Analysis and Simulation of Scientic Networks
-
Upload
felix-puetsch -
Category
Science
-
view
136 -
download
1
description
Transcript of Analysis and Simulation of Scientic Networks
Institute of Theoretical Physics
University of Cologne
Diploma Thesis
Analysis and Simulation
of Scienti�c Networks
Felix P�utsch
�
July 14, 2003
425
453
482492
529
251
181
530
80
199202
222
336
41
3987
27
223
224
371477
102
105
134
338
359
supervised by Prof. D. Stau�er
Hereby, I con�rm (according to the Pr�ufungsordnung of July 12, 1996, x20(5)) having
composed this diploma thesis alone, using no other than the mentioned sources and tools.
Citations have been marked.
Hiermit versichere ich gem�ass x20(5) der Pr�ufungsordnung vom 12. Juli 1996, dass ich
diese Diplomarbeit alleine erstellt und keine anderen als die angegebenen Quellen und
Hilfsmittel verwendet habe. Zitate wurden kenntlich gemacht.
Though the mountains divide
And the oceans are wide
It's a small world after all
R. M. and R. B. Sherman
Contents
1. Introduction 7
1.1. Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2. Six Degrees of Separation . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3. Small World E�ect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4. Science Collaboration Networks . . . . . . . . . . . . . . . . . . . . . . . 8
2. Network Models 9
2.1. Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1. Small World E�ect . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.2. Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.3. Scale-Free Behavior . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2. Regular Lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3. Erd}os-R�enyi Random Networks . . . . . . . . . . . . . . . . . . . . . . . 11
2.4. Watts-Strogatz Small-World Networks . . . . . . . . . . . . . . . . . . . 11
2.5. Barab�asi-Albert Network Model . . . . . . . . . . . . . . . . . . . . . . . 11
2.6. New Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3. Empirical Collaboration Network 15
3.1. Typology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1.1. Citation Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1.2. Collaboration Graph . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2. Building the net . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.1. Proceeding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2.2. Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3. Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1. Authors per paper . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.2. Connections per author . . . . . . . . . . . . . . . . . . . . . . . 18
3.3.3. Double vs. unique links . . . . . . . . . . . . . . . . . . . . . . . 19
3.3.4. Cluster sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.4. Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5
Contents
3.4.1. Erd}os-R�enyi random graphs . . . . . . . . . . . . . . . . . . . . . 21
3.4.2. Watts-Strogatz small-world networks . . . . . . . . . . . . . . . . 21
3.4.3. Barab�asi-Albert networks . . . . . . . . . . . . . . . . . . . . . . 22
4. Spin models 23
4.1. Leadership e�ect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1.1. Ising model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1.2. Phase transition . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1.3. Degree distribution . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1.4. Spin ip model . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2. Cluster limited Ising models . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2.1. Proceeding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.2. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.3. Bias adjusti�cation . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.4. Linear Relationship . . . . . . . . . . . . . . . . . . . . . . . . . 29
5. Barab�asi-Albert network models 31
5.1. Modi�ed Barab�asi-Albert model . . . . . . . . . . . . . . . . . . . . . . 31
5.2. Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2.1. Isolated clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2.2. Merging clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2.3. scale-free behavior . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6. Conclusion 37
A. Acknowledgements 39
B. Source code 41
B.1. Ising model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
B.2. Spin ip model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
B.3. Modi�ed Barab�asi-Albert model . . . . . . . . . . . . . . . . . . . . . . 51
C. Figures 55
Bibliography 55
6
1. Introduction
1.1. Networks
In recent times, hearing the word network
immediately arouses the idea of physically
wired networks as those formed by tele-
phone lines or computer links. Though, net-
work is a concept a good deal more general
than only this.
Mathematically spoken, networks are
graphs, i.e. a set of nodes (of whatever
kind) connected by edges (links, connec-
tions) between certain pairs.
This abstraction has been known for a
long time. Probably the �rst paper of graph
theory was written by Euler [1], the so-
called \bridge problem of K�onigsberg". Eu-
ler discusses whether or not it is possible
to make a round walk, passing of each of
K�onigsberg's nine bridges exactly once (�g-
ure 1.1).
The concept of networks can be applied
to lots of theoretical or experimental sub-
jects [2{4], nodes being people [5], Internet
servers [6], scientists [7, 8] or others, the
range of links comprises e-mails [9], friend-
ships [5], citations [10, 11] and more.
Thus, there are numerable di�erent kinds
of networks, physical ones (e.g. hard wired)
as well as logical (e.g. dependencies) or
social ones (e.g. contacts, friendships),
stretching out to topics far from wired net-
works [9, 12]. The area is under vigorous
research. Good reviews can be found in [2{
4, 13, 14].
1.2. Six Degrees of Separation
Out of personal experience, nearly every-
body has been confronted with what we call
small world e�ect. There are numerous ex-
amples:
At a party, we �nd out to know some
stranger we just started talking to by only a
few middle-persons (or technically spoken
we are only separated from him by a low
degree). E.g., he could be our street neigh-
bors' colleague's son. We hear \My god,
how world is small".
Rumors are another example. We are as-
tonished to experience the pace at which
they spread. After a few hours and thus
only a few possibilities of telling rumors to
others, whole city seems to know.
Milgram [15] made an experiment on
this. He instructed a set of people to try
to send a letter to some stranger, only by
using personal contacts. He found out that
an astonishing short chain of social links is
needed for this task, which entered in every-
day's language as Six Degrees of Separa-
tion. Recently, this has been reviewed on
a more popular basis by a German weekly
newspaper [5].
1.3. Small World E�ect
Six Degrees of Separation is only one man-
ifestation of a more general principle: the
small world e�ect [16{18].
Observations of many real-world net-
7
1. Introduction
Figure 1.1.: K�onigsberg bridge problem: Is it possible to make a round trip, passing each bridge exactly
once? [1]
works in computer science, biology, chem-
istry, linguistic, sociology, etc. have re-
vealed a crucial di�erence from regular lat-
tices.
Regarding average (or sometimes max-
imum) path lengths on such networks, we
would expect to see an increase linearly with
the number of nodes. Instead, we examine
distances growing logarithmically with sys-
tem size.
This behavior is not only an amus-
ing e�ect but has far-spreading conse-
quences [19]. Prominent examples are In-
ternet's stability against attacks [20], dis-
ease spreading [21, 22] or path �nding
strategies [23, 24].
1.4. Science Collaboration
Networks
In context of science, the network between
scientists as nodes of the graph is of par-
ticular interest. This network belongs to
the group of social ones, with humans as
nodes. Unlike many other forms of social
relationships, that are mostly quite diÆcult
to capture objectively, the �eld of published
papers is very widespread covered by the
Science Citation Index [25] and so easily
available to research.
8
2. Network Models
There are many types of networks com-
peting to describe observations made in
socio-physics. After discussing which mea-
surements describe a given networks struc-
ture, we will give a short overview about
what we think to be the most important
ones and discuss advantages and possible
disadvantages.
2.1. Measurements
2.1.1. Small World E�ect
As illustrated in the introduction, we are
interested in the correlation of network size
and average (maximum) path lengths. We
investigate if there is linear, logarithmic or
other behavior. In case of a logarithmic one,
the net is said to show the Small World
E�ect.
2.1.2. Clustering
In friendship network, we �nd friends of one
person often to be friends themselves. This
is true for most social networks and even
other ones. Links are not spread randomly
but arranged in clusters.
To describe thing mathematically, we in-
troduce a clustering coeÆcient C
i
of a node
i describing the portion of m established
links between all k
i
next neighbors com-
pared to the maximum possible number of
M =
�
m
2
�
, i.e.
C
n
=
m
M
:
This value is averaged over all vertices to
give a clustering coeÆcient C for the whole
network.
Typical values experienced are far above
results expected for random networks [2, p.
50].
2.1.3. Scale-Free Behavior
A third observation regarding social net-
works is its distribution of degrees. Regard-
ing frequency of vertices of given coordina-
tion numbers, we do not �nd an exponential
but a power law [26].
In all, we have three possible means
to classify networks. Many social graphs
show small path lengths, high clustering
and scale-free behavior.
2.2. Regular Lattices
The simplest form of a lattice is a symmet-
rical formation of nodes connected by edges
between all pairs (or all pairs of adjacent)
nodes as shown in �gure 2.1a,b. Reasons
to choose this linking are e.g. to simulate
neighborship in a town etc.
Such network show a high degree of clus-
tering, as wished. The average path lengths
are very long, though and scale with sys-
tem size. So, small world behavior cannot
be found which makes the model inappro-
priate for our needs.
9
2. Network Models
Figure 2.1.: network types: a, b regular lattices, c random network, c scale-free graph [14]
10
2.5. Barab�asi-Albert Network Model
2.3. Erd}os-R�enyi Random
Networks
Random networks are the extremum on the
other side of the spectrum. A number of
nodes is wired by pure chance, i.e. we throw
the dices to select two nodes and place an
edge between them.
Such graphs have been �rst proposed by
Solomono� and Rapoport [27] and have
been extensively studied by Erd}os and R�enyi
[28]. Actual results have been reviewed in
[29]. A typical result can be seen in �g-
ure 2.1c.
We �nd that average path lengths behave
logarithmically with network size. While
this is as desired for small world simulation,
obviously there is no clustering.
2.4. Watts-Strogatz Small-World
Networks
The idea is plausible to try combining both
presented models to sum up their corre-
sponding advantages. A big step towards
this goal was done by Watts and Strogatz
[16].
Their model starts with a circular graph
that is regularly wired (�gure 2.2). Step
by step, edges are chosen by chance and
rewired to an arbitrary destination node.
Thus, a small fraction of links are long-
range ones. To illustrate, this could be
habitants of a street of neighbors having
relationships with far-away relatives.
At �rst, the model seems to ful�ll our de-
sires. It shows small average path lengths as
well as high clustering. Looking at the de-
gree distribution, i.e. the frequency of nodes
Figure 2.3.: Barab�asi and Albert [30]
of a certain degree, we �nd strong di�er-
ences from real-world data as there is no
scale-free behavior.
