Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Brunel University London
Department of Mathematics
Complex Networks with NodeIntrinsic Fitness:
On Structural Properties andContagious Phenomena
Doctoral Dissertation of:Konrad Hoppe
2014
• i
Abstract
Complex networks is a vibrant research field and has received much attention over thelast decade. Central to this area is the question of how networks around us are con-structed. The essential notion of network research is that these systems are assembledin a decentralised way, thus no central agent is planning the network beforehand. De-spite this lack of central coordination, many networks present intriguing universalities,such as broad degree distributions, in the form of power-laws. The subject of study inthis thesis is a class of networks that are constructed by a node intrinsic variable, calledfitness. The way these networks grow could be called a rich-get-richer mechanism. Thefitter a node is, the more likely it is to acquire new connections inside the network. Sev-eral aspects that are directly connected to these networks are explored in this thesis. Inthe first part, the properties of growing networks that are driven by fitness are investi-gated and it is shown that the introduction of growth leads to a topological structure thatis different from its static counterpart. In the subsequent chapter, percolation on fit-ness driven networks is studied. The results give insights into possible mechanisms thatcan stabilise systems. Furthermore, the theory can be used to identify vulnerable struc-tures around us. In the following chapter, the world trade network is discussed. Thisnumerical investigation highlights possible improvements to the methodology to makestatistical analysis more robust. That chapter is followed by an analysis of time-varyingnetworks. Time-varying networks represent an interesting construct that allows a for-mulation of stochastic processes on the same time-scale as the evolution of the networkitself. This possibility is highly relevant to the investigation of epidemics, for instance.In the last chapter, a study of a system of clusters and their self-organised formation ispresented.
CONTENTSCONTENTS
CHAPTER 1 PREFACE PAGE V
1 Bibliographical Note v
2 Acknowledgement vi
CHAPTER 2 INTRODUCTION PAGE 1
1 Networks - The Lingua Franca of Science? 1
2 Examples of Networks 3
Seven Bridges of Königsberg — 3 • The Medici — 4 • The 9/11 Terrorists’ Network —6
3 Universal Structures of Complex Networks 7
4 Models of Networks 10
5 Stochastic Processes on Networks 15
Targeted Attacks and Random Failures — 16 • Epidemic Spreading — 16 • OpinionFormation — 19 • Percolation Phenomena on Complex Networks — 20
CHAPTER 3 A NETWORK GROWTH MODEL WITH NODE INTRINSIC FITNESS PAGE 23
1 Introduction 24
2 The Model 26
3 The Design Problem 30
4 Scale-free Networks 33
ii
• iii
5 Fitness Correlations 36
6 Conclusion 38
CHAPTER 4 PERCOLATION OF FAILURES ON COMPLEX NETWORKS PAGE 40
1 Introduction 41
2 Models 43
Static Model — 44 •Dynamic Model — 49
3 Percolation 52
4 Results 56
5 Conclusions 59
CHAPTER 5 THE WORLD TRADE NETWORK PAGE 60
1 Introduction 61
2 The Static Fitness Model 63
3 The Choice of a Data Source 66
4 Definition of Fitness 68
5 Intertemporal Structure of the WTN 70
6 Static Structure of the WTN 73
7 Conclusion 78
8 Countries Covered in this Study 80
CHAPTER 6 MUTUAL SELECTION IN TIME-VARYING NETWORKS PAGE 81
1 Introduction 82
2 Model 85
Degree Distribution — 85 • Epidemic Spreading — 88 • Consensus Formation — 93
3 Discussion 95
• iv
CHAPTER 7 CLUSTER FORMATION DYNAMICS BASED ON FITNESS PAGE 96
1 Introduction 97
2 Model 99
3 Results 103
Results for a Null Model without Fitness — 103 • Results of Fitness Based Cluster Dy-namics — 104
4 Concluding Remarks 110
CHAPTER A COMPUTATIONAL METHODOLOGY PAGE 114
Chapter 1Preface
1 Bibliographical Note
Some aspects of this work have been published in:
I. E. Smolyarenko, K. Hoppe, G. J. Rodgers. Network growth model with intrinsic vertexfitness. Physical Review E 88, 012805 (2013).
K. Hoppe, G. J. Rodgers. Mutual selection in time-varying networks. Physical Review E88, 042804 (2013).
K. Hoppe, G. J. Rodgers. Percolation on fitness-dependent networks with heterogeneousresilience. Physical Review E 90, 012815 (2014).
K. Hoppe, G. J. Rodgers. A microscopic study of the fitness-dependent topology of theworld trade network. Physica A: Statistical Mechanics and its Applications 419(1): 64-74(2015).
v
Preface 2. Acknowledgement • vi
2 Acknowledgement
Undertaking this PhD was a great experience and I have enjoyed learning so many newaspects during this time. There are many people without whom this would not have beenpossible.
First, I would like to thank my supervisor Professor Geoff Rodgers. He has been of sucha great support during my entire time of study and always stood behind me during theadverse periods of my PhD curriculum. Moreover, I am very thankful for the great op-portunities with which he has provided me. I am very proud to be a student of such adistinguished researcher and I look up to his poise with which he guides any audienceand situation.
I also want to thank my parents, Heidelotte Hoppe and Manfred Hoppe who have alwaysloved and supported me. This work would not be possible without them. The processof obtaining a doctorate degree is certainly longer than the core time of study and myparents have prepared me excellently to go this entire path. I am infinitely thankful forall their devotion.
A special acknowledgement goes to my girlfriend Emma Haddi, who was always patientwith me and very sensitive to the special circumstances of having a PhD student for apartner (which are plenty). Her love and support made me stay focussed and helped meto refuel during our times together. I love you, Emma.
My thanks also go out to the Department of Mathematics at Brunel that provided me witha generous stipend for the time of my doctoral research. In particular, I want to mentionhere Kay Lawrence. Her kind and helpful way always made me feel very welcome.
I also want to acknowledge the kind hospitality of Professor Bosiljka Tadic, whom I visitedin Ljubljana during my PhD studies. She has given me many interesting insights into thefield of dynamical systems and I very much enjoyed our conversations.
Also the staff members of the Graduate School at Brunel University should be mentionedhere. I am very thankful for their recognition and support during my time as a doctoralresearcher. They have made the experience at Brunel very joyful. I would like to thank es-pecially Diane Smith. She was always very welcoming and helpful and has kindly offeredto proofread this work, which is so much appreciated because of my persistent resilienceagainst British rules of punctuation.
Chapter 2Introduction
In this first chapter, the general concept of networks is introduced. The historical back-
ground of network research is illustrated by several examples that are found in the real
world. Despite the apparent complex and chaotic structure of many networks, some uni-
versal aspects of complex networks are highlighted, followed by various models that can
give rise to these universal patterns. The chapter closes with a survey of various types of
stochastic processes that can be embedded on complex networks.
1 Networks - The Lingua Franca of Science?
Complex networks is a research field that has drawn knowledge from two directions: the
mathematical treatment of graphs and the many particle systems in physics. The social
sciences have identified complex networks theory as a viable tool to understand many
phenomena around us, such as the spread of rumours, the cascading failure of institu-
tions and the contagion of diseases. Networks are commonly described by a set of nodes
and a set of edges that connect the nodes. These sets are also the main objects of study
in graph theory. While classical graph theory investigates various properties of graphs
that could be relatively small, such as the possibility of different colourings [7], ways of
traversing graphs during search [75], or forming round-tours [48], the study of complex
1
Introduction 1. Networks - The Lingua Franca of Science? • 2
networks that is rooted in physics aims to understand phenomena on almost exclusively
very large, and heterogeneous structures. Problems in this area are more probabilis-
tic and need different approaches compared to problems of classical graph theory. For
instance, while graph theoretical problems can often be cast in terms of the adjacency
matrix of a network, the structure of complex networks is often understood in terms of
sets of differential equations. The particular strength of complex networks research lies
in the ability to describe local interactions in a much richer way than is possible by us-
ing lattices or cellular automata. In particular, the strong heterogeneity of connectivity
patterns that appears to be omnipresent in the networks that surround us can easily be
modelled using networks, rather than lattices. Also other features, such as the tendency
of clustering in social networks, can easily be embedded into networks. In section 4
of this chapter, several constructive models that give rise to large scale heterogeneous
structures are presented.
