) at any time instant, value functions satisfy
the famous Bellman equations
   
(15) V π (x) = π(a | x)P (y | x, a) R(x, a) + γV π (y)
a y

and
  
(16) V ∗ (x) = max P (y | x, a) R(x, a) + γV ∗ (y) .
a
y
5
Later on, a more general definition will be given for the state of the system: the state
will be a vector of state variables in the fMDP description. For that reason, the boldface
vector notation is used here already.
Learning and Representation 467

Most algorithms that solve MDPs build upon some version of the Bellman
equations.

4.4. ε-Markov Decision Processes

An important observation in RL is that near-optimal policies can be found in

varying environments, where states and actions may vary up to some extent.
This is the so called ε-MDP model family first introduced in [55] and later
elaborated in [116]. Then we can use tabulated systems if the controller
can execute the actions with ε precision. This precision can be achieved,
e.g., by our SDS controller that executes speed-field tracking. Thus, we
will be ready with the insertion of the robust controller into RL if we can
transcribe the traditional state–action formulation of RL into state–desired
state description.

4.5. Formal description of event learning

Similarly to most other RL algorithms, the event-learning algorithm also

uses a value function, the event-value function E : X × X → R. Pairs of
states (x, y) and (x, yd ) are called events and desired events, respectively.
For a given initial state x, let us denote the desired next state by yd . The
ed = (x, yd ) state sequence is the desired event, or subtask. E π (x, yd ) is the
value of trying to get from actual state x to next desired state yd and then
upon arriving to the next state y, which could be different from the desired
state yd and then following policy π afterwards:
  
(17) E π (x, yd ) = P (y | x, yd ) R(x, yd ) + γV π (y)
y

One of the advantages of this formulation is that one may – but does not
have to – specify the transition time: Realizing the subtask may take more
than one step for the controller, which is working in the background.
The value of E π (x, yd ) may be different from the expected discounted
total reward of eventually getting from x to yd . We use the former defin-
ition, since we want to use the event-value function for finding an optimal
successor state. To this end, the event-selection policy π E : X × X → [0, 1]
is introduced. π E (yd | x) gives the probability of selecting desired state yd
in state x. However, the system usually cannot be controlled by “wishes”
468 A. Lőrincz

(desired new states), decisions have to be expressed in actions. This is done

by the action-selection policy (or controller policy) π A : X × X × A → [0, 1],
where π A (a | x, yd ) gives the probability that the agent selects action a to
realize the transition x → yd .6
An important property of event learning is the following: only the
event-selection policy is learned (through the event-value function) and the
learning problem of the controller’s policy is separated from event learning.
From the viewpoint of event learning, the controller’s policy is part of the
environment, just like the transition probabilities.
Alike to (14), let the state value function of policies π E and π A be

∞  
πE ,π A t 
V (x) := E γ rt  x = x0 .
t=0

The event-value function corresponding to a given action selection policy

can be expressed by the state value function:

E ,π A    A   E A 
Eπ x, yd = π a | x, yd P (y | x, a) R(x, y) + γV π ,π (y) ,
a y

and conversely:

E ,π A
   E A
Vπ (x) = π E yd | x E π ,π (x, yd ).
yd

From the last two equations the recursive formula

E ,π A    A 
(18) Eπ x, yd = π a | x, yd P (y | x, a)
a y
  
  πE ,π A
 
× R(x, y) + γ π E
z |y E
d
y, z d

zd

can be derived. For further details on the algorithm, see [116].

6
Control vectors will enter the description later and the notation is changed accord-
ingly. In what follows, actions can be taken from a discrete set, but may be modified by
the robust controller and may take continuous values, ineffective for the RL description.
In the RL formalism, we thus keep the summation over the actions.
Learning and Representation 469

4.5.1. Robust controller in event learning. Our robust controller can

be directly inserted into event-learning by setting
  
  1 if a = at xt , ytd ,
(19) πtA a| xt , ytd =
0 otherwise,
 
where at xt , ytd denotes the action that represents the actual combination
of labels determined by the SDS controller. Note that the action space is
still infinite.


Corollary 3 [116]. Assume that the environment is such that y  P (y |

x, a1 )−P (y | x, a2 ) ≤ Ka1 −a2  for all x, y, a1 , a2 .7 Let ε be a prescribed
number. For sufficiently large Λ and sufficiently small time steps, the SDS
controller described in (19) and the environment form an ε-MDP.

Consequently, the theorem on the near-optimality of the value function

detailed in [116] applies.
In what follows, we shall use the traditional state and action description
for the MDP tasks. It is worth noting that all considerations can be
transferred without any restriction to the state an desired state description,
i.e., to the event-learning formalism.

4.5.2. Exact Value Iteration. Consider an MDP (X, A, P, R, xs , γ). The

value iteration for MDPs uses the Bellman equations (15) and (16) as an
iterative assignment: It starts with an arbitrary value function V0 : X → R,
and in iteration t it performs the update
  
(20) Vt+1 (x) := max P (y | x, a) R(x, a) + γVt (y)
a
y∈X

for all x ∈ X. For the sake of better readability, we shall introduce vector
notation. Let N := |X|, and suppose that states are integers from 1 to
N , i.e. X = {1, 2, . . . , N }. Clearly, value functions are equivalent to N -
dimensional vectors of reals, which may be indexed with states. The vector
corresponding to V will be denoted as v and the value of state x by vx .
We shall use the two notations interchangeably. Similarly, for each a let us
define the N -dimensional column vector ra with entries rax = R(x, a) and
7
Note that the condition on P (x, ., y) is a kind of Lipschitz-continuity.
470 A. Lőrincz

N × N matrix P a with entries Px,y

a = P (y | x, a). With these notations,

(20) can be written compactly as

 
(21) vt+1 := max ra + γP a vt .
a∈A

Here, max denotes the componentwise maximum operator.

It is also convenient to introduce the Bellman operator T : RN → RN
that maps value functions to value functions as
 
T v := max ra + γP a v .
a∈A

As it is well known, T is a max-norm contraction with contraction factor

γ: for any v, u ∈ RN ,

T v − T u∞ ≤ γv − u∞ .

Consequently, by Banach’s fixed point theorem, exact value iteration (which

can be expressed compactly as vt+1 := T vt ) converges to a unique solu-
tion v∗ from any initial vector v0 , and the solution v∗ satisfies the Bell-
man equations (16). Furthermore, for any required precision ε > 0, t ≥ 
log ε ∗ ∗
log γ v0 − v ∞ implies vt − v ∞ ≤ ε. One iteration costs O N · |A|
2

computation steps.

4.5.3. Approximate value iteration. In this section approximate value

iteration (AVI) with linear function approximation (LFA) in ordinary MDPs
is reviewed. The results of this section hold for AVI in general, but if we can
perform all operations effectively on compact representations (i.e. execution
time is polynomially bounded in the number of variables instead of the
number of states), then the method can be directly applied to the domain
of factorized Markovian decision problems, underlining the importance of
the following considerations [114].
Suppose that we wish to express the value function as the linear combina- 
tion of K basis functions hk : X → R X = {1, 2, . . . , N }, k ∈ {1, . . . , K} ,
where K ! N . Let H be the N × K matrix with entries Hx,k = hk (x). Let
wt ∈ RK denote the weight vector of the basis functions at step t. One can
substitute vt = Hwt into the r.h.s. of (21), but cannot do the same on the
l.h.s. of the assignment: in general, the r.h.s. is not contained in the image
space of H, so there is no such wt+1 that
 
Hwt+1 = max ra + γP a Hwt .
a∈A
Learning and Representation 471

Iteration can be put into work by projecting the r.h.s. into w-space: let
G : RN → RK be a (possibly non-linear) mapping, and consider the iteration
 
(22) wt+1 := G [ max ra + γP a Hwt ]
a∈A

with an arbitrary starting vector w0 .

Lemma 6 [114]. If G is such that HG is a non-expansion, i.e., for any

v, v ∈ RN ,
HG v − HG v ∞ ≤ v − v ∞ ,
then there exists a w∗ ∈ RK such that
 
w∗ = G [ max ra + γP a Hw∗ ]
a∈A

and iteration (22) converges to w∗ from any starting point.

Note that if G is a linear mapping with matrix G ∈ RK×N , then the

assumption of the lemma is equivalent to HG∞ ≤ 1 [114].

4.5.4. Convergent projection. In this section, one of the possibilities for

projection G is reviewed. For other possibilities and for the comparisons
between them, see [114].
Let v ∈ RN be an arbitrary vector, and let w = G v be its G -projection.
For linear operators, G can be represented in matrix form and we shall
denote it by G.

Normalized linear mapping: Let G be an arbitrary K ×N matrix, and define

its normalization N (G) as a matrix with the same dimensions and entries

+  Gi,j
N (G) i,j :=

,
( j  |Hi,j  |)( i |Gi ,j |)

that is, N (G) is obtained from G by dividing each element with the
corresponding row sum of H and the corresponding column sum of G.
All (absolute) row sums of H · N (G) are equal to 1. Therefore, (i)
H · N (G)∞ = 1, and (ii) H · N (G) is maximal in the sense that if the
absolute value of any element of N (G) increased, then for the resulting ma-
trix G , H · G ∞ > 1. Furthermore, this linear projection is convergent:
472 A. Lőrincz

Lemma 7 [114]. Let v∗ be the optimal value function and w∗ be the fixed
point of the approximate value iteration (22). Then

1
Hw∗ − v∗ ∞ ≤ HG v∗ − v∗ ∞ .
1−γ

According to the lemma, the error bound is proportional to the projec-

tion error of v∗ . Therefore, if v∗ can be represented in the space of basis
functions with small error, then this AVI algorithm gets close to the opti-
mum. Furthermore, the lemma can be used to check a posteriori how good
the basis functions are. One may improve the set of basis functions itera-
tively. Similar arguments have been brought up by Guestrin et al. [45], in
association with their LP-based solution.

4.6. Factored value iteration

MDPs are attractive because solution time is polynomial in the number of

states. Consider, however, a sequential decision problem with m variables.
In general, one needs an exponentially large state space to model it as an
MDP. So, the number of states is exponential in the size of the description of
the task. Factored Markov decision processes may avoid this trap because
of their more compact task representation.
The exact solution of factored MDPs is infeasible. The idea of repre-
senting a large MDP using a factored model was first proposed by Koller
and Parr [61], but similar ideas appear already in [13, 14]. More recently,
the framework (and some of the algorithms) was extended to fMDPs with
hybrid continuous-discrete variables [62] and factored partially observable
MDPs [98]. Furthermore, the framework has also been applied to structured
MDPs with alternative representations, e.g., relational MDPs [44] and first-
order MDPs [100].

4.6.1. Factored Markov decision processes. Assume that X is the

Cartesian product of m smaller state spaces (corresponding to individual
variables):
X = X1 × X2 × . . . × Xm .
For the sake of notational convenience assume further that each Xi has the
same size, |X1 | = |X2 | = . . . = |Xm | = n. With this notation, the size of
Learning and Representation 473

the full state space is N = |X| = nm . It is worth noting that all previous
derivations and proofs carry through to different size variable spaces.
A naive, tabular representation of the transition probabilities would
require exponentially large space (that is, exponential in the number of
variables m). However, the next-step value of a state variable often depends
only on a few other variables, so the full transition probability can be
obtained as the product of several simpler factors. For a formal description,
let us introduce a few notations:
For any subset of variable indices Z ⊆ {1, 2, . . . , m}, let X[Z] := × Xi ,
i∈Z
furthermore, for any x ∈ X, let x[Z] denote the value of the variables with
indices in Z. We may also use the notation x[Z] without specifying a full
vector of values x. In such cases x[Z] denotes
+ an element in X[Z]. For
single-element sets Z = {i} the shorthand x {i} = x[i] is appropriate.
A function f is a local-scope function if it is defined over a subspace
X[Z] of the state space, where Z is a (presumably small) index set. The
local-scope function
 f can be extended trivially to the whole state space by
f (x) := f x[Z] . If |Z| is small, local-scope functions can be represented
efficiently, as they can take only n|Z| different values.
Suppose that for each variable i there exist neighborhood sets Γi such
that the value of xt+1 [i] depends only on xt [Γi ] and the action at taken.
Then transition probabilities assume the following factored form

n
 
(23) P (y | x, a) = Pi y[i] | x[Γi ], a
i=1

for each x, y ∈ X, a ∈ A, where each factor is a local-scope function

(24) Pi : X[Γi ] × A × Xi → [0, 1] (for all i ∈ {1, . . . , m}).

We will also suppose that the reward function is the sum of J local-scope
functions:


J
 
(25) R(x, a) = Rj x[Zj ], a ,
j=1

with arbitrary (but preferably small) index sets Zj , and local-scope functions
 
(26) Rj : X[Zj ] × A → R for all j ∈ {1, . . . , J} .
474 A. Lőrincz

To sum up, a factored Markov decision process is characterized by the pa-

rameters {X  i : 1 ≤ i ≤ m}; A; {Rj : 1 ≤ j ≤ J}; {Γi : 1 ≤ i ≤ n}; {Pi : 1 ≤
i ≤ n}; xs ; γ , where xs denotes the initial state.

Functions Pi and Ri are usually represented either as tables or dynamic

Bayesian networks. If the maximum size of the appearing local scopes
is bounded by some constant, then the description length of an fMDP is
polynomial in the number of variables n.

The optimal value function is an N = nm -dimensional vector. To repre-

sent it efficiently, one should rewrite it as the sum of local-scope functions
with small domains. Unfortunately, in the general case, no such factored
form exists [45].

However, one can still approximate V ∗ with such expressions: let K be

the desired number of basis functions and for each k ∈ {1, . . . , K}, let Zk
be the domain set of the local-scope basis function hk : X[Zk ] → R. We are
looking for a value function of the form


K
 
(27) Ṽ (x) = wk · hk x[Ck ] .
k=1

The quality of the approximation depends on two factors: the choice of the
basis functions and the approximation algorithm. For given basis functions,
one can apply a number of algorithms to determine the weights wk .

4.6.2. Exploiting factored structure in value iteration. For fMDPs,

one can substitute the factored form of the transition probabilities (23),
rewards (25) and the factored approximation of the value function (27) into
the AVI formula (22), which yields


K  
m 
   
hk x[Ck ] · wk,t+1 ≈ max Pi y[i] | x[Γi ], a
a
k=1 y∈X i=1


J 
K 
   
· Rj x[Zj ], a + γ hk y[Ck ] · wk ,t .
j=1 k =1
Learning and Representation 475

By rearranging operations and exploiting that all occurring functions have

a local scope, one gets
& J
K
    
(28) hk x[Ck ] · wk,t+1 = Gk max Rj x[Zj ], a
a
k=1 j=1
  '

K     
+γ Pi y[i] | x[Γi ], a hk y[Ck ] · wk ,t
k =1 y[Ck ]∈X[Ck ] i∈Ck

for all x ∈ X. This update rule has a more compact form in vector notation.
Let
wt := (w1,t , w2,t , . . . , wK,t ) ∈ RK ,
let H be an |X| × K matrix containing the values of the basis functions and
let us index the rows of matrix H by the elements of X:
 
Hx,k := hk x[Ck ] .

Further, for each a ∈ A, let C a be the |X| × K value backprojection matrix

defined as
  
   
a
Cx,k := Pi y[i] | x[Γi ], a hk y[Zk ]
y[Zk ]∈X[Zk ] i∈Zk

and for each a, define the reward vector ra ∈ R|X| by


nr
 
rax := Rj x[Zj ], a .
j=1

Using these notations, (28) can be rewritten as

 
(29) wt+1 := G max ra + γC a wt .
a∈A

Now, all entries of C, H and r are composed of local-scope functions,

so any of their individual elements can be computed efficiently. This means
that the time required for the computation is exponential in the sizes of
function scopes, but only polynomial in the number of variables, making
the approach attractive. Unfortunately, the matrices are still exponentially
large, as there are exponentially many equations in (28). One can overcome
this difficulty by sampling.
476 A. Lőrincz

4.7. The sampling advantage

Let us select a random subset X ( ⊆ X of the original state space so that

(
|X| = poly(m), consequently, solution time will scale polynomially with m.
On the other hand, a sufficiently large subset should be selected so that
the remaining system of equations is still over-determined. For the sake of
notational simplicity, assume that the projection operator G is linear with
matrix G. Let the sub-matrices of G, H, C a and ra corresponding to X ( be
( ( (
denoted by G, H, C and (
a a
r , respectively. Then the following value update
 a 
(30) ( · max (
wt+1 := G ( a wt
r + γC
a∈A

can be performed effectively, because these matrices have polynomial size.