2.5. Barab�asi-Albert Network
Model
Barab�asi and Albert [30] started a new idea.
Their model consists of two ingredients:
growth and preferential attachment.
We start with a graph of m
0
= 3 vertices,
each one connected to each other. Now, in
each time step, we add a node that is con-
nected to others by m = 3 links. The new
node being one side of the links, the other
one is chosen at random from the existing
network. The probability of a vertex being
selected is proportional to the number of
links already attached to it.
To stay in the image: If you already have
lots of friends, you are more likely to get
new ones. \The rich get richer."
1
These rules result in a network (�g-
ure 2.1d) that is capable of reproducing
small-world behavior, as well as being scale-
free. Research has found good collapse with
1
\Whoever has will be given more, and he will
have an abundance." [31].
11
2. Network Models
(b) (c)(a)
Figure 2.2.: Watts-Strogatz network model: We start with a regular lattice (a) formed to a ring (b)
and re-wire a small fraction of links to random destinations (c) [18]
empirical networks, included e.g. the world
wide web [6]. Good introductions can be
found in [32, 33].
Clustering is present to a certain degree,
but still much too small regarding experi-
mental values.
2.6. New Approaches
Recently, new network models have been
developed to cope with inconveniences en-
countered with present ones.
Ravasz and Barab�asi [35] examined net-
works of a self-similar structure imitating
the idea of hierarchical organization in soci-
ology. Combining high clustering and scale-
free behavior, their model does not show
short path lengths, though.
Klemm and Egu
�
iluz [34]
2
developed an
auspicious model joining all three demands
in one network. The authors present a gen-
eralization of the Barab�asi-Albert model,
adding aging of nodes and some random
behavior.
There will be further research to be done
2
cf. also [36]
on this model to verify if it copes with re-
ality.
An overview of all models can be found
in �gure 2.4.
12
2.6. New Approaches
scale−free
high clustering
Klemm−Eguiluz
Barabasi−Albert
regular lattice
path lengths
short average
small world
Watts−Strogatz
hierarchical model
Ravasz−Barabasi
random network
Erdos−Renyi
Figure 2.4.: overview over recent network models (Erd}os and R�enyi [28], Watts and Strogatz [16],
Barab�asi and Albert [30], Klemm and Egu
�
iluz [34], Ravasz and Barab�asi [35])
13
2. Network Models
14
3. Empirical Collaboration Network
In context of science, the network be-
tween scientists as nodes of graph is of par-
ticular interest.
3.1. Typology
First, we want to deal with the de�nition
of a collaboration graph. As to the nodes,
we have the choice to identify each vertex
either with an author or with a paper.
The second possibility is also area of re-
search [37], but we think studying the rela-
tionship of scientists as the paper's authors
o�ers more insight in how research works.
So we will make each scientist a node of
our network.
As what concerns the edges, there are ba-
sically two possible choices|both covered
equally by the database used [25].
3.1.1. Citation Graph
We might chose to consider citations from
one author to another as links [10], thus
resulting in a directed graph.
Starting at a given paper, we can enlarge
our network by following links recursively
up to a certain depth, e.g. by depth-�rst
or width-�rst algorithms. Each new work
will cite several to many still un-included.
Roughly spoken, the number of publications
to include will raise exponentially with the
maximum depth chosen.
Quickly, we arrive at huge amounts of
data. Additionally, there is no canonical
end of the hunt for new links. In the ex-
treme case we could be caught in a giant
cluster containing all or nearly all of the pa-
pers ever published. We see no possibility to
narrow this down in a reasonable way with-
out fear of introducing arbitrary boundary
conditions.
3.1.2. Collaboration Graph
Second possibility to de�ne edges of a graph
is creating links by co-authorship in one or
several papers [7, 8]. If n scientists pub-
lish a paper together, they are connected
to each other by
�
n
2
�
edges.
As an additional advantage, we have the
choice to start with an arbitrary set of au-
thors, establishing links between them by
looking at all papers they are involved. This
will result in a graph of limited size.
Of course, we should think carefully
about reasonable selection, to avoid edge
e�ects. We will discuss this in the next sec-
tion.
3.2. Building the net
3.2.1. Proceeding
As solution, we choose the following pro-
ceeding: We start with one paper. As one
part of our work will be the comparison
of real world data to Barab�asi-Albert net-
works, we take the corresponding paper [30]
as center of investigation.
15
3. Empirical Collaboration Network
In order to determine the set of authors
we want to deal with, we select all 185 pa-
pers that cite this paper.
1
Secondly, we
construct a list of unique authors from all
these papers. A �rst approach delivers 559
scientists, whereof some turned out to be
identical but appearing in particular papers
with typos. We �nish with a set of 555
authors to whom we attribute consecutive
numbers.
The last step of the network creation pro-
cess consists in establishing links between
all these authors. This is done by selecting
one paper after the other and introducing
connections between each possible pair of
this paper's authors (i.e.
�
n
2
�
links for n au-
thors).
Eventually, this gives us a graph of 555
nodes representing scienti�c collaboration
in the area of Barab�asi-Albert networks.
The network size is relatively small com-
pared to all data in the Science Citation
Index [25] (approx. 10
7
papers). Studying
properties of this subnet, we hope getting
an insight to what leads to the structure
observed. Veri�cation with bigger networks
remains a task for the future.
3.2.2. Visualization
To get an idea of what we are dealing about,
we visualize the graph using a spring model
[38, 39]. In order to give manageable results
we remove a paper on the Human Genome
Project [40] with 274 authors. Brief exami-
nation yields that this is no harm, as scien-
1
We have to be careful not to mix citation data
from di�erent dates as new papers are continu-
ously added to the database. Base of our inves-
tigation is October 21
st
, 2002.
tists participating in this work did not co-
operate with others in our graph, and form
a big cluster on their own. The result is
shown in �gure 3.1.
3.3. Analysis
3.3.1. Authors per paper
authors frequency
1 37
2 69
3 47
4 21
5 6
6 4
7 1
Table 3.1.: Authors per paper
First thing we are interested in is the fre-
quency distribution of papers per author.
We expect to see many papers with few au-
thors and vice versa (table 3.1).
0
10
20
30
40
50
60
70
1 2 3 4 5 6
frequency
number of authors
~x^3.58*exp(-x/0.54)~x^-3.3
Figure 3.2.: frequency distribution of the number
of authors per paper
16
3.3. Analysis
net of cooperation
4582498
201
40
1
15 124
296457
30
345
422
4
209
307
393
6
470
245
274
272
7
27
223
224
371477
41
102
105
134
338
359
398
425
453
482492
529
251
181
530
9
383
401
480
12
271
463
13312
334
434
17
52
18
107
154
177
20
37
164
429
433
24
308
350
99
314
26
385
38
267
408
290
400
39
318
43
386
47
405
454
484
49
14550
212
449
51
122
5587
270
57
62
315
389
58
204258
365
461111
504
156
226
61
220452
339
68
79
404
69
215
332
478
74
502
77
375
498
80
199202
222
336
81
92
82
306
84
128
187
554
85
185
86
117
317
94
129
95
264
479
552
96275
346555
101
179
110
113
292 431
243
378 488
116
165
329
418
120
171
219
436
472
512
123344
125
259
130
205
136
494
138
153160
280 384257
142
213406
143
197
144
246
343
370
373374
147
440
485
151309
395
163
237252
234255
253
254
283
166341
377
491
381
167
302416
169
320
217
239
218
269
230
423493
232233
282
235
278 430
435
241
263515
265
444
273
533
279
316
285
486
536
299390
466
301
322
319
450
542
361
380
372
421
376
469506
391
411
451
524
Figure 3.1.: collaboration network with 555 nodes (plotted using GraphViz package [38, 39])
17
3. Empirical Collaboration Network
Indeed, the considered graph (�gure 3.2)
shows this behavior with one remarkable ex-
ception. There are much too few papers
written by only one author. This could be
due to the fact that collaboration helps in
science, but the more (scienti�c!) partners
you have, the slower gets your scienti�c out-
put as communication overhead increases.
In other words: establishing scienti�c re-
lationships with other authors is not easy.
You have to agree on the �eld of research,
coordinate your e�orts etc. Postulating
that cooperation with more scientists is al-
ways favorable, we can explain the statistics
by diÆculty of �nding new partners. This
even increases corresponding to the num-
ber of co-workers you already have, as ad-
ditional coordination is needed. The risk of
research overlap raises, too.
The power law predicted by Lotka [11]
with an exponent of �2 cannot be con-
�rmed. This could be due to insuÆcient
statistics for this test. Other recent stud-
ies of collaboration networks found an ex-
ponent of 2:1 or 2:4 [41] which is another
indication for statistical errors predominat-
ing our results of study.
3.3.2. Connections per author
Next, we study the number of connections
per author, which is the number of other
scientists an author ever published papers
with. This number is weighted by the num-
ber of papers, i.e. a coauthor with whom a
scientist published n papers contributes n
connections (table 3.2).
Again, we expect to see a frequency de-
crease with increasing number of connec-
tions. The experimental data (�gure 3.3)
links weighted unique
0 16 16
1 51 67
2 75 69
3 61 67
4 24 27
5 21 22
6 10 2
7 5 6
8 4 3
9 2
10 3
11 1
13 2 1
14 1
15 1
17 1 1
20 2
29 1
273 274 274
Table 3.2.: Connections per author
0.01
0.1
1
10
100
1 10
frequency
number of connections
x^-2.85x^-3.53
Figure 3.3.: frequency distribution of the num-
ber of connections per author
(black:weighted|grey:unique)
shows smaller frequencies for \isolated" au-
18
3.3. Analysis
thors that never publish with others as well
as for authors with only one coauthor. This
is comparable to the e�ect observed in the
last graph. The most productive seem to be
authors with two or three colleagues they
are working with.