The great interest in complex networks does not merely arise from these points, but
rather from the fact that almost everything that surrounds us can be described in terms
of a network. Obvious examples are the World Wide Web [4, 70] that is spanned by web-
pages, connected with hyperlinks. Also the physical internet that is spanned by routers
and cables has been investigated [52]. Other examples include the network of movie
actors [8], where nodes are actors and edges connect those that have acted in a movie
together. Similarly, scientific co-authorship networks have been investigated [8, 9, 90],
where nodes represent individual authors and edges connect academics that have co-
authored papers together. Other examples are transportation networks [12, 35, 63] and
the power-grid [2, 36]. More interesting than the mere description of the networks is
the analysis of processes that unravel on these. For instance, the prediction of disease
epidemics [35, 83, 120], and the spread of memes and information diffusion in social
networks [29, 126]. Furthermore, biological networks such as the structure and function
of the brain have been investigated [23] as well as protein encoding networks to better
understand the microscopic structure of our bodies and human diseases [118]. This in-
complete list of possible applications of complex networks research provides a glance
of the strength of this academic discipline. However, the application of networks is not
Introduction 2. Examples of Networks • 3
limited to these large scale structures. As initially mentioned, mathematical graph theory
can be utilised to solve a variety of problems, such as colouring and routing. Other appli-
cations of networks can be found in machine learning where artificial neural networks
are employed to mimic the human brain and understand complex data [106]. Finally,
computers perform elementary algebra by traversing trees which are a special class of
networks. The same structure is used for algorithms in mathematical logic. Thus, even
theorems can be proved using networks.
In summary, networks appear in a multitude of scientific disciplines and many problems
can be translated into graphs. The more recent history of complex networks, which are
classical graphs with the additional feature of being large and heterogeneous, has proven
extremely successful in tackling problems of the present century [65].
2 Examples of Networks
2.1 Seven Bridges of Königsberg
Problems related to graphs were explored early in the history of mathematics. A very fa-
mous problem is the Seven Bridges of Königsberg. Fig. 2.1a shows a map of Königsberg.
The task is to find a round tour through the city that visits every bridge exactly once.
Leonhard Euler proved in 1736 that this path does not exist [48]. Although the problem
is embedded in a geographic setting, its spatial dimension is irrelevant, because the po-
sition of the bridges does not matter. The only aspects that matter are the connecting
bridges. Therefore the problem can be represented as a graph, with the bridges as edges.
Fig. 2.1b illustrates this representation. Euler’s article is regarded as one of the earliest in
the field of graph theory and, the problem of finding an Eulerian tour has become a part
of standard theory.
Introduction 2. Examples of Networks • 4
(a)
#
#
#
#
(b)
Figure 2.1: (a)An ancient map of Königsberg with the river system in blue and the seven bridges markedin green. (b) Graphical representation of the bridges.
2.2 The Medici
The House of Medici was a very influential family in Florence that brought forth two
queens during the 15th and 16th centuries, as well as four popes in the course of the
15th-17th centuries. The family came to fame and political power for the first time un-
der Cosimo de’ Medici during the 14th century in Florence. The key to their sudden
prosperity was not their political dominance or wealth at that time, but rather their cen-
tral position in the network of marriages between different families [95]. Marriages be-
tween families have played a key role in arranging business deals and forming political
allies. Fig. 2.2 illustrates this network of inter-marriages. It is obvious from the figure
that the Medici inhabit a very central position in this network. This central position was
their advantage: many consolidations – commercial or political – between the families
of the Italian Renaissance were intermediated by the Medici. This graphical indicator
of importance in the network can also be quantified. Betweenness centrality of a node
measures the fraction of all shortest paths from all vertices to all others that pass through
that node. To make this more clear, define s i j (k ) as the number of shortest paths that go
from node i to node j through node k and s i j as the total number of shortest paths from
node i to j . For instance, the number of shortest paths from Guadagni to Salviati is 2,
and so is the number of paths that route via the Medici. Thus sGuadagni, Salviati = 2, and
sGuadagni, Salviati(Medici) = 2, while sGuadagni, Salviati(Albizzi) = sGuadagni, Salviati(Tornabuon) =
Introduction 2. Examples of Networks • 5
#
Acciaiuol
Medici
# Tornabuon
# Albizzi
# Ginori
#Peruzzi
#Barbadori
#Castellan
# Strozzi
#Ridolfi
#Pucci#
Bischeri
#Guadagni
#Salviati
#Pazzi
#Lambertes
Figure 2.2: Network of Florentine marriages in the 15th century. Nodes mark the families and edgesmarriages between the members of them.
1. The overall betweenness of a node is then
c (k ) =1
(N −1)(N −2)/2
∑
i 6=k 6=j
s i j (k )s i j
. (2.1)
The factor 2/(N −1)(N −2) normalises c (k ) to the unit interval. Interestingly the Medici
score the highest in terms of betweenness centrality with c (k ) = 0.522, while the next
highest centrality is found for the Guadagni with c (k ) = 0.255.
Another measure of importance is the number of adjacent edges of a node. This number
is called the node degree. Also in terms of the node degree, the Medici score top with six
adjacent edges, followed by the Strozzi and Guadagni with four adjacent edges each.
Given the subsequent rise of the Medici’s power in the 15th and 16th centuries, it is clear
that an investigation of the network structure of Florentine marriages provides viable in-
sights into future outcomes that lay beyond plain marriage patterns. The next subsection
will give some insights into a more recent network of interest, the network of terrorists
who were involved in the 9/11 attacks. Also in that case, simple ”back of the envelope
calculations” give useful insights into aspects that have later been identified by more
thorough investigations.
Introduction 2. Examples of Networks • 6
2.3 The 9/11 Terrorists’ Network
Shortly after the 9/11 attacks on the USA, information on the social ties of the terrorists
became publicly available. Nowadays, it is commonly known that Mohamed Atta was
one of the central planners of the attack. Fig. 2.3 illustrates the network of acquaintances
between the plane hijackers of 9/11. From that figure, it is clear that Mohammed Atta not
#
Wail Alshehri
#Wail Satam Suqami# Waleed Alshehri
# Abdul Aziz Al-Omari
2Marwan Al-Sheshhi
# Mohamed Atta2Fayez-Ahmed
2Mohand Alshehri
2Hamza Alghamdi
2Ahmed Alghamdi
4 Hani Hanjour
4 Khalid Al-Mihd4
Nawaf-Alhazmi
4 Salem Alhazmi
Ï Ziad Jarrah
ÏAhmed Al Haznawi
Ï
Saeed Alghamdi
Ï Ahmed Alhami
Figure 2.3: Network of terrorists involved in the 9/11 attacks. #: Flight AA #11 Crashed into the WorldTrade Centre. 2: Flight UA#175 Crashed into WTC. 4: Flight AA #77 Crashed into Pentagon. Ï: Flight UA#93 crashed in Pennsylvania.
only has a very central position within the network, but also serves as a intermediator
between the different teams that hijacked the four airplanes. Krebs [76] shows a larger
snapshot of this network with links also to other accomplices who provided money and
logistics. Considering this larger network, Mohamed Atta reserves a central role with
a betweenness centrality of 0.588 followed by Essid Sami Ben Khemais with a value of
0.252.
