Conference PaperPDF Available

On the Skewed Degree Distribution of Hierarchical Networks

Authors:
  • DeepMind and University of Liverpool

Abstract

In this paper, a prestige-based evolution process is introduced, which provides a formal framework for the study of linear hierarchies seen in human societies. Due to the deterministic characteristics of the proposed model, we are capable of determining equilibria in closed form. Surprisingly, these stationary points recover the power-law degree distribution as the shared property of the resulting hierarchal networks, explaining the prevalence of hierarchies in societies. This result sheds light on the evolutionary advantages of hierarchies.
On the Skewed Degree Distribution
of Hierarchical Networks
Bijan Ranjbar-Sahraei
Maastricht University, The Netherlands
Email: b.ranjbar@ieee.org
Karl Tuyls
University of Liverpool, United Kingdom
Email: k.tuyls@liverpool.ac.uk
Haitham Bou Ammar
University of Pennsylvania, United States
Email: haithamb@seas.upenn.edu
Gerhard Weiss
Maastricht University, The Netherlands
Email: gerhard.weiss@maastrichtuniversity.nl
Abstract—In this paper, a prestige-based evolution process
is introduced, which provides a formal framework for the
study of linear hierarchies seen in human societies. Due to
the deterministic characteristics of the proposed model, we are
capable of determining equilibria in closed form. Surprisingly,
these stationary points recover the power-law degree distribution
as the shared property of the resulting hierarchal networks,
explaining the prevalence of hierarchies in societies. This result
sheds light on the evolutionary advantages of hierarchies.
I. INTRODUCTION
To analyze the emergence of social networks, a variety of
mathematical models have been proposed. The earliest dates
back to the 1900’s, where Yule [1] studied the biological
evolution of species based on age and population data. Others,
e.g., Lotka [2] provided rules required for describing and
analyzing scientific publications. Resulting from these and
other studies, was the emergence of the power-law degree
distribution [3] as a shared common characteristic for a wide-
range of networks including but not limited to, the world wide
web, protein-protein interaction, airlines and social networks.
Given such a widely-shared characteristic, Barab´
asi and
Albert suggested a preferential attachment model for the
generation of scale-free graphs exhibiting a power-law degree
distribution [4]. As noted by Durret [5], the definition of
their process was rather informal. Since then, different precise
forms of the Barab´
asi-Albert model have been studied in
literature [6]. Though successful at recovering the power-law
degree distribution, these studies impose several restricting
assumptions on the underlying graph generating process. For
instance, such techniques typically adopt a binary attachment
model, in which two nodes are either connected or not [4], [7].
On the other hand, the existence of hierarchical relation-
ships is another shared common characteristic for a wide-range
of networks. Research has shown that human physique (e.g.,
stature) and body hormones play a crucial role in enabling
dominance in the society. Contrary to animal societies which
base hierarchies on dominance, however, human societies
replace dominance by “prestige” to construct reciprocal rela-
tionships between leaders and followers [8]. Thus, evolutionary
considerations of real-world networks suggests the emergence
of scale-free behavior (i.e., networks exhibiting a power-law
degree distribution) in networks as a resultant of hierarchal
attachment processes which are not reflected through current
preferential attachment models. Apart from this modeling
restriction, another problem inherent to existing binary models
lies in their explanatory capabilities. For instance, they fail to
manifest connection strengths between individuals; a property
being at the core of behavioral emergence in real networks [9].
To provide more realistic modeling outcomes, in this paper,
we contribute by proposing a deterministic hierarchal graph at-
tachment process for prestige-based human societies. Contrary
to preferential attachment models, our approach only assumes
hierarchal connections between individuals, thus bridging the
modeling gap to real-world evolutionary networks. Among
many advantages, our deterministic setting enables the deriva-
tion of the degree distribution in closed-form. Performing
this derivation recovers, surprisingly, the power-law degree
distribution as the main property of the resultant hierarchal
networks, which explains the prevalence of such hierarchies
in societies.
II. NOTATI ON S
A network is described via a graph, G= (V,W), con-
sisting of a set of Nnodes (or vertices) V={v1,...,vN}
and an N×Nadjacency matrix W= [wij ]where non-
zero entries wij indicate the weighted connection from vj
to vi. In this paper we consider the undirected graphs which
have symmetrical W. The neighborhood Nof a node vi
is defined as the set formed by its connected vertices, i.e.,
N(vi) = jvj:wij >0.