Statistical data in the area of highly con-
nected authors shows a truncated power
law. The exponent of approximately �2:85
falls well in the region bounded by analysis
of other scale-free networks (www: around
2:3 [6, 26, 42]). The sharp or exponen-
tial cuto� at very high connection numbers
has been reported for other networks, too
[7, 43]. Mossa et al. [44] o�er an expla-
nation using a model with limited (local)
information on the network. Surely, no sci-
entist knows all others, so this could lead
to the observed e�ect.
3.3.3. Double vs. unique links
We are interested how things change when
we cease weighting connections by num-
ber of papers published together, i.e. we
only take into account how many other
unique scientists a researcher published pa-
pers with. Results can be found in table 3.2.
We see that despite the di�erent num-
bers, results are qualitatively the same. Sci-
entists working together with two other au-
thors are the most productive.
Depending on whether your glass is half
full or half empty there are two contrary
explanations:
1. It is common practice that you name
persons as authors of your work that
did not contribute to it, out of a feeling
of debt, may it be sponsors or others.
2. Science lives from cooperation. Work-
ing together on one subject increases
scienti�cal output whilst reducing er-
rors.
The author of this paper will not judge.
3.3.4. Cluster sizes
size frequency
1 16
2 28
3 15
4 13
5 2
6 3
7 2
8 2
9 2
10 1
26 1
274 1
Table 3.3.: Cluster sizes
Our last focus is on subnets of science
that exist in our net of collaboration. Au-
thors group into several clusters by connec-
tions established between them. We inves-
tigate the frequency of clusters of a given
size. Our expectation is getting a frequency
increase for growing cluster sizes up to a
peak, and then a decay as clusters grow
very big, comparable to the statistics we
saw already.
The experimental data (table 3.3, �g-
ure 3.5) shows this behavior, but with one
surprise: although the most frequent cluster
size is 2 due to a big number of publications
19
3. Empirical Collaboration Network
338
359
398
425
453
482
492529
251
181
530
80
199202
222336134
7
27223
224
371477
41
102
105
1
15
124
296
457
30
345
422
8 341
377
491
381
285
486
536
458
249
8
201
40 6
470
245274 272
10
160
280
384
257
299
390
466 151
309
395
142
213
406
153
120
171
219436
472
512138
317
2
3
4
5
26
a b
a b
a b
ab
b
a
ba
c
6
b
a
117
77375
498
230423
493 167
302
416
86
253
254
283
282
255163
237
252
234
144246
343
370
373
374
58204
258
365 461
111504
156
226
9
166
5762
315
38918
107
154
434
61220
452
339
84
128
187
554
13
312334
177
9
383
401
480
235 278
430 435
69
215
332
4784
209
307
393
74
502 391
411169
320101
179 372421
26
385
319
301
322 217
23949
145232
233
123
344 451
524
218269
450
449 12271
463
241
263515
68
79 404
212542
147 440
485
55
87
270
376
469
506
50
95
533
17
52361
380 85
185
273494 386
43
39 318
265
444 82306
125
259 143
197
110
7
129
113
292431
243
378488
38
267
408
290
400
94
136
264479
552 96
275
346555
47
405
454484
24
308
350
99314
8192
279
316
51
122 130
205
116
165329
418
20
37164
429
433
Figure 3.4.: Collaboration clusters, ordered by size of participants. The shaded box no. 458 represents
my professor, D. Stau�er.
20
3.4. Comparison
0.01
0.1
1
10
1 10
frequency
cluster size
x^-2.61
Figure 3.5.: frequency distribution of the cluster
size
with two authors, most scientists maintain
collaboration with three others.
The can only be due to scientists be-
ing member of research groups involved in
di�erent themes, thus connecting di�erent
clusters formed by single papers. This can
directly be veri�ed by the graphical repre-
sentation (�gure �gure 3.4) of the clusters,
ordered by size.
3.4. Comparison
We have collected some statistical �gures
to express the structure of the network in
concern. Now we want to �nd out whether
classical or current network models bear
similar results.
3.4.1. Erd}os-R�enyi random graphs
Connections per author In a random
graph the distribution is a binominal one,
i.e. the probability of a node with k con-
nections in a net with N nodes is
P (k) /
N � 1
k
!
p
k
(1� p)
N�1�k
:
In the limit of large N this approaches a
Poisson distribution around the expectation
value hki = pN
P (k) / e
�hki
hki
k
k!
:
This is contrary to the power-law statistic
observed in the collaboration network.
Cluster sizes For random graphs, perco-
lation theory predicts that the cluster size
distribution shows an exponential decay for
big cluster sizes [2, 45].
Again, the considered network show
rather a power-law decay than an exponen-
tial one.
Unsurprisingly, the structure of scienti�c
collaboration di�ers basically from that of
a random network.
3.4.2. Watts-Strogatz small-world
networks
Connections per author The degree
distribution of Watts-Strogatz small-world
networks is similar to that of a random
graph [2]. It has a peak and decays ex-
ponentially for large connection numbers,
contrary to the collaboration network.
Cluster sizes The usual case in Watts-
Strogatz networks is re-wiring of only a
small portion of links. Thus, the network
stays well connected, mostly forming one
giant cluster.
21
3. Empirical Collaboration Network
This model does not describe scienti�c
collaboration as well.
3.4.3. Barab�asi-Albert networks
Connections per author Barab�asi-
Albert networks show a vertex degree
distribution as P (k) / k
�3
[30, 46].
In our science collaboration network we
found out exponents of 2:85 resp. 3:53
which is only a slight deviation.
Cluster sizes In Barab�asi-Albert net-
works, new sites are added with links to
already existing nodes. Consequently only
one giant cluster forms. Obviously this dif-
fers crucially from the net of co-authorship.
22
4. Spin models
4.1. Leadership e�ect
4.1.1. Ising model
In 1925, Ising [47] published a paper on a
model of spin interaction that later became
very famous. The idea to this had been
given to him by his teacher Lenz [48], so it is
sometimes referenced as Lenz-Ising model.
1
The idea is to consider spins (e.g. on a
square lattice) and an interaction Hamilto-
nian
H = �
X
i6=j
J
i;j
S
i
S
j
where S
i
are the spins and J
i;j
is a matrix
describing the interaction forces. Usually
we consider the case
S
i;j
=
(
J : i; j nearest neighbors
0: else,
i.e. only allow equal interaction between
nearest neighbors (J > 0 for ferromagnetic
behavior).
In the following chapter, we investigate
how such a model behaves on our con-
structed collaboration network.
1
A generalization of the Ising model is the Potts
model [49, 50]. Instead of Ising spins with two
possible states +1 and �1, Potts allows k � 2
di�erent spin values. The Hamiltonian is
H = �J
X
i6=j
Æ
i;k
;
i; j being nearest neighbors.
Applying it to the scienti�c network, I �nd re-
sults very similar to those of Ising's model.
Figure 4.1.: Ising [47]
We use a Metropolis [51] Ising model,
i.e. probabilities for a single spin ip of
p / e
�
�E
=k
B
T
if �E > 0 and 1 otherwise.
To determine �E we sum up spins of all
vertices connected to a given node. In prin-
ciple, we have the choice between two pro-
ceedings:
� consider only unique links between two
nodes
� count a connection several times ac-
cording to the number of links, i. e.
the number of papers the correspond-
ing authors published together.
Both possibilities have been examined.
2
2
change the switch NODOUBLE in line 17 of source
23
4. Spin models
4.1.2. Phase transition
The results of both experiments show qual-
itative similarity. We observe a rounded
phase transition at about
k
B
T
=J = 0:8 op-
posed to a value of about 2:3 on the regular
square lattice.
A closer look reveals that decay of mag-
netization with raising temperature is ex-
ponential. This result corresponds to re-
search on Ising models on Barab�asi-Albert
networks by Aleksiejuk, Ho lyst, and Stauf-
fer [52], who also found an exponential law.
Anyhow, the critical temperatures found
by me and by Aleksiejuk et al. [52] di�er
by more than one order of magnitude. This
can easily be explained by di�erent coor-
dination numbers in both networks. The
collaboration graph holds a maximum of 14
neighbors of a single vertex, the graph of
Aleksiejuk et al. [52] exceeds this by sev-
eral orders of magnitude. This makes it far
more \diÆcult" to break the ferromagnetic
bonds, resulting in a higher critical temper-
ature.
4.1.3. Degree distribution
Dorogovtsev et al. [53] studied random
graphs with given degree distributions P (k)
of a vertex of degree k. They deduced an
estimate for the critical temperature of an
Ising model on such networks as
J
k
B
T
c
=
1
2
ln
hk
2
i
hk
2
i � 2hki
!
:
Considering the collaboration network as
a random graph with given degree distribu-
tion (table 4.1), we use their formula and
code in section B.1
degree several unique
0 16 16
1 51 67
2 75 69
3 61 67
4 24 27
5 21 22
6 10 2
7 5 6
8 4 3
9 2
10 3
11 1
13 2 1
14 1
15 1
17 1 1
20 2
29 1
Table 4.1.: Degree distribution
get
T
c
=J = 2:91, counting only unique links
between scientists. Using all links, we �nd
T
c
=J = 5:46.
Both values are far from critical temper-
atures observed in simulation. This is a
strong clue towards the statement that col-
laboration networks are crucially di�erent
from random networks, even with the same
degree distribution.
4.1.4. Spin ip model
Following a suggestion of Ho lyst
3
, we can
determine the importance of most con-
nected authors of our collaboration network
by successive ipping of most connected
3
personal correspondence, cf. [52]
24
4.2. Cluster limited Ising models
spins and pinning them in their new posi-
tion.
In other words: After some time of equili-
bration, we chose the author who has most
connections to others, and change his/her
spin permanently to a value of �1, opposite
to all others (at T = 0, or nearly all others
else). Subsequently, we allow the system
to relax some time, after which we perma-
nently ip the second most connected spin,
and so on.
0
50
100
150
200
250
0 50000 100000 150000 200000 250000 300000
M
t
uniquemultiple
Figure 4.2.: Ising model with successive spin ips.