Introduction 3. Universal Structures of Complex Networks • 7
1762
1455
939
1234
13561536
12261478
1443
1264
1072
986
1756
9651041
17771654
1094
12681778
1694
931
1545
912
1408
1679
1425
1743 1671
13321357
1910
1385
905
1212
1364
962
1114
1844
1510
1121
13721295
1664
1776
1855 1514
1692
606
1767
484
15741602
1210
987
1348
949
1169
10421681
1011
1428
1690
1496
1876
1657629
910
1770
1311
1526
649
1294
10081397
34
14641155
1053
1660
464
1031
929
1266
1087
1461
1034
1701
1384
1476
1140
1419
14051792
996
1822
1477
10951202
1318
1565
1113
1142
427
1766
1080
3071
1656
1684
3230
1534
1853
1122
14461045948
549
1358
28143304331528212695 30922729 2948
327
219
6
147
95
154
3080
2766
2809
2769 32293095
2726
2710
3286 304228512967
3294 26633394
3414
3010
3436
3388
2963
2961
2884
3153 28072789
3139
3231
2978
1469
1831
1037
1052
1354
937
1109641
1745
914
637
1270
1785
483
1829
1062
1106
1010
1338
1228
1709
940
1296
1189
12751035
1183
1529
1738
1070
1630
1699
917
567
1064
1650
16961343
15991897
9151531
1808
1001
148715931427
1012
1237
15621223
526
1025
17081579
1203
1422 1641
1073
18811023
1038
1549
991
1161
1908
1678 1050
961
1445
1561
973
1171
1558
1651
166
1055
18651670
1564 1441
1527
1150
1058
1197
1838
923
1830
1674
1507
596
1236
1304
1482
1749
1633
1624
1355
1634
1308
19041260
1306
1667
1739
1869
1213
985
1090
1672
1089
1005
1233
1188
944
1815
1249
1139
1299
11411906
12251310
1748
13623989
4031
4011
943
1131
1415 1541
1259
977
30782925
2864
2698
28673363
3426
3054
2913
2969
28802910
2920
2669
2909
2754
2919
33242694
3085
3065
2852
2897
30042928
2915
2706
2705
2835
28963090
3228
2738
3237
30443202
2746
2679
2810
279
192
305111
220
44 262 4952
241
255
205
162
226
149
319
343
333
293
89
20
112
93
337
310
32137
216
167
243312
115
2
28
14
326
3036
3430
3193 2794
33032854
280627613213
28913123
3137
2981
2731
3400
3023
3243
3114
2887
3362 2977
3025
32342924
2953
2907
3135
2743
3257
3084
27563220
28742917
3184
3129
2763
2850
31192683
30522908
3364
2673
2745
3104
3390
4007
4001
3987
4016
4006
40124032
3990
4028
4029
3996
4037
4036
3985
3993
4015
3980
4008
4024
4023
4035
3988
3995
4034
4000
4018
40174014
3983
39813998
4003
4021
39823994
3986
4030
4026
4009
39974019
3991
3984
4027
4013
4022
4020
4038
4002
4004 40104033
3999
39924005
4025
97
253
110
201
264
306
182
91
328
78
193
273218 259
8195
41
19116
151140
174
289
144
17
138
214
321
46296
177
263
68
102
99
86 278 70
131
35
4
245
181
275
230
143
227
225
30032740
2885
2704
3290
2838
2750
280033872755
2815
2777
2911 3232
3284
3096
2717
3226
28273108
33482666
2985
3040
3142
2912
3399
29743417
3047
2872
1157
173
1854
899
1486
13831395
31493195
2943
27572986
3339
2966
3022
2664
30643014
2945
2773
3238
2873 2993
3029
3433
11341000
1840
16401168
1474
17591065
1481
1167
1244974
1241
1127
1258
1472
1159
1727
17031525
1366
1093
1051
1658
1711
1615
968
1342
913
1901
970
1847
1162
1247
1794
1022
1429
1575
1807
1248
1103
1406 1503
1246
1828
1166
1850
933
1251
1435
1646
918
1731
1144
1433
1686
1725
964
1500
15041453
14731814
14121147
1566
924
1418
1217
17811444
1436
1790
1713
15691281
13241170
1152
3435
2931
33842929
2994
2681
3158
2987
3293
3166
2916
30982960
29442866
2782
2839
3033
2927
2906
2905
28773056
26763174
3163
30732988
3250
2890
3026
26802940
2749
2786
2741
2689 3301
3000
2862295658
18581779
1151
1544
1860
1077
1321
1637
1048
1165
1765
1276
1317
1528
1484
1265
1834
1301
14111033 1283
9071262
1687
1555
1098
1967
1360
13861145
1698
1497
1825
911
1325
1013
1096
1511
9501885
1817
1911
12971626
1805
1084
1852 1883
1434
1387
1905
1521
1585
1851
1369
18431029
1519
1460
1513
1389
1239
1560
1135
1274
11101407
1164
1540
1806
1178
1195
920
1775
951
1337
1327
1774
2871
3109
3337
2783
2984
28592675
3015
2715
3219
3189
3259
2667
2932
2686
3369
2955
3242
3255
2844
2784
3249
3196
3264
3187
3032
3256
2843
2709
3048
2816
3341
3395
3208
31383087
33582702
33652762
285627953067
3329
271829213058
3144
2772
3099
2918
2865
32633402
30463167
2831
32173346
2690
2696
3246
2739
3379
2700 3221
2847
3190
3292
3312
2801
31563081
3425
3374
3275
2771
2858
3359
3398
2954
2685
3061
3389
3405
32393351
3265
3201
3247
2869
198
2661
3035
2939
1002
1027
1105
1252
1394
1627
3179
3355
2849
3280 3347
3113
3434
3386
3397
2793
3258
3150
3094
2973
3106
3344
3051
3120
3391
3252
3116
3101
2742
3082
3021 33502991
2828
1208
36832672853
3169 3070
3320
2765
3297
2719
31322833
552
1546956
384
10161403
3198
3088
2780
29383111
3140
3299
2875
3076
3261
2936
2712
3170
3272
2753
3288
2881
2687
2785
3403
2965
2876
30093188
33673160
3411
30172894
3011
3235
27372989
2716
3360
33322796
31723378
2937
31823145
3079
3396
3060
2778
3381
3277
3316
29923206
2674
3102 3416
27702970
3146
3152
3204
3185
3115
3419
2976
3330
3306
3212
2720
3016
3353
29503194 2735
3110
2768
27582727
3307
2803 3349
3327
28703393
3043
3171
30502895
31072711
3136
2791 3210
2723
3415
3075
3352
3328
2845
28682832
3276
3254
2995
3385
33572811
3323
3024
3041
2790 30693211 3161
3343
3421
26623118 3175
3157
3253
2665
2848
3093
3034 27443191
33733319
3418
3002
2830
2730
3151
32033018
2882
3072
2697
2878
2836
3133
2747
3066
3305
2983
30632820
2935
2721
2691
3091
2824
2671
2922
3271
3310
2933
2997
3308
3401
3427
3376
3028
3424
3338
2926
3266
3283
2959
2982
2879
3382
3309
2703
2670
3423
30082972
3218
3053
3314
3354
2699
3200
3311
3141
286131312767
2930
3126
3322
3372
3336
3180
3216
3128
2949
3083 2886
2975
3368 3012
90
3325
3318
33
3251
42
2834
2889
2776
2774
2732
3127
2902
2901
3406
3240
34092781
3321
3270
2904
2998 3342
3005
3186
2914
327430073205
33452813
3302
2892
3057
3154
2775
2764
2964
3245
2968
3432
3164
3278
3224
3049
2802
2980
2883
2863
2722
2748
3296
2979
3100
3291
3335
2855
3038 2707
33403331
32892825 3027
3030
3392
2946
3130
2837
2888
3227
2798 3300
3117
30063165
27253410
2752
990
2826
2677
3361
3269
2899
3317
34133059
3333
3045
3428
3143
2759
3407
2788
3268
28572817
2792
2952
1111
1253
1138
1368
979 1499
1218
1071
1043 955
1788
17261505
3031
1552
1666
3183
2842
16421326
124010301430
1758
29711450
3147
2799
3105
3074
3122
3055
2805
3037
3125
1543
1349
1616
1177
1882 1677
930
1506
1878
1763
909
1695
1870 18901801
1254
364
399501
441
564
393
476
351
3112
3377
2797
2672
2701 3039
3262
26842823
3199
3422
27083298
2840
3225
3260
2733
3209
3281
3148
3287
3013
2819
33132846
3192
28042958
3420
32223375
3168
2962
3380
3197
3279
3248
32233177
2812
2942
31243162
2941
3233
3103
3295
2724
3207
3020
2688
3366
2751
2787 3215
2760
3019
2734
3097
3404
34313062
299630892951
2923