It can be shown that the solution from sampled data is close to the true
solution with high probability. The proof is closely related to the method
presented in [34, 35] with the important exception the infinity-norm and
not the L2 -norm is needed here. For the details of the proof, see [114].
There are attractive properties of factored RL and there are reassuring
experimental findings from the point of view of neuroscience.
1. State space does not grow exponentially with the number of variables
in factored RL. Sampling methods overcome the problem that the
number of equations may scale with the number of variables in the
exponent.
2. Neuroscience indicates that different components for RL and different
forms of RL are exploited in the brain [101, 81, 118, 27, 60, 63, 29],
including
(a) prediction of immediate reward,
(b) state value estimation,
(c) state action value estimation,
(d) estimations of the values of desired states,
(e) temporal difference learning,
and so on.
Learning and Representation 477

4.8. The symbol learning problem: Core of the learning task

We have seen that once factors – as defined earlier – are given, then sev-
eral sources of combinatorial explosion can be ignored in the factored RL
framework. However, for a general learning system, these factors should be
found or extracted without prior knowledge.
This issue is closely related to the so-called ‘symbol grounding prob-
lem’ [47] of cognitive philosophy. According to the symbolic model of the
mind [37, 92] we have an ‘autonomous’ symbolic level. In our wording it
means that the factors are given to us. The problem is to connect these sym-
bols to sensory information, that is to ground them to the physical world,
which is a long-standing and enigmatic problem in cognitive science.
Our factored RL model shares one particular property with the symbolic
model of mind: states are described in very few bits (symbols), but in the
background there is a huge number of other variables that the factored RL
model neglects. Symbol manipulation (∼‘action’) in our case means that a
few bit information can be dealt with even if a large part of the available
information carried by the neglected variables is omitted. We suggest that
although the symbol-grounding problem may have its own merits and may
be important in supervised training scenarios, but the core of the learning
problem is not the grounding of the symbols, but the forming and learning
of the symbols, i.e., the distillation of the few bit descriptors. If this ‘symbol
learning problem’ is feasible, i.e., if this task can be solved in polynomial
time, then we can close the loop, because the other components of the
learning problem, such as the separation of independent groups of factors
and the optimization of factored RL have polynomial time solutions and
combinatorial explosion may vanish.
One might worry about the very existence of the symbols, or about
the existence of constructive polynomial time algorithms yielding these
symbols. The fact that our brain seems to have such symbols and we
can learn to manipulate them does not provide further cues about the two
candidates, i.e., symbol grounding and symbol learning. Below, we provide
an information theoretic direction aiming to formulate the symbol learning
problem.
478 A. Lőrincz

4.9. Information theoretic direction for the symbol learning prob-

lem

In this paper, we have argued that both representation and goal-oriented

behavior can be reduced to bits and tables, respectively. For example,
the cross-entropy method used for the optimization of L0 norm provides
the indices of the sparse components and the continuous values of the
actual components is then determined. In addition, the robust controller
can put to work by using labels; the indices of the actuators are sufficient
for robust control. Factored RL can take advantage of these sparse and
discrete transformations, provided that a few bits – that we call symbols –
are satisfactory for the description of the state.
Let us see an example. Consider the symbol ‘chair’. Many properties
of any instantiations are neglected by this symbol. For example, the chair
could be an armchair, or it could be the moon, like in the trademark of
DreamWorks, or clouds are appropriate chairs for cherubs. Symbol ‘chair’ is
a ‘low-entropy’ variable, but it gives no hint about the actual manifestation.
The actual manifestation, on the other hand, is a ‘high entropy’ variable.
All materializations of symbol chair share a few common characteristics,
e.g., we can sit onto the chair or we can stand up from the chair. Note
that neither of these are ‘actions’ in the sense that they provide little if
any information about the usage of the actuators. Instead, they can be
characterized as desired events with certain probabilities that we can make
these events happen. These desired events are also ‘low-entropy’ variables,
because many of the details of the events are again neglected. From the
point of view of RL, we are interested in the following two questions:
1. Is there any partitioning of the observations such that the transition
probabilities between the low-entropy variables are good approxima-
tions of the transition probabilities between the high-entropy manifes-
tations? In other words, can we claim that there are symbols, which
are useful for reinforcement learning?
2. If such symbols exist then can we find them, or are they inherited?
For the sake of argument, assume that the answer to the first question is
affirmative, so the second question is relevant. If we happen to inherit the
symbols then it seems wise to focus on the symbol grounding problem and
to search for the underlying genetical structure. Having found it, we might
be able to copy the method and develop more efficient machines, or prewire
higher level symbols into the brain. On the other hand, if we learn the
Learning and Representation 479

symbols through experiences, then the symbol grounding problem vanishes.

Furthermore, we should look for the algorithm, which is so successful and
efficient already at infancy that we do not even notice when it is at work.
Recent advances that extend extreme graph theory to the information
theoretic domain use the above terms [117] and may fill in the gap by
providing rigorous formulation to the conditions of the ‘symbol learning
problem’ for the hardest case, the case of extreme graphs.8 If this can be
done and if a given environment satisfies the related conditions, then symbol
learning may be accomplished in polynomial time in that environment,
because constructive polynomial time algorithms do exist for such problem
types [4, 40, 41].
It is important to note that symbol grounding and symbol learning are
not incompatible with each other. Furthermore, in some cases these learn-
ing methods can rely on each other. Take communication as an example:
symbols of communicating agents should be matched in two ways: (i) the
symbols they learned should be matched against each other and (ii) the sig-
nals the agents use should refer to the matched symbols. The combination
of the reconstruction network and the cross entropy method seems efficient
in solving such tasks, including the case that the observations of the agents
may originally differ [46]. Because the interlaced nature of symbol learning
and symbol grounding could be crucial for the understanding of the emer-
gence of language, it is relevant whether our concepts enable both learning
methods. The model is promising in this respect: learning rule (10) shows
bottom-up learning that works in a self-organizing manner, whereas the
same rule admits supervisory top-down training, which is made apparent
by the form of (9).

5. Summary

It seems that the core problem of learning concerns symbols. One should
either learn symbols or should ground the symbols if those are inherited.
Here, symbols are low entropy variables and represent high entropy instan-
tiations. The main concern is if there are low entropy variables such that
the transition probabilities between the low entropy variables determine
the transition probabilities between the high entropy manifestations or not
8
For extreme graphs, the number of edges is proportional to the square of the number
of the vertices, which is not typical in nature.
480 A. Lőrincz

and whether the low entropy variables can be found in polynomial time or
not. Recent advances in extreme graph theory [117] and related construc-
tive polynomial time algorithms [4, 40, 41] might translate to a rigorous
formulation for this problem. If that succeeds, then the conditions for poly-
nomial time algorithms for the symbol learning problem can be established.
We have also argued that both symbol learning and symbol grounding may
have their own merits and that they are compatible with each other.
This issue is relevant from the point of view of learning. If symbols
can be found then the optimization of the desired events in reinforcement
learning becomes possible. Some of the events can be combined and such
low complexity combinations can be very efficient in problem solving [113].
It is also shown in the cited work that low-complexity combinations can be
found by means of the same cross entropy method that we introduced into
our model for the learning of sparse representations [73].
Learning algorithms described in this paper were selected by our soft
constraints, the experimental findings in neuroscience and cognition. The
bonus of the approach is that – in certain cases – there are reasonable
matches between the algorithmic components and some neural substrates
[70, 72, 75, 77].

Acknowledgement. The author is highly indebted to Gábor Szirtes for his

help in the serialization of the cognitive map of the soft and hard constraints
of this review. This research has been supported by the EC FET ‘New
Ties’ Grant FP6-502386 and NEST ‘PERCEPT’ Grant FP6-043261 and
Air Force Grant FA8655-07-1-3077. Any opinions, findings and conclusions
or recommendations expressed in this material are those of the authors and
do not necessarily reflect the views of the European Commission, European
Office of Aerospace Research and Development, Air Force Office of Scientific
Research, Air Force Research Laboratory.
Learning and Representation 481

References

[1] P. Abbeel and A. Y. Ng, Apprenticeship learning via inverse reinforcement learning,
in: D. Schuurmans, R. Geiner and C. Brodley, editors, Proceedings of the 21st
International Conference on Machine Learning, pages 663–670, New York, NY,
2004. ACM Press.
[2] K. Abed-Meraim and A. Belouchrani, Algorithms for joint block diagonalization,
in: Proceedings of EUSIPCO, pages 209–212, 2004.
[3] D. Ackley, G. E. Hinton and T. Sejnowski, A learning algorithm for Boltzmann
machines, Cognitive Science, 9 (1985), 147–169.
[4] N. Alon, R. A. Duke, H. Lefmann, V. Rödl and R. Yuster, The algorithmic aspects
of the regularity lemma, Journal of Algorithms, 16 (1994), 80–109.
[5] F. Attneave, Some informational aspects of visual perception, Psychological Review,
61 (1954), 183–193.
[6] F. R. Bach and M. I. Jordan, Beyond independent components: Trees and clusters,
Journal of Machine Learning Research, 4 (2003), 1205–1233.
[7] F. R. Bach and M. I. Jordan, Finding clusters in Independent Component Analysis,
in: Proceedings of ICA2003, pages 891–896, 2003.
[8] D. H. Ballard, G. E. Hinton and T. J. Sejnowski, Parallel visual computation,
Nature, 306 (1983), 21–26.
[9] H. B. Barlow, Sensory Communication, pages 217–234, MIT Press, Cambridge,
MA, 1961.
[10] A. Barto, Discrete and continuous models, International Journal of General Sys-
tems, (1978), 163–177.
[11] A. P. Batista and W. T. Newsome, Visuo-motor control: Giving the brain a hand,
Current Biology, 10 (2000), R145–R148.
[12] J. Baxter, A. Tridgell and L. Weaver, Machines that learn to play games, chapter
Reinforcement learning and chess, pages 91–116, Nova Science Publishers, Inc.,
2001.
[13] C. Boutilier, R. Dearden and M. Goldszmidt, Exploiting structure in policy con-
struction, in: Proceedings of the 14th Fourteenth International Joint Conference on
Artificial Intelligence, pages 1104–1111, 1995.
[14] C. Boutilier, R. Dearden and M. Goldszmidt, Stochastic dynamic programming
with factored representations, Artificial Intelligence, 121(1–2) (2000), 49–107.
[15] R. I. Brafman and M. Tennenholtz, A near-optimal polynomial time algorithm for
learning in certain classes of stochastic games, Artificial Intelligence, 121(1–2)
(2000), 31–47.
[16] R. I. Brafman and M. Tennenholtz, R-max – a general polynomial time algorithm
for near-optimal reinforcement learning, Journal of Machine Learning Research, 3
(2002), 213–231.
482 A. Lőrincz

[17] L. Buşoniu, R. Babuška and B. De Schutter, Multi-agent reinforcement learning:

A survey, in: Proceedings of the 9th International Conference on Control, Automa-
tion, Robotics and Vision, pages 527–532, 2006.
[18] E. Candes and J. Romberg, Quantitative robust uncertainty principles and op-
timally sparse decompositions, Foundations of Computational Mathematics, 6
(2006), 227–254.
[19] E. Candes, J. Romberg and T. Tao, Robust uncertainty principles: Exact signal
reconstruction from highly incomplete frequency information, IEEE Transactions
on Information Theory, 52 (2006), 489–509.
[20] J. F. Cardoso, Multidimensional independent component analysis, in: Proceedings
of ICASSP, volume 4, pages 1941–1944, 1998.
[21] O. Carter, D. Presti, C. Callistemon, Y. Ungerer, G. Liu and J. Pettigrew, Med-
itation alters perceptual rivalry in Tibetan Buddhist monks, Current Biology, 15
(2005), R412–R413.
[22] Y.-H. Chang, T. Ho and L. P. Kaelbling, All learning is local: Multi-agent learning
in global reward games, in: Advances in Neural Information Processing Systems
16, 2004.
[23] S. Choi, A. Cichocki, H.-M. Park and S.-Y. Lee, Blind source separation and
independent component analysis, Neural Information Processing – Letters and
Reviews, 6 (2005), 1–57.
[24] J. J. Chrobak, A. Lőrincz and G. Buzsáki, Physiological patterns in the hippo-
campo-entorhinal cortex system, Hippocampus, 10 (2000), 457–465.
[25] P. Comon, Independent Component Analysis, a new concept? Signal Processing,
Elsevier, 36(3) (April 1994), 287–314. Special issue on Higher-Order Statistics.
[26] V. Conitzer and T. Sandholm, AWESOME: A general multiagent learning algo-
rithm that converges in self-play and learns a best response against stationary
opponents, Machine Learning, 67 (2007), 23–43.
[27] N. D. Daw, Y. Niv and P. Dayan, Uncertainty-based competition between pre-
frontal and dorsolateral striatal systems for behavioral control, Nature Neuro-
science, 8 (2005), 1704–1711.
[28] P.-T. de Boer, D. P. Kroese, S. Mannor and R. Y. Rubinstein, A tutorial on the
cross-entropy method, Annals of Operations Research, 134 (2004), 19–67.
[29] M. R. Delgado, Reward-related responses in the human striatum, Annals of the
New York Academy of Sciences, 1104 (2007), 70–88.
[30] D. C. Dennett, Consciousness explained, Little Brown, Boston, MA, 1991.
[31] D. W. Dong and J. J. Atick, Statistics of natural time varying images, Network
Computation in Neural Systems, 6 (1995), 345–358.
[32] D. W. Dong and J. J. Atick, Temporal decorrelation: A theory of lagged and
nonlagged responses in the lateral geniculate-nucleus, Network Computation in
Neural Systems, 6 (1995), 159–178.
[33] D. Donoho, Compressed sensing, IEEE Transactions on Information Theory, 52
(2006), 1289–1306.
Learning and Representation 483

[34] P. Drineas, R. Kannan and M. W. Mahoney, Fast monte carlo algorithms for
matrices i: Approximating matrix multiplication, SIAM Journal of Computing,
36 (2006), 132–157.
[35] P. Drineas, M. W. Mahoney and S. Muthukrishnan, Sampling algorithms for l2
regression and applications, in: Proceedings of the 17th Annual SODA, pages 1127–
1136, 2006.
[36] D. J. Field, What is the goal of sensory coding?, Neural Computation, 6 (1994),
559–601.
[37] J. A. Fodor, Methodological solipsism considered as a research strategy in cognitive
psychology, Behavioral and Brain Sciences, 3 (1980), 63–109.
[38] T. Fomin, T. Rozgonyi, Cs. Szepesvári and A. Lőrincz, Self-organizing multi-
resolution grid for motion planning and control, International Journal of Neural
Systems, 7 (1997), 757–776.
[39] M. Franzius, H. Sprekeler and L. Wiskott, Slowness and sparseness lead to place,
head-direction and spatial-view cells, PLoS Computational Biology, (8), 2007,
doi:10.1371/journal.pcbi.0030166.
[40] A. M. Frieze and R. Kannan, The regularity lemma and approximation schemes
for dense problems, in: Proceedings of the 37th Annual IEEE Symposium on Foun-
dations of Computing, pages 12–20, 1996.
[41] Alan Frieze and Ravi Kannan, A simple algorithm for constructing szemerédi’s
regularity partition, Electronic Journal of Combinatorics, 6 (1999).
http://www.emis.ams.org/journals/EJC/Volume 6/PDF/v6i1r17.pdf.
[42] C. Fyfe and R. Baddeley, Finding compact and sparse-distributed representations
of visual images, Network Computation in Neural Systems, 6 (1995), 333–344.
[43] C. G. Gross, G. S. Yap and M. S. A. Graziano, Coding of visual space by premotor
neurons, Science, 266 (1994), 1054–1057.
[44] C. Guestrin, D. Koller, C. Gearhart and N. Kanodia, Generalizing plans to new
environments in relational MDPs, in: Proceedings of the 18th International Joint
Conference on Artificial Intelligence, 2003.
[45] C. Guestrin, D. Koller, R. Parr and S. Venkataraman, Efficient solution algorithms
for factored MDPs, Journal of Artificial Intelligence Research, 19 (2002), 399–468.
[46] V. Gyenes and A. Lőrincz, Co-learning and the development of communication,
Lecture Notes in Computer Science, 4668 (2007), 827-837.
[47] S. Harnad, The symbol grounding problem, Physica D, D 42 (1990), 335–346.
[48] D. A. Henze, L. Wittner and G. Buzsáki, Single granule cells reliably discharge
targets in the hippocampal CA3 network in vivo, Nature Neuroscience, 5 (2002),
790–795.
[49] G. E. Hinton and R. R. Slakhutdnikov, Reducing the dimensionality of data with
neural networks, Science, 313 (2006), 504–507.
[50] Y. K. Hwang and N. Ahuja, Gross motion planning – a survey, ACM Computing
Surveys, 24(3) (1992), 219–291.
484 A. Lőrincz

[51] A. Hyvärinen, Independent component analysis for time-dependent stochastic

processes, in: Proceedings of ICANN, pages 541–546, Berlin, 1998. Springer-Verlag.
[52] A. Hyvärinen and U. Köster, FastISA: A fast fixed-point algorithm for independent
subspace analysis, in: Proceedings of ESANN, Evere, Belgium, 2006.
[53] S. Ishii, H. Fujita, M. Mitsutake, T. Yamazaki, J. Matsuda and Y. Matsuno, A re-
inforcement learning scheme for a partially-observable multi-agent game, Machine
Learning, 59(1–2) (2005), 31–54.
[54] W. James, The Principles of Psychology, 1890, p. 488
http://www.archive.org/details/theprinciplesofp01jameuoft
[55] Zs. Kalmár, Cs. Szepesvári and A. Lőrincz, Module-based reinforcement learning:
Experiments with a real robot, Machine Learning, 31 (1998), 55–85.
[56] M. Kawato, H. Hayakawa and T. Inui, A forward-inverse model of reciprocal
connections between visual neocortical areas, Network, 4 (1993), 415–422.
[57] M Kearns and S. Singh, Near-optimal reinforcement learning in polynomial time,
in: Proceedings of the 15th International Conference on Machine Learning, pages
260–268, San Francisco, CA, 1998. Morgan Kaufmann Publishers Inc.
[58] F. Kloosterman, T. van Haeften and F. H. Lopes da Silva, Two reentrant pathways
in the hippocampal-entorhinal system, Hippocampus, 14 (2004), 1026–1039.
[59] B. J. Knowlton and L. R. Squire, The learning of categories: parallel brain systems
for item memory and category knowledge, Science, 10 (1993), 1747–1749.
[60] B. Knutson and G. E. Wimmer, Splitting the difference: How does the brain code
reward episodes?, Annals of the New York Academy of Sciences, 1104, (2007),
54–69.
[61] D. Koller and R. Parr, Policy iteration for factored MDPs, in: Proceedings of the
16th Conference on Uncertainty in Artificial Intelligence, pages 326–334, 2000.
[62] B. Kveton, M. Hauskrecht and C. Guestrin, Solving factored MDPs with hybrid
state and action variables, Journal of Artificial Intelligence Research, 27 (2006),
153–201.
[63] D. Lee and H. Seo, Mechanisms of reinforcement learning and decision making in
the primate dorsolateral prefrontal cortex, Annals of the New York Academy of
Sciences, 1104 (2007), 108–122.
[64] H. Lee, A. Battle, R. Raina and A. Y. Ng, Efficient sparse coding algorithms, in:
B. Schölkopf, J. Platt and T. Hoffman, editors, Advances in Neural Information
Processing Systems 19, pages 801–808. MIT Press, Cambridge, MA, 2007.
[65] D. A. Leopold and N. K. Logothetis, Activity changes in early visual cortex reflect
monkeys’ percepts during binocular rivalry, Nature, 379 (1996), 549–553.
[66] D. A. Leopold, M. Wilke, A. Maier and N. K. Logothetis, Stable perception of
visually ambiguous patterns, Nature Neuroscience, 5 (2002), 605–609.
[67] N. K. Logothetis and J. D. Schall, Neuronal correlates of subjective visual-percep-
tion, Science, 245 (1989), 761–763.
Learning and Representation 485

