Wilfrid Laurier University

Scholars Commons @ Laurier

Theses and Dissertations (Comprehensive)

2008

Web Search Algorithms and PageRank

Laleh Samarbakhsh
Wilfrid Laurier University

Follow this and additional works at: http://scholars.wlu.ca/etd

Part of the Theory and Algorithms Commons

Recommended Citation
Samarbakhsh, Laleh, "Web Search Algorithms and PageRank" (2008). Theses and Dissertations (Comprehensive). Paper 872.

This Thesis is brought to you for free and open access by Scholars Commons @ Laurier. It has been accepted for inclusion in Theses and Dissertations
(Comprehensive) by an authorized administrator of Scholars Commons @ Laurier. For more information, please contact [email protected].
 NOTE TO USERS

This reproduction is the best copy available.

UMI
1*1 Library and
Archives Canada
Bibliotheque et
Archives Canada

Published Heritage Direction du

Branch Patrimoine de I'edition
395 Wellington Street 395, rue Wellington
Ottawa ON K1A0N4 Ottawa ON K1A0N4
Canada Canada

Your file Votre reference

ISBN: 978-0-494-38720-7
Our file Notre reference
ISBN: 978-0-494-38720-7

NOTICE: AVIS:
The author has granted a non- L'auteur a accorde une licence non exclusive
exclusive license allowing Library permettant a la Bibliotheque et Archives
and Archives Canada to reproduce, Canada de reproduire, publier, archiver,
publish, archive, preserve, conserve, sauvegarder, conserver, transmettre au public
communicate to the public by par telecommunication ou par Plntemet, prefer,
telecommunication or on the Internet, distribuer et vendre des theses partout dans
loan, distribute and sell theses le monde, a des fins commerciales ou autres,
worldwide, for commercial or non- sur support microforme, papier, electronique
commercial purposes, in microform, et/ou autres formats.
paper, electronic and/or any other
formats.

The author retains copyright L'auteur conserve la propriete du droit d'auteur

ownership and moral rights in et des droits moraux qui protege cette these.
this thesis. Neither the thesis Ni la these ni des extraits substantiels de
nor substantial extracts from it celle-ci ne doivent etre imprimes ou autrement
may be printed or otherwise reproduits sans son autorisation.
reproduced without the author's
permission.

In compliance with the Canadian Conformement a la loi canadienne

Privacy Act some supporting sur la protection de la vie privee,
forms may have been removed quelques formulaires secondaires
from this thesis. ont ete enleves de cette these.

While these forms may be included Bien que ces formulaires

in the document page count, aient inclus dans la pagination,
their removal does not represent il n'y aura aucun contenu manquant.
any loss of content from the
thesis.

Canada
 WEB SEARCH ALGORITHMS
AND
PAGERANK

by

Laleh Samarbakhsh

(BSc, Sharif University of Technology, 2005)

THESIS

Submitted to the Department of Mathematics

in partial fulfillment of the requirements for

Master of Science in Mathematics

Wilfrid Laurier University

2008

Copyright ©Laleh Samarbakhsh, Wilfrid Laurier University

 Abstract

The mathematical theory underlying the Google search

engine is the PageRank algorithm, first introduced by Sergey

Brin and Lawrence Page, the founders of Google. A ranking

of web pages is made considering many criteria. PageRank

exploits the graph structure of the web. The web's hy-

perlink structure forms a massive directed graph, where

the web pages are presented as nodes and hyperlinks as

edges. The PageRank equation finds a score by solving a

recursive equation which calculates the PageRank vector.

The PageRank vector is the stationary distribution of an

ergodic Markov chain. The Perron-Frobenius theorem en-

sures that the primitive matrix produced by this massive

Markov chain will converge to a unique stationary distri-

bution. The PageRank vector existence is guaranteed since

the so-called Google matrix is stochastic and has all entries

positive.
ii ABSTRACT

In a recent work by Litvak, Scheinhardt and Volkovich

[14], a mathematical model is presented that explains an in-

teresting relation between PageRank values and in-degrees

in power law graphs. They analytically prove that in power

law graphs, the tail distributions of PageRank and in-degree

differ only by a multiplicative factor.

We survey the mathematics of the PageRank algorithm,

and study the work of Litvak et. al. We implement a

PageRank calculator and expose different graphs to our

calculator. For various power law graphs, we show that

the ranking of the nodes by PageRank will be the same as

the ranking given by in-degree. We give a counterexam-

ple for graphs which are not power law. For these graphs,

the ranking derived from PageRank is different from the

ranking derived from the in-degree values.

K e y w o r d s : graphs, directed graphs, PageRank, Google

matrix, Markov chains, random walk, power law graph,

binary tree
 Acknowledgements

I would like to express my sincere thanks to my thesis

supervisor, Dr. Anthony Bonato. He has been very helpful

and provided me with constant supervision all throughout

my Master's here at Laurier. I would especially like to

thank him for providing financial support and encouraging

me to take part in AARMS summer school in 2007, where

I accomplished part of the requirements for my degree.

Special thanks to all my graduate course instructors Dr.

Marc Kilgour, Dr. Joe Campolieti, Dr. Zilin Wang and

Dr. Roderick Melnik from whom I learnt and strengthened

my knowledge in different aspects of Mathematics. I would

like to thank Dr. George Lai, Dr. Zilin Wang, and Dr. De-

jan Delic for serving on my thesis committee. I am always

grateful to Dr. Sydney Bulman-Pleming for his contribu-

tion to my growth graduate student.

iv ACKNOWLEDGEMENTS

Last but not least, I would like to thank both my family

and my friends for their never-ending emotional support

which is essential for my work.

 Contents

Abstract i

Acknowledgements iii

List of Figures vii

Chapter 1. Introduction 1

1.1. Motivation 1

1.2. Graph Theory 5

1.3. Linear Algebra 9

1.4. Markov Chains 12

Chapter 2. The PageRank Algorithm 17

2.1. Introduction and Motivation 17

2.2. Random Walks on Graphs 19

2.3. The Google Matrix 21

2.4. Another View of PageRank 26

2.5. The World's Largest Matrix Computation 27

2.6. Implementation of a PageRank Calculator 31

vi CONTENTS

Chapter 3. PageRank in Power Law Graphs 33

3.1. Regularly Varying Random Variables 36

3.2. The Relationship between In-degree and

PageRank 39

3.3. Stochastic Equations 44

3.4. Numerical Experiments and Conclusion 49

Chapter 4. PageRank and In-degree 53

4.1. Introduction 53

4.2. Binary Trees 56

4.3. Calculating the Stationary Distribution 59

4.4. Proofs of Main Results 71

4.5. Conclusions for Binary Trees 77

4.6. PageRank of Random and Power Law Graphs 77

Chapter 5. Conclusion and Future Work 81

Appendix 85

Bibliography 89
 List of Figures

1.1 A strongly connected digraph. 7

2.1 A basic search engine model. 18

2.2 A directed graph G. 29

4.1 The binary tree T2(3) with its 0-1 labelling. 57

4.2 The binary tree T2(3) with a^j labelling. 58

4.3 An arbitrary binary tree 60

4.4 The directed binary tree T 2 (4). 67

4.5 PageRank versus in-degree for T2(4). 69

4.6 PageRank versus in-degree distributions for the

binary tree T2 (4). 70

4.7 The general form of a binary tree with the labeled

nodes. Here xr^ denotes the i-th node on the r-th

row. 72

4.8 The location of some of the nodes used in the

proof of Theorem 4.3.1. 74

viii LIST OF FIGURES

4.9 PageRank versus in-degree in a random digraph

with 100 nodes. 78

4.10 PageRank versus in-degree in a power law digraph

with 1,200 nodes. 79
 CHAPTER 1

Introduction

1.1. Motivation

With the rapid growth of the world wide web, informa-

tion retrieval presents increasing theoretical and practical

challenges. With the massive amount of information enter-

ing the world wide web every moment, it becomes harder

and harder to retrieve information from the web. That is

why the presence of a search engine is as vital as the exis-

tence of the web itself. Since the birth of the web, it has

been a central discussion in the web research community

to design faster, more efficient, and more accurate search

engines.

The most popular search engine currently is Google.

The mathematical theory behind the Google search engine

is the PageRank algorithm, which was introduced by Sergey

B r i n a n d L a w r e n c e P a g e [3], t h e f o u n d e r s of G o o g l e . In

1998, Brin and Page were PhD students. They took a leave

of absence from their Ph.D. to focus on developing their

I
2 1. INTRODUCTION

Google prototype. Their pioneering paper described the

PageRank algorithm, which is used to this day by Google

to generate its rankings.

A search engine consists of key components: a crawler,

and indexer, and a query engine [2]. The crawler collects

and stores data from the web. Data is stored in an indexer

which extracts information from the data collected from the

crawler. The query engine responds to queries from users.

As part of the query engine, a ranking algorithm, ranks web

pages in order of their relevance to the query. The ranking

is achieved by the assignment of a score to each web page.

PageRank is a ranking algorithm of web pages and uses

the link structure of the web. The web's link structure

forms a directed graph where the web pages are represented

as nodes and links as directed edges. A page is considered

"important" if it is pointed to by other important pages.

The following PageRank equation finds a score by solving

the iterative equation:

irT = TTT(aS + (l-a)J),

 1.1. MOTIVATION 3

where J is the matrix of all l's whose order equals the

number of pages. The matrix S is the stochastic matrix as-

sociated to the directed adjacency matrix of the web graph.

The parameter a is called the teleportation factor, a con-

stant between 0 and 1 which is normally assumed to be

around 0.85, and TT is the PageRank vector [13]. PageRank

will be discussed in detail in Chapter 2.