3178
2818
3214
26783001
2822
2713
3412
26932903
3326
3282
2860
3134 3121
2714
26923173
3370
2900
2808
3273
28412990
2668
3429
3334
31593371
3086
3241
3356
3155
27363068
2829
32853408
2728
2682
2999
2947
3077
2893
3236
2779
32443383
2898
3176
2934
2957
674630
677
678
618
574
589 624
668590
594 663 655
667616
671
610584
389 563
639
679420
373
683
391
471
370
392
520451
438
461
591
524
646601
614
645
423
669
513
604
385
407
652
505
477
486
380
381
543
509
424
551
620
414
638
609
608587
485
459586
607
613
602
634585
422
369
480
435
673
469
457
518
644554 357
516666
403410474
349
531
396
372
519
356
354361
1802
1601
1137
1333
1798
1568
401
1319
958
1097
466 437
522
562
468
429
379
411
530
377
443
1363
901
1648
1493
1892
1880
1468
1378
1186
1118
453532
467
386
454
499
572
529
448
470
571
568
383
382
498
406
527
525
408431348
503
358512
500
487425
560456
1733
938
903
1216
1629521
1895
992488
350
378
419
447
482
395
514
497
430
542
412
374
440
656 511
545
1314
394
371
9631837
462
1548
409
1371
1784
494
398
375
1232
1452
1533
1747536
352
404493
1787
479
550
416
496537
559465 561
376
400
492
475
1645
16731625
1773
1490
1200366
1631
1740
1424 1693
1889
3631353
517
976
538
1760
428
1154
566
1057
1179936 1896
1313
1300 580
1606
16071744
6511400
3531320
1512
1374
648
635
659
642
647
581
628
576
598
573
684
611
575
625
603
605
619
588
633
664597
626
681
632
662
670
582
578
599
600
577661
658
653
672
636
654
617
579
631
680
685
682
657
627
643 583 650
640
615
595
675
665
660
623
612
621
622
592
547
533 355539
481
455
450
418
359
504
489
413
478
534
546
570
449
433
541528
676
460
540367
434
387
473
444
523
507
432
495
458
490
426
421491
405
510
360
593
558
535
508
397
388515
439
402 506
417
446
436
544
553
472
463557442
415569
556
445
365
452
362
390
502
548
555
565
2336
2372
2279
1940
22232138
2068
2384
2510
2471
2518
2187 2451
1945
2327
2643
2196 2199
2224
2649
243619262399
21172285
220323192629
2508
23512102
25472239
26052163
2039
2283
2042
22642250
1955
26082343
19322642
2133
2589
22982289
20382254
2267
21762588
22802368
2174
24912180
2295 2498
209
2493
2525
2641
2416
1923
2310
248026372325
2226
2432
23962235
2029
2209
2362
2031 2452
19442557
2270
2477 2484
19862237
3454
20442303
2092
2147
2554
2563 230025061989 2058
263124072098
2478
2165
3925
21642213
257426062353 2056
2311
2613 2499
2476
2391
2258
2021
2576
2055
1963
2556
2579
2136
246725522532
2591
21792306
26512322
2517
25142296
2474
2566
2345
2522
2465
2330
2445
2544
2616
1987
21272653
2135
2599
2533
2259
23982128
2134
2572
2594
2496
2144
2048
19522047
2249
2473
2502
2232
2367
2425
2529
20992305
2242
2355
1964
2272
2635
2379
1913
23661965
21822612 2186
2141
2205
22112066
2318
26102175
2244
2586
2381
2090
2492
2073
2323
2464
2340
20302363
26252220
2290
2059
2507
21722188
2123
26152630
2590
2218
2410
20332409
1946 2369
2331
2600
2604
2131
2088
2326
2655
2408
2638
2374 2423 2564
2139 23543723
1962
2578
2356
1085
2275
3961
2184
21502206
2078
2271
2240
1938
2064
2601
2118
2201
1917
25932104
2370
20742500
34872395
2233
1983
2069
2278
2611
2560
25611966
2551
2376
2334
2095
2646
24332623
2550
26242210
2040
2526 2060
2266
2229
2142
19852324
2309
2542
2414
2602
2287
2462
2084
2429
2392
2536
2193
2580
2082211421072618
136
2265
2358
2085
2145
2453
2406
2036
1974
2256
2313
23122214
2230
2389
2400
2413
19542268
2417
2291
2531
2571
19822472
2581
2143
19692361
2228
22771930
2166
2540
2292
2338
2071
2567
25012153
20872592
2569
2420
2194
2191
2457 2148
2512
2007
2026
2195
2293
1912
2315
2597
21372302
2385
2215
2057
1994
2609
1993
2607
2468
2470
2596
2543 1941
1925
2347
2189
2079
2530
2288
2443
2269
2208
2065
2387
2524
2427
2350
2449
2479
2252
1981
2652
2274
2390
2110
2434
2415
2598
2062
19481959
2555
233220281920
2183
21112081
23772294
1577
2125
2132
2344
24381465
1947
2509
2463
2151
2149 25822241
2072
2032
2161
25372617
2475
2461
2458
22472394
2155
2568
202
292
2222
2096
2185
2603
215
2177
2198
207520031939
1916
1970
23042101
22212519
2419
23332246
2002
2273
2053
21692516
2511
23282054 2010
20522282
2243
2200
2020
2460
2154
2339
2299
2112
24462005
2257
1971
24282103
2121
1979
208319532077
2495
22762109
2261
2386
2573
2482
1918 2348
261920632093
2520
1984
2553
2485
2404
24302115
2469
2212
2253
2190
2559
2045
2108
21242352
2654
21401943
20432549
23082122
2359
2418
1929
2307
2539
25212037
24892046 2575
2086 2329
2504
2216
1997
2546
228
281
283
183
251
127
159
76309
36
108
139
135
164
3138
178
207288
15839
130
320
222
84
311
121
197284
184
117
45
29
165
314
274
133
294260
206
96
57
15373
77
65
69
132
221
23110
10540
106
2113
2450
1956
2535 23012439
1915
2217
2100
24242051
2558
263220412105
1922
2120
21602587
25702146
2156
2341
23881949
22512130
2614
2455
2660
2337
2659
2027
2004
2494
21292070
2357
1933
2412
2238
2181
2405
2402
21522360
2373
2197
2119
2577
2091
226019601934
25622089
2015
2380
2456
19372227
2008
237
190
51
83
240
124
191
160
5
48
340
54
339
88
302
157
307
254
47
24
92
175
101
299
267
152
269
71
155
81
23
61
286
100
249
330
187
266
80
180
194
53 301
94
242
204
300
346
129347
317
2291
316
168
196
126
336
200
9
271
56
285
122
26252
325
297
203
282188
98
176 236113
21
199109
318
313
291
332
323
280
148
247
161185
79
303224
334
338
248
170
345
60
315
322
169
67
304
246
156
103
308
128
329
119
213
87
146
7
120
324
125 27
22
235
208
25
30
277
50
290
257
258
341
276
62
344
85272
186172
104
13
261
223342
239
118
244
238
55
141
75
142
82331 134
16
295
123
211
212
72
66
59
265
270
250268
63
3
298
232
234
256
335
179
233
904
171 1224
287
145
15
43
74
12
163
210
11
18
37
114
0
64
189
217
150
3181
2342
2080
2656
2317
2162
21261977
2316 25342505
2231
2448
1990
2173
2411
2371
2523
2628
2023
2397
2503
1978
2314
2286
2158
19362335
1975
2440
24972403
2248
2049
22362644 2035
2488
787
848
817 849
770
851
837
731 724
706
695726
842
784
853
847
760 834
773694
752
692
766
696 713745
761
701762
710 812
723
721
814693
844
734778
826756 727
700
711738
795
794
774
688815
820
780
781
810
718
754
800
827
686
805
719
824
697
801830
705
823
828
759737
777
845831846
722
821
715
751690
829
704749
716832
775
740
730
755
783
720
797
748
758
804
809
764
687
753
807
769
838
880
856
843 739
792
835
741709
757
763
855
808
793852
717
779
765
803
811
890
798
869
735
840
708
728
703
772
747
819
707
767
822712
833789
796
702
725
691
806
839
732
791
714
689 799
841
836
790
750
744
818
746
786
854
813
816 736
742768
733
699
785788
850
782
802
743
771
865
871
895
858
881
879
868
870
729
892
891
698
825