[68] A. Lőrincz, Forming independent components via temporal locking of reconstruc-

tion architectures: A functional model of the hippocampus, Biological Cybernetics,
75 (1998), 37–47.
[69] A. Lőrincz and G. Buzsáki, Two-phase computational model training long-term
memories in the entorhinal-hippocampal region, Annals of the New York Academy
of Sciences, 911, (2000), 83–111.
[70] A. Lőrincz, Gy. Hévı́zi and Cs. Szepesvári, Ockham’s razor modeling of the matri-
some channels of the basal ganglia thalamocortical loop, International Journal of
Neural Systems, 11 (2001), 125–143.
[71] A. Lőrincz, V. Gyenes, M. Kiszlinger and I. Szita, Mind model seems necessary
for the emergence of communication, Neural Information Processing – Letters and
Reviews, 11 (2007), 109–121.
[72] A. Lőrincz, M. Kiszlinger and G. Szirtes, Model of the hippocampal formation ex-
plains the coexistence of grid cells and place cells, http://arxiv.org/abs/0804.3176,
2008.
[73] A. Lőrincz, Zs. Palotai and G. Szirtes, Spike-based cross-entropy method for re-
construction, Neurocomputing, 2008, (in press).
[74] A. Lőrincz, I. Pólik and I. Szita, Event-learning and robust policy heuristics,
Cognitive Systems Research, 4 (2003), 319–337.
[75] A. Lőrincz and Z. Szabó, Neurally plausible, non-combinatorial iterative indepen-
dent process analysis, Neurocomputing, 70 (2007), 1569–1573.
[76] A. Lőrincz, B. Szatmáry and G. Szirtes, Mystery of structure and function of
sensory processing areas of the neocortex: A resolution, Journal of Computational
Neuroscience, 13 (2002), 187–205.
[77] A. Lőrincz and G. Szirtes, Autoregressive model of the hippocampal representation
of events, in: Proceedings of IJCNN2009, (in press).
[78] L. Margolin, On the convergence of the cross-entropy method, Annals of Operations
Research, 134 (2005), 201–214.
[79] B. L. McNaughton, F P. Battaglia, O. Jensen, E. I. Moser and M.-B. Moser, Path
integration and the neural basis of the ćognitive map,́ Nature Reviews Neuroscience,
7 (2006), 663–678.
[80] T. C. Mills, Time Series Techniques for Economists, Cambridge University Press,
Cambridge, 1990.
[81] P. R. Montague, S. E. Hyman and J. D. Cohen, Computational roles for dopamine
in behavioural control, Nature, 431 (2004), 760–767.
[82] G. Neu and Cs. Szepesvári, Apprenticeship learning using inverse reinforcement
learning and gradient methods, in: Proceedings of the 24th Conference on Uncer-
tainty in Artificial Intelligence, pages 295–302. AUAI Press, 2007.
[83] A. Y. Ng and S. Russell, Algorithms for inverse reinforcement learning, in: Pro-
ceedings of the 17th International Conference on Machine Learning, pages 663–670,
San Francisco, CA, 2000. Morgan Kaufmann Publishers Inc.
486 A. Lőrincz

[84] G. Nolte, F. C. Meinecke, A. Ziehe and K. R. Müller, Identifying interactions in

mixed and noisy complex systems, Physical Review E, 73 (2006), doi: 051913.
[85] B. A. Olshausen and D. J. Field, Emergence of simple-cell receptive field properties
by learning a sparse code for natural images, Nature, 381 (1996), 607–609.
[86] B. A. Olshausen and D. J. Field, Sparse coding with an overcomplete basis set:
A strategy employed by V1? Vision Research, 37 (1997), 3311–3325.
[87] W. X. Pan, R. Schmidt, J. R. Wickens and B. I. Hyland, Dopamine cells respond
to predicted events during classical conditioning: Evidence for eligibility traces in
the reward-learning network, Journal of Neuroscience, 25 (2005), 6235–6242.
[88] B. Póczos and A. Lőrincz, Independent subspace analysis using geodesic spanning
trees, in: Proceedings of the 22nd International Conference on Machine Learning,
pages 673–680, New York, NY, USA, 2005. ACM Press.
[89] B. Póczos, Z. Szabó, M. Kiszlinger and A. Lőrincz, Independent process analysis
without a priori dimensional information, Lecture Notes in Computer Science, 4666
(2007), 252–259.
[90] B. Póczos, B. Takács and A. Lőrincz, Independent subspace analysis on innovations,
in: Proceedings of ECML, pages 698–706, Berlin, 2005. Springer-Verlag.
[91] T. Poggio, V. Torre and C. Koch, Computational vision and regularization theory,
Nature, 317 (1985), 314–319.
[92] Z. W. Pylyshyn, Computation and cognition: Issues in the foundations of cognitive
science, Behavioral and Brain Sciences, 3 (1980), 111–169.
[93] R. P. N. Rao and D. H. Ballard, Predictive coding in the visual cortex: A functional
interpretation of some extra-classical receptive-field effects, Nature Neuroscience,
2 (1999), 79–87.
[94] P. Redgrave and K. Gurney, The short-latency dopamine signal: a role in discov-
ering novel actions?, Nature Reviews Neuroscience, 7 (2006), 967–975.
[95] A. D. Redish, F. P. Battaglia, M. K. Chawla, A. D. Ekstrom, J. L. Gerrard, P. Lipa,
E. S. Rosenzweig, P. F. Worley, J. F. Guzowski, B. L. McNaughton and C. A.
Barnes, Independence of firing correlates of anatomically proximate hippocampal
pyramidal cells, Journal of Neuroscience, 21 (2001), 1–6.
[96] M. Rehn and F. T. Sommer, A network that uses few active neurones to code
visual input predicts the diverse shapes of cortical receptive fields, Journal of
Computational Neuroscience, 22 (2007), 135–146.
[97] P. Reinagel and R. C. Reid, Temporal coding of visual information in the thalamus,
Journal of Neuroscience, 20 (2000), 5392–5400.
[98] B. Sallans, Reinforcement Learning for Factored Markov Decision Processes, PhD
thesis, University of Toronto, 2002.
[99] K. Samejima and K. Doya, Multiple representations of belief states and action
values in corticobasal ganglia loops, Annals of the New York Academy of Sciences,
1104 (2007), 213–228.
[100] S. Sanner and C. Boutilier, Approximate linear programming for first-order MDPs,
in: Proceedings of the 21th Annual Conference on Uncertainty in Artificial Intelli-
gence (UAI), pages 509–517, 2005.
Learning and Representation 487

[101] W. Schultz, Getting formal with dopamine and reward, Neuron, 36 (2002), 241–
263.
[102] W. B. Scoville and B. Milner, Loss of recent memory after bilateral hippocampal
lesions, Journal of Neurology, Neurosurgery and Psychiatry, 20 (1957), 11–21.
[103] P. Spronck, M. Ponsen, I. Sprinkhuizen-Kuyper and E. Postma, Adaptive game ai
with dynamic scripting, Machine Learning, 63(3) (2006), 217–248.
[104] L. R. Squire, Memory and hippocampus: a synthesis of findings with rats, monkeys
and humans, Psychological Review, 99 (1992), 195–231.
[105] H. Stögbauer, A. Kraskov, S. A. Astakhov and P. Grassberger, Least dependent
component analysis based on mutual information, Physical Review E, 70, 2004.
[106] Z. Szabó, B. Póczos and A. Lőrincz, Cross-entropy optimization for independent
process analysis, in: Lecture Notes in Computer Science, 3889 (2006), 909–916.
Springer, 2006.
[107] Z. Szabó, B. Póczos and A. Lőrincz, Separation theorem for K-independent sub-
space analysis with sufficient conditions, Technical report, 2006,
http://arxiv.org/abs/math.ST/0608100.
[108] Z. Szabó, B. Póczos and A. Lőrincz, Undercomplete blind subspace deconvolution,
Journal of Machine Learning Research, 8 (2007), 1063–1095.
[109] Cs. Szepesvári, Sz. Cimmer and A. Lőrincz, Neurocontroller using dynamic state
feedback for compensatory control, Neural Networks, 10 (1997), 1691–1708.
[110] Cs. Szepesvári and A. Lőrincz, Approximate inverse-dynamics based robust control
using static and dynamic feedback, in: J. Kalkkuhl, K. J. Hunt, R. Zbikowski and
A. Dzielinski, editors, Applications of Neural Adaptive Control Theory, volume 2,
pages 151–179. World Scientific, Singapore, 1997.
[111] Cs. Szepesvári and A. Lőrincz, An integrated architecture for motion-control and
path-planning, Journal of Robotic Systems, 15 (1998), 1–15.
[112] I. Szita and A. Lőrincz, Learning Tetris using the noisy cross-entropy method,
Neural Computation, 18(12) (2006), 2936–2941.
[113] I. Szita and A. Lőrincz, Learning to play using low-complexity rule-based policies:
Illustrations through Ms. Pac-Man, Journal of Artificial Intelligence Research, 30
(2007), 659–684.
[114] I. Szita and A. Lőrincz, Factored value iteration converges, Acta Cybernetica,
accepted (2008). http://arxiv.org/abs/0801.2069.
[115] I. Szita and A. Lőrincz, Online variants of the cross-entropy method,
http://arxiv.org/abs/0801.1988v1, 2008.
[116] I. Szita, B. Takács and A. Lőrincz, Epsilon-mdps: Learning in varying environ-
ments, Journal of Machine Learning Research, 3 (2003), 145–174.
[117] T. Tao, Szemerédi’s regularity lemma revisited, Contributions to Discrete Mathe-
matics, 1 (2006), 8–28.
[118] S. C. Tanaka, K. Doya, G. Okada, K. Ueda, Y. Okamoto and S. Yamawaki, 3,4,
Prediction of immediate and future rewards differentially recruits cortico-basal
ganglia loops, Nature Neuroscience, 7 (2004), 887–893.
488 A. Lőrincz

[119] G. Tesauro, Temporal difference learning and TD-gammon, Communications of the

ACM, 38(3) (1995), 58–68.
[120] F. J. Theis, Uniqueness of complex and multidimensional independent component
analysis, Signal Processing, 84(5) (2004), 951–956.
[121] F. J. Theis, Blind signal separation into groups of dependent signals using joint
block diagonalization, in: Proceedings of ISCAS, pages 5878–5881, 2005.
[122] F. J. Theis, Towards a general independent subspace analysis, in: Advances in
Neural Information Processing Systems 19, pages 1361–1368, 2007.
[123] R. Vollgraf and K. Obermayer, Multi-dimensional ICA to separate correlated
sources, in: Advances in Neural Information Processing Systems 14, pages 993–
1000. MIT Press, 2001.
[124] S. Yu and J. Shi, Multiclass spectral clustering, in: Proceedings of ICCV, pages
313–319, 2003.

András Lőrincz
Department of Software Technology and
Methodology
Eötvös Loránd University
Pázmány Péter sétány 1/C
Budapest Hungary H-1117
e-mail: [email protected]
BOLYAI SOCIETY Handbook of Large-Scale
MATHEMATICAL STUDIES, 18 Random Networks
pp. 489–530.

Chapter 12
Telephone Call Network Data Mining:
A Survey with Experiments∗

MIKLÓS KURUCZ, LÁSZLÓ LUKÁCS, DÁVID SIKLÓSI,

ANDRÁS A. BENCZÚR, KÁROLY CSALOGÁNY
and ANDRÁS LUKÁCS

We survey some results of social network modeling and analysis relevant for
telephone call networks and illustrate these results over the call logs of major
Hungarian telephone companies. Our unique data sets include millions of users,
long time range, and sufficiently strong sociodemographic information on the
users. We explore properties that give stronger intuition on how contacts within
real social networks arise, and suggest properties unexplained by current network
evolution models.
We cover four areas: link prediction that investigates measures of connection
strength between members of the community; geographic contact distribution
modeling that explains how small world networks arise and span large physical
distances; high-level partitioning and clustering that identifies top level organiza-
tional principles of the network; finally classification in the presence of a network
with strong influence between connected nodes.
Our key findings are summarized as follows. Firstly, the strength of the
contacts within our telephone call networks is best characterized by the cosine
similarity of the neighborhood of a pair of users. Secondly, the call graph within
a small region such as Hungary is strongly distorted by large cities but otherwise
follows a law similar to larger networks such as US blogger friendship that appear
in the literature. Thirdly, spectral graph partitioning, a traditional method
performs very well in combination with some recent social network clustering
heuristics and is capable of identifying the top level structure of the call graph.
Finally, graph stacking, a recent method of semi-supervised classification that
constructs features based on the neighboring labels, proves to be a powerful
tool to classify the elements of the network. We illustrate this by the example
of classifying churn, i.e. those users who discontinue their participation in the
network.
∗
This work was supported by grants OTKA NK 72845 and NKFP-07-A2 TEXTREND.
490 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

1. Introduction

There has been a considerable growth of interest in the properties of net-

works with a particular focus on the evolution of the contacts, the analysis
of communities within networks or the classification of network objects.
Networks underline all aspects of our life including friends, social contacts,
computers and even brain cells or protein interaction in bacteria. Several
surveys cover recent results: Barabási [8], Newman [66] or Scott [78] to
name a few.
The purpose of investigating, measuring and modeling social networks
may include the analysis of community formation or information spread
within the network. The network of telephone communication contacts
is particularly important in practice. Telephone call network models may
serve the purposes of user segmentation or the selection of communities with
desired or undesired properties. A desired community may be one with
high ADSL penetration where new ADSL lines or other advanced services
are likely sold with success to those members who have no subscription yet.
An undesired community may be one with high recent churn rate where a
campaingn may have to be designed to keep the members in the service.
Other applications include viral marketing analysis [76] and other means
of enhancing marketing communication by also relying on the spread of
information within the social network.
In this paper we survey four kinds of applications illustrated by measure-
ments over a more than two million Hungarian telephone users’ data. First
in Section 2 we compare measures for the strength of the connection be-
tween members of the network by performing a link prediction experiment.
Then in Section 3 we investigate the geographic location as a predictor of
proximity in the social network. Next in Section 4 we describe the character-
istics of clusters that can be algorithmically found by measuring both graph
properties and external sociodemographic parameters such as geographic
location. Finally in Section 5 we rely on the strength of the connection of
pairs of users to classify churn, i.e. those users who discontinue their use of
the service or, in other words, quit the network.
The telephone graph appears less frequently in publications of the data
mining community compared to the social network of bloggers [56, and
references therein] or the World Wide Web [30, and many others]. Few ex-
ceptions include a theoretical analysis of connected components and eigen-
values [3, 26, 27] and several machine learning methods for churn prediction
Telephone Call Network Data Mining: A Survey with Experiments 491

on real data [7, 83, etc.]. Closest to our results are the structural inves-
tigations of mobile telephone call graphs [64, 69, 70] and the sketch-based
approximate k-means clustering of traffic among AT&T collection stations
over the United States [28]; for the latter result however the underlying
graph is much smaller (20,000 nodes) and the main goal was to handle the
time evolution as an additional dimension.
Our experiments are performed on the call graph of more than two
million Hungarian landline telephone users [13], a unique data of long time
range with sufficiently rich sociodemographic information on the users. Our
results differ from prior work on network clustering and characterization in
two aspects:
• In our evaluation we deploy external sociodemographic parameters
such as geographic location in addition to graph properties.
• Our problems are larger, sometimes by several orders of magnitude,
than previously reported: our graph has near 50,000,000 edges, which
poses challenges even for quite a few of more recent algorithms. Im-
proved hardware capabilities require new algorithms and lead to new
empirical findings in the paper.
The rest of the paper is organized into four sections, each describing one
of the four main topics. While the sections partly build on one another,
they can be read separately. For this purpose each section starts with a
specific introduction and a survey of the relevant results of the field. While
the first two sections contain new experimental results, the last two build
on our recent work [57] and [16].

1.1. Data sets

Before presenting the main results, we describe the data sets that we use
for illustration and show its main graph theoretic parameters. Our graphs
obey the generally observed properties of social networks: they have power
law degree distribution [9, 10, 17], small diameter [82, 4] and consist of a
single giant component [3].
In this paper we use two data sets for illustration. The first large data set
covers a time range of 8 months and a large portion of Hungarian landline
users. After aggregating calls between the same pairs of callers we have
obtained a graph with n = 2, 100, 000 nodes and m = 48, 400, 000 directed
edges out of which 10,800,000 is bidirectional. As the other small set we
492 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

use data from a small Hungarian landline telephone service provider. For a
time range of 12 months, after aggregating calls between the same pairs of
callers we have obtained a graph with n = 73, 000 nodes and m = 1, 360, 000
directed edges. The first data is used in our link prediction, geographic
modeling and clustering experiments while the second contains all additional
features suitable for churn prediction. The summary of the data sets is in
Fig. 1, left.
Bidirectional edges are crucial in some of our applications since they
show mutual connection compared to a one-directional call to e.g. a public
service number. When considering bidirectional edges only, approximately
30,000 users (1.5%) become isolated from the giant component of the large
graph.

Fig. 1. Left: Main parameters of the call graphs. Right: Distribution of the user
number by settlements in the data; the capital Budapest of near 600,000 users is
trimmed.