The node’s degree, deg(vi), is given by the cardinality of its
neighborhood. The strength of a node is of major importance
in hierarchical networks. Next, we define three concepts: (1)
relative strength (2) strength observation and (3) absolute
strength.
Relative Node Strength: The relative strength of jth node
with respect to ith node with i>jis denoted by Ψi(vj)and
represents the sum over all edge weights between jth node
and every kth node where k < i. Namely,
Ψi(vj) =
i1
X
k=1
wjk , i > j. (1)
In other words, when node iis monitoring node jwith
j < i, it just observes those connections from other ks to j
which k < i.
Strength Observation The Strength observation of the ith
node is denoted by the vector
~
Ψi= [Ψi(v1),Ψi(v2),...,Ψi(vi1)]T
with cardinality i1.
Absolute Node Strength: The absolute strength of ith
node is defined as
Ψ(vi) =
N
X
k=1
wik.(2)
III. NET WORK DYNAM IC S
Here, we propose a dynamical process which captures the
edge dynamics of a complete network. Let ω={wij |∀i, j =
1,2, ..., N, i > j}denote the state vector of the process, where
each state variable wij corresponds to the weight of the link
between the jth and ith node. In the very general case, one
considers the rate of changes in wij as a function of all state
variables
˙wij =f(ω)(3)
In this paper, however, the focus is on hierarchical networks
in which for any i>j,˙wij is a function of wij itself and the
strength observation of the ith node ~
Ψi
˙wij =fΨ(wij ,~
Ψi), i > j (4)
In other words, the dynamics of the linking strength between
iand jis independent of any other node lwhere lis higher
than ior jin the hierarchy.
Using fΨfrom Equation 4, and sorting the state variables
wij increasingly (based on Ni+j), the overall dynamic process
can be written as
˙ω=d
dtw21, . . . , wN(N1) T
=hfΨ(w21,~
Ψ2), . . . , fΨ(wN(N1),~
ΨN)iT
Next, we introduce a possible strategy for fΨ(.), namely
f(P)
Ψ(·)which represents the Prestige-based attachment (PA)
models.
IV. PRESTIGE-BA SED ATTACHMENT MODEL
The overall strength of node iin establishing connection
with every other jth node is assumed to be limited and sums-
up to 1. The prestige-based attachment model can be formally
derived as follows. Let
˙wij =f(P)
Ψ(wij ,Ψi(vj),
~
Ψi
)
and
f(P)
Ψ(wij ,Ψi(vj),
~
Ψi
) = Ψi(vj)
~
Ψi
wij , j < i (5)
where ||.|| denotes the first norm.
By studying the dynamic process proposed in Equation 5,
it can be easily seen that ˙wij , i > j is a function of every wkl
for k, l < i. Without loss of generality we assume w(P)
11 = 1,
such that
Ψ2(v1)=1 (6)
and w(P)
ii = 0 for every i > 1. It is straightforward to compute
the equilibrium point of this system
w(P)
ij =Ψi(vj)
~
Ψi
(7)
The equilibrium point (7) explains that the connection
strength between node iand node jdepends on the connection
strength among nodes iand jand every kth node with
k < max(i, j). To illustrate, imagine Nagents who are all
connected to each other, and continuously the agents with
higher order share their available resources with agents with
lower order. The strength of link between iand jshows
the amount of resources which are transmitted from ito j.
According to (7) the 2nd individual shares all of her resources
with 1st individual (i.e., w21 =1
1= 1). The 3rd individual
shares one third of her resources with the 2nd individual and
two third of it with the 1st individual (i.e., w32 =1
1+2 =1
3
and w31 =2
1+2 =2
3). With the same respect, ith individual
shares portions of her resources with each of the jindividuals
where j < i, while those with lower order receive more. We
call this a prestige-based model as the lower orders reflect a
kind of prestige in the group and high prestige agents receive
more than agents with lower prestige.
An immediate result of (7) is that
i
X
j=1
w(P)
ij =
i1
X
j=1
w(P)
ij =
i1
X
j=1
Ψi(vj)
~
Ψi
=
~
Ψi
~
Ψi
= 1.(8)
Next, we study the amount of resources each individual
receives in such prestige-based network (captured by node’s
strengths), and also compute the distribution of node strengths.