883
875
885
867
866
874
873
884
872
859
887
877
894
889 861
876
878
893
776
886
864863
882888
1309
969
971
1732
1489
8981454
17041751
1081
1582
1638
150212611316
16611820
1099
984
1706
1102
1018
1508
1907
1020
1893
1595
1032
1894
1382
1396
1108
1848
1373
9191783
1303
942
12821824
1215
1515
18711812
1649
954
1066
1492
1273
1729
1466
1581
1705
120611741410
1494
1417
860
1221
928
1451
1479
1120
994
1143
1245
1007
1235
935
1263
900
1818
1061
1676
1136
1307
1119
952 999
1653 17961652
1293
1659
1750
1209
1683
906
1539
1867
934
17371340
1524
1375
10041632 1491
1416
1721
1076
12501597
1795
1112
1401
1700
1344
1877
1242
1771
1557 921
1149
1600
16431746
1735
1572
1689
175213651079
1184
1439
155615321789
1181
16231198
1530
1287 1547
966
1302
1810
1538
1821
1017
1668
1068 1791
16081420978
1230
162118861717
92513671827
12691598
1059
1537
1800
1554
9271047
1467
16751049
1463
1861
17541891
1377
1146
993
1222
1523
1753
980
1520
1604
16031390
1833
1589
1078
1613
11851399
1823
1028
1291
1341
1714
1799
1267
1757
15161816
16971780
9951594
18461702 1440
1272
1312
1639
1480
1160
1180
1173
9531509
12551614
1409
1685 1835
1329
1813
1196 18631437
1797
1193
9221730
1056
15901074
897
14701535
1909
11821485
1026 18111609
967
1423
1636
1067
10211426
1129
1728
1381
1414
1764
1715
1404
975
12981462
1583
101513791438
902
1315
1176
9811115
1884
1592
1550
10821100
1655
1292
1691
1279
10441755
1036
1857
1345
1596
1257
1841988
15871432
1903
1576 1322
18731862
9471278 1644
1819
10541003
1584107
1567
1866
17191421
1803
1588
1187
1328
1393
1710
1620
1688 1075
1553
1334
908
1243
1086
1339
1290
1116
1734
1091
17241088
1518
17611128
1712
1006
1107
1619
1153
1380
1336
1665
1101
1335
1156
18791736
1617 1839
1707
11991211
1610
1471
1662
1793
1163
12891132
1826
1456
1351
1125
1457
1849
1842
15591431
1172
1330
1117
1551
1663
1352
1391
1207
1376
1092
1570
1864
1622
1522
1669
946
1612
1563
17681124
12381305
1175
1845
1191
11261804
1888 9161741
13701288
9831285
1024
1809
1271982
1039
960 959
1449
1580 1498
1398
1130
1040
1256
18681832
1402
1083
1898
1874
1605
1722
1856
1517 1501
1359
19021214 1201
1331
1723
1123
1447
1361
1769
1542
1205
1219
997
1483
1442
1133
1220
17721388
1900
16181571
9981323 926
1458
932
1280
1782 1836
1346
1718
972
1069
1488
1063
1046
1887
989
1586
1194
945
896
1104
18991350
1573
1611
1628
1459
1192
1014
1060
1204
1277
1448
1475
1786
1680
1190
1591
1495
11581392
941
1875
1742 1720
1148
1872
1859
1019
1229
1286
1635
1284
1647
1227
1231
1009
1578
1413
9571682
1347
1716
3604
3547
3847
3805
3484
3979
3888
3795
3438
379635223480
3496
3950
3830
3575
3552
3563
3928
3512
3548
3492
3755
34773970
3760
3944
3817
3573
3831
3520
3916
3908 37653661
3672
3461
3788
3533
39433651
3456
372135773621
3693
3938
36563971
3586
3873
3877
3517
3545
3734
3449
3880
3633
3495
3968
3841
3797
3948
3851
3756
3741
3540
3692
3609
3677
3750
3501 3525
3684
3440
3550
3886
3674
3592
3962
3872
3556
3625
3689
3839
3779
3780
3842
3889 3626
3618
3667
3683
3712
2515
2320
2650
2444
1992
2012
2022
2061
2159
2013
1980
2014
1931
2620
2483
2622
2595
2426
2207
2527
20762204
2486
2466
24212437
19962565
2255
2401
2106
2490 1924
2513
1988
2170
2349
2168
3589
37753688
3451
3570
3733
3504
3551
3530
3861
3954
2262
1942
2442
2019
2393
2454
2067
2584
2383
2545
2234
36003561
34533650
3798
3729
38533437
3904
3444
3742
3748
3935
3820 3856
3566
3709
3849
3673
3974
3804
3640 3790
3731
857
3488
3762
3528 3452
3557
3527
3670
39023514
3578
3845
3471
3603
3658
3462
3934
3736
3503
38913559
3913
3635
3574
3681
3858
3838
3702
3737
3824
39263687
3884
3860
3611 3605
3966
3642
3596
3697
3800
3828
862
3448
3491
3707
3518
3497
3785
3730
3450
35423627
3485
3475
3703
3929
3576
3930
3852
3654
2375
2621
2034
2263
2528 2548
1961 2634
2321
2422
2219
2017
23652627
1951
1919
2202
1998
22252583
19352364
21782431
2459
2382
2541
2441
2116
1973
19722167
2481
2009
1976
1928
1921
2018
22451968 2094
2487
2538
2171
2640
2024
2281
2157
24472001
1999
1995
2378
20062633
19141958
24352050
2025
1957
20972657
2585
20002645
20161950
264826392011
2192
1927
2297
2346
2636
26262647
2658
2284 1991
3757
3579
3519
3465
3476
3708 37743975 3701
3675
3720
3967
3616
3641 3766
3664
3447
3546
3580
3899
3571
3894
3679
3933
36463786
3623
3620 34583535 3782
3792
3510
3668
3761
3459
36023771
3881
3524
3634
3630
3949
3727
3759
3624
3836
3490
3837
3478
3783
3791
35873500
3920
3655
3776
3696
3960
3809
34683768
3493
381036483937
3499
3855
3822
3969
3509
3976
3631
3711
35343821
3529
3614
3818
34603866
3483
3506
3924
3591
3593
3833
3826
3951
3878
3896
3719
37143569 3718
3903
3486
3558
390039473643
3958
3945
3882
3572
3829
3931
3671
3564
3724
3507
3794
3690
3909
3906
3706
3584
3793
3758
3622
3738
3764
3781
3568
3532
3752
3443
3816
3946
3554
3834
3956
3777
3843
3466
3972
3823
3763
3918
3957
3653
3441
3963
37873537
3811 3705
38873678
3767
3694
38143964
3939
3789 3905
3923 34463863
3910
3806
3585
3751
3965
3536
3531
3745
3682
3652
3749
3665
3645
3606
3464
360138933474
3489
3676
3582
3915
3716
3638
3647
3632
3439
3747
3607
3615
3901
3932
3636
3867
39213599
3715
3508
3740
3663
3907
38763936
3482
3865
3629
3553
3827
3498
36083815
3835
3713
3977
3735
3802
3952
3844
3801
3588
3848
3799
3699
3513
3538
3892
3644
3912
3874
3728
3541
3862
3753
3710
3698
3473
3725
3567
3657
3917
3739
3868
3595
3819
3743
3857
3685
3472
38403773 3598
34453850
3871
3769
37783666
3442
3927
3825
39553539
3869
3583
3494
3612
35493662
37703469
3511
3784
3555
3726
38953617
3515
3463
3870
3812
3680
3754
3973
3803
3890
3808
3660
3941
3859
36103953
3942
38073922
3875
3846
3885
3854
3732
3879
3746
3914
3919
35623649
3639
3695 34813516
38323940
36863457
3521
3704
3560
3717
39593467
3744
3691
3897
3911
38643700
3526
3669
3813
3502
35053470
3455
3722
35973590
3628
3613
3898
3479
3581
3883
3594
3523
35443978
3565
3619 3637
36593543
3772
Figure 2.4: Snapshot of the network of friendships on Facebook (source: Stanford Large Network DatasetCollection)
This is therefore another example of the importance of network metrics that shed light
on aspects far beyond the scope of topological insights. The network of terrorists and
further consequences of its topology have also been investigated in [99]. In particular, it
is important to note that an understanding of the underlying network/hierarchy struc-
ture is essential for a successful anti-terror operation. The impact of different network
structures on deliberate attempts to disturb the function of a network is discussed later
in this chapter.