Information on the name and geographic location of the settlement (city,

village, suburban district) of the nodes is used in several of our experiments
and models. Settlement sizes (Fig. 1, right) follow a distribution very close
to lognormal with the exception of a very heavy tail of Hungary’s capital
Budapest of near 600,000 users. In a rare number of cases the data consists
of subpart names of settlements resulting in a relatively large number of
settlements with one or two telephone numbers; since the total number of
such nodes is negligible in the graph, we omit cleaning the data in this
respect.
Telephone Call Network Data Mining: A Survey with Experiments 493

2. Link Prediction

In this section we provide a link prediction experiment that, given an

observed period of usage, predicts pairs of users (edges, or links) that will
appear in a future time period. By such an experiment we may investigate
node similarity measures for finding important connections and ignoring
“accidental” unimportant ones. Our methods and findings will also be
used in future sections: we use them as input weights for graph clustering
(Section 4) and graph based information propagation (Section 5).
Link prediction is to our best knowledge first investigated by Liben-
Nowell and Kleinberg [60] with very similar motivations in an earlier paper
of Newman [65]. They consider a wide variety of methods based on the
neighborhood and the ensemble of paths corresponding to the pair of nodes
in question. Next we describe a selection of these methods including those
that performed best in their experiments. Then we will show link prediction
measurements over the telephone call graph.
Several algorithms were designed to evaluate node-to-node similarities
in networks that can be used to give alternate, similarity based weights
to node pairs. We refer to [60] for an exhaustive list of the available
methods ranging from co-citation to more complex measures such as max-
flow/min-cut-based similarities defined in [61]. These weights are used in
applications outside the link prediction area: [40, and many more] apply
them to improve clustering quality; co-citation is for example first used in
[41] as an elementary step of trust propagation.
Similarity in a telephone call graph is best characterized by the undi-
rected graph since communication is typically bidirectional regardless of the
actual caller–callee direction. We also have a choice to use cost, duration or
number of calls as a weight of a pair of users, or we may ignore weights and
consider an unweighted graph. We will compare both the input directed
graph, its transpose by changing the direction of each edge, or the undi-
rected version arising as the union of the previous two graphs. We will refer
to the three variants as directed, reversed and undirected versions. For an
edge weight function d : V × V → R we use d− (u, v) = d(v, u) for the re-
versed and d↔ = d + d− for the undirected version. We extend this notion
for an arbitrary similarity measure sim (u, v) computed over edge weights d
and compute sim− (u, v) over d− and sim↔ (u, v) over d↔ .
In the discussion below we identify scalability as the main challenge for
computing node similarities. Computing all pairs’ similarities is computa-
494 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

tionally challenging even for our networks of a few million nodes since the
entire quadratic size similarity matrix would occupy several Terabytes. No-
tice that the experiments of Liben-Nowell and Kleinberg [60] were conducted
on much smaller data. As one possibility we may calculate similarity only
for existing edges. The resulting scheme downweights unimportant edges
but is unable to add “uncaught contacts” to the network. As a possible
solution to finding the potential strong relationship between pairs of nodes
not connected by an edge, we may find all pairs with weight above a given
threshold by fingerprinting techniques; these techniques will however be
specific to the given similarity measure.

2.1. Neighborhood based methods

The first broad class of measures for the strength of the connection between
two nodes u and v depend on the strength of the overlap between the
neighborhood of u and v. Next we define the measures of cocitation,
Jaccard, Adamic/Adar and cosine similarities.
The cocitation or common neighbors coc (u, v) is defined as the number
of common in-neighbors of u and v. This measure turned out most effec-
tive for Web spam classification [14]. By the notation of edge directions,
coc− (u, v) denotes the bibliographic coupling (nodes pointed to by both u
and v) and coc↔ (u, v) is the total number of (undirected) common neigh-
bors. We may also define a variant of cocitation that is downweighted by
degree as coc (u, v)/d(u) · d(v).
The Jaccard coefficient Jac (u, v) is the ratio of common neighbors within
all neighbors. If we let Γ(u) denote the neighbors of u, then
   
Jac (u, v) =  Γ(u) ∩ Γ(v) / Γ(u) ∪ Γ(v)

For a weighted graph we may divide the total weight of edges leading to
common neighbors by the total weight of edges from u and v. Unfortunately
this measure does not correlate the pairs of weights ux and vx for common
neighbors x. Due to this problem we observe a particularly poor perfor-
mance in the case when we have a single strongly related neighbor x of u
and y of v and the Jaccard similarity is 0. If edges uy and vx receive an “ac-
cidental” low weight, the Jaccard coefficient however immediately becomes
very high while the actual similarity remains very low.
Cosine similarity fixes the above problem of the Jaccard coefficient.
We consider the row of the adjacency matrix corresponding to node u as
Telephone Call Network Data Mining: A Survey with Experiments 495

vector u. The cosine similarity of nodes u and v is simply cos(u, v) = uT v.

We may similarly define cos− (u, v) over the transpose matrix and cos↔ (u, v)
over the sum.
Adamic and Adar [1] define a measure that downweights high degree
common neighbors as they may occur simply by chance. The Adamic/Adar
measure is defined as
  
AdamicAdar (u, v) 1/ log  Γ(z) .
z∈Γ(u)∩Γ(v)

Simple neighborhood based edge weighting schemes already pose com-

putational challenges for large networks since filling the quadratic size sim-
ilarity matrix is infeasible. Next, we describe the min-hash fingerprint of
Broder et al. [21] to identify all pairs with weight above a given threshold.
Based on the min-hash fingerprint and embedding, more complex approxi-
mation of related measures such as cosine is described in [24].
The fingerprint of node u under a random permutation1 π of all nodes
is defined as the minimum neighbor of u in the ordering of π:
 
fingerprintπ (u) = min π(u ) : u ∈ Γ(u) .

For two nodes u and v the fingerprints coincide if and only if the minimum
of Γ(u) ∪ Γ(v) under π belongs to Γ(u) ∩ Γ(v), hence the probability of
this event is equal to the Jaccard similarity of the nodes. By generating
a sufficiently large number of fingerprints (in practice 100–10000) we may
approximate Jac (u, v) as the fraction of the fingerprints of u and v that
coincide.

2.2. Multi-step propagation: Methods based on path ensembles

Advanced node similarity measurement methods are capable of using a part

or all of the entire path ensemble connecting the given pair of nodes and not
just the neighborhood that corresponds to length 2 paths. In this section
we briefly introduce such methods and the efficient algorithms [38, 77] for
approximately computing them.
1
In fact π does not have to be random: the weaker so-called min-wise independence
requirement suffices.
496 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

Path ensemble measures became widespread with the success story of

Google’s PageRank [19, 71] and other hyperlink-based quality measures
[52, 18]. Since its introduction in 1998, PageRank remains he prominent
example of a path ensemble measure as it is defined as a certain multi-step
generalization of the degree defined below. In fact, PageRank is best known
as a quality measure based on the recursive reasoning that the importance
of a node is high if it is pointed to by several important nodes. Personalized
PageRank, a variant of PageRank dating back to the original paper of Page
et al. [71], is however capable of measuring the strength of the connection
between a node or a weighted set of nodes and another node.
Next we introduce notation for (personalized) PageRank. Let A denote
the stochastic matrix corresponding to the random walk on the network, i.e.

1/ outdeg (i) if host i points to j,
Aij =
0 otherwise.

The PageRank vector p = (p1 , . . . , pN ) is defined as the solution of the

following equation [19]:


N
(1) pr = (1 − c) · pv Avu + c · r,
v=1

where r = (r1 , . . . , rN ) is the teleportation distribution and c is the

teleportation probability with a typical value of c ≈ 0.15. We get the
PageRank if we set all ri to 1/N ; for general r we get PageRank personalized
on r. If r = χ(w) consisting of all 0 except for node w where χw (w) = 1, then
we personalize on the single vertex and let PPRw denote the corresponding
vector.
As we will see, variants of the PageRank of u personalized on v de-
fine similarity measures of u and v. These values are however even more
expensive to compute for all u, v than the neighborhood based measures
of the previous subsection. Below we describe two reformulations of the
PageRank equation that yield scalable approximation algorithms for sev-
eral related measures. The Monte Carlo simulation procedure of Fogaras
and Rácz [39] is a general method to estimate random walk based path en-
semble measures. The Dynamic Programming algorithm [48] gives rise to
approximation algorithms [77] that can also be used for estimating several
weighted neighborhood values similar to those of [12].
Telephone Call Network Data Mining: A Survey with Experiments 497

The first PageRank reformulation was noticed independently by [37, 48].

The (personalized) PageRank of a vertex is equal to the probability of a
random walk terminating at the given vertex where the length is given by a
geometric distribution: we terminate at step t with probability c·(1 − c)t . To
justify this, notice that PPRw (and PageRank in general) can be rewritten
as a power series
∞

(2) PPRw = χ(w) · c(1 − c)t · At .
t=0

The term χ(w)At corresponds to a random walk of length t starting at w

and c · (1 − c)t is the probability of termination. The above equation also
explains why PPRw (u) is a path ensemble based similarity of u and w: we
enumerate all paths from w to u by giving exponentially decreasing weight
to long paths.
As described by Fogaras and Rácz [39], equation (2) can be approxi-
mated by randomly generating paths with length according to the geometric
distribution c(1 − c)t . They empirically observe that 1000 samples suffice
for a good quality approximation even in large graphs. For algorithmic
details and error analysis we refer to [39].
A second, equivalent reformulation of the path summing formula (2) is
the Decomposition Theorem proved by Jeh and Widom [48] stating that a
node’s personalized PageRank vector is expressible with the average per-
sonalized PageRank vector of its out-neighbors giving extra weight to the
node itself:

(3) PPRu = cχu + (1 − c) · PPRv /d+ (u).
v : (uv)∈E

As observed by Sarlós et al. [77], the above equation is the right choice for
computing all PPR values above certain threshold. The other altenative
is the path summation formula (2); however there we accumulate all error
when entering a large in-degree node and hence we must compute partial
results fairly exact. The dynamic programming equation (3) in contrast
averages all partial results into a new PPRu and because of averaging we do
not amplify error at large in-degrees. In particular we may safely discard
all partial PPRu (vi ) values below threshold for further computations since
the total error will remain below the threshold in (3).
Given the path summation reformulation (2) of personalized PageRank
we may define several variants of weighting neighbors at distance k. We may
498 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

define reachability and exact reachability by dk (u, v)reach = 1 if v is reach-

able from u by a walk over k edges, 0 otherwise, respectively dkexact (u, v) = 1
if v is reachable from u in exactly k steps and over no shorter paths, 0 other-
wise. We may use the number and the weighted number of such walks in the
definition: dknum (u, v) is the number of walks over k edges that reach from
u to v and dkwnum (u, v) is the probability of reaching v when starting at u
and at each step choosing a random neighbor with probability proportional
to the outgoing edge weights.
One example of the generalization of path summation is the histori-
cally earliest path ensemble measure of Katz [51] dating back to the fifties
defined as
∞

Katzβ (u, v) = β t · dknum (u, v);
t=1

weighted Katz measure arises if we replace dknum by dkwnum . These measures

can be approximated by both of the above methods.
More complex path ensemble measures arise as the multi-step variants of
cocitation and the Jaccard coefficient. Jeh and Widom [47] define SimRank
as a multi-step generalization of downweighted cocitation as follows:

Sim(k) (u1 , u2 )
⎧   
⎨(1 − c) · Sim(k−1) (v1 , v2 )/ d− (u1 )d− (u2 ) if u1 = u2
=
⎩1 if u1 = u2 .

In an alternative formulation [77] SimRank equals the total weight of pairs

of walks

v1 = w0 , w1 , . . . , wk−1 , wk = u

v2 = w0 , w1 , . . . , wk−1


, wk = u

that both end at u and one of them comes from v1 while the other one from
v2 . The weight of the pair of walks is the expected (1 − c) meeting distance
as defined in [47]:
 
(4) (1 − c)k / d− (w1 ) · · · d− (wk ) · d− (w1 ) · · · d− (wk ) ;
Telephone Call Network Data Mining: A Survey with Experiments 499

notice that we get cocitation back for k = 1. Fogaras and Rácz [38] describe
XJaccard as the weighted sum of Jaccard coefficients of the distance k
neighborhoods as follows:

XJac (u, v) = (1 − c)k Jac(k) (u, v)

where Jac(k) (u, v) is the Jaccard similarity of the distance k neighborhood of

u and v. These measures can be approximated in a similar way of PageR-
ank by path sampling in equation (4) [38]; more complex space optimal
algorithms are also described in [77] for certain SimRank variants.

2.3. Link prediction experiments

In order to illustrate the strength of the methods from the previous two
subsections, we set up the following link prediction experiment. We compute
the similarity measures based on the first four month (training period) of
the large data (Fig. 1). We use these similarity measures as a prediction
for the next four month period (test period). Given a threshold value, we
measure precision and recall as
 
 {edges above threshold} ∩ {actual edges at months 5–8}
Precision =   ;
 {edges above threshold}
 
 {edges above threshold} ∩ {actual edges at months 5–8}
Recall =   .
 {actual edges at months 5–8}

We also introduce weighted recall to bias quality measures towards correctly

identifying heavy weight edges and penalizing less for low weight ones. By
letting we denote the weight of an edge in the second (test) four month
period we let

{we : e has weight above threshold}

WRecall =
.
e we

Results of the link prediction experiment are shown in Fig. 2 in terms of

precision–recall (top) and precision–weighted recall (bottom) curves. The
curves are obtained by varying the threshold. Weighted recall is significantly
higher in all cases, indicating that heavy weight edges are easier to predict;
the relative order of the quality of the predictions however remains the same
500 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

Fig. 2. Precision-recall curves of the link prediction experiment trained on the first
4 months and tested on the last 4 months of the large graph. Top graphs:
precision–recall curves. Bottom graphs: precision and weighted recall curves.
Precision is over the vertical axis in both cases. Left graphs: curves corresponding to
measures that exclude the (logarithmic) weight of the training period. Right graphs:
curves for measures that all include a logarithmic edge weight term.

for both curves. Similarly to the findings of Liben–Nowell and Kleinberg

[60], variants of Katz performed best among path ensemble measures while
cosine among neighborhood measures. Due to the limitations of the visu-
Telephone Call Network Data Mining: A Survey with Experiments 501

alization, we could not place all variants in Fig. 2. We only show Jaccard
and Adamic/Adar in addition to the above neighborhood based measures
as well as preferential attachment, a trivial baseline method defined by the
product of the degrees of the two node in question.
A key difference in our experiment compared to [60] is that we predict all
edges in the test period, not just new edges. This distinction is visible when
comparing the left and right graphs of Fig. 2. The left hand side graphs
show lower quality because those u–v similarity measures do not take into
account whether u and v are connected by an edge or not. The measures on
the right hand side count the existence of an edge between u and v either as
a part of the Katz measure, or else directly add it to neighborhood measures
(common neighbors, cosine, Jaccard). Best performance is obtained when
the logarithm of the edge weight (in time duration) is added; these measures
are shown in Fig. 2, right.
Since in our task we also have to predict edges that already existed in
the training period, aggregated time duration of an edge turns out to be a
very strong predictor in itself. This measure is outperformed only for high
recall ranges when neighborhood based measures are capable of identifying
additional new edges in the network. In this range Katz performs very well.
We also draw attention to the importance of edge weights. In the graphs
we use weights in two places. Time duration values are on one hand entries
in the vectors that define the weighted cosine measure; on the other hand
they modify the path probabilities in weighted Katz. Weighted cosine turns
out to be the best neighborhood based measure while weighted Katz is only
defeated by logarithmic weight plus weighted cosine at certain ranges of
recall.
To draw a conclusion, we have surveyed a number of node similarity
measures based on both neighborhood overlap and entire path ensembles
and sketched some scalable algorithms and techniques to approximate them.
Having analyzed precision-recall curves of a link prediction experiment, we
have observed best performance for weighted cosine and Katz similar to the
findings of [60]. In the next sections we will also use these measures as
alternative weights to the edges as an input to further processing.
502 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

3. Network Topology: Navigation in the Small World

The so-called “small world” phenomenon was first observed in social net-
works by Milgram [63] in the United States and found an average distance
of six steps, later referred to as “six degrees of separation”. Social and other
networks exhibit low diameter as demonstrated by several results. As an ex-
ample, in [4] the diameter of the World Wide Web is measured. As another
one, in [25] the low diameter of a wide class of networks obeying degree dis-
tribution constraints such as power law distribution is proved. Telephone
call networks fall into this category; in this section we observe low distances
and efficient navigation in our graphs.
The first graph model explaining the small world phenomenon is de-
scribed by Watts and Strogatz [82] [54, 53] the model that uses local infor-
mation only in each step. Notice that Milgram’s experiment [63] did not
only show low average distances but also a capability of network nodes to
find these paths based solely on their local information on the network. In
fact information is not entirely local, we must have at least certain global
information on the target node. It is easy to see that if all we can determine
whether a given node is the target or not, we have no choice other than to
perform a random walk until we reach the target, an algorithm that visits
each node several times on average. In Milgram’s experiment among others
the geographic location of the target was given to the node. By using this
information, an algorithm may for example select the current node’s neigh-
bor geographically closest to the target, a clear advantage over a random
walk.
Kleinberg’s celebrated small world model [54] describes the following
network with a geographic path finder algorithm. We obtain a small world
graph in a d dimensional attribute space by placing nodes in a d-dimensional
grid, connecting all pairs within a constant distance and adding long range
contacts with probability proportional to r−d where r is the distance. In a
recent extension, Kumar et al. [56] observe over the network of bloggers that
nodes are not distributed evenly over the grid and r −d should be replaced
by t(r)−1 where t(r) is the number of users of distance at most r.
In the following we show measurements for the distribution of the dis-
tance between pairs of contacts and fit the results to Kleinberg’s model.
We conduct this experiment over the graph with nodes formed by the users
in the large data of Fig. 1; in Fig. 3 we show the geographic location of
Telephone Call Network Data Mining: A Survey with Experiments 503

Fig. 3. Left: The geographical distribution of the customer network within Hungary
with axes showing latitude and longitude. Right: the western region avoiding the
capital Budapest at latitude 47 ◦ 28’ 19” N and longitude 19 ◦ 03’ 01” E selected for the
experiment.

these users. Since the capital Budapest locates roughly 1/3 of all nodes
that are hence of geographic distance 0, we have selected the western region
of near 750,000 users shown in the right of Fig. 3. By Kleinberg’s model
the distribution of the distance of a given node from its neighbors is inverse
polynomial; for a two-dimensional area as in Fig. 3 the exponent is −2. Our
measurements shown in Fig. 4 justify this model as follows. On the left
hand side we see a noisy behavior due to large cities and in particular the
capital; the distribution fits slightly better to r −1 but the quality is poor.
The quality of the fit however significantly improves if we remove the effect
of the capital: the western region with no city containing 50,000 or more
users fits r −2 very well.
We also describe a path finder experiment where for each intermediate
node we greedily select the neighbor geographically closest to the target
node. Unfortunately we have no information other than location, hence
we say that the path terminates if it reaches the settlement of the target.
In Fig. 5 we show the number of steps required by this greedy routing
algorithm to find the target for a set of 1,000,000 randomly selected pairs
of users. We see a fast exponential decay in the number of paths required
beyond distance 10. The distances we measure are close to the “magical”
six in Milgram’s experiment despite of the fact that our network is much
smaller and we are satisfied with simply reaching the settlement location
of the node. Notice however, that we are unable to use information other
than location in intermediate steps and thus for example we never move to
a node within the same settlement. In a practical scenario, in contrast, a
504 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

Fig. 4. Left: The number of edges as a function of the distance between its endvertices.
Right: The same measurement over the western region (Fig. 3) to filter out the effect of
the capital. Both figures contain the data transformed to test linear fit with functions
x−1 and x−2 . Horizontal axes show distance in the coordinates of Fig. 3.

participant may know a neighbor who have relatives near the target area
and the walk may advance very close after a local step that is seemingly
useless in the model. In this sense the participants in Milgram’s experiment
were able to use much richer data for their routing decision.