A. Analysis of Node’s Strength
Here we determine a closed form solution for the sum over
the strength of every jth node from the perspective of the ith
node, where j < i.
Lemma 1: In the prestige-based attachment model, the
sum of the relative node strengths of every jth node from
perspective of the ith node, where j < i is as following
K(i) : k~
Ψik= 2i3.
Proof: The above lemma can be proved by using induc-
tion:
Initial Step: According to Equation (6) we have k~
Ψ2k=
Ψ2(v1)=1. Therefore, K(i)holds for i= 2.
Inductive Step: Let
K(i1) : k~
Ψi1k= 2i5,
and also note that Ψi(vj) = Ψi1(vj) + w(P)
(i1)j. Therefore
we can write
k~
Ψik=
i1
X
j=1
Ψi(vj)
=Ψi(vi1) +
i2
X
j=1 Ψi1(vj) + w(P)
(i1)j
=
i1
X
j=1
w(P)
(i1)j+k~
Ψi1k+
i2
X
j=1
w(P)
(i1)j
By using K(i1) and Equation 8, we’ll get
k~
Ψik=
i1
X
j=1
Ψi(vj) = 1 + 2i5 + 1 = 2i3(9)
Therefore, K(i)holds for every i, concluding the proof.
B. Analysis of Edge Weights
We can compute the edge weight between ith node and
jth node as follows.
Lemma 2 (Edge Weight): For the weighted graph G,
evolved with PA model, ith node is connected to jth node
with an edge of weight
K(i) : w(P)
ij =1
2i2
ij
Y
k=1
2i2k
2i2k1,j < i. (10)
Proof:
The validity of Equation 10 can be proved for each iand
for every j < i using induction.
Initial Step: The second node is connected to the first node
with w(P)
21 = 1, meaning that K(2) holds.
Inductive Step: Now assume that
K(i1) : w(P)
(i1)j=1
2i4
ij1
Y
k=1
2i2k2
2i2k3
holds for every j < i 1. For computing the edge weight
between ith and jth node, recall that Ψi(vj) = Ψi1(vj) +
w(P)
(i1)j. By using (7) and Lemma 1, it can be seen that:
Ψi(vj) = Ψi1(vj) + w(P)
(i1)j
= (2i5)w(i1)j+w(P)
(i1)j
= (2i4)w(i1)j(11)
Using Equations 7, 11 and Lemma 1, the edge weight
between ith and jth nodes can be written as
w(P)
ij =Ψi(vj)
Pi1
k=1 Ψi(vk)=1
2i2
ij
Y
k=1
2i2k
2i2k1
for j < i 1. Therefore, K(i)holds for every i, concluding
the proof.
Before, computing the distribution of strengths for the PA
model, we present the following proposition providing the
relative strength of the jth node from the perspective of the
ith for every i>j (i.e., Ψi(vj)) in closed form.
Proposition 1 (Relative Node Strength): For the weighted
graph G, evolved according to the PA model, the strength of
the jth node from perspective of the ith node is given by
K(i) : Ψi(vj) = Qi
k=j+2 2k4
2k5for j < i 1
Ψi(vj) = 1 for j=i1.(12)
Proof: Again, induction can be used to prove the validity
of Equation 12. Starting with the initial step we get
Initial Step: From Equation 6, the strength of the first
node from the perspective of the second node is Ψ2(v1)=1.
Besides, using Lemma 2 we can deduce that
Ψ3(v1) = w(P)
11 +w(P)
21
3=2
3.
Therefore, K(2) holds. For the inductive step we proceed as
follows
Inductive Step: Assume that following holds.
K(i1) : Ψi1(vj) = Qi1
k=j+2 2k4
2k5for j < i 2
Ψi1(vj)=1 for j=i2.
For computing Ψi(vj), consider Ψi(vj) = Ψi1(vj) +
w(P)
(i1)j. Using Equation 7 and Lemma 1, we can show that
for every j < i 1
Ψi(vj) = Ψi1(vj) + w(P)
(i1)j=
i
Y
k=j+2
2k4
2k5
Besides using Equation (8), Ψi(vj)= 1 for j=i1.
Therefore, K(i)holds for every iand the proof is concluded.