3 Universal Structures of Complex Networks
The networks around us are of astonishing size and complexity, see Fig. 2.4 for an exam-
ple of a snapshot of Facebook. However, there are also a number of interesting regular-
ities in these structures. To illustrate this concept, some features of three different net-
works are investigated in this section: a graph of the world wide web provided by Google,
the collaboration network in the area of condensed matter on arXiv1, and the network of
publicly shared circles on Google+2. All data has been obtained from the Stanford Large
Network Dataset Collection3.
Many networks are so called small-world networks. These networks are characterised in
1www.arxiv.org2https://plus.google.com3https://snap.stanford.edu/data/
www.arxiv.orghttps://plus.google.comhttps://snap.stanford.edu/data/
Introduction 3. Universal Structures of Complex Networks • 8
terms of their average path length L. That is the number of steps it takes on average to
walk from one node to another, whereby steps can only be made along edges. A network
is identified as a small world if the average path length growths logarithmically with its
size [125], thus
L ∝ log N , (2.2)
where N is the number of nodes. The small-world property is sometimes also known as
the six degrees of separation, coined by Guare [62]. Another property of interest is the
local clustering coefficient c i of a node. The clustering coefficient measures how close
the neighbours of a node are to being a clique. A clique is a set of nodes that is fully
connected, such that every possible edge exists. Formally, the local clustering coefficient
is defined as
c i =1
k i (k i −1)
�
�
�
¦
e j k : v j , vk ∈Ni , e j k ∈ E©
�
�
� (2.3)
whereNi is the set of vertices in the direct neighbourhood of node i . Furthermore, E =
{e i ,j : i , j = 1, . . . , N } denotes the set of all edges and V the set of all nodes with V =
{vi : i = 1, . . . , N }. For an undirected network, the concept can be accordingly defined,
but needs renormalisation by factor 2. Fig. 2.5 illustrates the local clustering coefficient
graphically. The average clustering coefficient is then given by
i
#
# #
(a)
i
#
# #
(b)
i
#
# #
(c)
Figure 2.5: Example configurations of a node’s neighbourhood. Solid edges mark present edges, dashededges possible ones. The local clustering coefficient is for the different configurations: (a) c i = 0, (b) c i = 1,and (c) c i = 1/3.
〈c 〉=N∑
i=1
c iN
. (2.4)
Real-world networks show significantly more clustering than random networks. The
most simple random network model was proposed by Erdős and Rényi [46] and serves
Introduction 3. Universal Structures of Complex Networks • 9
as a null model for many network concepts. The model is described by a fixed number of
nodes and edges. Each edge exists with probability p , such that the existence of edges are
statistically independent from each other. In this case the clustering coefficient is sim-
ply 〈c 〉= p , since the probability that pairs of edges exist in the neighbourhood does not
depend on any neighbourhood linked property. Clustering describes the cliquishness of
a network. In terms of friendship networks, high clustering implies that the friends of
someone are themselves friends of each other. This transitivity of friendships is likely
in social networks and thus, real-world networks exhibit much higher clustering than
pure random graphs. Tab. 2.1 summarises these network measures for three different
networks. Another feature that can be observed in real-world networks is that they have
N M 〈c 〉 DiameterGoogle web graph 875, 713 5, 105, 039 0.5413 21arXiv cond-mat 23, 133 93497 0.6334 14Google+ circles 107, 614 13673453 0.4901 6
Table 2.1: Central properties of the three networks under investigation, with N as the number of nodes,M the number of edges, 〈c 〉 the average clustering coefficient, and the diameter as the longest shortest pathin the network.
broad degree distributions, to be more precise: power-law degree distributions [31]. The
degree distribution p (k ) is the probability that a randomly chosen node has k neigh-
bours. In the case of a power-law, p (k ) has the form
p (k )∝ k−a . (2.5)
Power-laws appear in many areas of interaction, also in the three networks that serve as
examples in this section; see Fig. 2.6. It is not entirely clear why interaction leads regu-
larly to power-law distributed properties. In fact there is an entire branch of literature
investigating this phenomenon [84]. For networks, one mechanism that is believed to
drive networks into the observed power-law degree distributions is preferential attach-
ment [8]. The notion of preferential attachment is that interactors with more existing
interactions are more likely to engage in new interactions. In terms of the network of
friendships, a person with many friends finds it easier to acquire new relationships, than
somebody who is entirely isolated. More details on this model can be found in the fol-
lowing section.
Introduction 4. Models of Networks • 10
1 10 100 100010-6
10-5
10-4
0.001
0.01
0.1
k
PHkL
(a)
2 5 10 20 50 1002001´10
-4
5´10-4
0.001
0.005
0.010
0.050
0.100
k
PHkL
(b)
1 10 100 1000 104
10-5
10-4
0.001
0.01
0.1
k
PHkL
(c)
Figure 2.6: Degree distributions of (a) the web graph (b) the collaboration network in the condensedmatter section on arXiv (c) the friendship network on Google+.
4 Models of Networks
In section 2, some examples of the usefulness of network studies have been illustrated.
While the focus was put on the existence of various observable artefacts, in this section
a number of models that lead to certain macroscopic patterns will be reviewed.
One of the first and simplest network models is the Erdös-Rényi (ER) Model [46]. Two
variants of this model exist: the G (N , M )model, that describes an undirected graph with
N nodes and M edges, and the G (N , p ) model, which elucidates a graph with N nodes
whereby each pair of nodes shares an edge with probability p . In both cases the number
of nodes is static and edges are deployed randomly. The resulting degree distribution is
binomial and can be approximated in the N →∞ regime as a Poisson distribution
PER(k ) =(N p )k e−N p
k !(2.6)
for the G (N , p )model. The relationship to the G (N , M )model is then described by M =�N
2
�
p .
The ER model is a useful null model to study networks, but it lacks several features that
are observed in real-world networks. One such feature is clustering. Networks around
us are highly clustered and have short average path lengths. A model that resembles this
feature is the Watts-Strogatz (WS) model [125]. The network consists of a fixed number
of nodes. All nodes are organised on a ring lattice. The central parameter of the model
Introduction 4. Models of Networks • 11
are N , k ∗ and p . N is the number of nodes, k ∗ the number of nearest neighbours on the
ring lattice that each node is connected to and p is the probability that an existing edge
gets rewired. p drives the structure of the network from a regular ring lattice for p = 0 via
a small-world network for 0 < p < 1 towards a random network at p = 1. The rewiring
takes place in a sequential process that considers each edge exactly once. More details
can be found in [125]. The degree distribution depends on the value of p . For p = 0, p (k )
is peaked at k ∗: p (k ) = δ(k − k ∗) where δ(x ) is the Dirac delta function. For p = 1, the
degree distribution is Poissonian, as in the ER model. Also for values of 0 < p < 1, the
degree distribution can be computed in terms of a series representation [14].
A further generalisation of the Erdös-Rényi model has been introduced in [24]. The
model by Caldarelli et. al. extends the ER model by introducing a vertex intrinsic variable
and an attachment kernel f (x , y ), which is defined as the probability that a pair of nodes
with intrinsic variables x and y share an edge in a network with N nodes. The hidden
variable is sometimes called fitness, which refers to its ability to attract edges inside the
network. The degree distribution p (k ) can be found modelling a more detailed quan-
tity, namely p (k |x ), the fitness conditional degree distribution. One possible approach
is to expand p (k |x ) into its contributions [21, Eq. (15)]. Another approach is to derive it
from first principles, noticing that the static nature of the network can be ignored for the
sake of the argument and edges deployed sequentially. In this case the evolution of the
fitness conditional degree distribution in a network with M edges and N nodes can be
described by
pM+1,N (k |x ) = pM ,N (k |x )[1−λ(x , N )]+pM ,N (k −1|x )λ(x , N ), (2.7)
whereλ(x , N ) is the probability that a node with fitness x receives an edge during an edge
deployment step. The details of this derivation are discussed later in Chapter 5, following
Eq. (5.5). The resulting fitness dependent node degree distribution is found equivalent
to [21] to be Poissonian:
pM ,N (k |x ) =e−k̄ (x )[k̄ (x )]k
Γ(k +1), (2.8)
where k̄ (x ) is the fitness conditional degree expectation. The exact form of the degree
Introduction 4. Models of Networks • 12
distribution of the resulting network depends then on the probability distribution of fit-
ness and the attachment kernel f (x , y ). The fitness model has also been analysed in
[111], where the inverse and direct problems have been addressed. That is the problem
of finding analytical forms of f (x , y ) and the fitness distribution ρ(x ), such that the re-
sulting network exhibits a given degree distribution. The solutions to these problems are
of high importance when the internal mechanisms of the network need to be understood
in more detail. The spreading of diseases can usually be expressed in terms of f (x , y ) and
ρ(x ), and solutions to the inverse and direct problem can give rise to these quantities.