Fig. 5. The distribution of the greedy distance routing steps for a random pair of
1M users.
Telephone Call Network Data Mining: A Survey with Experiments 505

4. Clustering Algorithms

Clustering covers a wide class of methods to partition a set of data in order to

locate relevant information by grouping and organizing similar elements in
an intelligible way. The purpose of clustering telephone users includes user
segmentation, selection of communities with desired or undesired properties
as e.g. high ADSL penetration or high recent churn rate. In a survey
Newman [67] observes that in social network research “particular recent
focus has been the analysis of communities”.
As a key focus in this section we justify the use of a top-down hierarchical
clustering in social networks that are small world power law graphs, in
particular in telephone call graphs. Small world implies very fast growth
of neighborhoods that strongly overlap; power law yileds high degree nodes
that locally connect a large number of neighboring nodes. Recent bottom-up
alternatives such as the SCAN algorithm [87] suffer from the phenomena of
an extremely large number of tiny communities that cannot be agglomerated
into a meaningful large scale structure.
The main difficulty in partitioning large social networks lies in the abun-
dance of community cores as seen in Fig. 6 or, in another terminology, tightly
knit communities (TKC). Several authors observe difficulties caused by the
TKCs: Lang [58] identifies them as the main obstacle for spectral parti-
tioning; Lempel and Moran [59] investigate hyperlink based ranking on the
Web; very recently [87] identifies hubs that bridge between several TKCs as
the main difficulty in network partitioning.
Several algorithms are proposed to identify community cores. Flake et
al. use network flows [35] or min-cut trees [36]. As a very recently proposed
solution, the method of Xu et al. [87] promises a partitioning of the network
based on the identification of community cores in an agglomerative method
that prefers core nodes and avoids bridges that connect more than one TKC.
In these methods however the cores identified are of small size on the global
scale and cannot yield information on the global structure. Xu et al. [87]
test their method in part on real graphs that are mere few 100 nodes and
in part on graphs generated for particular use for their algorithm. These
generated graphs are based on the construction of [67] that first determines
target clusters and then connects nodes within the same cluster with higher
probability than between two clusters but otherwise independent at random.
It is unclear whether this construction generates graphs with properties
similar to the real world networks.
506 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

Fig. 6. A 82-node (left) and a 317-node (right) subgraph of social networks, both with
two cores and several short tentacles.

We build clusters of telephone users based on the graph formed by their

calls over a period of several months. Since a comprehensive comparison of
clustering algorithms is beyond the scope of this report, we concentrate on
extracting the strength and effects of social contacts from the telecom logs.
In this work we use hierarchical spectral clustering, a widely used method
that fits particularly well with large data sizes. We also show some basic
facts about the behavior of clique percolation [29, 72] and SCAN [87] over
the telephone network: we show the size of the communities identified under
various edge weighting and thresholding schemes.

4.1. Two recent clustering methods

Before turning to our spectral clustering experiments we describe two recent

community detection algorithms and show that although they are reported
to work well for small networks, they are unable to identify homogeneous
large scale structures in our telephone call networks.
The clique percolation algorithm of Derényi and Palla et al. [29, 72] is
described in detail in Chapter 9. In this subsection we will not present
the deep theory behind the method but point interested readers to the full
chapter discussing the topic. Instead in a brief sketch we describe the algo-
rithm with emphasis only to implementational details for very large graphs.
Telephone Call Network Data Mining: A Survey with Experiments 507

Note that in [72] measurements over only a few thousand node graphs are
described. In another closely related publication on clique percolation [73]
a telephone call graph with over 4 million users is used as input and the
time evolution of the clusters is analyzed. In that paper however no details
are given on how to implement the algorithm for such a large scale problem,
the communities analyzed have less than 1000 users, and no global analysis
of the community size distribution is given. In order to easier relate all re-
sults in this paper we hence show some additional measurements on clique
percolation in this subsection.
Clique percolation grows communities from k-cliques as building blocks.
A k-clique is a complete subgraph over k nodes; we call two such cliques
connected if they share k−1 nodes. Clique percolation identifies the maximal
connected components of k-cliques as possibly overlapping clusters.
We describe a possible implementation of the clique percolation algo-
rithm we use that is based on the modification of the APRIORI frequent
itemset mining algorithm [2]. In one step described in Algorithm 1 we enu-
merate all (k + 1)-cliques for increasing values of k = 0, 1, . . . and store
them as lexicographically ordered sets in a trie. Given a (k + 1) element set
U ∪ u ∪ v with |U | = k − 1 and u and v of index higher than all elements in
U , this set is a (k + 1)-clique if and only if U ∪ u and U ∪ v are k-cliques and
there is an edge between u and v. This condition can easily checked since
U ∪ u and U ∪ v are both children of a level k − 1 element U of the trie of
k-cliques.

Algorithm 1. Construction of the trie O of (k + 1)-cliques from the trie I

of k-cliques.
O ← empty trie
for all sets U in the trie I on level k − 1 do
for all pairs of nodes u and v such that U ∪ u and U ∪ v ∈ I do
if u and v are connected by an edge then
add U ∪ u ∪ v to O
return O

Clique percolation in our experiment suffers from two problems due to

the scale of the input data. Firstly for dense graphs there are simply too
many cliques to enumerate: in the top table of Fig. 7 we see that, when run
on the small graph (Fig. 1), we have to discard the largest degree nodes. In
this mere 74,000 node graph there are 39M 3-cliques without filtering; for
4-cliques we ran out of memory at over 100M enumerated.
508 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

Fig. 7. Top: The size of the cliques for the small and large graphs of Fig. 1. For the
small graph the highest degree x nodes are discarded for different values of x.
Bottom: properties of clusters over the large graph with k = 3, 4 and 5 and the
histogram of their size distribution.

Secondly, it turns out to be hard to balance between an extreme large

number of small cliques that do not only pose computational challenges but
also connect most of the graph into a single cluster and a low number of
larger cliques that leave most of the nodes isolated or in very small clusters.
As seen in Fig. 7, bottom, in our large graph for k = 5 most of the nodes
remain in isolation while for k = 3 we are left with a giant component. The
best k = 4 choice places slightly more than half of the nodes in some cluster,
although many of these clusters are very small as seen in Fig. 7, bottomleft.

As another proposed solution, the assumption of Xu et al. [87] is that

there exist hub vertices in a network that connect or, in their terminol-
ogy, bridge many clusters. Therefore they define the SCAN algorithm that
selects pairs of vertices with a concentration of common neighbors as can-
didate intra-cluster nodes limited by parameter ε. Hubs, as opposed to
intra-cluster nodes, are then characterized by the distraction of neighbors.
Finally cores are formed by nodes that have at least μ neighbors within the
core.
Telephone Call Network Data Mining: A Survey with Experiments 509

The key step in the SCAN algorithm is the selection of edges between
pairs of nodes whose neighborhood similarity is above a threshold ε. In the
original algorithm of Xu et al. [87], with Γ(u) denoting the neighbors of u,
the similarity is measured as
  !  
σ(u, v) =  Γ(u) ∩ Γ(v) /  Γ(u)  Γ(v) .

For power law graphs, in particular for the Web graph in our experiments,
however the running time for computing σ(u, v) is very large due to the
huge neighborhood sets Γ(u) involved. Hence, we use the Jaccard similarity
   
Jac (u, v) =  Γ(u) ∩ Γ(v) / Γ(u) ∪ Γ(v)

that we approximate by 100 min-hash fingerprints [20] as described in

Section 2.1.
The SCAN Algorithm 2 (modified to use Jaccard similarity) proceeds
as follows. First it discards edges that connect pairs of dissimilar nodes
below threshold ε; these edges may bridge different dense regions [87]. Then
nodes with more than μ remaining edges are considered as community
cores. Finally, connected components of cores along remaining edges are
augmented by neighboring non-core nodes. The resulting components C
may overlap at these augmented vertices; these vertices are called hubs in
[87] since they provide connectivity across different communities.

Algorithm 2. Modified SCAN.

input: ε: similarity threshold of neighbors within same core, μ: size
threshold of neighborhood within core
output: list of communities
for all edges uv do
compute approximate Jac (u, v) by min-hash fingerprints
E  ← {(uv) : σ(u, v) ≥ ε}
V  ← {u : degE  (u) ≥ μ}
compute the connected components C of V  with edges E 
for all components C of C do
Add all vertices to C that are connected to C by edges of E 
return C

Tests with various parameter settings of the SCAN algorithm over the
large graph (Fig. 1) are shown in Fig. 8. As an overall evaluation we observe
that SCAN is unable to detect communities of size beyond a few tens in our
510 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

network. The better the parameter setting we use, the more communities
are found in the size range around 20 nodes. In more detail we compare
the Jaccard similarity as described in Algorithm 2 and the weighted cosine
similarity; due to computational constraints the latter is only computed for
the existing edges of the network. Even in this weaker setting weighted co-
sine identifies more meaningful communities. Best performance is obtained
by the smallest possible μ = 2 and small ε. If we increase either μ (Fig. 8,
left) or ε (Fig. 8, right), the number of communities identified decreases at
all size ranges.

Fig. 8. The distribution of the community sizes identified by the SCAN algorithm. Left:
the effect of varying μ for a low ε = .3 similarity value. Right: the effect of varying ε
for the weakest μ = 2 core size bound. Both figures contain the weighted cosine input
graph with the best parameter settings in addition to Jaccard as in Algorithm 2.

4.2. Cluster quality measures

Next we define the graph and sociodemographic based quality measures we

use for evaluating the output of a clustering algorithm. Let there be N
users with Nk of them in cluster Ck for k = 1, . . . , m. The
cluster ratio
is the number of calls between different clusters divided by i=j Ni · Nj .
The weighted cluster ratio is obtained

by dividing the total weight of edges
between different clusters by i=j wij Ni · Nj , where wij is the total weight
of the edges between cluster i and j. Modularity, a measure known to suit
social networks well [87] is defined as follows:
&  2 '
 d(Cs , Cs ) d(Cs , Cs )
(5) Q= − ,
M 2M
clusters s
Telephone Call Network Data Mining: A Survey with Experiments 511

where M is the total weight of the edges and d(X, Y ) is the weight of the
edges with tail in X and head in Y . Since modularity is not balanced by
the cluster size, we use normalized network modularity [80], defined as
& 2 '
 N d(Cs , Cs ) d(Cs , Cs )
Qnorm = − ,
Nk 2M M
clusters s

where M is the total weight of the edges and d(X, Y ) is the weight of
the edges with tail in X and head in Y . We remark that the authors
in [80] negate nomalized modularity compared to modularity; we stick to
their notation and use negative values of normalized modularity. In our
experiments normalized modularity turned out to be instable and we suspect
it may not be an appropriate measure for cluster quality.
Telephone users as nodes have rich sociodemographic attributes beyond
graph theory. We may measure clustering quality by the entropy and purity
of the geographic location or other external property within the cluster. By
using the notation of the previous subsection, let Ni,k denote the cluster
confusion matrix, i.e. the number of elements in cluster k from settlement i
and let pi,k = Ni,k /Nk denote the ratio within the cluster. Then the entropy
E and purity P are defined as

−1  Nk  1 
E= pi,k log pi,k and P = max Ni,k ,
log m N N i
k i k

where the former is the average entropy of the distribution of settlements

within the cluster and the latter measures the ratio of the “best fit” within
each cluster.

4.3. Spectral clustering: a brief history

Spectral clustering refers to a set of heuristic algorithms, all based on the

overall idea of computing the first few singular vectors and then clustering in
a low-dimensional (in certain cases simply one-dimensional [34]) subspace.
The standard Quadratic Integer Program for graph bisection is 14 xT Lx,
where L is the graph Laplacian and x is the ±1 cut indicator vector. In
order to avoid the trivial cut with all nodes on one side, we have xT e = 1
where e is a vector of all ones. When relaxing x to arbitrary real values
between −1 and +1, the optimum is known to be the second eigenvector
512 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

(the Fiedler vector) of L [34]. When however we relax indicator values to

be arbitrary n-dimensional vectors of norm 1, the resulting optimization
problem can be solved by semidefinite programming [58].
Variants of spectral partitioning dating back to the 1970’s fall into two
main branches as described in [6]. The first branch is initiated by the
seminal work of Fiedler [34] who separates data points into the positive
and negative parts along the principal axes of projection. His original idea
uses the second singular vector, the so-called Fiedler vector; later variants
[11, 5] use more vectors. Hagen and Kahng [45] are perhaps the first to use
the second smallest eigenvalue for graph partitioning of difficult real world
graphs.
The second branch of hierarchical spectral clustering algorithm divides
the graph into more than two parts in one step. While the idea of viewing
nodes as d-dimensional vectors after projecting the adjacency matrix into
the space of the top k singular vectors is described already by Chan et
al. [23], much later Zha et al. [88] introduce the use of k-means over the
projection. In our recent work [57] we compare the two branches and find
the superiority of using several spectral directions; in the rest of this section
we restrict our attention to this class of algorithms.
The formation of the input matrix to SVD computation from the de-
tailed call list strongly affects the outcome of clustering. Kannan et al. [50]
suggest modeling the input as a similarity graph rather than as a distance
graph, raising the question of how to turn the call information including the
number, total duration and price between a pair of callers into a similarity
measure. In addition to various ways of using cost and duration the node
similarity measures of Section 2 may also be used to reweight the input
graph.
Experiments on spectral graph partitioning either use the unweighted
or weighted Laplacian as the input to the singular value decomposition
procedure. The Laplacian is defined as D − A, where D is the diagonal
matrix whose i-th entry is the total edge weight at node i and A is the
adjacency matrix. The weighted Laplacian D−1/2 AD −1/2 is first used for
spectral bisection in [79, 31]. The Laplacian arises as the relaxation of the
minimum ratio cut [45]; the weighted Laplacian appears in the relaxation
of the normalized cut [79] and the min-max cut [31] problems. Weighting
strategies are discussed in more detail in [6, and references therein], and
Alpert and Kahng [5] empirically compared some of them. Unfortunately
these results deal with the netlist partitioning problem only. Since netlists
Telephone Call Network Data Mining: A Survey with Experiments 513

are hypergraphs, we may not directly use the findings of [5] but have to
remember the importance of comparing different graph weighting strategies.
When clustering the telephone call graph, the main practical problem
arises when the graph or a remaining component contains a densely con-
nected large subset. In this case spectral clustering often collects tentacles
loosely connected to the center [25] into one cluster and keeps the main
dense region unsplit [58]. While even the optimum cluster ratio cut might
have this structure, the disconnected cluster consists of small graph pieces.
We observe that each of these small pieces has a strong connection to certain
areas within the dense component and they would more naturally belong
to different subclusters of the dense component instead. In addition, a dis-
connected graph has a multiple maximal eigenvalue, meaning that we must
compute eigenvectors separately for each connected component. However,
if we treat each connected component as a separate cluster, we obtain an
undesired very uneven distribution of cluster sizes.

4.4. Small component redistribution heuristics

The key in using spectral clustering for power law graphs is our small cluster
redistribution heuristics described in the next subsection. After computing
a k-way split, we test the resulting partition for small clusters. First we try
to redistribute nodes to make each component connected. This procedure
may reduce the number of clusters. In a degenerate case we may even be
left with a single cluster; in this case the output is rejected and clustering
fails.

Algorithm 3. redistribute (C1 , . . . , Ck ): Small cluster redistribution

for all Ci do
Ci ← largest connected component of Ci
if |Ci | < limit · |C1 ∪ . . . ∪ Ck | then
Ci ← ∅
Outlier = (C1 − C1 ) ∪ . . . ∪ (Ck − Ck )
for all v ∈ Outlier do
p(v) ← j with largest total edge weight d(v, Cj )
for all v ∈ Outlier do
Move v to new cluster Cp(v)
return all nonempty Ci
514 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

Our subroutine to reject very uneven splits (Algorithm 3) proceeds as

follows. Given a split of a cluster (that may be the entire graph) into at
least two clusters C1 ∪ . . . ∪ Ck , we first form the connected components of
each Ci and select the largest Ci . We consider vertices in Ci − Ci outliers.
In addition we impose a relative threshold limit and consider the entire Ci
outlier if Ci is below limit.
Next we redistribute outliers and check if the resulting clustering is
sensible. In one step we schedule a single vertex v to component Cj with
d(v, Cj ) maximal, where d(A, B) denotes the number of edges with one end
in A and another in B. Scheduled vertices are moved into their clusters
at the end so that the output is independent of the order vertices v are
processed. By this procedure we may be left with less than k components,
and we will have to reject clustering if we are left with the entire input as a
single cluster. In this case we either try splitting it with modified parameters
or completely give up forming subclusters.

Algorithm 4. k-way hierarchical clustering

while we have less than cnum clusters do
A ← adjacency matrix of largest cluster C0
Project D−1/2 AD−1/2 into first d eigenvectors
For each node i form vector vi ∈ Rd of the projection
vi ← vi /vi 
(C1 , . . . , Ck ) ← output of k-means(v1 , . . . , v|C0 | )
Call redistribute(C1 , . . . , Ck )
Discard C0 if C0 remains a single cluster

In our benchmark implementation we give k, the number of subclusters

formed in each step, d, the dimension of the SVD projection and cnum,
the required number of clusters, as an input. Algorithm 4 then always
attempts to split the largest available cluster into k ≤ k pieces by k-means
after a projection onto d dimensions. Note that k-means may produce less
than the prescribed number k of clusters; this scenario typically implies the
hardness of clustering the graph. If, after calling small cluster redistribution
(Algorithm 3), we are left with a single cluster, we discard C0 and do not
attempt to split it further.
In our real life application we start out with low values of d and increase
it for another try with C0 , whenever splitting a cluster C0 fails. In this case
we may also decrease the balance constraint.
Telephone Call Network Data Mining: A Survey with Experiments 515

Notice the row normalization step vi ← vi /vi , and this step improves
clustering qualities for our problem. We also implemented column normal-
ization, its effect is however negligible.