Lemma 3 (Global Strength): For the weighted graph G,
evolved with the PA model, the global strength of the ith node
is Ψ(vi) = QN+1
k=i+2 2k4
2k5for i<N
Ψ(vi)=1 for i=N. (13)
Proof: We know that Ψ(vi) = ΨN(vi) + w(P)
iN for every
i<N. Using Equation 7 and Proposition 1, we have
Ψ(vi) = ΨN(vi) + ΨN(vi)
2N3
=2N2
2N3
N
Y
k=i+2
2k4
2k5
=
N+1
Y
k=i+2
2k4
2k5
for every i < N. Based on Equation (8) we have
Ψ(vN) =
N1
X
i=1
w(P)
Ni = 1.
This concludes the proof.
Finally, we can compute the strength distribution in a
closed form. In analysis of weighted networks, typically the
Distribution Function (DF) of node strengths is defined as
P(k) =
vi|∀i, k Ψ(vi)< k + 1
(14)
where Ψ(vi)defined in (2) denotes the strength of node vi.
To ease the analysis, in this work we make use of the Com-
plementary Cumulative Distribution Function (CCDF) defined
as:
Pc(k) =
vi|∀i, Ψ(vi)k
(15)
The following lemma clarifies the relation between the DF and
CCDF for a network with power-law Distribution.
Lemma 4 (Power-law Distribution): Consider a power-
law distribution in form of P(k) = ckα, where αis the
power-law exponent. The CCDF Pc(k)also follows a power-
law but with an exponent α1.
Proof: Can be easily seen by simple integration.
The following theorem provides the strength distribution
of a PA model.
Theorem 1 (Strength Distribution): For the weighted
graph Gevolved with the PA model, the distribution of the
global strength kfollows a power-law with exponent 3
P(k)k3.
For proving Theorem 1, we need the following lemma, first.
Lemma 5 (Fraction Product Series): Consider the follow-
ing product of fractions
L(i, N ) =
N+1
Y
k=i+2
2k4
2k5,
then γi1
2< L(i, N )< γ(i1)1
2, with γ=N1
Proof: We use the comparison test to compute the lower
and upper bounds of L(i, N ). Firstly, consider
Q(i, N ) =
N+1
Y
k=i+2
2k5
2k6(16)
Clearly, L(i, N)< Q(i, N )and L(i, N )Q(i, N) = 2N2
2i2.
Therefore, L(i, N )<q2N2
2i2γ(i1)1
2concluding the
upper-bound. To determine the lower bound, define
Q0(i, N ) =
N+1
Y
k=i+2
2k3
2k4(17)
It can be shown that L(i, N)> Q0(i, N )and
L(i, N )Q0(i, N) = 2N1
2i1. Therefore,
L(i, N )>r2N+ 1
2i1>r2N2
2iγi1
2.(18)
Using results of Lemmas 3, 4 and 5 we give the following
proof for Theorem 1.
Proof: The following lower and upper bounds can be
computed for the strength of the ith node
γi1
2<Ψ(vi)< γ(i1)1
2(19)
where γ=N1. From Equation (19), we have
Ψ(vi)k, for i∈ {1,2,3,...,γ2
k2}(20)
Pc(k) =
1,2,3,..., γ2
k2
'γ2k2(21)
Therefore,
Pc(k)k2(22)
Using Lemma 4, we have
P(k)k3,(23)
thus proof is concluded.
V. CONCLUSION
In this paper we proposed a dynamical model for prestige-
based hierarchical systems. Although, the dynamical system
was described using simple hierarchical rules, the derived sta-
tionary points had been shown to recover a power-law degree
distribution with exponent -3 (Theorem 1). Emergence of this
degree distribution despite the simple hierarchical structure
explains how hierarchical social structures have survived in
human societies. There are various interesting future directions
of this work. We plan to validate the attained results through
data gathered from real-world networks. Moreover, our model
is proposed in form of a dynamical process, which makes it
possible for the development of control strategies. The overall
idea would be the control of the evolution of the network to
arrive at specific network forms.
REFERENCES
[1] G. U. Yule, “A mathematical theory of evolution, based on the conclu-
sions of dr. jc willis, frs,Philosophical Transactions of the Royal Society
of London. Series B, Containing Papers of a Biological Character, pp.
21–87, 1925.