In Chapter 3, a further extension of the fitness model, introducing growth is studied.
Growth has been proven to be an important ingredient for many real-world models [42].
The fitness conditional degree distribution can be found with the same technique as
outlined above for the static case. This distribution is given by
p (k |x ) =1
1+λ(x )
�
λ(x )1+λ(x )
�k−1. (2.9)
Wherebyλ(x ) is similarly defined to that in the previous paragraph on static fitness driven
networks. More details of this model can be found in Chapter 3.
Another class of network models that have attracted significant amount of attention in
the complex networks literature are degree based models. Early work on this class was
motivated by bibliographic studies of cross-referencing between different publications
[102]. The cumulative advantage of having more citations in order to receive more fol-
lowing citations in the future was studied in more detail in [103]. However, Price’s inves-
tigations were not widely recognised until 1999. Barabási and Albert [8] reinvestigated
what is nowadays called the Barabási Albert (BA) model. The BA model is a growing net-
work model, that models the cumulative advantage by imposing that each new node that
arrives to the network connects preferentially to nodes with high degree. More formally,
define the probability that a new node connects to an existing node i as
πnew→i =k i
∑
j k j=
k i2t
. (2.10)
Introduction 4. Models of Networks • 13
k i is the degree of node i . Thus, the rate at which the degree of a node changes is
d k id t= k i /2t . (2.11)
Therefore, the final degree distribution is then found to be
p (k ) = 2k−3. (2.12)
This finding is the reason for the success of the BA model. Many real-world networks
exhibit power-law degree distributions [3]. Interestingly both features, growth and pref-
erential attachment, have to be present. Growth without preferential attachment leads
to an exponential distribution, while preferential attachment without growth leads to a
non-stationary distribution, ultimately resembling a complete graph, that is a network
with all possible edges present [8].
Next to the two classes of models that are built either exclusively by a hidden variable
or exclusively by a degree based mechanism, there exists also a third, a hybrid type.
Bianconi and Barabási [19] introduce a model which is motivated by the finding that
in the classical BA model old nodes had more chances to connect than young nodes
and will therefore have on average a higher degree. In real-world networks one can fre-
quently find nodes that arrive new to the network and are able to acquire large numbers
of new links in a short time. These out-performers must have some intrinsic quality
which drives this rapid growth. To account for this, a generalised attachment kernel that
considers this increased ability is introduced:
πnew→ik i x i
∑N (t )j=1 k j x j
, (2.13)
where x i is the fitness of node i and k i its degree. A generalisation of this model with
other approaches to link degree and fitness was introduced by Ergün and Rodgers [47].
Another form of combining the fitness model and the degree preferential attachment has
been investigated in [16]. Bedogne and Rodgers [16] introduce a model where one of two
things can happen at each time step. Either a new node with one link is added, whereby
Introduction 4. Models of Networks • 14
the target node is chosen degree preferentially as in the BA model, or a new edge between
two existing nodes is deployed, depending on the nodes’ fitness. The resulting degree
distribution is a pure power-law. The exponent depends only on the relative frequency
of node additions compared to edge addition steps.
Perra et al. [100] present a model that is characterised by dynamics on a much shorter
time scale than in the previously discussed models. The preceding models are either en-
tirely static or growth occurs gradually, one update at a time. The time-varying network
model in [100] is entirely different. Time varying networks are described by a fixed num-
ber of nodes and a varying number of edges. If the model is described in discrete time,
then at each time step each node becomes active with a node intrinsic probability and
connects to a fixed number of existing nodes inside the network. The motivation behind
this class of models is, for example, in the mobile phone network, where nodes represent
mobile phones and edges phone calls. If a phone initiates a call to another, then these
two nodes share an edge during that particular instance. Other examples include human
contact to spread diseases or even rumours. The particular appeal of a time-varying net-
work model lays in the possibility to formulate the evolution of the network on the same
timescale as the unfolding of stochastic processes on top of it. More details about this
class of networks can be found in Chapter 6.
The preceding network models are flat, in the sense that edges are just of one type and
binary. However, there exist a number of systems for which these assumptions do not
hold. Investigating the structure of link weights is important in a number of applica-
tions. Consider for instance the strength of a social tie, distinguishing long time friends
from random encounters. Another example forms the air-route network, where nodes
are airports and edges the flight routes between them. The weight of a link might de-
scribe for instance the number of seats that are available during a day on a given route.
This knowledge can be used to comprehend the contagion of infectious diseases or the
severity of service interruptions. Additionally, edges cannot just have different weights,
but can also be of different types. Continuing the airline network example from above,
cities are not only connected by air-routes, but also by streets and the railway system.
These different layers can either be represented as separate networks or as different types
Introduction 5. Stochastic Processes on Networks • 15
of links. Fig. 2.7 illustrates this notion. Also the network of international trade forms an
#N2
#N3
#N1
#N4
(a)
#N2
#N3
#N1
#N4
#N2
#N3
#N1
#N4
1st Layer
2nd Layer
(b)
Figure 2.7: Schematic illustration of multiplex networks. (a) Representation of multiplicity as differentedge types. (b) Representation of multiplicity as different network layers.
example of a multiplex network. Each node represents a country and the different edge
types indicate the traded amount of single product classes between pairs of countries.
Another example is the power supply network, that consists of different levels, such as
gas pipelines, high voltage power lines, water pipes, etc. Apart from merely describing
these structures, it is of central interest to understand how to prevent failures of these
networks and protect them against deliberate attacks. The study of these phenomena is
the subject of the following section.
5 Stochastic Processes on Networks
Many real-world phenomena can be described in terms of networks, such as trade rela-
tions, human contacts, collaborations, etc. Beyond a detailed description of these net-
works, it is of great importance to understand how these networks behave in the pres-
ence of adverse events and to what extent their function is compromised. In this section,
a number of stochastic processes that occur on networks are discussed.
Introduction 5. Stochastic Processes on Networks • 16
5.1 Targeted Attacks and Random Failures
Random networks described in the ER model and scale-free networks produced by the
BA model are very different with respect to their topological quantities. These differences
prove to be crucial when the networks are under targeted attack, or subject to random
failures. Albert, Jeong, and Barabási [5] present a numerical investigation of the conse-
quences for both type of networks. The scale-free (SF) network is very resilient against
random failures compared to the random network. The reason is the heterogeneous
contact pattern of the SF network. There is just a small number of hubs, that connect
large proportions of the network. Random failures are unlikely to occur at these hubs,
so a large fraction of the network can fail before the network’s function is affected. The
behaviour of the random network is entirely different. Most nodes have degree k ∼ 〈k 〉,
thus each node is equally important for the function of the network, and hence removing
nodes increases the diameter of the network linearly.
While the SF network is extremely resilient against random failures, it is prone to fail
quickly under targeted attack. An attacker who knows the hubs of the network that con-
nect large proportions can destroy the network’s function after a small number of re-
movals. In this case, the random network has the advantage. Since every node connects
on average the same fraction in the network, the additional knowledge of the attacker
does not contain any valuable information and thus the diameter increases in the same
way as in the case of random failures.
5.2 Epidemic Spreading
The study of epidemics has received attention from the mathematical community for a
long time. One of the earliest contributions has been made by the Swiss mathematician
Daniel Bernoulli. He investigated a mathematical model of smallpox in 1760 [17]. A com-
prehensive review of the coverage of infectious diseases in mathematics can be found for
instance in [66].