After 10
5
steps of equilibration, we
ip the most connected spin and
stick it to its new value. After some
relaxation of 10
4
steps, this step is
repeated. Network with multiple and
network with unique links used. Av-
eraged over 1000 runs. T = 0:2.
Results are shown in �gure 4.2. We ob-
serve two things:
1. Even after switching 20 most con-
nected spins, the system does not ip
in the opposite state with all spins
pointing down. In simulations of Alek-
siejuk et al. [52], less than 6 spins
were enough to ip a whole network
of 30; 000 nodes.
This is quite obviously due to the fact
that we don't have a contiguous graph,
but one consisting of di�erent clusters.
A spin ip in one cluster is not able to
a�ect spins in others.
In pictures of spins representing opin-
ions (yes/no, etc. [54]), this means a
few authors with view di�ering from
the broad mass of scientists are hardly
capable of changing the global opinion,
may they even be the most connected
(known) ones.
2. Allowing multiple links in our net,
we expect the magnetization to break
down much faster, as a ip of a spin is
able to in uence others in a stronger
way.
Yet, the simulation shows contrary
results. The graph containing only
unique links shows a much steeper de-
cay of magnetization (�gure 4.2).
Possible explanation is the fact, that
choosing most connected spins in a
network only having unique links picks
authors with connections to many
other authors, whereas in a network al-
lowing multiple links, there are as well
spins connected in a strong manner to
only few others.
It seems that in order to spread new
opinions, it is more advantageous to
have small in uence on many other
people than a big impact on only few
ones.
4.2. Cluster limited Ising models
The network we are looking at consists of
many distinct clusters of di�erent sizes (�g-
25
4. Spin models
ure 3.4). We may ask if they di�er regard-
ing their properties, or if they behave alike.
4.2.1. Proceeding
We split up the network into sub-nets, each
containing one cluster, numbered sequen-
tially, 2, 3a, 3b,. . . 26 (numbers and letters
as in �gure 3.4).
On each net, we run an Ising model
for temperatures from 0:1 to 6:9 in 0:2
steps. Each of these simulations runs for
10
6
steps, with magnetization measured ev-
ery 100 steps, thus resulting in 10
4
mea-
surements per run, to give good statistics.
4.2.2. Results
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
M/N
kT/J
Figure 4.3.: Ising model on the di�erent clusters
of a collaboration net, averaged over
10
4
measurements per temperature
and net.
Examining the results (�gure 4.3), we see
di�erent curves that all decay with raising
temperature, but show no apparent similar-
ities. We wonder why the curves seemingly
do not converge to zero but to �nite values.
Obviously, the network is so small that
macroscopic magnetization ips occur fre-
quently, even at moderate temperatures.
That means, expectation value of magne-
tization
4
at high temperatures is not zero,
but something around one!
4.2.3. Bias adjusti�cation
To validate this hypothesis, we simulate the
networks at very high temperature (
k
B
T
=J =
50), in order to determine M
1
= M(T =
1) (table 4.2).
net M
1
2 1.03
3a 1.71
3b 1.51
4a 1.59
4b 1.54
4c 1.55
5a 2.02
5b 1.96
6a 2.00
6b 1.94
7a 2.27
7b 2.23
8a 2.26
8b 2.29
9a 2.55
9b 2.55
10 2.54
26 4.25
Table 4.2.: Bias
4
all over this publication we consider the (un-
signed) value jM j as magnetization, not M !
Doing the latter leads to false results. E.g. at
low temperatures, averaging M over very long
times would give zero, as ips of the whole sys-
tem occur (though at very low probabilities).
26
4.2. Cluster limited Ising models
Using these values, we rescale our simu-
lation results from �gure 4.3 by use of the
scaling function
5
M
N
�!
M �M
1
N �M
1
;
where N is the total number of authors in
the cluster (�gure 4.4).
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
(M-M
∞)/
(N-M
∞)
kT/J
23a3b4a4b4c5a5b6a6b7a7b8a8b9a9b1026
Figure 4.4.: Same data as in �gure 4.3 but ad-
justed to a common scale from 0 to
1 by eliminating bias from random
uctuations.
We �nd a much cleaner image. Apart
from one exception (6b) all curves are par-
allel up to the M = 0:5-line, and even be-
yond there are very few crossings.
Each di�erent cluster can now be char-
acterized by the temperature # at which it
achieves M = 0:5. This leads to a sort of
\melting temperature" (table 4.3).
To get an idea of this temperature's
meaning, we sort the cluster's graphical rep-
resentations by # (�gure 4.5).
Is seems plausible that # is a measure
of coherence or connectiveness of a clus-
5
This function is a linear approximation that gives
1 for M ! N and zero for M !M
1
.
net #
2 1.7
3a 2.8
3b 1.5
4a 3.3
4b 2.1
4c 2.6
5a 4.0
5b 2.2
6a 4.8
6b 3.5
7a 2.8
7b 1.8
8a 2.5
8b 3.7
9a 2.2
9b 3.1
10 1.9
26 3.8
Table 4.3.: Melting temperatures
ter. Single bonds lead to lower melting tem-
peratures, fully connected subsets to higher
ones.
This o�ers a possible explanation why 6b
shows a di�erent behavior from all others
in �gure 4.4: This cluster consists of two
parts. One of them is a completely con-
nected set of �ve nodes, the other one a
single vertex. Both are linked by one single
edge. Probably, this \con ict of interest"
leads to the observed anomality.
Additionally, connectedness described by
# seems to be crucially di�erent from the
classi�cation given by standard clustering
coeÆcient. E.g. the net consisting of three
completely connected vertices 3a yields a
clustering coeÆcient of 1 but a low melting
27
4. Spin models
1.5
2.0
2.5
3.0
3.5
4.0
4.5
223
7
27
224
371
105
336222
202199
80
530
181
251
529492
482
453
425
398
359
338
134
102
41
477
504 111
461365
258204
58
156
466
390
299
488378
243
431292
113
110
226
422
345
30
457
296
124
15
1
554
187
128
84
257
384
280
160153
138
536
486
285
381
491
377
314
99
350
308
24
282
283
254
253
255
234
252
237
163
341166
418
329165
116
374
373
370
343
246144
542
450
433
429
16437
20
319
640
201
8
249
458
339
452
220 61
470
322
301
129 94
400
290
408
267
38
272274
245
Figure 4.5.: collaboration clusters from �gure 3.4, ordered by #
28
4.2. Cluster limited Ising models
temperature.
4.2.4. Linear Relationship
Surprisingly, we �nd a linear relationship be-
tween N and
E
=# (�gure 4.6). Thus, we
postulate
E
#
= aN � b
and conclude a formula for #:
#
calc
(E;N) =
E
aN � b
:
Fitting parameters to our measurements
(yielding a = 0:72; b = 0:89), this gives
good prediction of melting temperatures.
Results can be seen in �gure 4.7, as well
as a diagram showing errors being inferior
to 10% in most cases.
#
calc
=
E
aN � b
N!1
�!
hki
2a
:
We see that, in the limit of high N , the
melting temperature # is proportional to
the average number of edges per site hki =
2E
=N, a result known from mean �eld the-
ory.
29
4. Spin models
0
2
4
6
8
10
12
14
16
18
20
0 5 10 15 20 25 30
E/ϑ
N
0.76N-1.0
Figure 4.6.:
E
=# vs. N shows a surprisingly linear correlation.
1.5
2
2.5
3
3.5
4
4.5
5
1.5 2 2.5 3 3.5 4 4.5 5
ϑ c
alc
ula
ted
ϑ measured
a=0.72; b=0.89
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
2 4 8 16
∆ϑ
N
a=0.72; b=0.89
Figure 4.7.: # determined by #
calc
(E;N) =
E
aN�b
vs. measurements. E is the cluster's total number
of edges.
30
5. Barab�asi-Albert network models
5.1. Modi�ed Barab�asi-Albert
model
Network model of Barab�asi and Albert [30]
was introduced in section 2.5. We pointed
out that it shows rather good �ts with em-
pirical networks, but lacks support for dis-
jointed ones, as the algorithm only delivers
one giant cluster.
Thus, to cope with networks consisting
of several components, we must modify the
model. We chose a very simple approach:
In each step of adding nodes, we start a new
cluster of m
0
= 3 nodes with probability p.
Vertices added in consecutive time steps
can connect to any node in any component
respecting the same probability rule as in
the standard model.
� In the case m = 1, components can
only grow (isolated clusters), whereas
� in the case m > 1, new nodes are able
to connect two or more existing com-
ponents of the network (merging clus-
ters).
5.2. Simulation
In order to compare simulation results with
real-world data from a scienti�c collabora-
tion network [55], we let the network grow
up to the size of 555 nodes. For proper
statistics, this is repeated 10
4
times.
5.2.1. Isolated clusters
scale-free behavior In case of m = 1,
i.e. considering isolated clusters, we can be
sure to get scale-free behavior within the
distinct clusters, as the probabilities for at-
tachment of a new node to an existing one
are the same as in a single Barab�asi-Albert
network (modulo a constant factor due to
a new node having the \choice" between
di�erent clusters to connect to).
However, complete network is a priori not
necessarily scale-free, as total statistics is
made up by the sum of all scale-free sub-
networks or clusters. So, we have to focus
later on the question, if scale-free behavior
prevails.
cluster size distribution Next, we exam-
ine the number of clusters of di�erent sizes
(�gure 5.1). Obviously, we �nd that high
probability of starting a new net leads to
many smaller networks, whereas low values
privilege bigger networks. Yet, we make
an interesting observation: low probabili-
ties lead to a cluster-size distribution that
is not monotonic any more, but favors big
networks.
Looking at �gure 5.1 which shows the
number of points in clusters of a given size
instead of the sheer cluster count, makes
this more plausible.