The PageRank vector consisting of the PageRank of each

web page is the stationary value of a large ergo die Markov

Chain [3]. The Perron-Frobenius theorem is used to en-

sure that the so-called Google matrix associated with this

Markov Chain will converge to a stationary distribution

[15]. The Perron-Frobenius theorem supplies a unique nor-

malized positive dominant eigenvector, called the Perron

Vector, which is the PageRank vector of the Google ma-

trix.

In a recent work by Litvak, Scheinhardt, and Volkovich [14],

a mathematical model is presented that derives an interest-

ing relation between PageRank values and in-degrees of web

pages. They investigate why the PageRank and in-degree of

4 1. INTRODUCTION

web pages follow similar power laws in the web graph. Fur-

thermore, they analytically prove that in power law graphs,

the tail distributions of PageRank and in-degree differ only

by a multiplicative factor [14].

The aim of my thesis is to first survey the mathematics of

the PageRank algorithm, and then to investigate the recent

work of [14]. In Chapter 2, I introduce PageRank and

describe its key properties. I will implement a PageRank

calculator and expose different graphs to my calculator. In

Chapter 3, we summarize the work of [14], who proved that

the ranking of the nodes by PageRank in power law graphs

will be similar to their ranking via their in-degree values.

For binary trees, we show in Chapter 4 that the ranking

result from PageRank is different from the ranking of their

in-degree values.

What follows in this chapter is the background and def-

initions needed throughout my thesis. As we will see, the

mathematical study PageRank uses a blend of graph the-

ory, probability, and linear algebra.

 1.2. GRAPH THEORY 5

1.2. Graph Theory

This section gives a concise introduction to the graph

theory terminology used later in my thesis. For a general

reference in graph theory, see [6]. A graph G consists of

a nonempty vertex set V(G), and an edge set E(G) of 2-

element subsets from V(G). A graph is sometimes called

network, especially with regards to real-world examples.

More formally, we may consider E(G) as a binary rela-

tion onV(G) which is irreflexive and symmetric. We often

write G = (V(G), E(G)), or if G is clear from context, then

we write G = (V, E). The set E may be empty. Elements

of V are vertices, and elements of E are edges. Vertices are

occasionally referred to as nodes, while edges are referred

to as lines or links. We write uv for an edge {u,v}, and

say that u and v are joined or adjacent; we may as well say

that u and v are incident to the edge uv, and that u and

v are endpoints of uv. The most common way to visualize

a graph is by drawing a dot for each node and joining two

of these dots by a line if the corresponding two nodes form

an edge. By a non-empty graph, we mean a graph with at

least one edge.

6 1. INTRODUCTION

We allow graphs to have multiple edges, but no loops.

A simple graph is a graph without multiple edges. The

cardinality |V(G)| is the order of G, while |£?(G)| is its

size. For a node v G V(G), degc{v) is the degree of v in G,

namely the number of edges in G incident with v. A node

of degree 0 is isolated.

For a node re in a graph G, define the neighbourhood of

x, written NQ(X), to the nodes joined to x. For X C V(G),

NQ{X) is the union of the neighborhoods over nodes from

X. If X C V, then define the subgraph induced by X,

written G \ X (or as either (X)G or G*[X]), to be the graph

with nodes from X, with two nodes joined in G \ X if and

only if they are joined in G. A subgraph of G is a graph H

such that V(H) C V(G) and E(H) C E(G). A graph G is

called bipartite if V(G) admits a partition into two classes

such that every edge has its ends in different classes (hence,

nodes in the same partition class must not be adjacent).

A graph may be directed or undirected. A directed graph

or digraph is defined analogously as an undirected graph,

except that now E{G) need not be a symmetric binary

relation on V(G). The edges are written as ordered pairs,

 1.2. GRAPH THEORY 7

and are called directed edges, (u,v), where u is the head

and v is the £ai/. The directed edge (u, v) is then said

to be directed from u to v. All the previously mentioned

features and definitions can then be modified to directed

graphs. The in-degree of u, written deg~(u), is the number

of nodes v such that (v, u) are directed edges; the out-degree

deg + (w) is defined dually. Moreover, a directed graph is

called strongly connected if for each pair of nodes (vi,Vj),

there is a sequence of directed edges leading from Vi to Vj.

The directed graph in Figure 1.1 is strongly connected. In

FIGURE 1.1. A strongly connected digraph.

all of the above definitions, we will not mention G if it is

clear from the content.

One of the most important examples of a graph for us

is the web graph. It is the graph where the nodes represent

web pages, and the edges correspond to links between the

8 1. INTRODUCTION

pages. We write W for this graph, which is a real-world

evolving graph. We may consider W an undirected or di-

rected graph, depending on the context.

A key property of the web graph is the presence of a

power-law degree distributions. Given a graph G and a

non-negative integer fc, we define NkjG by

Nk,G=\{xeV(G):degG(x) = k}\.

The parameter Nk,G is number of nodes of degree k in G.

The degree distribution of G is the sequence (N^^G '• 0 <

k <t). The degree distribution of G follows a power law if

for each degree k,

Nk,G h_p
—— ~ k p,
t
for a fixed real constant j3 > 1. We say that (3 is the ex-

ponent of the power law. A graph whose degree distribu-

tion follows a power law is often referred to as a power law

graph. Power laws for the in-degree and out-degree distri-

butions may be defined in a similar fashion. The in- and

out-degree distributions of the web graph were observed to

follow power law in the experiments conducted by Broder

 1.3. LINEAR ALGEBRA 9

et al. [4], which sampled 200 million web pages and their

links. For additional reading on the web graph, the reader

is directed to the books [2, 5, 8].

1.3. Linear Algebra

Matrices and vectors will be denoted in bold. Further,

all vectors are column vectors unless otherwise stated. For

a matrix A, we use the notation a^ for the zj-entry of

A. An m x n matrix A is a non-negative matrix whenever

each a,ij > 0, and this is denoted by writing A > 0. The

notation A > B means that each a^ > bij. A matrix A is

positive when each a^ > 0, and this is denoted by writing

A > 0. More generally, A > B means that each a^- > b^.

A convenient representation of a graph is via its adja-

cency matrix. The adjacency 'matrix A(G) of a digraph G

is defined by

( 1 if^-eE(G9,
aij = <
I 0 otherwise.

If G is undirected of order n, then A(G) is an n x n symmet-

ric (that is, A(G) = A(G)T) matrix. Adjacency matrices

are non-negative.
10 1. INTRODUCTION

For an n x n matrix A, a scalar A for which

det(A - AI) = 0

is called an eigenvalue of A. A nonzero n x 1 vector x

for which A — Ax is the eigenvector of A for A. The pair

(A,x) is called an eigenpair for A. The set of all distinct

eigenvalues, denoted by <J(A) , is called the spectrum, of A.

The eigenvalues and eigenvectors are fundamental topics

in PageRank calculations. The adjacency matrix A(G) for

an undirected graph G is a real and symmetric matrix, and

hence, has n real eigenvalues Ai > 0, A2,..., An, which can

be ordered by their absolute values:

Ai = |Ai| > |A2| > . . . > |A„|.

(See, for example, [2].) The first (that is, largest in absolute

value) eigenvalue Ai is the radius of the spectrum, denoted

by p(A). The real number Ai is also called the dominant

eigenvalue.

We now state Perron-Frobenius theorem. A proof of this

important result may be found in [15]. A non-negative

matrix A is primitive if Am > 0 for some m > 0. The

 1.3. LINEAR ALGEBRA 11

1-norm (or taxicab norm) of x is defined as

n
Mli = £ T •I

•1=1

THEOREM 1.3.1 (Perron-Frobenius). If a matrix A > 0

is primitive, then each of the following assertions holds.

(1) r = p(A) > 0.

(2) There exists an eigenvector x > 0 such that A x =

rx.

(3) The Perron vector p is the unique vector satisfying

Ap = rp

and which is positive with 1-norm equal to 1. There

are no other non-negative eigenvectors for A regard-

less of the eigenvalue, except for the positive multi-

ples of p.

We will sometimes use limits of matrices. If (Mt) is a

sequence o f m x n matrices, and L is an m x n matrix, then

we write

lim Mt = L
t—>oo
12 1. INTRODUCTION

if for all 1 < i < m, 1 < j < n,

lim {Mt\3 = Lid.

,J
t—>oo

1.4. Markov Chains

Markov Chains provide a powerful framework for mod-

elling certain random processes. Our approach to ana-

lyze the PageRank algorithm in Chapter 2 will use Markov

chains. We give a brief discussion of Markov chains in this

section. For a general reference in probability theory, see

[11].

Fix n a positive integer. We denote ¥(A) the probability

of an event A in a probability space. A (discrete-time,

time-homogeneous, finite-state) Markov chain M consists

of a discrete-time random process (Xt : t G N) each with

codomain in the same finite set S — {ao,... ,an} with the

property that for all n > 1 and 1 < t < n,

F(Xt = at\Xt-i = o t _ i , . . . ,XQ = a 0 ) = F(Xt = at\Xt-i = <H-i)-

 1.4. MARKOV CHAINS 13

This definition expresses that the state Xt depends on the

previous state Xt-i, but is independent of how we actu-

ally arrived at Xt~\. In other words, the random process

does not remember the way it reached the state Xt-\. This

property is called Markovian or memoryless property for a

random process. It is important to note that the Markov

property does not imply that the state Xt does not depend

on the random variables XQ, X\,..., Xt-2- However, what a

Markovian property guarantees for Xt, is that any such de-

pendency on the past will be captured and recorded in the

value of Xt-\. In other words, only the present state giyes

any information of the future behaviour of the process. See

[16] for more background on Markov chains.