4.5. Spectral clustering: an experiment

In order to illustrate the importance of the choice of parameters and input

graph weights for clustering as well as the issues on the quality measurement
of the outcome, we present spectral clustering experiments over the large
graph of over 2M nodes and 10M bidirectional edges (for detailed data see
Fig. 1).

Fig. 9. Distribution of the edge weights across different clusters, for a spectral clustering
and a trivial clustering obtained by considering one settlement as one cluster.
The vertical axis contains the total edge weight in seconds and the horizontal axis shows
the number of cluster pairs with the given weight between them.

One may argue whether clustering reveals additional information com-

pared to the settlements themselves as “ground truth” clusters. We give
a positive answer to this question by showing that the distribution of the
total call duration across different clusters is optimal for those obtained by
spectral clustering. In Fig. 9 we form two graphs, one with a node for each
settlement and another with a node for each (spectral) cluster. The weight
of an edge between two such nodes is the total call duration between the
corresponding clusters. We observe that the edge weights follow a power law
distribution in both graphs. The graph obtained by spectral clustering has
a much smaller exponent and the edges across clusters have much smaller
weight, indicating an improved arrangement of weight mass for the spectral
clusters. In fact we use settlement information as an external validation
tool for our experiments and not as ground truth.
We have several choices to extract the network based on telephone calls
between users: we may or may not ignore the direction of the edges and
516 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

weight edges by number of calls, duration or price, the latter emphasizing

long range contacts. First of all we may try to use the total cost or duration
directly as a weight in the adjacency matrix. It turns out that the Lanczos
algorithm converges extremely slowly. While it converges within a maximum
of 120 iterations in all other cases, 900 iterations did not suffice for a single
singular vector computation with raw values. We therefore use 1 + log wij
where wij is either the total cost or duration between a pair of users i and j.
We also investigate a Jaccard and a cosine similarity based weight of
user pairs in line with the remark of Kannan et al. [50] who suggest node
similarities for input weights. These methods yield weights between 0 and 1,
and clustering in the reweighted graph has quality similar to the logarithm
of call cost or duration. For both similarity measures we use 1 + simij to
distinguish non-edges from low-weight edges.
The comparison of various input matrices and the branching factor k
is shown in Fig. 10. Most importantly, we notice the weighted Laplacian
D−1/2 AD−1/2 significantly outperforms the unweighted D − A in all re-
spects. Call length and call cost behaves similarly, and as expected, the
former yields geographically more homogeneous clusters by underempha-
sizing long distance calls, while the latter performing better for the cluster
ratio measure. The logarithm of the price or duration performs very close to
Jaccard reweighting with no clear winner. We also observe that the larger
the k, the better the performance in general, although for some settings
results remain stable, which we investigate in more detail next.
As suggested by [6, 62], more eigenvalues produce better quality cuts.
However the drawback of using more eigenvalues is a slowdown and an
increase in the storage occupied by the vectors. Hence a good balance
between the number d of eigenvalues and the branching k must be chosen.
In Fig. 11 we observe that we should choose k somewhere between 5 and 10
for this graph, and compute somewhat more eigenvectors.
We also notice a noisy behavior of normalized modularity in Fig. 11, bot-
tom right: counterintuitively and opposite of all other measures, best results
are obtained for large branching k regardless of the number of dimensions d
or even preferring low dimensions. Due to this unexpected behavior we sus-
pect normalized modularity may not be an appropriate measure for cluster
quality in social networks.
In Fig. 12 we see the effect of changing limit in Algorithm 3. Recall that
limit bounds the ratio of the smallest cut from the average. If this is very
large (100 in the Figure), we are left with no constraint. If however it is close
Telephone Call Network Data Mining: A Survey with Experiments 517

Fig. 10. Evaluation of various reweighting techniques over the adjacency matrix for
purity (left), entropy (center) and cluster ratio (right) of the arising clusters on the
vertical axis. Curves correspond to combinations of unweighted vs. weighted Laplacian
(NCut for normalized cut relaxation, as opposed to RCut, ratio cut relaxation), length
vs. cost based, and Jaccard vs. logarithmic weight input matrices. Results for k-way
partition with d = 30 for three values k = 2, 4 and 8 are shown on the horizontal axis.

to one, we enforce very strict balancing that deteriorates clustering quality.

The optimal values lie around 4. . . 6, and these values are also optimal for
running time. Very large values slow algorithms down by only marginally
reducing the largest cluster size after the costly SVD computation.
518 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

Fig. 11. Relation between dimensions d (vertical), branching k (horizontal) and quality
(darkness) for purity (top left), entropy (top right), cluster ratio (bottom left) and
normalized modularity (bottom right). The darker the region, the better the clustering
quality except for purity where large values denote good output quality.

We checked that the strict balancing constraints required for efficiency

have no negative effect on cluster quality. This is clearly seen for purity
and entropy in Fig. 12, top. Notice however the unexpected increase of
cluster ratio (middle left) for large values, which is due to the fact that
the densely connected near 600,000 Budapest users could be split only with
liberal balancing conditions as also seen in the table of largest remaining
cluster sizes in Fig. 12, middle right. While splitting Budapest has no effect
on purity or entropy, it adds a large number of edges cut in cluster ratio.
For this reason we repeated the experiment by removing Budapest users to
see no more negative effect of the strict balance constraint on the clustering
quality measures. Besides clustering quality, we also look at how “natural”
are the cluster sizes produced by the algorithms in Fig. 12, bottom. We
observe that intermediate values of limit produce log-normal-like distrib-
utions that have lower variance but are reminiscent to the distribution of
the settlement sizes. Also note that by discarding Budapest users, the huge
clusters trimmed from the diagramdisappear.
Telephone Call Network Data Mining: A Survey with Experiments 519

Fig. 12. Effect of size limits on clustering quality, k = 4 and d = 30 for purity, entropy
(top) and cluster ratio (middle left). The size of the largest and second largest
remaining cluster as well as the largest one after the removal of Budapest lines (middle
right). Distribution of cluster sizes for various parameters of the size limit (bottom).

5. Stacked Graphical Classification

In our final application for mining the telephone call network, we rely on
the strength of the connection of pairs of users to classify churn, i.e. those
users who discontinue their use of the service or, in other words, quit the
network. Churn prediction by machine learning methods on real data is
discussed by [83, 7, etc.], but these results do not exploit neighborhood
information embedded in the call graph.
520 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

Semi-supervised learning, a new field of machine learning surveyed e.g.

in [89] also exploits information from unlabeled data for learning. We focus
on a particular semi-supervised method called stacked graphical learning, a
meta-learning scheme in which a base learner is augmented by expanding
the features of one node with predictions on other related nodes in a graph
introduced recently by Kou and Cohen [55]. The methodology is used with
success for Web spam detection by Castillo et al. [22] who use the average
label of the neighbors as a new feature for the classifier.
We focus on the applicability of stacked graphical learning for classifying
telephone churn. Our assumption is that the attitude towards churn of a
node in a graph is similar to those of its neighbors. We generate features
by averaging known and predicted labels for similar nodes of the graph by
the measures in Section 2 to propose a set of churn classification methods
that combine graph based similarity to labeled nodes [14]. We compare
various measures based on neighborhood as well as path ensemble including
co-citation, cosine and Jaccard similarity described in detail in Section 5.2.
For the purposes of evaluation we consider these algorithms separately, by
performing one classification experiment for each feature.
Churn classification uses customer information (price package, time since
in service etc.) as well as traffic aggregates in various call zones and
directions. We use one year call detail record of the small data set (Fig. 1)
and all customer information up to a given time. The classification target
consists of users who will leave the service in the fourth month “in future”
(in a time period with no information available for the classifier). Due to
the sparsity of positive instances (below 1% churn in a month) and a large
amount of churn explained by external reasons such as the customer moves,
churn classification is a hard task with a baseline reaching F = 0.08. We
improve this baseline to 0.1 by means of stacked graphical learning. In the
industrial practice the goodness of the churn classification is measured by
the recall of the top list of 10% of the customers, i.e. they are willing to
involve a maximum of 10% of their customers in direct marketing campaigns
and want to maximize the potential churn reached. In this sense our baseline
classification has a recall of 40.8%, improved to 47% by stacked graphical
learning.
Telephone Call Network Data Mining: A Survey with Experiments 521

5.1. Label propagation: a broader outlook

Spreading certain information from known nodes with the help of network
similarity measures has several application besides churn: we may propa-
gate node labels such as trust and distrust in the case of Web spam [42] or
topics when classifying web documents. Several results [75, and the refer-
ences therein] demonstrate that classification accuracy can be significantly
increased by taking into account the class labels assigned to neighboring
nodes. These results in general suggest cocitation as the best similarity
measure for information spreading: this is the case for Web spam [14, 16]
as well as Web page topics [75].
Web spam is perhaps the most important application of a classification
problem for network objects: identifying and preventing spam is cited as one
of the top challenges for web search engines in [46, 81]. The target of Web
spammers is the search engine ranking scheme that, for all major search
engines, include sophisticated forms of anchor text and link analysis. Web
spam reaches the same level of sophistication: content as well as linkage
is manipulated in complex ways [43]. Spam hunters use a variety of both
content [33, 68] and link [44, 32, 85, 15, 14, 74, 86] based features to detect
Web spam. Recently, Catillo et al. [22] measured the combination of the
state of the art methods combined also with graph stacking.
Label propagation first appeared isolated from semi-supervised learning
simply as a ranking method. Trust and distrust propagation in trust net-
works originates in Guha et al. [41] for trust networks. In an even earlier
result [49], EigenTrust is defined as PageRank with weights that are trust
values. The first results extended initial trust or distrust judgments over
a small set of seed pages or sites to the entire web, such as trust [44, 86],
distrust [74, 32] propagation in the neighborhood or their combination [85]
as well as graph based similarity measures [14]. These methods are either
based on propagating trust forward or distrust backwards along the hyper-
links. Propagation is based on the idea that honest pages predominantly
point to honest ones, or, stated the other way, spam pages are backlinked
only by spam pages.

5.2. The stacked graphical learning framework

Stacked graphical learning consists of a two-phase procedure that can possi-

bly be iterated. First, the base classifiers are built to give a prediction p(u)
522 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

for all the unlabeled nodes u. Next, for each node v we construct new fea-
tures based on the above value p(u) for all neighbors u of v and the weight
of the connection between u and v. These new features are hence formed
by propagating the prediction towards network neighbors. All variants of
this scheme perform averaging over certain graph whose edges are possibly
weighted by node similarity values.

F-measure graph stacking method

×1000 none d coc coc− coc↔ Jac Jac− Jac↔ cosine PPR
iterations 1 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2
All 086 102 063 052 079 088 102 083 067 065 059 066 097 089 084 065 092 087
Nonchurn 161 155 141 142 197 200 114 121 254 265 153 147 175 158 267 280 277 257
sampled

Table 1. 1000 times the F-measure shown for different data sets and edge weights.

In our experiment over the small graph (Fig. 1) we form features based on
aggregated call cost duration in different cost segments, including daytime
and off-peak, weekday and weekend as well as local and different long-
distance call volumes. Part of the users perform calls via an alternate
provider by dialing a prefix, and these calls are aggregated similarly for
each user. We also use the pricing package information that also includes
a distinction of company and family lines as well as the start date of the
service usage. For a time range of 12 months, after aggregating calls between
the same pairs of callers we obtained a graph with n = 66, 000 nodes and
m = 1, 360, 000 directed edges.
We use the cost sensitive C4.5 implementation of the machine learning
toolkit Weka [84] with bagging. Since the running time on the full data
set was over 10 hours, we also compiled a smaller data set where a random
sample of non-churned users was dropped, resulting in 7, 151 users, but we
kept the entire graph. Results for the quality in F-measure are shown in
Table 1.
For a given unknown node u and edge weight function w (that may be
in or out-degree, cocitation, PageRank etc.), our algorithm selects the k
largest weight neighbors of u to generate a new feature, based on churned
nodes in this set. As in [14] we extract four different features from this
set of size k (or possibly less than k if u has less than k neighbors). Each
element v is either classified as churn with weight p(v) or else labeled churn
or nonchurn. In the latter cases we let p(v) be 0 and 1, respectively. Let
s and h be the sum of p(v) and 1 − p(v) for the selected neighbors v or u.
Remember s + h < k is possible if the neighborhood of u is smaller than k.
Telephone Call Network Data Mining: A Survey with Experiments 523

We define
 a weighted
 version s∗ and h∗ as the sum of w(uv) · p(v) and
w(uv) · 1 − p(v) for the same set of v.
We define our features as follows.
• Churn Ratio (CR): fraction of predicted churn, s/(s + h).
• Churn over Non-Churn (CON): total predicted churn divided by non-
churn in the top list, s/h.
• Churn Value Ratio (CVR): churn prediction weighted by similarity
to u, divided by the sum of similarity values under the appropriate
similarity to u, s∗ /(s∗ + h∗ ).
• Churn Value over Non-churn Value (CVONV): same as CON with
nodes weighted by their similarity to u, s∗ /h∗ .
We add the new feature defined by either of the above to the existing ones
and repeat the classification process with the extended feature set. Since
the features are unstable if the neighborhood N (u) is small, we also define
versions CR’, CON’, CVR’, CVONV’ by regressing towards the undecided
1/2 or 1 value:
 ! 
CR = 1/2 + (CR − 1/2) · 1 − 1/  N (u) ;
 ! 
CON = 1 + (CON − 1) · 1 − 1/  N (u) ;
 ! 
CVR = 1/2 + (CVR − 1/2) · 1 − 1/  N (u) ;
 ! 
CVONV = 1 + (CVONV − 1) · 1 − 1/  N (u) .

In our experiments we use CVR’ that also performed best in [16].

Our landline telephone churn classification measurements in Fig. 1
demonstrate that stacked graphical learning in combination with graph node
similarity methods improve classification accuracy. Due to the large number
of possible feature generation methods, the results are by no means com-
plete, but show a very good performance of cocitation and little actual use
of connections longer than two steps in the graph.
524 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

Conclusion

We have surveyed some results of social network modeling and analysis with
illustrations over the call logs of major Hungarian telephone companies with
millions of users, long time range, and sufficiently strong sociodemographic
information on the users. We have analyzed the results of link prediction,
clustering and classification results and compared the performance of various
node similarity measures in these tasks.
We have considered the telephone call network from the point of view of
data mining, a new discipline that builds on results from machine learning,
modeling and algorithmics with emphasis on data scale. Improved hardware
capabilities enable the production of huge data volumes. In order to scale
to the new demands, traditional problems require new algorithms and lead
to new empirical findings.
Our key results also point out to the importance of data mining method-
ologies in network analysis. A generic data mining process starts with the
appropriate choice for data preparation, cleansing and modeling. Given the
results provided by the final algorithm, we may have to reiterate: based on
the results of the first experiments we may have to completely revise our
models. The iterative data mining process cycle has been best illustrated
by selecting the appropriate graph weighting scheme for the data mining
problem: link prediction, clustering or classification.

References

[1] L. Adamic and E. Adar, Friends and neighbors on the Web, Social Networks, 25(3)
(2003), 211–230.
[2] R. Agrawal, H. Mannila, R. Srikant, H. Toivonen and A. I. Verkamo, Fast dis-
covery of association rules, in: U. M. Fayyad, G. Piatetsky-Shapiro, P. Smyth,
R. Uthurusamy (eds.) Advances in Knowledge Discovery and Data Mining, MIT
Press (1996), pp. 307–328.
[3] W. Aiello, F. Chung and L. Lu, A random graph model for massive graphs, in:
Proceedings of the 32th ACM Symposium on Theory of Computing (STOC) (2000),
pp. 171–180.
[4] R. Albert, H. Jeon and A. L. Barabási, Diameter of the world wide web, Nature,
401 (1999), 130–131.
[5] C. J. Alpert and A. B. Kahng, Multiway partitioning via geometric embeddings,
orderings, and dynamic programming, IEEE Trans. on CAD of Integrated Circuits
and Systems, 14(11) (1995), 1342–1358.
Telephone Call Network Data Mining: A Survey with Experiments 525

[6] C. J. Alpert and S. Z. Yao, Spectral partitioning: the more eigenvectors, the
better, in: DAC ’95: Proceedings of the 32nd ACM/IEEE conference on Design
automation, ACM Press, New York, NY, USA (1995), pp. 195–200.
[7] W. H. Au, K. C. C. Chan and X. Yao, A novel evolutionary data mining algorithm
with applications to churn prediction, IEEE Trans. Evolutionary Computation,
7(6) (2003), 532–545.
[8] A. Barabási: Linked, Perseus Publishing (2002).
[9] A. L. Barabási, R. Albert and H. Jeong, Scale-free characteristics of random net-
works: the topology of the word-wide web, Physica A, 281 (2000), 69–77.
[10] A. L. Barabási, R. Albert and H. Jeon, Mean-field theory for scale-free random
network, Physica A, 272 (1999), 173–187.
[11] E. R. Barnes, An algorithm for partitioning the nodes of a graph, SIAM Journal
on Algebraic and Discrete Methods, 3(4) (1982), 541–550.
[12] L. Becchetti, C. Castillo, D. Donato, S. Leonardi and R. Baeza-Yates, Link-based
characterization and detection of web spam, in: Proceedings of the 2nd International
Workshop on Adversarial Information Retrieval on the Web (AIRWeb) (2006).
[13] A. A. Benczúr, K. Csalogány, M. Kurucz, A. Lukács and L. Lukács, Sociodemo-
graphic exploration of telecom communities, in: NSF US-Hungarian Workshop on
Large Scale Random Graphs Methods for Modeling Mesoscopic Behavior in Biolog-
ical and Physical Systems (2006).
[14] A. A. Benczúr, K. Csalogány and T. Sarlós, Link-based similarity search to fight
web spam, in: Proceedings of the 2nd International Workshop on Adversarial In-
formation Retrieval on the Web (AIRWeb), held in conjunction with SIGIR2006
(2006).
[15] A. A. Benczúr, K. Csalogány, T. Sarlós and M. Uher, SpamRank – Fully auto-
matic link spam detection, in: Proceedings of the 1st International Workshop on
Adversarial Information Retrieval on the Web (AIRWeb), held in conjunction with
WWW2005 (2005), to appear in Information Retrieval.
[16] A. A. Benczúr, K. Csalogány, L. Lukács and D. Siklósi, Semi-supervised learning:
A comparative study for web spam and telephone user churn, in: Graph Labeling
Workshop in conjunction with ECML/PKDD 2007 (2007).
[17] B. Bollobás, O. Riordan, J. Spencer and G. Tusnády, The degree sequence of a
scale-free random graph process, Random Struct. Algorithms, 18(3) (2001), 279–
290.
[18] A. Borodin, G. O. Roberts, J. S. Rosenthal and P. Tsaparas, Finding authorities
and hubs from link structures on the world wide web, in: Proceedings of the 10th
World Wide Web Conference (WWW) (2001), pp. 415–429.
[19] S. Brin and L. Page, The anatomy of a large-scale hypertextual Web search engine,
Computer Networks and ISDN Systems, 30(1–7) (1998), 107–117.
[20] A. Z. Broder, On the Resemblance and Containment of Documents, in: Proceedings
of the Compression and Complexity of Sequences (SEQUENCES’97) (1997), pp.
21–29.
526 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