� For p = 0, we will see a graph /
Æ(555), as there is only one giant clus-
ter,
31
5. Barab�asi-Albert network models
1
10
100
1000
10000
100000
1e+06
10 100
fre
qu
en
cy
cluster size
0.010.040.10.40.8
100
1000
10000
100000
1e+06
10 100
no
de
co
un
t
cluster size
0.010.040.10.40.8
Figure 5.1.: Frequency of clusters (left) resp. number of nodes in clusters of a given size (right) vs.
cluster size at di�erent probabilities for a new net. Simulation was run 10
4
times with a
network growing up to 555 nodes. The curve for p = 0:01 is the one with the rightmost
peak; to the left follow the other p-values in descending order.
� for p = 1 a graph / Æ(m
0
= 3), be-
cause there are only embryonic sub-
nets.
� What we observe for 0 < p < 1 is the
transition between both extremes.
1
2
4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
exponent of pow
er
law
regio
n
probability for a new net
-exponente
2.25x
Figure 5.2.: Negative exponent of the power law
part of the curves in �gure 5.1 vs.
probability p for a new net. The line
corresponds to exponent = �e
2:25p
.
power law region For all p, we start with
a power law region, regarding the distribu-
tion of small and medium cluster sizes. The
exponent varies with network-birth proba-
bility p. In the semi-logarithmic plot in �g-
ure 5.2 it is shown, that the exponential
relation �e
2:25p
describes our data rather
well. Of course, this formula cannot be
true for general p as for p ! 1 we expect
m! �1!
Regarding empirical data from sections
3.2's collaboration graph, we �nd good
overlap with the model (�gure 5.3). In-
terestingly, the model is even able to ex-
plain facts formerly regarded as statistical
anomalies, as the observation of a giant
cluster of a size exceeding largely all oth-
ers in the network (section 3.2.2).
5.2.2. Merging clusters
Now, we modify the model by examining
m > 1. In this case, newly added vertices
develop several links to existing nodes (and
thus existing clusters), being able to con-
32
5.2. Simulation
0.01
0.1
1
10
1 10 100
fre
qu
en
cy
total authors in clusters of given size
collaboration networkp=4%’
Figure 5.3.: Semi-logarithmic plot comparing the simulation with p = 0:04 using the isolated clusters
model of �gure 5.1 and statistical data from a science collaboration network (section 3.2).
nect hitherto separated networks. In this
paper, we limit our considerations on the
standard Barab�asi-Albert case m = m
0
=
3.
Using di�erent p, we quickly recognize
that low and medium probabilities make
the simulations nearly always end up with a
single giant cluster containing all vertices.
Points of interest are higher p in the region
of 60{90%.
cluster size distribution Again, we plot
the total number of nodes contained in clus-
ters of a given size (�gure 5.4). For small
cluster sizes, we observe an non-uniform be-
havior of the graph. The explanation is as
follows: newly born clusters have a size of
m
0
= 3 and thus appear very often. Also,
cluster of sizes 4 or 7 are very probable,
whereas a cluster of size 5 is very rare, be-
cause it can only be formed by a new clus-
ter to which two new ones have connected
without gluing it to a second cluster.
In a semi-logarithmic plot (�gure 5.4), we
�nd a parabolic dependence for high cluster
sizes (i.e. a Gaussian distribution around a
mean depending on p). Appearently, the
merging clusters cannot cope with reality.
33
5. Barab�asi-Albert network models
1
10
100
1000
10000
100000
1e+06
2 4 8 16 32
fre
qu
en
cy
degree
p=60%p=80%
1
10
100
1000
10000
100000
1e+06
1e+07
0 5 10 15 20 25 30 35
fre
qu
en
cy
degree
p=60%p=80%
Figure 5.5.: Frequency of nodes with a certain degree. Simulation was run 10
4
times with networks
growing up to 555 nodes. Left plot is linear, right plot semi-logarithmic. M = 1 (isolated
clusters).
1
10
100
1000
10000
100000
1e+06
1e+07
2 4 8 16 32 64 128
fre
qu
en
cy
degree
p=1%p=40%
1
10
100
1000
10000
100000
1e+06
1e+07
0 20 40 60 80 100 120 140 160 180 200
fre
qu
en
cy
degree
p=1%p=40%
Figure 5.6.: Same plots as �gure 5.5 using M = 3 (merging clusters).
5.2.3. scale-free behavior
In �gure 5.5 we can see that there is no
pure scale-free behavior. There seems to
be power-law behavior for small degrees and
an exponential cuto� (�gure 5.5) at higher
values. Similar results have been observed
by Newman [7] for collaboration networks.
One could argue that this e�ect is due
to the fact that we do not plot the degree
distribution for single clusters, but for the
whole set of them. This demur only counts
at �rst sight, though. At p = 80% we have
several small clusters but virtually only one
giant cluster dominating the degree distri-
bution for high degrees. So, the fact of
averaging of many di�erent sized clusters
should manifest mainly in the area of small
degrees opposite to our observations.
Mossa et al. [44] o�er a possible expla-
nation for the exponential cuto� encoun-
tered. They use a model which attributes
to each node only a restricted knowledge
on the network, i.e. the vertex is not able
34
5.2. Simulation
10
100
1000
10000
100000
1e+06
1e+07
50 100 150 200 250 300 350 400
node c
ount
cluster size
p=70%p=80%p=90%
Figure 5.4.: Number of nodes in clusters of a
given size vs. cluster size at di�erent
probabilities for a new net. Simula-
tion was run 10
5
times with a net-
work growing up to 555 nodes.
to consider the whole graph's structure, but
only a subset according to its limited view.
M p ln(N
0
) k �.
1 1% 16.8 2.27 60.
1 40% 17.1 2.45 9.2
3 60% 18.1 3.1 5.1
3 80% 20.1 4.8 3.5
Table 5.1.: CoeÆcients for �gure 5.7
35
5. Barab�asi-Albert network models
1
10
100
1000
10000
100000
1e+06
4 8 16 32 64 128
frequency
degree
M=1; p= 1%M=1; p=40%M=3; p=60%M=3; p=80%
Figure 5.7.: Plots of �gure 5.5 and �gure 5.6 show good �t to exponentially truncated power laws
y = N
0
x
�k
e
�
x
=�
.
36
6. Conclusion
Dealing with real-world networks, scien-
tists found three properties predominating:
� short average path lengths (Small
World E�ect),
� scale-free behavior,
� high clustering.
Di�erent models were developed to cope
with this challenge, each having di�erent
advantages and disadvantages. The model
of Barab�asi and Albert [30] is a promis-
ing one, but lacks support for discontiguous
networks.
We constructed a network of co-
authorship with 555 authors. Only scien-
tists were chosen that cite a speci�c pa-
per [30]. The resulting net shows scale-free
characteristics but di�ers substantially from
accepted computer models' results.
Simulating Ising models on the network
reveals strong robustness against distur-
bances (spin ip experiment/leadership ef-
fect) and shows coherence with mean �eld
theory : We �nd the critical temperature of
subnets of our graph being proportional to
the average number of edges per site, in the
limit of a high node count.
In order to overcome the mentioned dis-
advantages of the Barab�asi-Albert model,
we developed a modi�ed version, allowing
formation of multiple clusters. We saw a
strong dependence of a node's edges count
on the network structure, separating two
cases: isolated clusters and merging clus-
ters.
Only the �rst case leads to results �t-
ting reality. Comparison with statistics from
our collaboration net shows similar behavior
and is even able to explain facts at �rst re-
garded as statistical anomalies as the obser-
vation of a giant cluster of a size exceeding
largely all others in the network. Even ex-
ponential cuto� of nodes with high degrees,
as encountered empirically, is reproduced.
Re-evaluating the model with a higher
number of authors would lead to better
statistics and greater reliability.
37
6. Conclusion
38
A. Acknowledgements
I would like to thank D. Stau�er
1
for giv-
ing me the idea of the subject and support-
ing me by comments and discussions during
my research.
Thanks to M. Abd-Elmeguid
2
for co-
judging this work.
My thanks for writing excellent software
goes to the authors of
� L
A
T
E
X, BibT
E
X, pdfL
A
T
E
X, dvips
� L
A
T
E
Xpackages: KOMAScript, natbib,
custom-bib, graphicx, listings, units,
hyperref, hypernat, colortbl, color
� SciTE
� gnuplot
� Ruby, C++, Perl, bash
� dotty, neato[39]
Thanks to A. Sindermann for supporting
my work by addressing computer network
problems.
Special thanks to K. Godthardt for sup-
porting me.
1
Institute of Theoretical Physics, University of
Cologne
2
II. Institute of Experimental Physics, University
of Cologne
39
A. Acknowledgements
40
B. Source code
B.1. Ising model
This C++ program simulates an Ising model on a given graph. In section 4.1.1 it was
used on our collaboration network.
// This program reads network data from a file and simulates
// an Ising model an this graph.
// Metropolis probabilities are used.
//
5 // Felix Puetsch <[email protected] -koeln.de >, 2003-01-22
#inc lude < i o s t r eam>
#inc lude < f s t ream>
#inc lude < s t d i o . h>
10 #inc lude < a s s e r t . h>
#def ine MAX INT 2147483647
#def ine MAX CONN 30
15 #def ine MAX NODE 600
#def ine NODOUBLE 1
// #define NOASSERT
20
us ing namespace s td ;
// === random number generator ================================
25 c l a s s Random f
p r i v a t e :
i n t s t a t e ;
pub l i c :
Random( i n t seed ) ;
30 i n t get ( ) f r e tu rn s t a t e �=65539 ; g // 16807
g ;
Random : : Random( i n t seed ) f
a s s e r t ( seed % 2 == 1) ;
35 s t a t e = seed ;
g
// === Vertex =================================================
41
B. Source code
40 c l a s s I s i n g ;
c l a s s Ver t ex f
p r i v a t e :
i n t number , conn count , s p i n ;
45 Ver t ex � ne i ghbou r [MAX CONN ] ;
I s i n g � i s i n g ;
pub l i c :
Ve r t e x ( I s i n g � i s , i n t nr ) ;
~ Ve r t ex ( ) ;
50 i n t ge tSp in ( ) f r e tu rn s p i n ; g
vo id s e t Sp i n ( i n t s ) f s p i n = s ; g
vo id addConn ( Ve r t ex � to , i n t nodoub le = 0) ;
i n t getNumber ( ) f r e tu rn number ; g
i n t getConnCount ( ) f r e tu rn conn count ; g
55 Ver t ex � getConn ( i n t i ) ;
i n t s imu l a t eS t ep ( ) ;
g ;
// === Ising ==================================================
60
c l a s s I s i n g f
pub l i c :
Random � rnd ;
i n t e n l i m i t [ 2�MAX CONN+1] ;
65 p r i v a t e :
i f s t r e am net ;
char b u f f e r [ 8 0 ] ;
i n t v count ;
Ve r t ex � v l i s t [MAX NODE ] ;
70 pub l i c :
I s i n g ( char � fname ) ;
vo id b u i l d n e t ( ) ;
vo id debug ( i n t nr ) ;
vo id r e s e t ( ) ;
75 vo id s imu l a t e ( double kT , i n t maxtime=�1, i n t s t e p t ime=0) ;
g ;
// === Vertex =================================================
// ... Struktors .............................................