The set of possible values S of M is called the state

space, and without loss of generality we will always consider

this to be { 1 , . . . , n}, where n is an integer. The transition

probability

Pid=F(Xt=j\Xt-i=i)

is the probability that the process moves from state i to

state j in one time-step. Using the Markovian property,

14 1. INTRODUCTION

every Markov chain can be uniquely expressed by a transi-

tion matrix defined as

/ \
Po,o Po,i p0,j
Plfi P\,l Pl j
P =

Pifi Pi,I PihJ

V /

Hence, the zj-entry in the matrix is the transition prob-

ability Pi j . The representation of a Markov chain via its

transition matrix makes it feasible to compute and predict

the distribution of the future states of the process.

A useful representation of a Markov chain is via a di-

rected weighted graph. The nodes correspond to the states,

and the weight on each directed edge is the positive transi-

tion probability of getting from the head state to tail state.

There is a directed edge (i,j) if and only if Pij > 0. A sto-

chastic matrix is a non-negative matrix in which each row

sum is equal to 1. Note that the transition matrix of every

Markov chain is a stochastic matrix (which follows from

the basic probability definitions). A stationary distribution

 1.4. MARKOV CHAINS 15

s of a Markov chain is a probability distribution (that is, a

vector whose sum of all entries equals 1) with the property

that

sT = sTP.

We can also express this by saying that s is an eigenvector

of P with eigenvalue 1, or s is a fixed-point of P .

Stationary distributions exist and are unique if the Markov

chain has a primitive transition matrix [2]. We refer to such

Markov chains as ergodic. Hence, if we consider an ergodic

Markov chain over a long period of time, the initial state

becomes increasingly forgotten, and the probability that we

are in state i approaches the ith component of s. As we will

see in Chapter 2, the PageRank vector corresponds to the

stationary distribution of a certain Markov chain.

 CHAPTER 2

The PageRank Algorithm

2.1. Introduction and Motivation

Information retrieval is the process of searching within a

collection of documents for a particular item of information.

The information you are looking for is normally called a

query. To retrieve information from the world wide web,

we need to first be able to model the web. The best way to

model a massive network like the web is by representing it

as a digraph. Each web page is a node of the graph and the

links between two nodes are directed edges. To perform a

search in this network, we should first be able to gather all

of the information about its link structure in a database,

and then classify and retrieve the query from this database.

A search engine consists of a crawler, indexer and a

query engine; see [13]. See Figure 2.1 for a simplified model

of a search engine.

17
18 2. THE PAGERANK ALGORITHM

Query
Crawler Indexer
Processor

' •

Ranking of the
Results

F I G U R E 2 . 1 . A basic search engine model.

The crawler performs frequent visits to the entire (or

a large part of the) world wide web. The crawler trav-

els from page to page to keep track of the existing links,

and more importantly to update our database with new

web pages and links. Imagine a backpacker who is walking

through every link, and upon arriving at every new web

page, writes down the address of the page and summarizes

its content. There are certainly web pages that have no

out-links. Hence, our backpacker will get stuck there. Af-

ter recording such pages (called dangling nodes), the back-

packer will step back as many steps needed to be able to

find a way out.

The links are classified once they are entered into the

indexer. Hence, now we can search for our query in the

 2.2. RANDOM WALKS ON GRAPHS 19

indexer. The search is done, and some number of pages

associated to the query are found.

After the search is done, let us say there are 200 pages

found in the end. The question becomes: how to rank

these pages? A useful way to display the 200 pages on the

result screen is to rank them by popularity. We therefore

need to find out how to rank web pages according to their

popularity. This is the role of the query engine. As we

will describe in Section 2.2, the PageRank algorithm is one

effective way to accomplish this ranking.

2.2. R a n d o m Walks on Graphs

Before we define PageRank, we make a short digression

to discuss random walks. A random, walk on a connected

graph G is a certain type of Markov chain defined by the

sequence of moves (over discrete time-steps) of a particle

between nodes of G. The location of the particle at a given

time-step is its state. In the uniform random walk, the par-

ticle may move from its current state to any of its neigh-

bouring nodes (with equal probability). A uniform random

walk on a graph of size n may be represented by a transition

20 2. T H E PAGERANK ALGORITHM

probability matrix P whose entries are p^-, where

Pij = <
I 0 otherwise.

Note that the transition probability matrix of a uni-

form random walk is stochastic; that is, the row sums are

equal to 1. We define an ergodic Markov chain to be one

whose transition matrix is primitive. An important the-

orem states that an ergodic Markov chain always has a

stationary distribution [19]. Hence, the stationary proba-

bility distribution exists for the uniform random walk if its

transition matrix is primitive, and is a probability vector

s T such that

s T P = sT.

The following theorem states sufficient conditions on G for

the stationary distribution to exist.

THEOREM 2.2.1. [2] Let G be a finite, connected, non-

bipartite graph. A random walk on G converges to a sta-

tionary distribution s T = (SJ), where

_ deg(i)
S%
~2\E(G)[
 2.3. THE GOOGLE MATRIX 21

Analogous result holds for uniform random walks on di-

rected graphs.

2.3. T h e Google Matrix

To define the Google matrix for an arbitrary graph G, we

consider the transition probability matrix for the uniform

random walk on G. Let n = \V(G)\ be the order of G and

apply a fixed enumeration from 1 to n to the nodes of G.

For the directed graph G, the matrix P i is defined by

I 0 otherwise.

The structure of the P i matrix guarantees that at every

node, the surfer will have equal probability to choose one

of the out-neighbours. If there is no out-link from i to j ,

then this probability is 0. In the web there always exist

web pages that do not link to any other web pages. These

nodes are called dangling nodes. If we assume the only

way to visit the web pages is by following the out-links,

then the surfer gets stuck at such nodes. To overcome this

problem, we manipulate the matrix P i in a way to bypass

the dangling nodes. Define P2 to be the matrix P i such

22 2. T H E P A G E R A N K ALGORITHM

that any zero rows are replaced with the vector with each

entry equal to - . Define the Google matrix (or PageRank

matrix) by
1—a
P = aP2 H J„,n,
n
where a; is a fixed real number in (0,1), and J n j n is the nxn

matrix of all l's. (We do not use the notation G for the

Google matrix, as G is reserved for graphs.) The constant

a, called the teleportation factor, is a parameter measuring

the frequency at which a surfer jumps to a new randomly

chosen web page, rather than following the out-links. We

now show why Google matrix is stochastic and primitive,

and hence, has a stationary distribution.

LEMMA 2.3.1. For a graph G with order n and P =

P(G*) equalling its Google matrix, then the following asser-

tions hold.

(1) The matrix P is stochastic.

(2) The matrix P is primitive.

Proof. For item (1), to show that P is stochastic, we

must show that the row sums in P are all equal to 1. For
 2.3. THE GOOGLE MATRIX 23

a fixed 0 < i < n, the row sum r,b equals:

Ti = ^ (a(P2kj+—^— (Jn,n)ij)
Kj<n ^ '

1—a

= " E ( P 2)*J + ( ! - « ) •
l<j<n

To find the value for J^ (^2)1^, we consider two cases.

Case i. Node i is dangling node.

In this case,

(P2)„ = i.
Hence,

= a V - + (1 - a)
l<j<n

= a + (1 — a) = 1.

Case £. Node i is not dangling.

In this case,

(P2)id = (Pi),,,.
24 2. T H E P A G E R A N K ALGORITHM

Hence,

n = a J2 (Pikj + ( l - a )
l<j<n

— a + (1 — a) = 1.

For item (2), since all entries of P are positive, P is primi-

tive. •

Lemma 2.3.1 demonstrates that the Google matrix P is

a transition probability matrix of an ergodic Markov chain.

The Markov chain associated to this matrix is called the

PageRank Markov chain, or the PageRank random, walk.

In this random walk, at any page, the surfer visits an out-

neighbour of that node with probability a and visits any

other node in G with probability 1 — a. In practice, the

parameter a is normally assumed to be around 0.85; see

[13].

We will now use the linear algebra preliminaries stated

in Chapter 1 of the thesis to prove an important theorem

about the PageRank random walk. The following theorem

 2.3. THE GOOGLE MATRIX 25

guarantees that with the described structure of Google ma-

trix, the PageRank random walk has a unique stationary

distribution, called the PageRank vector.

THEOREM 2.3.2. Fix a graph G. The PageRank Markov

chain with transition probability matrix P = P ( G ) con-

verges to a unique stationary distribution s.

Proof. Since P is positive and primitive, the PageRank

Markov chain is ergodic, and hence, converges to a station-

ary distribution s. To show that s is unique, we will use the

Perron-Frobenius theorem (See Theorem 1.3.1). By Theo-

rem 1.3.1, P has a unique positive and dominant eigenvalue

equal to 1. The corresponding eigenvector for this eigen-

value would be the vector s, where

s T P = sT. •

The vector s is the PageRank vector for the Google ma-

trix P . The entries in the PageRank vector are the PageR-

ank values for each node in the graph G (associated with

the fixed enumeration of V(G)). To calculate the PageRank

values, we need to find the stationary vector of the Google

26 2. T H E P A G E R A N K ALGORITHM

matrix. In Section 2.4 we discuss a practical method used

to calculate the PageRank, called the power method. In the

next section, we explain another approach to calculate the

PageRank vector first used by Brin and Page [3].

2.4. Another V i e w of PageRank

The original formula for PageRank due to [3], is a sum-

mation formula which calculates the popularity of the pages

by adding up the PageRank of all the pages pointing to this

web page. Let PR(Pi) denote the PageRank of the page

Pi and let In(Pi) denote the set of web pages that point to

Pi. The PageRank is then

(2.i) PR(P,)= J2 ^§r-

A problem is that the values for PR(Pj) are unknown. To

overcome this, we need to initialize all the web pages with

an equal PageRank value, and then transform equation

(2.1) into a recursive equation. Brin and Page assumed

that at the beginning all the pages have a constant PageR-

ank value of - , where n is the total number of pages in the

web graph. The iterative procedure calculates PageRank

 2.5. THE WORLD'S LARGEST MATRIX COMPUTATION 27

at (k + l)-th step as

(2.2) PRM(P,)= £ ^ p .

