MODELS

OF
NETWORKS
Scale Free
and
Preferential Attachment

ASM: Models of Networks 165

 • Power Laws
SCALE-FREE • Hubs
• Scale-Free Networks
AND • Degree Exponent

PREFERENTIAL • Preferential Attachment

• Barábasi-Albert Model
ATTACHMENT • Degree Dynamics

ASM: Models of Networks 166

 POWER LAWS

• A power law is a non-linear function of the degree of the form

P 𝑘 = 𝐴𝑘 −𝛾

• We can normalize the distribution (as with any probability distribution) and then it becomes

𝑘 −𝛾
P 𝑘 =
𝜁(𝛾, 𝑘𝑚𝑖𝑛 )

Where γ is the degree exponent and ζ(γ) is the Incomplete Riemann zeta-function, where we can use 1 as the
minimum value of the degree, then
∞

𝜁 𝛾, 𝑘𝑚𝑖𝑛 = ෍ 𝑘 −𝛾
𝑘=𝑘𝑚𝑖𝑛

ASM: Models of Networks 167

 POWER LAWS

• These distributions are scale-free in the sense that the rates of the probabilities are the same no matter the scale at
which we compare them, for example

𝑃(2) 𝑃(20)
= = 2−𝛾
𝑃(1) 𝑃(10)

• These distributions have fat-tails, in the sense that we may see more low degree and more high degree nodes than in
a Poisson network. The easiest way to see this is by log-transforming the distribution

• The log-transformed model can allow us to estimate the value of the degree exponent 𝛾 using a linear regression.
However, as we will see, we have to be a bit careful with this method and use Maximum Likelihood estimation
prefereably

ASM: Models of Networks 168

 POWER LAWS
• The log transformed models allow us to see the heavy tails of the scale free in contrast with the Poisson distribution

log 𝑃𝑘 = 𝐴 − 𝛾 log 𝑘

Poisson
Scale Free

ASM: Models of Networks 169

 UNIVERSALITY

There are many systems that share the scale-free property, we call this fact a universal network
characteristics:

• Wealth Distributions: Pareto distributions (1896)

• Frequency of words usage and Size of towns: Zipf’s law
• Law of anomalous digits: Bendford’s law
• Firm sizes, Stock Market Movements,…

ASM: Models of Networks 170

 SCALE-FREE NETWORKS

Power-law distributions have a series of peculiar properties with respect to their moments (mean, variance,…). If we
use the continuous formalism (approximate the sums to integrals), then

∞
1−𝛾 𝑘 𝑛−𝛾+1
< 𝑘 𝑛 > = (1 − 𝛾)𝑘𝑚𝑖𝑛
𝑛−𝛾+1 𝑘𝑚𝑖𝑛

Where n=1 is for the mean, n=2 for the variance, and so on…. The previous expression results in a convergent integra
only when

𝑛≤ 𝛾−1

Then the values of the exponent produce problems with these networks and let us stablish some of their
characteristics depending on it

ASM: Models of Networks 171

 DEGREE EXPONENT

Anomalous Ultra small-world Critical Small-World

𝛾 ≤2 (2,3) 3 ≥3

Average Degree diverges finite finite finite

Standard Deviation diverges diverges finite finite

log 𝑁 log 𝑁
Average Path Length constant ~ log log 𝑁 ~ ~
log log 𝑁 log 𝑘ത

ASM: Models of Networks 172

 ANOMALOUS NETWORKS

A degree sequence that can be turned into a simple graph is known as graphical.

Erdös and Gallai proposed a method to determine which sequences are graphical sequences. When it is applied to
scale-free networks, the number of graphical degree sequences with 𝛾 < 2 drops to zero, i.e. No degree distribution
in the anomalous regime can be turned into a simple graph

Graphical Non-Graphical

Degree N Degree N

2 3 2 2

3 1

ASM: Models of Networks 173

 HUBS

A power-law distribution implies that most of the nodes have a very low degree, while a few of them have a very large
value. These last are known as hubs

Direct integration of the probability distribution give us a relation between the natural cut-offs as

𝑘𝑚𝑎𝑥 = 𝑘𝑚𝑖𝑛 𝑁1/(𝛾−1) ≅ 𝑁 1/(𝛾−1)

Where the meaning of 𝑘𝑚𝑖𝑛 and 𝑘𝑚𝑎𝑥 is the same as with random networks. In the second step we have assumed
that 𝑘𝑚𝑖𝑛 = 1.