80
Ver t ex : : Ve r t ex ( I s i n g � i s , i n t nr ) f
conn count = 0 ;
i s i n g = i s ;
number = nr ;
85 g
Ver t ex : : ~ Ve r t ex ( ) f
cout << "~ Ve r t ex " << end l ;
g
90
vo id Ver t ex : : addConn ( Ve r t ex � to , i n t nodoub le ) f
i f ( nodoub le )
f o r ( i n t i =0; i<conn count ; i++)
42
B.1. Ising model
i f ( ne i ghbou r [ i ]==to ) r e tu rn ;
95 ne i ghbou r [ conn count++] = to ;
a s s e r t ( conn count < MAX CONN) ;
g
Ver t ex � Ver t ex : : getConn ( i n t i ) f
100 a s s e r t ( i < conn count ) ;
r e tu rn ne i ghbou r [ i ] ;
g
i n t Ver t ex : : s imu l a t eS t ep ( ) f
105 i n t sp insum=0;
f o r ( i n t i =0; i<conn count ; i++)
spinsum += ne ighbou r [ i ]�>ge tSp in ( ) ;
sp insum �= sp i n ;
i f ( i s i n g �>rnd�>get ( ) < i s i n g �>e n l i m i t [ sp insum+MAX CONN] )
110 s p i n �=�1;
r e tu rn s p i n ;
g
// === Ising ==================================================
115 // ... Struktors .............................................
I s i n g : : I s i n g ( char � fname ) f
cout << " I s i n g " << end l ;
rnd = new Random(1) ;
120 net . open ( fname ) ;
i f ( ! net . i s o p e n ( ) ) f
c e r r << " i npu t f i l e not found " << end l ;
e x i t ( 1 ) ;
g
125 b u i l d n e t ( ) ;
f o r ( f l o a t kT=50; kT<50 .1 ; kT+=0.2) f
c e r r << " s t a r t i n g s imu l a t i o n wi th kT=" << kT << end l ;
r e s e t ( ) ;
s imu l a t e (kT , 1 000000 , 1 00 ) ;
130 g
g
// ... Methods ...............................................
135 vo id I s i n g : : b u i l d n e t ( ) f
cout << " b u i l d n e t " << end l ;
f o r ( i n t i =0; i<MAX NODE ; i++) f
v l i s t [ i ] = new Ver t ex ( th i s , i ) ;
v l i s t [ i ]�> s e t Sp i n (1 ) ;
140 g
i n t from , to ;
whi le ( ! net . e o f ( ) ) f
net . g e t l i n e ( bu f f e r , 8 0 ) ;
i f ( s s c a n f ( bu f f e r , "%i �� % i ; " , & from , & to ) != 2) f
145 c e r r << " i npu t l i n e i g no r e d : ' " << b u f f e r << " ' " << end l ;
cont inue ;
g
43
B. Source code
a s s e r t ( from < MAX NODE) ; a s s e r t ( to < MAX NODE) ;
v l i s t [ from]�>addConn ( v l i s t [ to ] , NODOUBLE) ;
150 v l i s t [ to]�>addConn ( v l i s t [ from ] , NODOUBLE) ;
g
net . c l o s e ( ) ;
v count = 0 ;
f o r ( i n t i =1; i<MAX NODE ; i++)
155 i f ( v l i s t [ i ]�>getConnCount ( )>0) v count=i ;
e l s e v l i s t [ i ]�> s e t Sp i n (0 ) ;
c e r r << v count << " nodes " << end l ;
g
160 vo id I s i n g : : debug ( i n t nr ) f
Ver t ex � node = v l i s t [ n r ] ;
i n t nb = node�>getConnCount ( ) ;
cout << "node " << nr << " has " << nb
<< " conn e c t i o n s : " << end l ;
165 f o r ( i n t i =0; i<nb ; i++)
cout << node�>getConn ( i )�>getNumber ( ) << end l ;
g
vo id I s i n g : : r e s e t ( ) f
170 f o r ( i n t i =0; i<MAX NODE ; i++)
v l i s t [ i ]�> s e t Sp i n ( abs ( v l i s t [ i ]�>ge tSp in ( ) ) ) ;
g
// ... Simulation .............................................
175
vo id I s i n g : : s imu l a t e ( double kT , i n t maxtime , i n t s t e p t ime ) f
i f ( ! s t e p t ime ) s t e p t ime=maxtime ;
f o r ( i n t i=�MAX CONN; i<=MAX CONN; i++)
e n l i m i t [ i+MAX CONN] =
180 ( i n t ) (MAX INT � ( 2� exp (�2.� i /kT)�1) ) ;
i n t mag , t ime=0;
whi le ( time<maxtime ) f
f o r ( i n t s t e p =0; s tep<s t e p t ime ; s t e p++) f
mag = 0 ;
185 f o r ( i n t nr =1; nr<=v count ; n r++) f
a s s e r t ( v l i s t [ n r ] !=NULL) ;
mag += v l i s t [ n r ]�> s imu l a t eS t ep ( ) ;
g
g
190 t ime += s t ep t ime ;
cout << t ime << " " << kT << " " << abs (mag) << " " << mag << end l ;
g
g
195 // === main ===================================================
i n t main ( i n t argc , char �� argv ) f
a s s e r t ( c e r r << "debug mode on" << end l ) ;
a s s e r t ( a rgc==2) ;
200 I s i n g i s i n g ( argv [ 1 ] ) ;
r e tu rn 0 ;
44
B.1. Ising model
g
45
B. Source code
B.2. Spin ip model
This C++ program simulates an Ising model on a given graph. In regular time intervals,
the most connected spins (hard coded) are pinned to an up position. In section 4.1.4
this was used on our collaboration network.
// This program reads network data from a file and simulates
// an Ising model an this graph.
// additionally , every ... time steps the next most connected
// spin is flipped permanently .
5 // Metropolis probabilities are used .
//
// Felix Puetsch <[email protected] -koeln.de >, 2003-02-05
#inc lude < i o s t r eam>
10 #inc lude < f s t ream>
#inc lude < s t d i o . h>
#inc lude < a s s e r t . h>
#def ine MAX INT 2147483647
15
#def ine MAX CONN 40
#def ine MAX NODE 600
#def ine NODOUBLE 0
20
// #define NOASSERT
us ing namespace s td ;
25 // === random number generator ================================
c l a s s Random f
p r i v a t e :
i n t s t a t e ;
30 pub l i c :
Random( i n t seed ) ;
i n t get ( ) f r e tu rn s t a t e �=65539 ; g // 16807
g ;
35 Random : : Random( i n t seed ) f
a s s e r t ( seed % 2 == 1) ;
s t a t e = seed ;
g
40 // === Vertex =================================================
c l a s s I s i n g ;
c l a s s Ver t ex f
45 p r i v a t e :
i n t number , conn count , sp i n , s t i c k y ;
Ve r t ex � ne i ghbou r [MAX CONN ] ;
46
B.2. Spin ip model
I s i n g � i s i n g ;
pub l i c :
50 Ver t ex ( I s i n g � i s , i n t nr ) ;
~ Ve r t ex ( ) ;
i n t ge tSp in ( ) f r e tu rn s p i n ; g
vo id s t i c k I t ( ) ;
vo id s e t Sp i n ( i n t s ) f s p i n = s ; g
55 vo id addConn ( Ve r t ex � to , i n t nodoub le = 0) ;
i n t getNumber ( ) f r e tu rn number ; g
i n t getConnCount ( ) f r e tu rn conn count ; g
Ver t ex � getConn ( i n t i ) ;
i n t s imu l a t eS t ep ( ) ;
60 g ;
// === Ising ==================================================
c l a s s I s i n g f
65 pub l i c :
Random � rnd ;
i n t e n l i m i t [ 2�MAX CONN+1] ;
p r i v a t e :
i f s t r e am net ;
70 char bu f f e r [ 8 0 ] ;
i n t v count ;
Ve r t ex � v l i s t [MAX NODE ] ;
pub l i c :
I s i n g ( char � fname ) ;
75 vo id bu i l d n e t ( ) ;
vo id debug ( i n t nr ) ;
vo id r e s e t ( ) ;
vo id s imu l a t e ( double kT , i n t maxtime=�1, i n t s t e p t ime =0 , i n t s t a r t t im e
=0) ;
g ;
80
// === Vertex =================================================
// ... Struktors .............................................