The process initiates with setting PRo(Pi) — ^ for all pages

Pi. As discussed in the previous section, since the PageR-

ank Markov chain is ergodic, the eventual convergence of

the PageRank scores is guaranteed.

2.5. T h e World's Largest Matrix C o m p u t a t i o n

Cleve Moler, the founder of the well-known mathemati-

cal software matlab, cited PageRank as "The World's Largest

Matrix Computation" in [17]. At that time Google was ap-

plying the Power Method to a sparse matrix of order 2.7

billion. Now, it has at least 54 billion rows and columns!

(See [2].)

To find the PageRank vector, we should solve for the

eigenvector s such that

s T P = sT.

Since P is a dense massive matrix, a direct approach to the

calculations will not be feasible in general. To overcome

28 2. T H E P A G E R A N K ALGORITHM

the computational problems, the power method is used to

approximate the PageRank vector s. The algorithm works

as follows. Fix a directed graph G of order n.

(1) Initialize zo to be the stochastic vector with every

entry equal to 1/n.

(2) Define
zTk+1 = zpP = (zj )P*

The sequence (z& : k € N) consists of stochastic vectors,

since at every time step, we have the result of the product

of two stochastic matrices. It can be shown that

Inn zfc+i
K—>00

is the dominant eigenvector of the Google matrix; see [2].

From the Power method, we can approximate the PageR-

ank vector by taking powers of the Google matrix and mul-

tiplying it by zo- This amounts to simply summing up each

column of P and multiplying the sum by 1/n.

For completeness, we give an illustration of a PageRank

computation. Figure 2.5 shows a sample graph G with six

nodes. To find the PageRank vector of G. we first compute

 2.5. THE WORLD'S LARGEST MATRIX COMPUTATION 29

the various matrices required in the definition of the Google

matrix.

FIGURE 2.2. A directed graph G.

The P i matrix for G is

( 0 1/2 1/2 0 0 0 \

0 0 0 0 0 0
1/3 1/3 0 0 1/3 0
Pi
0 0 0 0 1/2 1/2
0 0 0 1/2 0 1/2
\ 0 0 0 1 0 0 j
30 2. THE PAGERANK ALGORITHM

The P2 matrix is

( 0 1/2 1/2 0 0 0 \

1/6 1/6 1/6 1/6 1/6 1/6

1/3 1/3 0 0 1/3 0

0 0 0 0 1/2 1/2

0 0 0 1/2 0 1/2

\ 0 0 0 1 0 0 /

If a = 0.85, then let a = a / 2 , b = (1 — a ) / 6 , c = a / 6 , and

d = a / 3 . Then the Google matrix of G is then

1 b a+b a+b b b b ^

6 + c 6 + c 6 + c 6+ c 6 + c b + c
b+d b+d b b b+d b

b b b b a+60+6
b b b a+b b a+b

\
b b b a+b b b

Using the power method, the approximate PageRank

vector (using the natural ordering {1,2,3,4, 5,6}, and with
two decimal places of accuracy) is

0.05 0.07 0.06 0.35 0.2 0.27

 2.6. IMPLEMENTATION OF A PAGERANK CALCULATOR 31

2.6. Implementation of a PageRank Calculator

In this section, a brief description of an implementa-

tion of PageRank is given on a graph with n nodes. The

PageRank calculator implemented for the thesis uses the

algorithm provided in [13].

(1) The vector piO is the initial vector, which we nor-

mally set to 1/n.

(2) H is the manipulated hyperlink matrix, P%.

(3) n is the size of the matrix or the web.

(4) alpha is the teleportation factor.

(5) e p s i l o n is the convergence tolerance; in the ac-

tual implementation, we set the total iteration steps

equal to 20.

(6) The vector a is the dangling node vector in which

an entry is 1 if its corresponding node is a dangling

node, and 0, otherwise.

%Implementation of PageRank c a l c u l a t o r
%using power method
function [pi,time,numiter]=

PageRank(piO,H,n,alpha,epsilon);
32 2. T H E P A G E R A N K ALGORITHM

rowsumvector=ones(l,n)*H';

nonzerorows=find(rowsumvector);

zerorows=setdiff(1:n,nonzerorows);

l=length(zerorows);

a=sparse(zerorows,ones(l,1),ones(l,1),n,1);

k=0;

residual=l;

pi=piO;

tic;

for ( i=0:20 )

%while(residual < epsilon)

prevpi=pi;

k=k+l;

pi=alpha*pi*H + (alpha*(pi*a)+l-alpha)

*((l/n)*ones(l,n));

residual=norm(pi-prevpi,1);

end;

numiter=k;

time=toc;

70save p i ;
 CHAPTER 3

PageRank in Power Law Graphs

PageRank roughly measures the popularity of a web

page based on its number of in-links. We discussed PageR-

ank in detail in Chapter 2, and we proved that it is the

stationary distribution of the PageRank random walk. We

now present recent work by Litvak et al. [14], who proved

that under certain assumptions PageRank and in-degree

distributions of a power law digraph obey a power law with

the same exponent. To prove this result, we model the re-

lation between PageRank and in-degree via a stochastic

equation. All the results described in this chapter come

from [14].

Studying the potential similarity between PageRank and

in-degree of the web pages is of particular importance be-

cause it provides ground for simpler, cheaper and less time-

consuming calculations. The matrix calculations performed

to estimate PageRank, are massive. However, it is straight

forward to find the in-degree of a web page.

33
34 3. PAGERANK IN POWER LAW GRAPHS

To study the behaviour of the PageRank tail, Laplace-

Stieltjes transforms are used. The Tauberian Theorem and

the theory of regularly varying variables are then applied

to a certain stochastic equation to prove analytically that

the tails of PageRank and in-degree distributions vary only

in a multiplicative constant. Hence, the PageRank and in-

degree distributions in power law graphs follow power laws

with the same exponent.

We begin by recalling the PageRank equation in its sum-

mation form. (See Equation 2.1 from the previous chapter.)

(3.1) PR(i) = a J2 ^ T ^ + (1 - ")•

J
jeN(i)

An interpretation of (3.1) is that the PageRank of node

i depends on the in-degree of i and PageRank of its in-

neighbours. However, it is important to note that while

the linear algebraic methods often used in PageRank lit-

erature work well for most PageRank computations, they

are not sufficient for analyzing the asymptotic properties

of the PageRank distribution. The mathematical approach

to PageRank analysis used in [14] stems more from applied

 3. PAGERANK IN POWER LAW GRAPHS 35

probability and stochastic operations research, than from

linear algebra.

In Donato et al. [7], Fortunato et al. [9] and Becchetti

and Castillo [1], experiments performed on the web graph

confirm the similarity in tail behavior of PageRank and in-

degree distributions. The exponent value (3 for the power

laws of the PageRank and in-degree distributions were found

in all cases to be around 1.1. Moreover, the cited experi-

mental studies have shown that the PageRank of the top

10% of the nodes always follows a power law with the same

exponent independent of the teleportation factor a.

In a power law distribution, there is a so-called 30-70

rule: the tail will cover 70 percent of the value of the distri-

bution. We will therefore compare the tail distribution of

PageRank and in-degree. In other words, we focus on tail

asymptotics for PageRank and its relation with in-degree.

Since we are only interested in the tail, we are looking into

web pages with high popularity or PageRank value, which

can be stated as

(3.2) F(PR> x),

36 3. PAGERANK IN POWER LAW GRAPHS

for a suitably large x, and where F(A) is the probability

of the event A in a probability space. Observe that (3.2)

defines the fraction of pages having PageRank greater than

x, where x is large. One way to analyze such a probability

is to find an asymptotic expression p{x) for which

,. F(PR > x) .,
lim — — — — - = 1.
x^oo pyx)

If such a p(x) is found, then p(x) and ¥(PR > x) are

asymptotically equal, and so we can approximate the tail

of PageRank by p(x).

3.1. Regularly Varying R a n d o m Variables

A real-valued function RV(x) is said to be regularly

varying of index j3 G M. if for every t > 0,

lim ™ = if.
z-»oo RV(X)

A real-valued function SV(x) is said to be slowly varying

if for every t > 0,

SV{tx)
lim = x
"cTTTT -
 3.1. REGULARLY VARYING RANDOM VARIABLES 37

A careful look at the above definitions leads us to a relation

between RV and SV functions: every regularly varying

function can be written as

RV(x)=x<3SV{x),

for some slowly varying function. We can also define the

regularly varying property for random variables as well as

functions. Recall that a(x) ~ b(x) if

a(x)
lim = L
T^T

A random variable X is said to be regularly varying with

index (3 if its distribution F(x) can be written as

1-F(x)~ x~pSV{x),

for some slowly varying function SV(x). The Laplace-

Stieltjes transform of X is

/( 5 ) = E[e-*],
38 3. P A G E R A N K IN P O W E R L A W GRAPHS

where s > 0 and E(Y) is the expectation of the random

variable Y. The n-th moment of X is written as

poo
£„ = / XndF(x).
Jo

By expanding / in a series at s = 0, the successive moments

of F can be obtained. Moreover, the n-th moment of X is

finite if and only if there exist coefficients £0, • • •, £,n such

that fn(s) — o(sn), as s —» 0. The following lemma states

this in a precise fashion.