Introduction 5. Stochastic Processes on Networks • 17
Before an underlying network structure was considered, the diseased population was
assumed to be perfectly mixed in the sense that an individual has equal probabilities
to connect to any other individual in the population to pass on or to receive the dis-
ease. This assumption allows one to describe the evolution of an epidemic in terms
of simple linear differential equations. A simple compartmental epidemic model is the
susceptible-infected-susceptible (SIS) model. In this model each node belongs to one of
the two compartments: infected I or susceptible S. If the population is assumed to be
perfectly mixed, then the evolution of the number of individuals in each compartment
is governed by
d s
d t=−βs i +γi and (2.14)
d i
d t=βs i −γi . (2.15)
where s = S/N and i = I /N are the fractions of susceptible, respectively infected indi-
viduals. β is the probability that a susceptible individual gets infected in a contact with
an infected individual and 1/γ is the time it takes to recover from the disease. Fig. 2.8
illustrates the development of the size of these two groups.
sHtL
iHtL
100 200 300 400 500t
0.2
0.4
0.6
0.8
1.0
Figure 2.8: Size of the two compartments: susceptible (S) and infected (I) in the SIS model with β = 1/10and γ= 9/100.
The figure shows that the number of individuals in each compartment reaches an equi-
librium value after a relatively short burn-in phase. A central quantity of epidemic re-
search is the reproduction number that measures the number of secondary infections
following a single infectious individual being introduced to the system. This number is
Introduction 5. Stochastic Processes on Networks • 18
defined as
R0 =β/γ. (2.16)
R0 describes the potential of the disease. In particular, if R0 < 1, the infectious individual
cannot generate an epidemic, since there are not enough secondary cases in order to
sustain the disease. If however R0 > 1, then the diseased population grows ad infinitum.
If R0 = 1, the disease will stay forever in the population, since every single case produces
one successive case. R0 = 1 marks the epidemic threshold that separates the two phases
of infinite growth of the infectious fraction and instantaneous immunity.
The preceding technology can easily be expanded to account for networks in which the
degree distribution is peaked at 〈k 〉 and decays exponentially [13]. For heterogeneous
networks, this is not the case. A typical node degree does not exist as it is possible to
find many nodes with k � 〈k 〉. It is therefore necessary to consider the different degree
classes separately, to compute the evolution of the epidemics. The variables of interest
are the fraction of infected nodes with degree k : i k (t ) = Ik /Nk , and the fraction of sus-
ceptible individuals in degree class k sk (t ) = Sk /Nk . The evolution of the SIS model is
then described by
d i kd t=βk skΘk (t )−γi k and (2.17)
sk = 1− i k . (2.18)
Θk is the density of infected neighbours around a node with degree k :
Θk (t ) =∑
k ′P(k ′|k )i k ′ (t ) (2.19)
In the absence of degree correlations, this quality becomes
Θk (t ) =Θ(t ) =1
〈k 〉
∑
k ′k ′P(k ′)i k ′ (t ). (2.20)
Notice, that the k independence holds only in the absence of degree correlations. With-
out degree correlations the conditional probability to find a node with degree k ′ in the
Introduction 5. Stochastic Processes on Networks • 19
neighbourhood of a node with degree k is given by
p (k |k ′) =k P(k )〈k 〉
. (2.21)
The epidemic threshold is found to be
β
γ≥〈k 〉〈k 2〉
. (2.22)
It has been illustrated that for many real-world networks that the degree distribution
follows a power-law p (k ) ∝ k−a with a typical exponent 2 < a < 3. Thus in the N →∞
limit, 〈k 2〉→∞ and therefore an epidemic threshold of zero is found, which implies that
a disease can never die out on a heterogeneous network. Real-world networks have finite
size and therefore a non-zero epidemic threshold can be computed, but nevertheless this
result shows that epidemics can spread over a heterogeneous network easily even with
very low transmission rates.
5.3 Opinion Formation
Opinion or consensus formation is a widely studied phenomenon in social sciences and
has received attention in the complex networks literature as well. The simplest models
are the voter model (VM), the invasion model (IM) and the link dynamics model (LD). All
these models comprise a set of agents that are characterised by a discrete, usually binary
variable s i =±1 that reflects their opinion on a certain topic, for instance Republican vs.
Democrat. Independent of the embedding geometry, the evolution of these three mod-
els is usually described in discrete time. In the voter model, at each time step a node is
chosen randomly and adopts the opinion of a randomly chosen neighbour. In the inva-
sion model, the randomly picked neighbour adopts the opinion of the initially chosen
node. In the LD model, an edge is chosen randomly and a randomly chosen node ad-
jacent to this edge adopts the opinion of the node on the other side of that edge. Early
work on voting and invasion can be found in [32]. If the community of agents is embed-
ded on a d -dimensional lattice, then the community reaches an ordered state, that is a
Introduction 5. Stochastic Processes on Networks • 20
situation where each agent has the same opinion, for d = 1, 2. Castellano, Vilone, and
Vespignani [26] derive results for voting on the small-world network (WS model) [125].
Interestingly, long range interaction induced by the rewiring in the WS model leads to
a faster ordering/agreement of the system only in finitely sized networks. In the ther-
modynamic limit of N →∞, the small-world network reaches a non-trivial equilibrium,
in which both opinions can coexist. Sood and Redner [113] investigate the time it takes
until a consensus on a heterogeneous network can be reached. This time will be called
consensus time TN . The consensus time on a regular d -dimensional lattice scales with
as TN ∝N 2 for d = 1 and TN ∝N ln N for d = 2 and TN ∝N for d > 2 [113]. This scaling
behaviour is found to be different on heterogeneous graphs. For power-law networks for
example, the scaling is found to be for p (k )∝ k−a as TN ∝N ln N for a = 3, N (2a−4)/(a−1)
for 2< a < 3 and TN ∝ (ln N )2 for a = 2 [113]. More results on the invasion model and the
link dynamics model can be found in [112].
5.4 Percolation Phenomena on Complex Networks
Percolation theory on complex networks is concerned with topological properties of oc-
cupying clusters on top of a network. Occupation of a single node can be understood as
a diseased state and percolation theory finds therefore applications in epidemic mod-
elling [91]. The topological properties of the percolating component that is described
by occupied sites and links between those can be understood in three distinct phases.
These phases are separated by the percolation threshold pc . For low occupation prob-
abilities p < pc , many small clusters exist. If the occupation probability is greater than
the threshold, i.e. p > pc , then the network is occupied by one giant component that
is globally connected. In terms of an epidemic, this state describes the situation when
a disease has broken out and a major part of the population is infected. At p = pc , the
system is critical and occupied clusters of all sizes exist.
To compute the properties of the giant component, define q as the probability that a
randomly chosen edge does not lead to the percolating component. This probability
q can be calculated self consistently. Consider randomly picking a node i 0 and follow
Introduction 5. Stochastic Processes on Networks • 21
one of its links randomly toward a node i 1. The link between i 0 and i 1 is only not part
of the percolating component if none of the other links attached to i 1 are not part of
the component. The overall probability for this event can be split into two parts: the
probability that the randomly picked link ends at a node with degree k and that none of
the other k −1 links belongs to the percolating component. In an uncorrelated network,
p (k |k ′) is independent of k ′ and the probability that none of the k − 1 links belongs to
the percolating component is just q k−1. Therefore q can be computed by
q =∑
k
k p (k )〈k 〉
q k−1. (2.23)
The construction above appears cumbersome at first glance, because of its construction
in terms of probabilities that a event does not occur. However, the readers can convince
themselves that the positive formulation would demand a more detailed analytic expres-
sion.
Define now the probability that a node belongs to the giant component as pG . Obviously,
1−pG =∑
k
p (k )q k . (2.24)
This probability can be computed by solving Eq. (2.23) and substituting for q . A trivial
solution to Eq. (2.23) is q = 1. Substituting q = 1 into Eq. (2.24) leads to pG = 0. There
exists however maximally one more non-trivial solution to Eq. (2.23). Define G (q ) =∑
kk p (k )〈k 〉 q