Ver t ex : : Ve r t ex ( I s i n g � i s , i n t nr ) f
85 conn count = 0 ;
s t i c k y = 0 ;
i s i n g = i s ;
number = nr ;
g
90
Ver t ex : : ~ Ve r t ex ( ) f
cout << "~ Ve r t ex " << end l ;
g
95 vo id Ver t ex : : addConn ( Ve r t ex � to , i n t nodoub le ) f
i f ( nodoub le )
f o r ( i n t i =0; i<conn count ; i++)
i f ( ne i ghbou r [ i ]==to ) r e tu rn ;
n e i ghbou r [ conn count++] = to ;
100 a s s e r t ( conn count < MAX CONN) ;
47
B. Source code
g
Ver t ex � Ver t ex : : getConn ( i n t i ) f
a s s e r t ( i < conn count ) ;
105 r e tu rn ne i ghbou r [ i ] ;
g
vo id Ver t ex : : s t i c k I t ( ) f
s t i c k y = 1 ;
110 s p i n = �1;
c e r r << number << " s t i c k e d . " << end l ;
g
i n t Ver t ex : : s imu l a t eS t ep ( ) f
115 i f ( s t i c k y !=0) r e tu rn s p i n ;
i n t sp insum=0;
f o r ( i n t i =0; i<conn count ; i++)
spinsum += ne ighbou r [ i ]�>ge tSp in ( ) ;
sp insum �= sp i n ;
120 i f ( i s i n g �>rnd�>get ( ) < i s i n g �>e n l i m i t [ sp insum+MAX CONN] )
s p i n �=�1;
r e tu rn s p i n ;
g
125 // === Ising ==================================================
// ... Struktors .............................................
I s i n g : : I s i n g ( char � fname ) f
cout << " I s i n g " << end l ;
130 rnd = new Random(1) ;
net . open ( fname ) ;
i f ( ! net . i s o p e n ( ) ) f
c e r r << " i npu t f i l e not found " << end l ;
e x i t ( 1 ) ;
135 g
b u i l d n e t ( ) ;
double kT=0.2 ;
f o r ( i n t i =1; i <=100; i++) f
c e r r << " s t a r t i n g s imu l a t i o n wi th kT=" << kT << end l ;
140 r e s e t ( ) ;
// the following is not beautiful , but quick :-)
s imu l a t e (kT , 1 00000 , 1 000 ) ;
v l i s t [27]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 1 00000 ) ;
145 a s s e r t ( v l i s t [27]�> ge tSp in ( )==�1);
v l i s t [329]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 1 10000 ) ;
v l i s t [116]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 1 20000 ) ;
150 v l i s t [223]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 1 30000 ) ;
v l i s t [251]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 1 40000 ) ;
v l i s t [237]�> s t i c k I t ( ) ;
48
B.2. Spin ip model
155 s imu l a t e (kT , 1 0000 , 1 000 , 1 50000 ) ;
v l i s t [491]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 1 60000 ) ;
v l i s t [365]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 1 70000 ) ;
160 v l i s t [7]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 1 80000 ) ;
v l i s t [418]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 1 90000 ) ;
v l i s t [381]�> s t i c k I t ( ) ;
165 s imu l a t e (kT , 1 0000 , 1 000 , 2 00000 ) ;
v l i s t [199]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 2 10000 ) ;
v l i s t [398]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 2 20000 ) ;
170 v l i s t [15]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 2 30000 ) ;
v l i s t [492]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 2 40000 ) ;
v l i s t [461]�> s t i c k I t ( ) ;
175 s imu l a t e (kT , 1 0000 , 1 000 , 2 50000 ) ;
v l i s t [371]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 2 60000 ) ;
v l i s t [249]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 2 70000 ) ;
180 v l i s t [80]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 2 80000 ) ;
v l i s t [486]�> s t i c k I t ( ) ;
s imu l a t e (kT , 1 0000 , 1 000 , 2 90000 ) ;
g
185 g
// ... Methoden ...............................................
vo id I s i n g : : b u i l d n e t ( ) f
190 cout << " b u i l d n e t " << end l ;
f o r ( i n t i =0; i<MAX NODE ; i++) f
v l i s t [ i ] = new Ver t ex ( th i s , i ) ;
v l i s t [ i ]�> s e t Sp i n (1 ) ;
g
195 i n t from , to ;
whi le ( ! net . e o f ( ) ) f
net . g e t l i n e ( bu f f e r , 8 0 ) ;
i f ( s s c a n f ( bu f f e r , "%i �� % i ; " , & from , & to ) != 2) f
c e r r << " i npu t l i n e i g no r e d : ' " << b u f f e r << " ' " << end l ;
200 cont inue ;
g
a s s e r t ( from < MAX NODE) ; a s s e r t ( to < MAX NODE) ;
v l i s t [ from]�>addConn ( v l i s t [ to ] , NODOUBLE) ;
v l i s t [ to]�>addConn ( v l i s t [ from ] , NODOUBLE) ;
205 g
net . c l o s e ( ) ;
v count = 0 ;
f o r ( i n t i =1; i<MAX NODE ; i++)
49
B. Source code
i f ( v l i s t [ i ]�>getConnCount ( )>0) v count=i ;
210 e l s e v l i s t [ i ]�> s e t Sp i n (0 ) ;
c e r r << v count << " nodes " << end l ;
g
vo id I s i n g : : debug ( i n t nr ) f
215 Ver t ex � node = v l i s t [ n r ] ;
i n t nb = node�>getConnCount ( ) ;
cout << "node " << nr << " has " << nb
<< " conn e c t i o n s : " << end l ;
f o r ( i n t i =0; i<nb ; i++)
220 cout << node�>getConn ( i )�>getNumber ( ) << end l ;
g
vo id I s i n g : : r e s e t ( ) f
f o r ( i n t i =0; i<MAX NODE ; i++)
225 v l i s t [ i ]�> s e t Sp i n ( abs ( v l i s t [ i ]�>ge tSp in ( ) ) ) ;
g
// ... Simulation .............................................
230 vo id I s i n g : : s imu l a t e ( double kT , i n t maxtime , i n t s t ep t ime , i n t s t a r t t i me
) f
i f ( ! s t e p t ime ) s t e p t ime=maxtime ;
f o r ( i n t i=�MAX CONN; i<=MAX CONN; i++)
e n l i m i t [ i+MAX CONN] =
( i n t ) (MAX INT � ( 2� exp (�2.� i /kT)�1) ) ;
235 i n t mag , t ime=s t a r t t i m e ;
whi le ( time�s t a r t t ime<maxtime ) f
f o r ( i n t s t e p =0; s tep<s t e p t ime ; s t e p++) f
mag = 0 ;
f o r ( i n t nr =1; nr<=v count ; n r++) f
240 a s s e r t ( v l i s t [ n r ] !=NULL) ;
mag += v l i s t [ n r ]�> s imu l a t eS t ep ( ) ;
g
g
t ime += s t ep t ime ;
245 cout << t ime << " " << kT << " " << mag << end l ;
g
g
// === main ===================================================
250
i n t main ( i n t argc , char �� argv ) f
a s s e r t ( c e r r << "debug mode on" << end l ) ;
I s i n g i s i n g ( " net . t x t " ) ;
r e tu rn 0 ;
255 g
50
B.3. Modi�ed Barab�asi-Albert model
B.3. Modi�ed Barab�asi-Albert model
This Ruby program creates modi�ed Barab�asi-Albert models with a given �nal size. In
section 5 this was used on our collaboration network.
#!/home/fxp/bin/ruby -w
$P = 0 .8
$RUNS = 1E0 . t o i
5 $M = 1
$ c l u s t e r i n i t i a l s i z e = 3
$MAXINT = 2147483647�2+1
10
c l a s s Random
@@ibm = 1
def rnd (max=n i l )
@@ibm �= 65539 # 16807 # 65539
15 @@ibm &= $MAXINT
max ? @@ibm�max/$MAXINT : @@ibm
end
end
20 c l a s s Ve r t ex
a t t r r e a d e r : c onn e c t i o n s
a t t r a c c e s s o r : n r
@@ t o t a l v e r t i c e s = 0
de f i n i t i a l i z e
25 @connec t i on s = [ ]
@@ t o t a l v e r t i c e s += 1
@nr = @@ t o t a l v e r t i c e s
end
de f i n s p e c t
30 " I 'm node nr . #f@nr g connected to #f@connec t i on s . c o l l e c t f j c j c . n r g .
j o i n (" , " ) g . "
end
de f a d d l i n k ( p a r t n e r )
@connec t i on s <<= pa r t n e r
end
35 de f connect ( p a r t n e r )
a d d l i n k ( p a r t n e r )
p a r t n e r . a d d l i n k ( s e l f )
end
end
40
c l a s s BA net
de f i n i t i a l i z e ( prob new=0)
@p new = ( prob new � $MAXINT) . t o i
@nodes = [ ]
45 @ke r t e s z = [ ]
@r = Random . new
100 . t imes f @r . rnd g
s t a r t n ew n e t
51
B. Source code
end
50 de f s i z e
@nodes . l e n g t h
end
de f s t a r t n ew n e t
max = @nodes . l e n g t h
55 $ c l u s t e r i n i t i a l s i z e . t imes f @nodes << Ver t ex . new g
$ c l u s t e r i n i t i a l s i z e . t imes f j i j
o r i g = i + max
de s t = (( i +1) % $ c l u s t e r i n i t i a l s i z e ) + max
@nodes [ o r i g ] . connect ( @nodes [ d e s t ] )
60 @ke r t e s z << o r i g << de s t
g
end
de f i n s p e c t
@nodes . c o l l e c t f j node j node . i n s p e c t g . j o i n ("nn") + "nn#f@ke r t e s z .
i n s p e c t g"
65 end
de f add node
i f @r . rnd < @p new
s t a r t n ew n e t
e l s e
70 @nodes << new node = Ver t ex . new
l = @ke r t e s z . l e n g t h
$M. t imes f
de s t = @ke r t e s z [ @r . rnd ( l �1) ]
new node . connect ( @nodes [ d e s t ] )
75 @ke r t e s z << @nodes . l eng th �1 << de s t
g
end
end
de f c h e c k s ub t r e e ( node , s ub t r e e )
80 r e t u r n i f ! node . nr
s ub t r e e << node . nr
node . nr = n i l
node . c onn e c t i o n s . each f j n j ch e c k s ub t r e e (n , s ub t r e e ) g
end
85 de f a n a l y s i s
@nodes . each f j node j
next i f ! node . nr
t r e e = [ ]
c h e c k s ub t r e e ( node , t r e e )
90 # p t r e e . s o r t
# [ 1 ] [ 2 , 3 ] [ 4 , 5 , 6 , 7 ] [ 8 . . . ] . . .
bucket = t r e e . l e n g t h
i f ! $ s t a t i s t i k [ bucket ]
$ s t a t i s t i k [ bucket ] = 1
95 e l s e
$ s t a t i s t i k [ bucket ] += 1
end
g
end
100 end
52
B.3. Modi�ed Barab�asi-Albert model
$ s t a t i s t i k = fg
$RUNS. t imes f j i j
$ s t d e r r . p r i n t "#f i g . . . " i f i %100==0
105 mynet = BA net . new ($P)
beg in
mynet . add node
end wh i l e mynet . s i z e < 555
# p mynet
110 mynet . a n a l y s i s
g
$ s t a t i s t i k . k ey s . s o r t . each f j l j p r i n t f "%3 i %6.4 f nn" , l , $ s t a t i s t i k [ l
]g # . t o f /$RUNS g
53
B. Source code
54
C. Figures
The �gures found in this publication were
made by myself, except the following:
� �gure 1.1 was found on http://-
www.math.colostate.edu/~betten/-
courses/M501/combi.html.