LEMMA 3.1.1 ([14]). The n-th moment of X is finite if

an
and only if there exist real numbers £0 = 1 d £i> • • • > £n;

such that
n p

s—>0 •«•—' ?!
i=0

The above coefficients & may be uniquely found. If £n <

co, then we write:

fn(s) = (-ir+i(f(s)-J2^(-s^
\ i=0

Later, we will use fn(s) further to discuss the tail proper-

ties of the distribution. There exist an important relation

 3.2. THE RELATIONSHIP BETWEEN IN-DEGREE AND PAGERANK 39

between asymptotic behaviour of a regularly varying distri-

bution and its Laplace-Stieltjes transform. The following

theorem, used throughout this chapter, makes this relation

precise.

THEOREM 3.1.2 (Tauberian Theorem; [14]). For n G N

and if£n < oo7 (3 = n + r), and rj G (0,1), then the following

are equivalent

(1) / n ( s ) ~ ( - 1 ) ^ ( 1 - pySV{% ass^O

(2) 1 - F(x) ~ x-PSV(x), as x -> oo.

The proofs of the above lemma and theorem may be

found in [14]. Theorem 3.1.2 plays an important role in

finding the relation between asymptotic distributions of

PageRank and in-degree.

3.2. T h e Relationship between In-degree and

PageRank

We now describe the relationship between PageRank

and in-degree. We consider equation (3.1), but make some

important simplifying assumptions. These assumptions will

enable us to model this relation by focusing on the influence

of in-degree without considering other factors. Naturally,

40 3. PAGERANK IN POWER LAW GRAPHS

these assumptions are not realistic, but in further discus-

sions we try to reduce them by generalizing the model ob-

tained. Rewrite equation (3.1) in the following form of a

distributional identity with the random variable R:

M
(3.3) R = a^2-Rj + (l-a),
Lb
3=1

where = represents a distributional identity and M is the

in-degree of the considered random page. The assumptions

we make are as follows.

(1) Let R represent the PageRank of a randomly chosen

page. One of our goals in this chapter is to deter-

mine the distribution of the random variable R.

(2) Fix d > 1 and assume that it is the number of out-

going links for all pages. Hence, out-degree is equal

for all nodes.

(3) The dangling node effect is neglected. That is, we

do not consider the effect of pages without outgoing

links.

(4) The random variables R and M are independent.

That is, the in-degree distribution and PageRank

 3.2. THE RELATIONSHIP BETWEEN IN-DEGREE AND PAGERANK 41

distribution of a random page have independent dis-

tributions (which is not the case, in general).

(5) All Rj's are independent and have the same distri-

bution as R and hence, R = 1 constitutes the unique

solution of the equation (3.3).

The equation (3.3) has the same form as the original PageR-

ank formula as in equation (3.1).

We will now find the in-degree distribution for a ran-

domly chosen web page. Although it is well-known that

the in-degree distribution of the web graph follows power

law (see for example, [2]), we need to be able to formally

describe this random variable for our analysis. We use the

theory of regularly varying random variables. The in-degree

of a randomly chosen page is modeled by a non-negative

integer-valued, regularly varying random variable which is

distributed as N(X). In particular, the random variable X

is regularly varying with index (3 and N{x) is the number

of Poisson arrivals on the time interval [0, x\. For more de-

tails on Poisson processes and their application in Markov

chains, see Sections 8.2 and 8.3 of [12].

42 3. PAGERANK IN POWER LAW GRAPHS

The variable N(x) is a "discretization" of the random

variable X. In this way, we guarantee that while in-degree

has a power law distribution, it only takes integer values

and hence, we do not have to put any restrictions on X.

In Theorem 3.2.1 below, we will prove that N(X) is also

regularly varying with the same index as X, and so fol-

lows a power law with the same exponent. First, let Fx

and FN(X) be the distribution functions of X and N(X),

respectively. Let / and <fi be their corresponding Laplace-

Stieltjes transforms.

THEOREM 3.2.1 ([14]). The following are equivalent.

(1) 1 - Fx(x) ~ x-PSV(x), as x -> oo

(2) 1 — F/v(x) ~ x~^SV(x), as x —> oo

We give a brief sketch of the theorem. We first need a

technical lemma.

LEMMA 3.2.2 ([14]). Let fn(s) and(j)n(s) be the Laplace-

Stieltjes transforms of X and N(X), respectively. Then

fn(s) = o(sn) if and only if (f>n(s) = o(sn).

 3.2. THE RELATIONSHIP BETWEEN IN-DEGREE AND PAGERANK 43

While we omit the proof of the lemma, here is an infor-

mal sketch of its proof. One shows that the corresponding

moments of X and N(X) always exist. It may be shown

that since we fixed the out-degree of all pages to be equal

to d, then the average in-degree would also equal d. That

is, E[X] = d and similarly, E[N(X)] = d. The final step

in the proof of Lemma 3.2.2 is to consider the generating

function of N(X) and derive its Laplace-Stieltjes transform

in terms of the Laplace-Stieltjes transform of the random

variable X.

We now sketch a proof of Theorem 3.2.1.

Proof of T h e o r e m 3.2.1. We only prove that (1)

implies (2). Prom Theorem 3.1.2 in the previous section,

we have that

1 — Fx{x) ~ x~^SV(x), as x —> oo

implies that

fn(t) ~ ( - 1 ) ^ ( 1 - 0)1?SV Q , as t -* 0

where n is the largest integer smaller than /?, and T is the

gamma function. Since 0(s) = f(t)7 by Lemma 3.2.2 we

44 3. PAGERANK IN POWER LAW GRAPHS

have that fn(s) ~ o(sn) where t(s) = 1 — e~s ~ s. By

Lemma 3.1.1 and Theorem 3.1.2, we have that

1 — FN(X)(X) ~ x _ / ? 5F(a:) as x —> oo. D

The model of Litvak et al. for the number of incoming

links of a randomly chosen web page works well, since it de-

scribes an in-degree distribution which follows a power law

with finite expectation and a non-integer exponent (3 > 1.

Having obtained the distribution for in-degree and PageR-

ank, we will now proceed to retrieve the main stochastic

equation for the relation between PageRank and in-degree

and compare their tail distributions in the next section.

3.3. Stochastic Equations

Using the discussion in the previous two sections, we can

now reformulate the equation (3.3) as follows:

N(X)
(3.4) R = a Y; -jRj + {l-a),
CX
3=1

where a € (0,1) is the teleportation factor, d > 1 is the

fixed out-degree of each page, and N(X) describes the in-

degree of a randomly chosen page in terms of the Poisson

 3.3. STOCHASTIC EQUATIONS 45

arrivals on a regularly varying random variable X which

represents time. The stochastic equation (3.4) adequately

represents the PageRank distribution and its relation with

the in-degree distribution. We can now apply analytical

methods to study the tail behaviour.

The main idea of the analysis is to apply the Laplace-

Stieltjes transforms of X and R. By using the Tauberian

Theorem, we may prove that R is regularly varying with

the same index as X. By the Theorem 3.2.1, this then

guarantees similarity in the tail behaviour of the PageRank

R and the in-degree N{X).

The first step is to write the Laplace-Stieltjes transform

of the PageRank distribution R in terms of the probability

generating function of N(X). Let GN(X) be the generating

function of N(X). By applying Laplace-Stieltjes transform

46 3. PAGERANK IN POWER LAW GRAPHS

to the definition of R in (3.4) we have that

N{X)
r{s) = E[e- ] = sR l a a \]
E{e-< - ^}E exp | — s— Ri
~d i=i

exp
= e
—s(l—a)
r4E^ i=l
F(N(X) = k)

oo
s(l-a) r\s-X) F{N(X) =k)
k=i
d

Ki-")(S N(X) a
r s
~d

Note that for all i, we have Ri = R, and that is how the

second equality in the above set of equations is obtained.

can
In the following corollary, we prove that GJV(X) be ex-

pressed in terms of the Laplace-Stieltjes transform of X.

COROLLARY 3.3.1 ([14]). I/GN(X) ^ the generating func-

tion of N(X) and f is the Laplace-Stieltjes transform of X,
then we have that

&N(X)(s) = / ( l - S)
 3.3. STOCHASTIC EQUATIONS 47

Proof. By definition of the generating function,

GN(X)(S) = E[sNW]
roo
= / E[sN®]dFx(t)
Jo
roo
= / e-^dFxit)
Jo
= f(l-s). •

The above corollary leads us to an important conclusion for

the Laplace-Stieltjes transform of R. By Corollary 3.3.1, as

CJV(X)(S) = / ( I — s) w e obtain that

(3.5) r(s) = f(l-r(^s))e-<1-a\

As in the previous section where the distribution for in-

degree was calculated, here we perform the analysis by

showing the correspondence between the existence of the

n-th moments of X and R. The independence of N(X)

and the Rj's is heavily used. For example, using this in-

dependence, we can take the expectation from both sides

48 3. PAGERANK IN POWER LAW GRAPHS

of (3.4). Similar to the result of Lemma 3.1.1, we can re-

formulate it and show that

(3.6) fn(s) = o{sn) if and only if rn{s) = o(sn).

Now we can present the final theorem in this chapter in

which the observed correlation between in-degree and PageR-

ank distributions is explained in power law graphs. The

proof of the theorem can be found in [14].

THEOREM 3.3.2. The following are equivalent.

(1) 1 — FN(X){X) ~ x~^SV(x), as x —> oo In particular,

N(X) is a regularly varying random variable with

index j3.
<
(2) 1 — FR{X) ~ d0 ^a/3dx~f3SV(x), as x —> oo. In partic-

ular, R is a regularly varying random variable with

the same index (3.

Theorem 3.3.2 shows that the asymptotic behaviour of

PageRank and in-degree are similar in power law graphs,

since they both follow a power law with equal exponents

(they differ only by the multiplicative factor J^_ Pd).