[2] A. J. Lotka, “The frequency distribution of scientific productivity.”
Journal of Washington Academy Sciences, 1926.
[3] R. F. i. Cancho and A. H. Fern´
andez, “Power laws and the golden
number,Problems of general, germanic and slavic linguistics, pp. 518–
523, 2008.
[4] A.-L. Barab´
asi and R. Albert, “Emergence of scaling in random
networks,” Science, vol. 286, no. 5439, pp. 509–12, Oct. 1999. [Online].
Available: http://www.ncbi.nlm.nih.gov/pubmed/10521342
[5] R. Durrett, Random Graph Dynamics (Cambridge Series in Statistical
and Probabilistic Mathematics). New York, NY, USA: Cambridge
University Press, 2006.
[6] B. Bollob´
as and O. Riordan, “Coupling scale-free and classical random
graphs,” Internet Mathematics, vol. 1, no. 2, pp. 215–225, 2003.
[Online]. Available: http://projecteuclid.org/euclid.im/1089229509
[7] D. J. Watts and S. H. Strogatz, “Collective dynamics of ‘small-world’
networks,” nature, vol. 393, no. 6684, pp. 440–442, 1998.
[8] M. E. Price and M. Van Vugt, “The evolution of leader–follower
reciprocity: the theory of service-for-prestige,” Frontiers in human neu-
roscience, vol. 8, 2014.
[9] A. Barrat, M. Barth´
elemy, and A. Vespignani, “Modeling the evolution
of weighted networks,” Physical Review E, vol. 70, no. 6, p. 066149,
2004.
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
We describe the service-for-prestige theory of leadership, which proposes that voluntary leader–follower relations evolved in humans via a process of reciprocal exchange that generated adaptive benefits for both leaders and followers. We propose that although leader–follower relations first emerged in the human lineage to solve problems related to information sharing and social coordination, they ultimately evolved into exchange relationships whereby followers could compensate leaders for services which would otherwise have been prohibitively costly for leaders to provide. In this exchange, leaders incur costs to provide followers with public goods, and in return, followers incur costs to provide leaders with prestige (and associated fitness benefits). Because whole groups of followers tend to gain from leader-provided public goods, and because prestige is costly for followers to produce, the provisioning of prestige to leaders requires solutions to the “free rider” problem of disrespectful followers (who benefit from leader services without sharing the costs of producing prestige). Thus service-for-prestige makes the unique prediction that disrespectful followers of beneficial leaders will be targeted by other followers for punitive sentiment and/or social exclusion. Leader–follower relations should be more reciprocal and mutually beneficial when leaders and followers have more equal social bargaining power. However, as leaders gain more relative power, and their high status becomes less dependent on their willingness to pay the costs of benefitting followers, service-for-prestige predicts that leader–follower relations will become based more on leaders’ ability to dominate and exploit rather than benefit followers. We review evidential support for a set of predictions made by service-for-prestige, and discuss how service-for-prestige relates to social neuroscience research on leadership.
Article
Systems as diverse as genetic networks or the World Wide Web are best described as networks with complex topology. A common property of many large networks is that the vertex connectivities follow a scale-free power-law distribution. This feature was found to be a consequence of two generic mech-anisms: (i) networks expand continuously by the addition of new vertices, and (ii) new vertices attach preferentially to sites that are already well connected. A model based on these two ingredients reproduces the observed stationary scale-free distributions, which indicates that the development of large networks is governed by robust self-organizing phenomena that go beyond the particulars of the individual systems.
Article
The theory of random graphs began in the late 1950s in several papers by Erdos and Renyi. In the late twentieth century, the notion of six degrees of separation, meaning that any two people on the planet can be connected by a short chain of people who know each other, inspired Strogatz and Watts to define the small world random graph in which each site is connected to k close neighbors, but also has long-range connections. At about the same time, it was observed in human social and sexual networks and on the Internet that the number of neighbors of an individual or computer has a power law distribution. This inspired Barabasi and Albert to define the preferential attachment model, which has these properties. These two papers have led to an explosion of research. While this literature is extensive, many of the papers are based on simulations and nonrigorous arguments. The purpose of this book is to use a wide variety of mathematical argument to obtain insights into the properties of these graphs. A unique feature of this book is the interest in the dynamics of process taking place on the graph in addition to their geometric properties, such as connectedness and diameter.