� �gure 2.1 was found in [14].
� �gure 2.3 was found on http://-
www.science.nd.edu/physics/-
Faculty/barabasi.html.
� �gure 2.3 was found on http://-
www.phys.psu.edu/~ralbert/.
� �gure 2.2 was found in [18].
� �gure 4.1 was found on http://-
www.physik.tu-dresden.de/itp/-
members/kobe/isingphbl/.
55
C. Figures
56
Bibliography
[1] L. Euler. Solutio problematis ad geometriam situs
pertinentis. Commetarii Academiae Scientiarum
Imperialis Petropolitanae 8 (1736).
[2] R. Albert and A.-L. Barab
�
asi. Statistical me-
chanics of complex networks. Rev. Mod. Phys. 74,
47{97 (2002).
[3] S. N. Dorogovtsev and J. F. F. Mendes. Evo-
lution of networks. Adv. Phys. 51, 1079{1187
(2002).
[4] A.-L. Barab
�
asi. Linked. The New Science of Net-
works (Perseus, Cambridge, Massachusetts, 2002).
[5] K. Rieger. Sind Sie mit Marlon Brando befreun-
det? Die Zeit 44, BL21 (1999).
[6] A.-L. Barab
�
asi, R. Albert, and H. Jeong.
Scale-free characteristics of random networks: the
topology of the world-wide web. Physica A 281,
69{77 (2000).
[7] M. E. J. Newman. Scienti�c collaboration net-
works. Network construction and fundamental re-
sults. Phys. Rev. E 64, 01631 (2001).
[8] M. E. J. Newman. Scienti�c collaboration net-
works. Shortest paths, weighted networks, and cen-
trality. Phys. Rev. E 64, 01632 (2001).
[9] H. Ebel, L.-I. Mielsch, and S. Bornholdt.
Scale-free topology of e-mail networks. Phys. Rev.
E 66, 035103 (2002).
[10] S. Redner. How popular is your paper? an empir-
ical study of the citation distribution. Europ. Phys.
J. B 4, 131{134 (1998).
[11] A. J. Lotka. The frequency distribution of scien-
ti�c productivity. J. Wash. Acad. Sc. 16, 317{323
(1926).
[12] F. Liljeros, C. R. Edling, et al. The web of
human sexual contacts. Nature 411, 907 (2001).
[13] S. N. Dorogovtsev and J. F. F. Mendes. Evo-
lution of networks: From Biological Nets to the
Internet and WWW (Oxford University Press, Ox-
ford, 2003).
[14] S. H. Strogatz. Exploring complex networks. Na-
ture 410, 268{276 (2001).
[15] S. Milgram. The small world problem. Psych.
Today 61 (1967).
[16] D. J. Watts and S. H. Strogatz. Small world.
Nature 393, 440{442 (1998).
[17] D. Watts. Small worlds: the dynamics of net-
works between order and randomness (Princeton
University Press, Princeton, New Jersey, 1999).
[18] M. E. J. Newman. Models of the small world|a
review (2000). ArXiv:cond-mat/0001118.
[19] M. Gladwell. The Tipping Point. How little
things can make a big di�erence. (Little Brown
and Company, Boston, Massachusetts, 2000).
[20] R. Cohen, K. Erez, D. ben Avraham, and
S. Havlin. Resilience of the internet to random
breakdowns. Phys. Rev. Lett. 85, 21, 4626{4628
(2000).
[21] A. L. Lloyd and R. M. May. How viruses spread
among computers and people. Science 292, 1316{
1317 (2001).
[22] M. E. J. Newman, S. Forrest, and
J. Balthrop. Email networks and the spread
of computer viruses. Phys. Rev. E 66, 035101
(2002).
[23] B. J. Kim, C. N. Yoon, S. K. Han, and H. Jeong.
Path �nding strategies in scale-free networks.
Phys. Rev. E 65, 027103 (2002).
[24] J. M. Kleinberg. Navigation in a small world.
Nature 406, 845 (2000).
[25] Isi web of science (2002). http://www.isinet.com/-
isi/products/citation/wos/index.html.
[26] R. Albert, H. Jeong, and A.-L. Barab
�
asi. Di-
ameter of the world-wide web. Nature 401, 130
(1999).
[27] R. Solomonoff and A. Rapoport. Connectivity
of random nets. Bull. Math. Biophys. 13, 107{227
(1951).
[28] P. Erd
}
os and A. R
�
enyi. On random graphs. Pub.
Mathem. 6, 290{297 (1959).
[29] M. E. J. Newman. Random graphs as models of
networks (2002). ArXiv:cond-mat/0202208.
57
Bibliography
[30] A.-L. Barab
�
asi and R. Albert. Emergence of
scaling in random networks. Science 286, 509{512
(1999).
[31] St. Matthew. New Testament. In The Bible,
chap. 12, 13 (God, ca. 10AD).
[32] D. Cohen. All the world's a net. New Scientist
174, 24{29 (2002).
[33] A.-L. Barab
�
asi and E. Bonabeau. Scale-free net-
works. Sci. Amer. 3, 50{59 (2003).
[34] K. Klemm and V. M. Egu
�
iluz. Growing scale-free
networks with small-world behavior. Phys. Rev. E
65, 057102 (2002).
[35] E. Ravasz and A.-L. Barab
�
asi. Hierarchical or-
ganization in complex networks. Phys. Rev. E 67,
026112 (2003).
[36] K. Klemm and V. M. Egu
�
iluz. Highly clustered
scale-free networks. Phys. Rev. E 65, 036123
(2002).
[37] A. F. J. van Raan. Fractal dimension of co-
citations. Nature 347, 626 (1990).
[38] AT&T. Graphviz|open source graph drawing
software (2002). http://www.research.att.com/-
sw/tools/graphviz/.
[39] S. C. North. Drawing graphs with neato. AT&T
(2002).
[40] J. C. Venter, M. D. Adams, E. W. Myers,
et al. The sequence of the human genome. Sci-
ence 291, 1304{1351 (2001).
[41] A.-L. Barab
�
asi, H. Jeong, Z. N
�
eda, E. Ravasz,
A. Schubert, and T. Vicsek. Evolution of the so-
cial network of scienti�c collaborations. Physica A
311, 590{614 (2002).
[42] A.-L. Barab
�
asi. The physics of the web
(2001). http://www.physicsweb.org/article/-
world/14/7/09.
[43] L. A. N. Amaral, A. Scala, M. Barth
�
el
�
emy,
and H. E. Stanley. Classes of small-world net-
works. Proc. Nat. Acad. Sci. USA 97, 21, 509{512
(2000).
[44] S. Mossa, M. Barth
�
elemy, H. E. Stanley, and
L. A. N. Amaral. Truncation of power law be-
haviour in "scale-free" network models due to in-
formation �ltering. Phys. Rev. Lett. 89, 208701
(2002).
[45] D. Stauffer and A. Aharony. Introduction to
Percolation Theory (Taylor and Francis, London,
1994), second ed.
[46] A.-L. Barab
�
asi, R. Albert, and H. Jeong. Mean
�eld theory for scale-free random networks. Phys-
ica A 272, 173{187 (1999).
[47] E. Ising. Beitrag zur Theorie des Ferromag-
netismus. Zeitschr. f. Phys. 31, 253{258 (1925).
[48] W. Lenz. Beitrag zum Verst�andnis der mag-
netischen Erscheinungen in festen K�orpern. Phys.
Zeitschr. 21, 613{615 (1920).
[49] R. B. Potts. Some generalized order-disorder
transformations. Proc. Cambridge Phil. Soc. 48,
106 (1952).
[50] F. Y. Wu. The Potts model. Rev. Mod. Phys. 54,
235{268 (1982).
[51] N. Metropolis, A. W. Rosenbluth, and A. H.
Teller. Equation of state calculations by fast
computing machines. J. Chem. Phys. 21, 6, 1087{
1092 (1953).
[52] A. Aleksiejuk, J. A. Ho lyst, and D. Stauffer.
Ferromagnetic phase transition in Barab�asi-Albert
networks. Physica A 310, 260{266 (2002).
[53] S. N. Dorogovtsev, A. V. Goltsev, and J. F. F.
Mendes. Ising model on networks with an arbi-
trary distribution of connections. Phys. Rev. E 66,
016104 (2002).
[54] K. Kacperski and J. A. Ho lyst. Phase transi-
tions as a persistent feature of groups with leaders
in models of opinion formation. Physica A 287,
631{643 (2000).
[55] F. P
�
utsch. Analysis and modeling of science col-
laboration networks. Adv. Compl. Sys. (2003).
Submitted on May 6
th
.
58