 3.4. NUMERICAL EXPERIMENTS AND CONCLUSION 49

3.4. Numerical Experiments and Conclusion

How do we verify a power law behaviour in practice?

It is not always simple to plot, measure, or numerically

identify power law distributions. A well-known technique

is to plot the so-called log-log graph of the distribution.

More precisely, we plot the degree distribution in logarith-

mic scale and expect to obtain a straight line. Experiments

conducted by Newman in [18] suggests that since we are

focusing on tail distributions, we should plot the fraction

of quantities which are not less than a certain value. In

particular, we should plot the complementary cumulative

function instead; that is,

l-F(x)=F(X >x),

rather than to plot the histogram. In this way, we will have

a more concentrated plot.

Another issue is that if a distribution X follows power

law with exponent (3 such that 1 — F(x) ~ Cx~^, where C is

a constant, then the corresponding histogram has exponent

50 3. P A G E R A N K IN P O W E R L A W GRAPHS

(3+1. Thus, the plot of 1 — F{x) on logarithmic scale will

have a smaller slope than the original plot of the histogram.

Computation of the correct slope from real-world data

is also an important part of the numerical analysis. Gold-

stein et al. in [10] suggest that using an MLM (Maximum

Likelihood Method) is advantageous over the standard least

square fit method, since the former provides us with a more

robust estimation of the power law exponent. The calcula-

tions based on MLM yield a slope of —1.1 which confirms

that both in-degree and PageRank have power laws with

the same exponent (3 — 1.1.

In the results retrieved from the experiments using web

data, Litvak et al. focus on the right tail behaviour of the

PageRank distribution. The result is that in a log-log plot,

both in-degree and PageRank distributions plot as parallel

lines for all values of the teleportation factor, as long as

we focus on large PageRank values. In fact, comparing

PageRank and in-degree does depend on the teleportation

factor. However, the PageRank distribution of the top 10%

of web pages obeys a power law with the same exponent as

in the in-degree, independent of the teleportation factor.

 3.4. NUMERICAL EXPERIMENTS AND CONCLUSION 51

Despite their results, the Litvak et al.'s model however,

lacks the realistic dependencies between the PageRank val-

ues of the pages sharing a common neighbor. This is why

the exact value of the multiplicative constant provided in

Theorem 3.3.2 does not fit the results from their web crawls.

Further work would be to reduce the assumptions made

in Section 1.2 so that the generalized model can capture

mainly the dangling node effect and the dependencies be-

tween PageRank values of the pages pointing to one certain

web page.

In conclusion, Litvak et al. showed that in power law

graphs, PageRank and in-degree follow the same power law

distribution which varies only in a multiplicative constant.

In the next chapter, we provide examples of graphs where

the PageRank and in-degree do not follow similar tail dis-

tributions.
 CHAPTER 4

PageRank and In-degree

4.1. Introduction

In this chapter, we supply some examples complement-

ing the findings of [14]. Before we begin, let us have a quick

review of the materials discussed in Chapter 2 on PageR-

ank. The PageRank vector for a digraph G is calculated

by first calculating the PageRank matrix

(4.1) P = P(G) = a P 2 + ^—^-3n,n,

n

where J n / n is the nxn matrix of all l's and a £ (0,1) is the

teleportation constant. The matrix P defined on the left

hand side of the equation above is the Google matrix (or

PageRank matrix). The PageRank matrix is positive and

stochastic, and therefore, is the transition matrix for some

Markov chain.

The Markov chain attributed to the PageRank matrix

converges to a stationary distribution s. This convergence

53
54 4. PAGBRANK AND IN-DEGREE

is guaranteed as it is an ergodic Markov chain. Since s

is the dominant eigenvector of the transition probability

matrix of this Markov chain,L) we have that

The vector s is called the PageRank vector, whose ith en-

try is the PageRank of the ith node of the graph (according

to some fixed enumeration of the nodes). Hence, to calcu-

late the PageRank vector of a graph, we should find the

stationary distribution of the Google matrix P in (4.1).

A good approximation to the PageRank vector can be

evaluated using the Power method, discussed earlier in Sec-

tion 2.3. For this method, we start with an initial (arbitrary

but fixed) non-negative, non-zero vector so, and then define

(4.2) sj+1 = sf

= (sJ)P*-

After a sufficient number of iterations (normally 20 to 50

in practice; see [3]), s approximates the PageRank vector.

The iterative process in (4.2), presents a useful alternative

 4.1. INTRODUCTION 55

for calculating the s. In (4.2), there are two steps: first, we

raise the Google matrix P to a power t and then multiply it

by the vector so- If we take so to be the vector of all 1 's, then

this multiplication will give the column sum of the matrix

P . Hence, the PageRank vector is simply the column sum

of the limiting vector in the powers of the Google matrix

(which is later on normalized to ensure it is stochastic).

The Google matrix, however, is a dense matrix and the

Power Method calculations involving matrix multiplication

become increasingly costly as higher powers are formed. An

alternative is to only work with the sparse matrix P2. In

this case, the stationary distribution of the uniform random

walk is computed (not PageRank).

Litvak et al. [14] introduced numerical methods and a

new model that proves that with certain assumptions, in

power law graphs, the PageRank and in-degree distribu-

tions are similar. This result is interesting and of practical

importance because PageRank calculations are costly when

compared to the computation of in-degree. (To find the in-

degree of the ith node, simply find the ith column sum of

the adjacency matrix. The adjacency matrix of the web

56 4. PAGERANK AND IN-DEGREE

graph is sparse.) Litvak et al. [14] proved that this result

is true for power law graphs, but not for arbitrary graphs.

The main goal of the coming sections is to provide exam-

ples of graphs whose PageRank and in-degree distributions

are distinct.

4.2. Binary Trees

A tree is a connected, acyclic digraph; a rooted tree has

a distinguished node called the root. A binary tree is a

rooted tree in which every node other than the leaves have

in-degree equalling 2. For a fixed i e N , the it\\ row of a

binary tree consists of those nodes which are connected to

the root by a directed path of length exactly i — 1. Define

^ ( r ) to be a binary tree with r rows.

There are several interesting properties for binary trees.

For instance, the set of nodes of T2(r) may be identified

with a set of finite 0-1 sequences (or strings), with the root

representing the empty sequence. Figure 4.1 displays such

a binary string labelling.

Our goal is to calculate the PageRank for every node

in the binary tree, and then compare the ranking with the
 4.2. BINARY TREES 57

J0

FIGURE
A
00 01 10 11

4.1. The binary tree T2(3) with its 0-1 labelling,

in-degree of the nodes. As we will see for small examples,

the PageRank and in-degree distributions of binary trees

do not correlate. We conjecture that this holds in general

for all binary trees. For larger examples, while we do not

prove this directly, we offer evidence for this conjecture

by proving that the stationary distribution of the uniform

random walk on the binary tree does not correlate with the

in-degree distribution.

We first need some notation for T2(r). This will help to

quickly recognize on which row each node is located. Let

Xij denote the i-th node on the j - t h row of the binary tree.

The Figure 4.2 shows such a labelling for T2(3).

Although the proof of the following lemma is folklore,

we include it for completeness.

58 4. PAGERANK AND IN-DEGREE

A
1,1

X
2,lf£ ilX2,2

y y y y
3,1 3,2 3,3 3,4

FIGURE 4.2. The binary tree T2(3) with xitj labelling.

LEMMA 4.2.1. Fix an integer r > 1.

(1) For 0 < i < r, the number of nodes on the i-th row

(assuming the root to be the 1st row) of the binary

treeT2{r), is 2i~l.

(2) The binary tree T2{r) has order n — 2r — 1.

We note that all throughout this chapter we assume the

binary tree to be a full binary tree, meaning that all of the

leaves are on the same level and every non-leaf node has

two children.

Proof: For item (1), we perform induction on i. For the

base step of the induction, consider the first row of nodes in

T2(r). As i = 1, hence, 2 l _ 1 = 2° = 1. But there is exactly

one node in the first row.

 4.3. CALCULATING THE STATIONARY DISTRIBUTION 59

The induction hypothesis assumes that on row i, we have

2'™1 nodes. Moving to the row i + 1, every node in row i

has two children (since the binary tree is full) and so the

number of nodes on row i + 1 is twice the number of nodes

on row i:

#(nodes on row i-\-\) = 2 x #(nodes on row i)

= 2 x 2i_1

= 2\

For item (2), the total number of nodes in a full binary tree

T ^ r ) , is counted by adding up the total number of nodes

on each row. Hence:

\T2(r)\ = 2 1 + 2 2 + 23 + . . . + 2 r
r—1

= E2"
i=0
= 2r-l. •

4.3. Calculating the Stationary Distribution

Calculating the exact PageRank vector for the binary

tree would require us to compute all the powers of the dense

60 4. PAGERANK AND IN-DEGREE

Google matrix. As we are presently not able to perform this

calculation, we decided instead to calculate the stationary

distribution of a uniform random walk on the binary tree.

As PageRank is the stationary distribution of the uniform

random walk with teleportation, our results are suggestive

of the actual PageRank values. The proofs in this chapter

are original work.

AA:AA
FIGURE 4.3. An arbitrary binary tree

A binary tree is depicted in Figure 4.3. The adjacency

matrix (namely P i ) for the binary tree has the following

 4.3. CALCULATING THE STATIONARY DISTRIBUTION 61

structure.
/ 0 0 0 0 ^

1 0 0 ...

1 0 0 ...
Pi =
0 1 0 ...

0 1 0 ...

The matrix P i contains a nice pattern: starting from the

second row, every two consecutive rows are equal. The next

pair of rows results by a single shifting of the previous pair

of rows to the right. For X^r), this pattern continues until

the l's reach the (2 r _ 1 )-th column of the matrix. Since the

leaves of the tree have zero in-degree, the matrix will have

a rectangular block of zeros of size (2 r — 1) x 2 r _ 1 on its

right side. The relative simplicity of this pattern allows a

rigorous analysis of the uniform random walk on T2(n).