This equation implies that for a network with 106 nodes, a minimum degree of 1 and a degree exponent of 2.1, the
maximum degree will be of 284,804… way different from a random network!

ASM: Models of Networks 174

 HUBS

Sometimes scale-free and degree correlation functions have a conflict in simple networks. For a neutral network, the
expected number of links between two nodes with degrees 𝑘 and 𝑘′ is

𝑘𝑝𝑘 𝑘′𝑝𝑘′
𝐸𝑘𝑘′ = 𝑁
𝑘ത

Which can be larger than 1 if the degrees are too large. In particular, there is a boundary value, known as the
structural cut-off from which this happens

1/2
𝑘𝑠 ≈ 𝑁 𝑘ത

Then:

• For 𝛾 ≥ 3, there is no structural cut-offs since it is always that 𝑘𝑚𝑎𝑥 < 𝑘𝑠

• For 𝛾 < 3, there are nodes whose degree is 𝑘𝑠 < 𝑘 < 𝑘𝑚𝑎𝑥 . These networks become disassortative not because of
the generating process, this phenomenon is known as structural disassortativity.

ASM: Models of Networks 175

 HUBS

In real networks will generally not follow a perfect power law. We will find deviations in the low degree and high
degree regions, where we will usually find less observations than the ones predicted by a pure power law. This is
known as low degree saturation and high degree cut-off, and can be modelled using

−𝛾
𝑘
𝑝𝑘 = 𝑎 𝑘 + 𝑘𝑠𝑎𝑡 exp −
𝑘𝑐𝑢𝑡

The presence of these terms does not imply that the scale-free nature of the network is broken:

• No property is affected by the low degree saturation

• The high degree cut-off limits the divergence of the second moment < 𝑘 2 >, which can be seen by direct
integration

It is true, however, that when we fit this model, we usually lose statistical significance which indicates that we are
either not fitting the true form or indicates the presence of outliers.

ASM: Models of Networks 176

 HUBS

The ultra-small world is also a consequence of the existence of hubs. In this regime the the average path
length grows as

𝑑ҧ ≈ log log 𝑁

which is significantly smaller than the growth for a random network.

It occurs because hubs reduce the path length, since they are in the middle of all (most of) the paths and
tend to connect many nodes just through it.

ASM: Models of Networks 177

 PREFERENTIAL ATTACHMENT

Price (1965, 1976) proposed a model adopting ideas from Simons (1955) in the analysis of the citation network
between researchers. His main idea was that an article would gain citations over time proportionally to the number of
citations it had at a given time.

In the recent literature this link formation process is known as preferential attachment and is used to describe the
Barabási-Abert model, a process that has two main ingredients

• The system grows with time, i.e. new nodes are added to the network
• The existing objects grow at rates proportional to their size, a property that ensures a rich-get-richer effect

We may allow for a fitness attribute (and not just degree) as a source of the generation of links, this is the Bianconi-
Barabàsi model and allows degrees not to be purely “age” dependent (as happens in real situations)

ASM: Models of Networks 178

 PREFERENTIAL ATTACHMENT

• If we remove preferential attachment we add nodes, but they connect randomly to any other node in
the network. In this case the degree distribution is that of an exponential family (not exactly a
Poisson network since those models are static). These distributions do not support hubs and we lack
the rich-get-richer effect.

• If we remove growth we lose the dynamics, by adding links at each iteration the endpoint of the
network evolution is just a complete graph

• In any case, just as we present them, these network models fail in the prediction of the clustering
coefficient. This may be solved by allowing for a non-linear growth and/or changing the number of
nodes and/or links added at each iteration.

ASM: Models of Networks 179

 PREFERENTIAL ATTACHMENT

Bolobás and Riordan showed that in a preferential attachment scheme:

• If we add only one link per iteration, the network consists of a single component with diameter
proportional to log 𝑁
• If we add more logthan
𝑁
one link per iteration, the network consist of a single component with diameter
proportional to
log log 𝑁

We may describe these models with 𝑝𝑘 ≈ 2𝑚1/𝛽 𝑘 −𝛾 , where 𝑚 is the number of links created per iteration
and 𝛽 is called the dynamical exponent. This exponent is 1/2 in the Barabási-Albert model, and since

1
𝛾= +1
𝛽

It results that the Barabási-Albert model is always for 𝛾 = 3, i.e. they describe the critical regime

ASM: Models of Networks 180