[21] A. Z. Broder, M. Charikar, A. M. Frieze and M. Mitzenmacher, Min-wise inde-

pendent permutations, Journal of Computer and System Sciences, 60(3) (2000),
630–659.
[22] C. Castillo, D. Donato, A. Gionis, V. Murdock and F. Silvestri, Know your neigh-
bors: web spam detection using the web topology, Proceedings of the 30th annual
international ACM SIGIR conference on Research and development in information
retrieval (2007), pp. 423–430.
[23] P. K. Chan, M. D. F. Schlag and J. Y. Zien, Spectral k-way ratio-cut partitioning
and clustering, in: DAC ’93: Proceedings of the 30th international conference on
Design automation, ACM Press, New York, NY, USA (1993), pp. 749–754.
[24] M. Charikar, Similarity estimation techniques from rounding algorithms, Proceed-
ings of the thiry-fourth annual ACM symposium on Theory of computing (2002),
pp. 380–388.
[25] F. Chung and L. Lu, The average distances in random graphs with given expected
degrees, Proceedings of the National Academy of Sciences of the United States of
America, 99(25) (2002), 15,879–15,882.
[26] F. Chung, L. Lu and V. Vu, Eigenvalues of random power law graphs, Annals of
Combinatorics (2003).
[27] F. Chung, L. Lu and V. Vu, Spectra of random graphs with given expected degrees,
Proceedings of National Academy of Sciences, 100 (2003), 6313–6318.
[28] G. Cormode, P. Indyk, N. Koudas and S. Muthukrishnan, Fast mining of massive
tabular data via approximate distance computations, in: ICDE ’02: Proceedings of
the 18th International Conference on Data Engineering, IEEE Computer Society,
Washington, DC, USA (2002), p. 605.
[29] I. Derényi, G. Palla and T. Vicsek, Clique percolation in random networks, Physical
Review Letters, 94 (2005), 49–60.
[30] C. H. Q. Ding, X. He and H. Zha, A spectral method to separate disconnected
and nearly-disconnected web graph components, in: KDD ’01: Proceedings of the
seventh ACM SIGKDD international conference on Knowledge discovery and data
mining, ACM Press, New York, NY, USA (2001), pp. 275–280.
[31] C. H. Q. Ding, X. He, H. Zha, M. Gu and H. D. Simon, A min-max cut algorithm for
graph partitioning and data clustering, in: ICDM ’01: Proceedings of the 2001 IEEE
International Conference on Data Mining, IEEE Computer Society, Washington,
DC, USA (2001), pp. 107–114.
[32] I. Drost and T. Scheffer, Thwarting the nigritude ultramarine: Learning to identify
link spam, in: Proceedings of the 16th European Conference on Machine Learning
(ECML), Lecture Notes in Artificial Intelligence, vol. 3720, Porto, Portugal (2005),
pp. 233–243.
[33] D. Fetterly, M. Manasse and M. Najork, Spam, damn spam, and statistics –
Using statistical analysis to locate spam web pages, in: Proceedings of the 7th
International Workshop on the Web and Databases (WebDB), Paris, France (2004),
pp. 1–6.
[34] M. Fiedler, Algebraic connectivity of graphs, Czechoslovak Mathematical Journal,
23(98) (1973).
Telephone Call Network Data Mining: A Survey with Experiments 527

[35] G. Flake, S. Lawrence and C. L. Giles, Efficient identification of web communities,

in: Sixth ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, Boston, MA (2000), pp. 150–160.
[36] G. W. Flake, R. E. Tarjan and K. Tsioutsiouliklis, Graph clustering and minimum
cut trees, Internet Mathematics, 1(4) (2003), 385–408.
[37] D. Fogaras, Where to start browsing the web? in: Proceedings of the 3rd Interna-
tional Workshop on Innovative Internet Community Systems (I2CS), Lecture Notes
in Computer Science (LNCS), vol. 2877/2003, Springer-Verlag, Leipzig, Germany
(2003), pp. 65–79.
[38] D. Fogaras and B. Racz, Practical Algorithms and Lower Bounds for Similarity
Search in Massive Graphs, IEEE Transactions on Knowledge and Data Engineer-
ing, 19(5) (2007), 585–598. Preliminary version appeared at WWW 2005.
[39] D. Fogaras, B. Rácz, K. Csalogány and T. Sarlós, Towards Scaling Fully Personal-
ized PageRank: Algorithms, Lower Bounds, and Experiments, Internet Mathemat-
ics, 2(3) (2005), 333–358. Preliminary version from the first two authors appeared
in WAW 2004.
[40] M. Girvan and M. E. Newman, Community structure in social and biological
networks, Proc. Natl. Acad. Sci. USA, 99(12) (2002), 7821–7826.
[41] R. Guha, R. Kumar, P. Raghavan and A. Tomkins, Propagation of trust and
distrust, in: Proceedings of the 13th International World Wide Web Conference
(WWW) (2004), pp. 403–412.
[42] Z. Gyöngyi and H. Garcia-Molina, Spam: It’s not just for inboxes anymore, IEEE
Computer Magazine, 38(10) (2005), 28–34.
[43] Z. Gyöngyi and H. Garcia-Molina, Web spam taxonomy, in: Proceedings of the 1st
International Workshop on Adversarial Information Retrieval on the Web (AIR-
Web). Chiba, Japan (2005)
[44] Z. Gyöngyi, H. Garcia-Molina and J. Pedersen, Combating web spam with
TrustRank, in: Proceedings of the 30th International Conference on Very Large
Data Bases (VLDB), Toronto, Canada (2004), pp. 576–587.
[45] L. W. Hagen and A. B. Kahng, New spectral methods for ratio cut partitioning
and clustering, IEEE Trans. on CAD of Integrated Circuits and Systems, 11(9)
(1992), 1074–1085.
[46] M. R. Henzinger, R. Motwani and C. Silverstein, Challenges in web search engines,
SIGIR Forum, 36(2) (2002), 11–22.
[47] G. Jeh and J. Widom, SimRank: A measure of structural-context similarity, in:
Proceedings of the 8th ACM International Conference on Knowledge Discovery
and Data Mining (SIGKDD) (2002), pp. 538–543.
[48] G. Jeh and J. Widom, Scaling personalized web search, in: Proceedings of the 12th
World Wide Web Conference (WWW), ACM Press (2003), pp. 271–279.
[49] S. D. Kamvar, M. T. Schlosser and H. Garcia-Molina, The EigenTrust algorithm for
reputation management in P2P networks, in: Proceedings of the 12th International
World Wide Web Conference (WWW), ACM Press, New York, NY, USA (2003),
pp. 640–651.
528 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

[50] R. Kannan, S. Vempala and A. Vetta, On clusterings – good, bad and spectral, in:
IEEE:2000:ASF (2000), pp. 367–377.
[51] L. Katz, A new status index derived from sociometric analysis, Psychometrika,
18(1) (1953), 39–43.
[52] J. Kleinberg, Authoritative sources in a hyperlinked environment, Journal of the
ACM, 46(5) (1999), 604–632.
[53] J. Kleinberg, Navigation in a small world, Nature (2000), p. 845.
[54] J. Kleinberg, The Small-World Phenomenon: An Algorithmic Perspective, in: Pro-
ceedings of the 32nd ACM Symposium on Theory of Computing (2000).
[55] Z. Kou and W. W. Cohen, Stacked graphical models for efficient inference in markov
random fields, in: SDM 07 (2007).
[56] R. Kumar, J. Novak, P. Raghavan and A. Tomkins, Structure and evolution of
blogspace, Commun. ACM, 47(12) (2004), 35–39.
[57] M. Kurucz, A. A. Benczúr, K. Csalogány and L. Lukács: Spectral clustering in
telephone call graphs, in: WebKDD/SNAKDD Workshop 2007 in conjunction with
KDD 2007 (2007).
[58] K. Lang, Fixing two weaknesses of the spectral method, in: NIPS ’05: Advances in
Neural Information Processing Systems, vol. 18. Vancouver, Canada (2005).
[59] R. Lempel and S. Moran, The stochastic approach for link-structure analysis
(SALSA) and the TKC effect, Computer Networks, 33(1–6) (2000), 387–401.
[60] Liben-D. Nowell and J. Kleinberg, The link prediction problem for social networks,
in: Proceedings of the 12th Conference on Information and Knowledge Management
(CIKM) (2003), pp. 556–559.
[61] W. Lu, J. Janssen, E. Milios and N. Japkowicz, Node similarity in networked
information spaces, in: Proceedings of the Conference of the Centre for Advanced
Studies on Collaborative research (2001), p. 11.
[62] J. Malik, S. Belongie, T. Leung and J. Shi, Contour and texture analysis for image
segmentation, Int. J. Comput. Vision, 43(1) (2001), 7–27.
[63] S. Milgram, The small world problem, Psychology Today, 2(1) (1967), 60–67.
[64] A. A. Nanavati, S. Gurumurthy, G. Das, D. Chakraborty, K. Dasgupta, S. Mukher-
jea and A. Joshi, On the structural properties of massive telecom graphs: Findings
and implications, in: CIKM (2006).
[65] M. Newman, Clustering and preferential attachment in growing networks, Physical
Review E, 64(2) (2001), 25,102.
[66] M. Newman, The Structure and Function of Complex Networks, SIAM Review,
45(2) (2003), 167–256.
[67] M. Newman, Detecting community structure in networks, The European Physical
Journal B – Condensed Matter, 38(2) (2004), 321–330.
[68] A. Ntoulas, M. Najork, M. Manasse and D. Fetterly, Detecting spam web pages
through content analysis, in: Proceedings of the 15th International World Wide
Web Conference (WWW), Edinburgh, Scotland (2006), pp. 83–92.
Telephone Call Network Data Mining: A Survey with Experiments 529

[69] J. Onnela, J. Saramaki, J. Hyvonen, G. Szabo, D. Lazer, K. Kaski, J. Kertesz

and A. Barabasi, Structure and tie strengths in mobile communication networks,
Proceedings of the National Academy of Sciences, 104(18) (2007), 7332.
[70] J. Onnela, J. Saramaki, J. Hyvonen, G. Szabo, M. de Menezes, K. Kaski,
A. Barabasi and J. Kertesz, Analysis of a large-scale weighted network of one-
to-one human communication, New Journal of Physics, 9(6) (2007), 179.
[71] L. Page, S. Brin, R. Motwani and T. Winograd, The PageRank citation ranking:
Bringing order to the web, Tech. Rep. 1999-66, Stanford University (1998).
[72] G. Palla, D. Ábel, I. J. Farkas, P. Pollner, I. Derényi and T. Vicsek, K-clique
percolation and clustering, in this volume (2008), 369–408.
[73] G. Palla, A. Barabasi and T. Vicsek, Quantifying social group evolution, Nature,
446(7136) (2007), 664–667.
[74] PR10.info: BadRank as the opposite of PageRank (2004),
http://en.pr10.info/pagerank0-badrank/ (visited June 27th, 2005).
[75] X. Qi and B. D. Davison, Knowing a web page by the company it keeps, in:
Proceedings of the 15th Conference on Information and Knowledge Management
(CIKM) (2006).
[76] M. Richardson and P. Domingos, Mining knowledge-sharing sites for viral market-
ing, in: KDD ’02: Proceedings of the eighth ACM SIGKDD international conference
on Knowledge discovery and data mining, ACM Press, New York, NY, USA (2002),
pp. 61–70.
[77] T. Sarlós, A. A. Benczúr, K. Csalogány, D. Fogaras and B. Rácz, To randomize or
not to randomize: Space optimal summaries for hyperlink analysis, in: Proceedings
of the 15th International World Wide Web Conference (WWW) (2006), pp. 297–
306. Full version available at
http://www.ilab.sztaki.hu/websearch/Publications/
[78] J. Scott, Social Network Analysis: A Handbook, Sage Publications (2000).
[79] J. Shi and J. Malik, Normalized cuts and image segmentation, IEEE Transactions
on Pattern Analysis and Machine Intelligence (PAMI) (2000).
[80] M. Shiga, I. Takigawa and H. Mamitsuka, A spectral clustering approach to opti-
mally combining numerical vectors with a modular network, in: KDD ’07: Proceed-
ings of the 13th ACM SIGKDD international conference on Knowledge discovery
and data mining, ACM, New York, NY, USA (2007), pp. 647–656.
[81] A. Singhal, Challenges in running a commercial search engine, in: IBM Search and
Collaboration Seminar 2004. IBM Haifa Labs (2004).
[82] D. J. Watts and S. Strogatz, Collective dynamics of ‘small-world’ networks, Nature,
393(6684) (1998), 440–442.
[83] C. P. Wei and I. T. Chiu, Turning telecommunications call details to churn predic-
tion: a data mining approach, Expert Syst. Appl., 23(2) (2002), 103–112.
[84] I. H. Witten and E. Frank, Data Mining: Practical Machine Learning Tools and
Techniques, second edn, Morgan Kaufmann Series in Data Management Systems,
Morgan Kaufmann (2005).
530 M. Kurucz, L. Lukács, D. Siklósi, A. A. Benczúr, K. Csalogány and A. Lukács

[85] B. Wu, V. Goel and B. D. Davison, Propagating trust and distrust to demote web
spam, in: Workshop on Models of Trust for the Web. Edinburgh, Scotland (2006).
[86] B. Wu, V. Goel and B. D. Davison, Topical TrustRank: Using topicality to combat
web spam, in: Proceedings of the 15th International World Wide Web Conference
(WWW), Edinburgh, Scotland (2006),
[87] X. Xu, N. Yuruk, Z. Feng and T. A. J. Schweiger, Scan: a structural clustering
algorithm for networks, in: KDD ’07: Proceedings of the 13th ACM SIGKDD
international conference on Knowledge discovery and data mining, ACM Press,
New York, NY, USA (2007), pp. 824–833.
[88] H. Zha, X. He, C. H. Q. Ding, M. Gu and H. D. Simon, Spectral relaxation for k-
means clustering, in: T.G. Dietterich, S. Becker, Z. Ghahramani (eds.) NIPS, MIT
Press (2001), pp. 1057–1064.
[89] X. Zhu, Semi-supervised learning literature survey, Tech. Rep. 1530, Computer
Sciences, University of Wisconsin-Madison (2005).

Miklós Kurucz, László Lukács,

Dávid Siklósi, András A. Benczúr,
Károly Csalogány and András Lukács
Data Mining and Web Search Research
Group
Informatics Laboratory
Computer and Automation Research
Institute
Hungarian Academy of Sciences
e-mail: [email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
 Index

c/j-graph, 28 axon, 283, 285, 294–300, 304, 315, 365

k-clique adjacency graph, 371, 375–378,
380–382 Barabási–Albert (BA) model, 26–28, 30,
k-clique percolation, 369–397 70, 72, 85, 144, 171, 172, 230,
k-clique template, 371–373, 375, 376, 256, 316
384, 388, 389 basic decomposition, 182, 185, 193, 198
k-core, 87–90, 370 Bayes’ rule, 418
x log x property, 192 betweenness centrality, 246, 249
BFGS optimization method, 428, 433
abscisic acid, 260 binomial random graph model, 15, 19,
action, 33, 277–279, 281, 282, 284, 288, 21
291–293, 448, 460–462, 464, Bollobás–Janson–Riordan (BJR) model,
466–469, 473, 476, 477 29–31, 34, 77, 89, 99, 102, 308–
action-perception cycle, 278, 284 310
Adamic/Adar, 494, 495, 501 Bolzano–Weierstrass theorem, 419
adjacency matrix, 209, 332, 336–338, Bordetella bronchiseptica, 265, 267
342, 343, 346, 347, 349–351, Bordetella pertussis, 265, 267
354, 355, 357, 358, 494, 512, bosonic processes, 233
514, 516, 517 boundary, 30, 120, 122, 132, 133, 137,
aging, 257, 401, 430, 434 163, 198, 252, 277, 283, 288,
Akaike information criterion, 337–339, 301, 306, 311, 313, 401, 463
349, 355 box covering, 143, 160–162
allocortex, 277, 281, 292, 293, 298, 299, branch, 148, 152, 153, 164, 294, 300, 375
306 branching process, 31, 40–44, 46, 48–
almost sure convergence, 177 50, 52–54, 56, 60, 63, 64, 69,
analytic phase, 290 71, 74, 76, 77, 79–82, 84–89,
ARMA model, 454, 458 99, 100, 146–148, 154, 164, 171,
asset graph, 401 175, 177, 189–192, 315, 375,
assortativity, 400 389
asymptotic degree distribution, 179 branching random walk, 34, 81
attractor landscape, 277, 278 bridge vertex, 361
attributed graph, 409
autoregressive model, 454
532 Index