Consider the P 2 matrix for T2(r). Recall that the P 2

matrix is just P i without zero rows. In the binary tree, the

root or the node x\^_ is the unique dangling node, where the

random surfer would become stuck in the uniform random

walk. Hence, we will assume that the root is pointing to

62 4. PAGERANK AND IN-DEGREE

all other nodes in the graph; this assumption turns P 2 into

a stochastic matrix. The P2 matrix for the binary tree is

therefore,

' 1/n 1/n .. . 1/n »

1 0 0 ...

1 0 0 ...
M = P2 =
0 1 0 ...

0 1 0 ...

V ; • ; • • /

where n = \V(T2(r))\. Throughout, let

M
6 — J n . l ~~

w
We now state the main results of this section

THEOREM 4.3.1. Let H be the P 2 matrix of the binary

tree X ^ r ) . Fix k a positive integer. Define [Hfc] • to be the

sum of the column corresponding to the node xp^. For all

 4.3. CALCULATING THE STATIONARY DISTRIBUTION 63

k > 1 and 1 < p < r,

where 1 < i < j < 2P~1. In particular, the column sums of

any two nodes on the same row in T2{r) are equal.

THEOREM 4.3.2. Let H be the P2 matrix of the binary

tree T2(r). For all k > 1, 1 > i > j > 2v~l and 1 < p <

r — 1, we have that

[H }Pjl = 2[H }p+ij + - [ H ]i ; i;

ft/

where n = 2r — 1.

Note that -[H]i,i does not depend on either p or j . We

defer the proofs of Theorem 4.3.1 and 4.3.2 until the fol-

lowing section. We have however, the following corollary.

COROLLARY 4.3.3. Let s be the stationary distribution

vector of the P2 matrix of the binary tree T2(r).

(1) For any two nodes on the same row of T2(r), the
corresponding entries in s are equal.

(2) For all 1 < p < r — 1, the entry of s corresponding

to any node on the p-th row is approximately twice

64 4. PAGERANK AND IN-DEGREE

the entry corresponding to any of the nodes on the

(p + l)-st row of s.

Proof: For the proof of (1), by definition we have that

(4.3) s T = lim e r H t .
t—»oo

Now apply [•]-,• to both sides of (4.3), representing the sum

of the j - t h column (or as in this case, the j-th element of

the vector) on both sides of the limit:

[sr],- = [lime^H'b
t—>oo

= nm[eTHt]i
t—*oo

= l-limlH*],-.

Note that [s]j = [e^H*]j represents the j-th. element of the

vector. By Theorem 4.3.1, since H is the P2 matrix of the

binary tree T2(r), for all t and for 1 < i, j < 2 P ~ 1 ,

[H]p,i = [H ]p,j.

But the column sum [H*]P)i represents the stationary value

for the node xPti: namely [S]2P-I+J_I; similarly [H*]pj is the

stationary value for the node xpj, namely [ S ^ P - I + J - I - Hence,

 4.3. CALCULATING THE STATIONARY DISTRIBUTION 65

the corresponding (2P~1 + i — l)-th and (2P~1 + j — l)-th

entries in the stationary vector are the same. As i and j

were arbitrary, any two nodes on the same row have equal

entries in s. The proof of item (1) follows.

For the proof of item (2), since s is the stationary dis-

tribution of the P 2 matrix of the binary tree, we can apply

Theorem 4.3.2 to H. As in the proof of the previous corol-

lary item, we use the definition of s, given in (4.3) for the

entries i and j corresponding to the nodes lying on two

consecutive rows p and p + 1:

[s T ], = [limeTHt]i
t—>oo

= lim^H*],

Using part (1) of the corollary, by Theorem 4.3.2, and the

fact that the jth node is located on the next row right after,

we can write for any fixed t > 0 and for 1 < i, j < 2P~1,

[H ]Pj = 2 x [H ]p+ij -\—[H l^i.

66 4. PAGERANK AND IN-DEGREE

But the column sum [H*]P)j represents the stationary value

for the node x p j , and similarly [H L ] p + ij is the stationary

value for the node xp+i,j: any node on the (p + l)-th row.

Hence, the corresponding (2 P _ 1 + i — l)-th entry is approx-

imately twice the (2P"X + j — l)-th entry in the stationary

vector. As i and j were arbitrarily chosen, any two nodes

on the two consecutive rows will have this property in s. •

By Corollary 4.3.3, we see that the stationary distri-

bution decreases from the largest value at the root to the

smallest value at the leaves. This is analogous to a power

law, since the leaves are the most abundant nodes. How-

ever, the in-degree distribution has either values 0 or 2. In

particular, the stationary distribution for the random walk

and the in-degree distribution are quite different. We con-

jecture that an analogous difference occurs when comparing

the PageRank distribution with the in-degree distribution.

Before moving on to the next section, let us make this

conjecture more plausible by comparing the PageRank and

in-degree for T 2 (4). The graph of T2(4) is given in Fig-

ure 4.4.
 4.3. CALCULATING THE STATIONARY DISTRIBUTION 67

FIGURE 4.4. The directed binary tree T2(4).

After calculating the adjacency matrix P i , we change

the row one zero values to 1/15 to recover the dangling

node x\ \. The result of this step is the P2 matrix given by

/ 1/15 1/15 1/15 . . . \

1 0 0
1 0 0
P2 =
0 1 0
0 1 0

V J
Assuming the teleportation factor a to be equal to 0.85

and using the power method with t = 20 iterations, we

68 4. PAGERANK AND IN-DEGREE

calculate the PageRank vector to be approximately

(0.28 0.14 0.14 0.06 0.06 0.06 0.06 0.02

0.02 0.02 0.02 0.02 0.02 0.02 0.02)

Therefore, the nodes in T 2 (r) can be ranked by PageRank

as

( 1 2 2 3 3 3 3 4 4 4 4 4 4 4 4),

which implies that the root is the highest ranked page (as

expected). Moreover, all nodes on the same row, have equal

rank.

As we can see in Figure 4.4, the in-degree of all the nodes

of 12(4) is 2, except for the leaves which have 0 in-degree.

The in-degree vector, written I D , has its first seven entries

equal to 2 (corresponding to the non-leaf nodes) and the

next eight elements equal to 0 (corresponding to the leaves

 4.3. CALCULATING THE STATIONARY DISTRIBUTION 69

of the tree):
/ 2 \

2
ID =
0

W
Thus, the in-degree ranks the pages as

( 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2),

implying that the first seven pages in the graph (that is,

all the non-leaf nodes) have equal ranking. All the leaves

come in second position.

Q. 0.4
CO
i_ • * • • PageRank
O)
Q) indegree
= 0.3
a>
2
D)
<D

•g 0.2
x x
(A
i 0.1 X X X X
to
OH
0>
en
_fc X X X X X X X

CL 0 5 10 15
Nodes in Binary Tree

FIGURE 4.5. PageRank versus in-degree for T 2 (4).

70 4. PAGERANK AND IN-DEGREE

In Figure 4.6, the values for PageRank and in-degree

are plotted in one graph, so we can see the difference be-

tween their rankings. As it is evident from the figure, the

PageRank and in-degree in this example do not correlate:

PageRank follows a rough power law, while the in-degree

is a step function.

1 ' 1 1 1 1 1 '

m PageRank
-

•1 0.15 0.2 0.35

PageRank value

j in—degree

0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
in-degree value

FIGURE 4.6. PageRank versus in-degree distributions

for the binary tree T2(4).

This is more clearly sketched through the plot of the

PageRank distribution of the nodes versus the in-degree

distribution of the nodes, given in separate histograms in

Figure 4.6.
 4.4. PROOFS OF MAIN RESULTS 71

4.4. Proofs of Main Results

Consider the nodes on two consecutive rows of T2(r). Let

us say the first row starts with the node 2^1 and the second

row starts with the node 2^+1,1. Note that we already know

there are 2 l _ 1 nodes on the i-th row and 2l nodes on the

(i + l)-st row. It is essential to know which column in the

adjacency matrix the node Xij will be presented by. Always

counting from top to bottom and left to right, the node x^j

is the j'-th node on the i-th row. So, counting the nodes,

we have total of 2*_1 — 1 nodes before the i-th row begins;

adding it up with the j nodes until we reach x-ij, we will

have that the node xi}j is the {{2l~l — 1) + j)-th node in

T2(r). Hence, we now know that the in-degree, and the

stationary distribution of the node x,-hj will be presented

through the (2* _1 + j — l)-th column of the corresponding

adjacency matrices.

Before stating the proof for the theorems, we need lem-

mas from linear algebra:

72 4. PAGERANK AND IN-DEGREE

A
Xri>(1<i<2M)

Xr+1i,(1o<2r)

F I G U R E 4.7. The general form of a binary tree with the

labeled nodes. Here xr^ denotes the i-th node on the r-th
row.

L E M M A 4.4.1. For every matrix A, and a vector

fai\
OC2
a =

\ On J
 4.4. PROOFS OF MAIN RESULTS 73

we have that

1
a i \ ' ^

a2 OL2
A = [AilAal-.-IA*]

\ ®n J \®n J
= aiAi +a2A2 +... +anAn,

where A$ stands for the i-th column of A.

LEMMA 4.4.2. For all matrices A and B (with appropri-

ate sizes for matrix multiplication), the multiplication can

also be evaluated as follows:

A B = A [ B i | B 2 | . . . |B„] = [AB!|AB 2 |...|AB n ].