Callaway–Hopcroft–Kleinberg– clustering coefficient, 246, 247, 253, 255–

Newman–Strogatz (CHKNS) 257
model, 28, 30, 74 CoCoMac, 327, 345
cancer, 370, 395, 397–398 cognition, 277, 279, 284, 447, 449, 461,
cellular neural network, 293, 294 480
cerebral cortex, 301, 325, 326 cognitive functions, 285, 301, 327
chaos, 306 collaboration network, 401, 426–430
churn, 489, 490, 492, 505, 519–523 coloured neighbourhood metric, 102,
citation network, 409, 411, 412, 415, 418, 103
426, 428, 431, 432, 436 colouring, 117, 123, 125
classification, 125, 433, 489, 490, 494, combinatorial explosion, 445–449, 453,
519–524 460, 465, 477
semi-supervised, 520 community, 21, 91, 145, 340, 346, 356,
stacked graphical, 521–523 359–362, 369, 370, 372, 386,
clique, 21, 34, 81–83, 246, 247, 356, 359, 388, 392, 393, 396, 397, 399,
361, 369–403, 506–508 401, 402, 489, 490, 505–507,
clique percolation, 81, 383, 384, 391, 509, 510
506, 507 community finding, 369, 371, 383, 390,
cluster, 15, 16, 32, 33, 82, 104, 125, 391
130, 143, 144, 150, 160, 161, community graph, 395, 396, 399, 400
163, 222, 224, 243, 246, 247, component, 15–85, 117, 119–121, 126,
250–253, 255–257, 285, 304, 128–135, 137, 146, 148–151,
316, 325–328, 330, 331, 337, 158, 159, 164–166, 229, 231,
338, 342, 346, 352, 354, 358, 239, 240, 242, 245, 246, 248,
361, 362, 369–373, 375, 377– 250–253, 258, 260–262, 266,
380, 382, 383, 386, 388–393, 268, 280, 285, 288, 296, 304,
400–402, 457, 489–493, 505– 314, 337, 361, 369, 375, 377,
508, 510, 511, 513–519 378, 380, 381, 436, 437, 448,
cluster growing, 160, 161 451–457, 459, 462, 476–478,
cluster ratio, 510, 516–519 480, 491, 507, 509, 513, 514
clustering, 15, 16, 32, 33, 82, 104, 144, component exploration process, 41
243, 246, 247, 250, 251, 253, compressive sampling, 445, 450
255–257, 316, 338, 352, 362, conditional distribution, 182, 199, 200,
370, 371, 391, 392, 400, 402, 336, 337
457, 489, 491–493, 505–518, configuration model, 23, 25, 27, 68, 69,
524 80, 86, 144–146, 152, 155, 158,
clique percolation, 506–508 215, 219
quality measures, 510–511 configuration space, 23–25, 465
cluster ratio, 510 conjugate gradient method, 428
entropy, 511 connectional similarity, 328, 329
network modularity, 510 continuous degree approximation, 207,
purity, 511 215, 216, 218
SCAN, 508–510 convergence
spectral, 511–518 almost sure, 177
Index 533

convergence in probability, 177, 180 depletion zones, 207, 221, 224

cortical areas, 326–328, 361 diameter, 17, 21, 84–87, 122, 130, 301,
dorsal visual, 361 315, 317, 327, 346, 352, 361,
sensorimotor, 361 436, 491, 502
ventral visual, 361 diffusion-limited regime, 203, 207, 213,
cortical column, 296, 298, 299, 326 215, 218, 220, 221, 225
cosine, 489, 494, 495, 501, 510, 516, 520, direct problem of evolving networks, 413
522 directed k-clique, 386–391, 393
coupling, 22, 42, 44, 46, 54, 64, 82, 89, directed k-clique percolation, 386–391,
303, 494 393
Coverage, 117, 121 directed links, 345, 386–388
critical exponents, 149, 159, 166, 380 directedness, 296, 328, 329, 387
critical point, 73–75, 77, 79, 81, 82, disc, 117, 118, 121, 122, 127
84, 88, 149, 151–153, 166, 369, disc percolation, 309
371–374, 377, 378, 384, 388– discrete dynamic model, 269
392 disk, 313
criticality, 33, 143, 151–153, 158, 164, disorder, 143–160, 402
166, 277, 285, 290–292, 295, dominated convergence, 186
301, 304 dopamine, 278, 286
cross-entropy method, 478 double Pareto function, 432, 433
cut metric, 96–98, 100, 102, 103 Dubins’ model, 72, 74, 76
cyclic closure, 400–402 dynamic network model, 264, 269

data analysis, 346, 426 ECoG, electrocorticogram, 281–283,

data mining, 244, 334, 489–530 288, 290, 291, 293
declarative memory, 448, 459 edge weights, 143, 248, 383–385, 393,
degree cutoff, 214, 217, 219 397, 493, 498, 501, 515, 522
degree distribution, 17, 27, 33, 92, 104, edit distance, 91, 92, 94, 99
144, 145, 147, 150, 152, 162, EEG, electroencephalogram, 281, 283,
163, 165, 178, 179, 204, 209, 293, 297, 301, 302, 304, 327
211, 214–217, 229–231, 246, electrocorticogram, 281–283, 288, 290,
247, 250, 251, 253, 256, 257, 291, 293
305, 310, 311, 313, 317, 326, electroencephalogram, 281, 283, 293,
395, 396, 400, 429, 436–438, 297, 301, 302, 304, 327
440, 491, 502 entorhinal cortex, 298, 458
exponential, 440 entropy, 145, 209, 450, 452, 455–457,
power-law, 436, 437 478–480, 511, 517–519
degree sequence, 15, 17, 23–27, 29, 31, of clustering, 511
33, 68, 69, 76, 86, 145, 172 epidemic models, 205
degree-degree correlations, 208, 209, Erdős–Rényi random graph model, 143,
211, 213–215, 219, 229, 230 144, 307, 326, 339, 369, 370
densification, 436 event learning, 460, 462, 463, 467–469
density correlations, 203, 207, 222–225 evolving network model, 256
density of the point process, 174 excitation, 288
534 Index

experimental data, 245, 268, 328, 397 giant component, 15, 28, 31, 34–36, 41,
exponential, 29, 50, 58, 61, 121, 152, 43–47, 61, 62, 67–70, 72–74, 76,
155, 159–161, 163, 174, 182, 77, 79, 81, 82, 84, 85, 92, 146,
197, 250, 288, 294, 304, 311– 147, 149, 150, 164, 297, 304,
313, 316, 317, 333, 373, 390, 369, 370, 377, 379–382, 491,
395, 427, 440, 446, 448, 472, 492, 508
473, 475, 476, 497, 504 giant connected component, 229, 250,
exponentially expanding graph, 311, 316 251
Gibbs sampling, 336
factored reinforcement learning, 445 Gilbert model, 118, 128, 129, 131
factored value iteration, 472 global properties, 84, 95, 100, 192
feedback, 246, 250, 253, 254, 266, 267, goodness of a kernel-based model, 413
277, 281, 285–288, 290, 306, grandmother cell, 291
307, 316, 452, 461, 462 graph partitioning, 330, 489, 512, 513
fermionic processes, 203, 206, 207, 209, graph reverse engineering, 243
233 graph stacking, 489, 521–523
Fiedler vector, 512 graphon, 32, 95–98
fingerprint, 494, 495, 509 greedy optimization, 333, 337, 338, 355
min-hash, 495, 509 ground space, 30
fixed-point equation, 419 growing m-out random graph, 27, 28, 72
fMRI, functional magnetoresonance
imaging, 281, 282, 301, 304, half-planes, 123, 125
328 Hebbian learning, 286
focal closure, 400–402 heterogeneous mean-field theory, 203,
forest fire model, 431, 436–438 207, 209, 233
fractal, 143, 151, 153, 156, 158, 160–166 Hilbert transform, 290
fractal dimension, 143, 151, 153, 158, hippocampal formation, 285, 448, 458–
160–162, 166 460
fractal networks, 160 hippocampus, 298, 410, 459
fragmentation, 196 historical ordering of vertices, 176
fringe distributions, 179 hop distance, 154, 155, 159, 161, 163
functional magnetoresonance imaging, hub, 227, 231, 252, 253, 301, 302, 305,
281, 282, 301, 304, 328 403, 505, 508, 509

Galton–Watson branching process, 40, immune, 264–266, 398

44, 77, 79, 80 in-degree, 25, 246, 247, 250, 261, 305,
gene expression, 240, 242, 244, 245, 258, 311, 313, 317, 393, 412, 428,
259, 398 430–440, 497
general branching process, 84, 175, 177, independent component analysis, 280,
189 454
generating function, 19–21, 37, 38, 73, independent process analysis, 454
74, 147–149, 152, 164, 198, 200, independent subspace analysis, 454, 455
373–377 inhibition, 240, 241, 262, 266, 290, 306
geographic properties, 502, 505, 515, 518
Index 535

inhomogeneous clustering model, 15, 32, 350, 355, 410, 415, 417, 421–
82 426, 432, 434, 435, 452
inhomogeneous graph, 34 limit distribution, 177
inhomogeneous hypergraph model, 32, linear weight function, 177, 179, 196
82 link prediction, 489, 490, 492–501, 524
inhomogeneous random graph model, local properties, 95, 177, 192
15–18, 26, 29, 69, 84, 85, 89, log-likelihood, see likelihood
308 Lotka’s law, 17
integration
neural, 326 macaque monkey, 327, 332, 345
intensity, 29, 77, 83, 118, 119, 121, 123, macroscopic, 281, 282, 289, 291, 293,
127, 128, 134, 135, 137, 281, 298, 300, 302, 307
283, 312, 313, 383–385, 393 magnetoencephalogram, 281, 293, 302,
intentionality, 279, 285 304, 327
interactome, 241, 243, 269 Malthusian parameter, 177, 190–192,
inverse dynamics, 448, 461–465 197, 198
inverse problem of evolving networks, Malthusian process, 190–192
409, 410, 413, 415 Markov chain, 21, 174, 175, 196, 197,
irreducible kernel, 78, 79, 83, 84, 87 334–337, 346–348
Markov chain Monte Carlo method, 329,
Jaccard, 494, 495, 498, 499, 501, 509, 334–337, 346, 349, 355
510, 516, 517, 520 Markov decision process, 472, 474
jump rate, 175 Markov process, 197
Markovian pure birth process, 174, 189
K-set, 281, 283, 284, 289, 305, 306 matrix
Katz, 498, 500, 501 adjacency, 209, 332, 336–338, 342,
kernel, 29–31, 33, 42, 77, 78, 82–84, 87, 343, 346, 347, 349–351, 354,
89, 96, 100, 102, 308, 309, 410– 355, 357, 358, 494, 512, 514,
441, 457 516, 517
kernel family, 33, 82 directed Laplacian, 338, 342
kernel function, 410–441 Laplacian, 337, 342
kernel-based networks, 412, 413 maximum likelihood estimation, 333,
Kesten–Stigum theorem, 192 343
Kleinberg, 28, 33, 73, 81, 493, 494, 500, maximum likelihood method, 410, 415,
502, 503 417, 422, 424, 434, 452
Kullback–Leibler divergence, 338 MCMC (Markov chain Monte Carlo)
method, 329, 334–337, 346,
Laplace transform, 174, 183 349, 355
large-scale network, 15, 303 MEG, magnetoencephalogram, 281,
LCD (Linearized Chord Diagram) 293, 302, 304, 327
model, 26–28, 32, 34, 70, 72, membership vector, 331, 334
76, 80, 81, 85, 89, 147, 317 mesoscopic dynamics, 292
likelihood, 253, 257, 282, 294, 298, 332– metabolic network, 241, 247, 248, 251,
336, 339, 341, 343, 345, 348– 253, 257
536 Index

metabolome, 241, 243 optimal policy, 466

Metropolis–Hastings algorithm, 335 optimization, 33, 157, 261, 333, 337, 338,
microscopic, 209, 260, 277, 281–283, 345, 355, 362, 363, 419, 421,
285, 291, 294, 298, 307 425, 427, 428, 430, 433, 450–
midbrain, 301 453, 477, 478, 480, 512
Milgram, 502–504 goal function, 362, 363
min-hash, 495, 509 stochastic, 362
minimum spanning tree, 143, 156, 157 order parameter, 149, 277, 282, 283, 285,
modularity, 253, 283, 360, 391, 510, 511, 291, 304, 371, 380–382, 385,
516, 518 386, 390, 391
module, 244, 251, 253, 277, 283, 288, overfitting, 332, 334, 337, 341, 342, 350
289, 294, 300, 327, 369, 370, overlapping communities, 392
384, 392, 394, 397, 398, 403
PageRank, 495–499, 521, 522
Nelder–Mead optimization method, 428 paleocortex, 298
Nelder–Mead optimization methods, partial differential equation, 197, 283,
430 378, 379
neocortex, 277, 281, 292–295, 298–302, partial ordering, 173
306–312, 316 partition metric, 101–103
network dynamics, 410 patent network, 428, 430, 432, 433, 435–
network evolution, 257, 410, 417, 489 437
network modularity, 510 path length, 156, 159, 160, 248, 252,
networks 255–257, 261, 329, 346, 352,
geographic properties, 502–505, 361
515–518 percolation, 34, 69, 81, 82, 117–119,
small world, 502–505 128–129, 132–138, 143, 146–
stochastic reconstruction, 328 152, 157–159, 165, 166, 303,
stochastic representation, 361 307, 309, 310, 369–397, 506–
telephone call graphs, 489–530 507
neural connections, 297, 328 percolation threshold, 129, 146, 147,
heteromodal, 361 158, 165, 373, 375, 376, 381,
indirect, 348 384
long distance, 326 Perron vector, 338
nonexistent, 362 Perron–Frobenius theorem, 338, 416
nonreciprocal, 362 phase cone, 301
neuromodulator, 281, 286 phase transition, 15–21, 34–39, 68–77,
neuron doctrine, 326 81, 90, 146, 149, 166, 205, 277,
neuronal diversity, 326 284, 285, 288, 290, 291, 293,
neuropercolation, 280, 303, 307 295, 300, 303, 306, 307, 310,
null spike, 277, 291 377, 380
Nyquist limit, 453 pioneer neurons, 296, 297, 302, 304
point process, 38, 174, 183, 189
olfaction, 285, 292
optimal path, 143, 154–160
Index 537

Poisson process, 29, 44, 49, 77, 82, 118, resolution limit, 391
119, 121, 123, 125, 128, 134, response, 245, 250, 251, 254, 258, 260,
137, 138, 209, 309, 312, 313 263–266, 280, 281, 283, 288–
Poisson–Dirichlet distribution, 198 291, 398
population, 34, 164, 173, 189–191, 205, robust control, 448, 460, 465, 467–469,
242, 258, 263, 280, 282–286, 478
288, 289, 291, 292, 296, 298, robustness, 73, 253, 464
302, 303, 306, 326, 327, 451 root locus, 281, 283, 285
power method, 416, 417 rooted ordered tree, 172, 173, 176
power-law, 15, 17, 27, 33, 172, 204, 216– rooted planar tree, 172
218, 222, 247, 249, 251, 257, rota, 124
282, 290, 292, 293, 295, 301,
311, 316, 317, 390, 402, 429, scale-free networks, 204
432, 436, 437 scaling, 155, 158–160, 165, 166, 198, 200,
PPI network, 395–397 204, 219, 293, 312, 338, 369,
preafference, 279, 291 381, 382, 390, 395, 437
preferential attachment, 26, 27, 33, 72, scaling exponent, 200
171, 172, 231, 257, 304, 311, SCAN, 505, 506, 508–510
317, 396, 430, 432–434, 436, second correlation function, 183
437, 439, 440, 501 segregation, 207, 221, 223–225, 326, 329,
preferentiality, 301–304 361
probabilistic cellular automata, 303, 304 neural, 326
progeny, 173, 189–191, 196, 200 semi-supervised, 489, 520, 521
protein interaction graph, 241 semidefinite programming, 512
purity, 511, 517–519 sensors, 34, 121, 123, 124, 126, 454
of clustering, 511 separation theorem, 457
pyramidal cell, 298, 300, 326, 410 signal transduction network, 242, 248,
249, 251, 253–255, 260, 263
quality measures, 341, 496, 499, 510, 518 similarity, 90, 91, 103, 143, 282, 293,
quasi-randomness, 326 295, 298, 301, 329, 397, 402,
quasi-static approximation, 214, 215, 489, 493–497, 499, 501, 509,
217, 226 510, 512, 516, 520–524
Adamic/Adar, 495
random characteristic, 189–191 cosine, 495
random regular graphs, 23, 25 Jaccard, 494
rate code, 453 Katz, 498
reaction-diffusion processes, 203, 205, of nodes, 493–501
233 simulated annealing method, 428
real-world network, 15–18, 27, 32, 34, singular value decomposition, 337, 338,
91, 92, 172, 311, 396 513
reciprocity, 328 SIS model, 205
refractory period, 283 small-world model, 17, 85, 144, 203, 204,
reinforcement learning, 445–447, 459, 206, 207, 225, 250, 293, 294,
478, 480 328, 329
538 Index

social networks, 15–17, 165, 252, 370, transcriptome, 241–243

398, 399, 401, 402, 489–491, transition probability, 338, 466, 473, 479
502, 505, 506, 510, 516, 524 tree growth model, 171
sparse representation, 448–453, 461, 480 tree networks, 207, 230–233
spatiotemporal pattern, 277, 279 tree subgraph distance, 99
spectral, 290, 292, 402, 489, 505, 506, tree-indexed Markov field, 194
511–518
speed-field tracking, 448, 462, 465, 467 uncertainty, 245, 259, 328, 329, 345, 361,
spontaneous activity, 300, 307 362
spread-out model, 34 uncorrelated configuration model, 215,
stacked graphical, 519–523 219, 220, 224, 229, 231
stationary network, 422, 424 uncorrelated networks, 214–216, 218–
stochastic domination, 55, 57 220, 225, 228, 229, 231, 232
stocks, 370, 401 uniformly grown Z-out graph, 28, 76
stomatal closure, 261, 262 uniformly grown random graph, 28, 30,
stopping time, 174, 176 73
stratification, 434 universality, 149, 252
Strogatz, 16, 17, 28, 32, 73, 85, 144, 502
strong disorder, 143, 155–158 vertex space, 30
strongly connected component, 246, 252, voter model, 402
261, 266 vulnerability, 70, 72
subgraph distance, 94, 95, 99
subthreshold fluctuation, 301 Watts, 16, 17, 32, 70, 85, 144, 502
subtree, 99, 171, 173, 177, 178, 193, 194 weak disorder, 155, 156, 160
survival probability, 44, 69, 71, 76, 77, Web spam, 494, 520–521
80–83 weight function, 171, 172, 174, 175, 177–
synapse, 286, 294, 297–299, 302, 306, 181, 185, 191, 196–198, 200,
307, 326, 328, 410 493, 523
Szemerédi’s Regularity Lemma, 98, 329, weighted k-clique percolation, 383
330, 363, 365 weighted links, 383
wireless networks, 118
telescopic product, 195
total weight of a tree, 154–157, 175, 176 Yule-tree, 197
transcriptional regulatory graph, 241,
242