P r o o f of T h e o r e m 4.3.1: We shall proceed by induc-

tion on the parameter k, which is the power of the matrix

H. For k — 1, the theorem holds since in H 1 , all columns

have the same sum (equal to 2 + - ) , except for the leaves of

the binary tree which are total of 2 r ~ 1 nodes with column

sum equal to 1/n.

74 4. PAGERANK AND IN-DEGREE

We now assume that the theorem holds for k = i and

move forward to prove it for k = i + 1 . Using Lemmas 4.4.1
and 4.4.2, we have that

[H ]rj = [H • H]rti

= [H fc ] r+1)2l _ 1 + [H']r+i>2i + l/n[H*]i

Similarly,

fH*+l [H -H] r j
r,3

[H'U^-x + lH'Ui^ + l/nlH*]!

x A
r+1,2i-1 r+1,2j-1

F I G U R E 4.8. The location of some of the nodes used in

the proof of Theorem 4.3.1.
 4.4. PROOFS OF MAIN RESULTS 75

However, if the nodes xr^ and xT)j are assumed to be on

the same row (which is the case, since 1 < i < j < 2r'~1

and the total number of nodes on the r-th row is 2 r _ 1 ) , the

nodes xr+\^, a^+i^+i, xr+ij and xr+itj+i are on the same

row. See Figure 4.8.

Now, using the induction hypothesis, we have that:

[H ]r+l,2i-l = [H ]r+l,2i = [H ]r+l,j = [H jr+ij+i.

Hence,

[H + i ], r ) i = [ H + i ] r j - .

The final step of the induction is carried out and hence, for

all k > 1,

[H*]r>< = [H fc ] rj . D

One interesting point to consider is that, not only are

the above sums equal, but also the value for each element

on row r, is approximately twice the value for the elements

on row r + 1. To verify this, we prove Theorem 4.3.2.

76 4. PAGERANK AND IN-DEGREE

Proof of Theorem 4.3.2: Using the results of Theo-

rem 4.3.1 we have that

(4.4)[H fc+1 ] p , = pHfc]p+i,2i-i + [ H f c U i i a + l / n [ i r ] i 1 i .

But since both nodes xp+\^i~i and xp+i^i are located on

the same row, they have equal column sums:

[H ] p +i,2i-l = [H ]p+i,2i.

Hence,

(4.5) [H* +1 ] Pii = 2 x [ H V i , 2 i + l M H i ] i , i .

By Theorem 4.3.1 applied to each row, all the nodes

will have the same column sum value. By considering the

equations (4.4) and (4.5), and not considering the effect

of the root (which is carried as a constant all along the

equations) it is seen that the column sum for the nodes on

the pth row is approximately twice the column sum of each

of the nodes on the (p + l)-th row. •

 4.6. PAGERANK OF RANDOM AND POWER LAW GRAPHS 77
4.5. Conclusions for Binary Trees

In Sections 4.3, we calculated the stationary distribu-

tion of the uniform random walk on the binary tree. This

vector, while not equal to the PageRank vector, is closely

related to it. We proved in Theorems 4.3.1 and 4.3.2 that

the values of the stationary distribution for the nodes on

each row is the same and it reduces to approximately half

for every row we move away from the root. This behaviour

is suggestive of a power law degree distribution. On the

other hand, all nodes of the binary tree have in-degree 2,

except the leaves which have in-degree 0. The binary tree

is, therefore, an example which shows that in-degree and

the stationary distribution of the uniform random walk are

not correlated. We conjecture that such a difference ex-

ists between in-degree and the PageRank distributions for

binary trees.

4.6. PageRank of R a n d o m and Power Law Graphs

We provide some experimental results from simulations

of both random digraphs and power law graphs. The results

here corroborate the theoretical ones correlating PageRank

78 4. PAGERANK AND IN-DEGREE

and in-degree in power law graphs described in previous

sections of this chapter.

We first consider random digraphs. A random digraph

has n nodes, and the probability of an edge between two

distinct nodes occurs independently with probability 1/2.

See Figure 4.9 for a randomly sampled digraph with 100

nodes. It may be proven that as the number of nodes n

tends to infinity, we have a binomial distribution for in-

degree. This follows from the fact that the degree of a node

is asymptotically concentrated on n / 2 . (See Theorem 3.11

in [2]-)

30

20

10

0
0.C 107 0.008 0.009 0.01 0.011 0.012 0.01:

30
20
10
0
) 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

FIGURE 4.9. PageRank versus in-degree in a random

digraph with 100 nodes.
 4.6. PAGERANK OF RANDOM AND POWER LAW GRAPHS 79

However, the actual web graph has a power law, not

binomial, degree distribution. We therefore include the

PageRank distribution of a power law (undirected) digraph

(produced using the freely available software Pajek). In the

histogram in Figure 4.10, the distributions of PageRank

and in-degree for a power law digraph with 1,200 nodes

is plotted. As is evident, both distributions follow similar

power laws.

300 ~i r -l r

200

100

0 J I L k_> I I- . I 1_

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

FIGURE 4.10. PageRank versus in-degree in a power law

digraph with 1,200 nodes.
 CHAPTER 5

Conclusion and Future Work

We surveyed the mathematics of PageRank, which is a

link-based ranking algorithm measuring the popularity of

nodes in a digraph. The study of PageRank presented in

this thesis combines graph theory, Markov chains, stochas-

tic calculus, and statistics. In Chapter 2, we defined PageR-

ank and summarized its key properties. We implemented

a PageRank calculator in Matlab. With this calculator,

we experimented with different graphs and compared their

PageRank and in-degree distributions. In Chapter 3, we

studied the recent work of Litvak et al. [14]. They proved

(under certain assumptions) that in power law graphs, the

PageRank and in-degree distributions follow power law dis-

tributions with the same exponent. In Chapter 4, we con-

sidered binary trees as a counterexample to the assertion

that PageRank and in-degree possess similar distributions.

The analysis of the PageRank of the class of binary trees is

81
82 5. CONCLUSION AND FUTURE WORK

significant, since it demonstrates that in general we cannot

correlate PageRank with in-degree.

Several problems remain open related to the work de-

scribed in this thesis. We list two few such problems here,

which we will consider in the future.

(1) Derive the PageRank distribution of binary trees for

all orders. We conjecture that such a distribution

follows a power law, with PageRank decreasing as

we move further from the root node. All nodes

on the same row should have the same PageRank

value. More generally, we would like to compute

the PageRank of m-ary trees, where m > 2 (in these

digraphs, all nodes except the leaves have constant

in-degree equalling m).

(2) Livak et al. [14] made certain unrealistic assump-

tions in order to rigorously analyze PageRank using

stochastic equations. For example, they assumed

that all nodes have constant out-degree. They also

assumed that the PageRank of nodes pointing to a

similar page are independently distributed (in fact,

 5. CONCLUSION AND FUTURE WORK 83

web pages that point to a similar page have corre-

lated PageRank distributions). Can their analysis

be generalized if these assumptions are removed?

 Appendix

We include the original code used in the computational

results in this thesis.

function M = MatRead (pjklnput)

f i d = fopen(pjklnput);
A = fgetl(fid);
v = sscanf(A, "/.*s 0 / 0 d');
for i = l:v+l
A = fgetl(fid);
end;
for i = l:v
i;
L = fgetl(fid);
M(i,:) = str2num(L);

end;

fcloseCfid);

85
86 APPENDIX

function ID =InDegree (X,n)

"/Calculates the indegree or column sum

°/0of a random H-matrix of size n.

ID = sum(X);

% Now to normalize ID:

ID = (ID - mean(ID))/max(abs((ID-mean(ID))));

ID = (ID + l)/2;

save ID;

end

function Inl = Initial (n)

"/Generates the initial column matrix

%for the PageRank algorithm for size n.

Inl = zeros(l,n);

for i=l:n
Inl(i)=l/n;
end%for
 APPENDIX

%return Inl;

save Inl;

function J = RandomSample(n);

"/oGives a random H-matrix of s

X = randint(n,n);

for j=l:n

for i=l:n

if i==j

X(i,j)=0;

end°/0if

end%for

end°/0f o r

y = sum(X,2);

for j=l:n

for i=l:n

J(i,j)=X(i,j)/y(i);

end%for

end°/0f o r

save J;
88 APPENDIX

°/0 Plots the histogram for both InDegree and

% PageRank for T_2(4)

hold on;

X=[0; 1; 2; 3; 4; 5; 6; 7; 8; 9; 10; 11; 12; 13; 14];

PR_15_1 = [0.2755; 0.1402; 0.1402; 0.0648; 0.0648;

0.0648; 0.0648; 0.0231; 0.0231; 0.0231; 0.0231; 0.0231;

0.0231; 0.0231; 0.0231];

ID_15_1 = [2; 2; 2; 2; 2; 2; 2; 0; 0; 0; 0; 0; 0; 0; 0];

ID_15_2 = (1/49) * ID_15_1;

subplot(2,1,1); hist(PR_15_l);

subplot(2,1,2); hist(ID_15_2);
 Bibliography

[1] L. Bechetti,C. Castillo, The distribution of PageRank

follows a power law only for particular values of the

damping factor, In Proceedings of the 15th International

Conference on World Wide Web, 2006.

[2] A. Bonato, A Course on the Web Graph, Graduate

Studies in Mathematics 69, American Mathematical So-

ciety, Providence, Rhode Island, 2008.

[3] S. Brin, L. Page, Anatomy of a large-scale hypertextual

web search engine, In: Proceedings of the 7th Interna-

tional World Wide Web Conference, 1998.

[4] A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S.

Rajagopalan, R. Stata, A. Tomkins, J. Wiener, Graph

structure in the web, Computer Networks 33 (2000)

309-320.

[5] F.R.K. Chung, L. Lu, Complex Graphs and Networks,

American Mathematical Society, U.S.A., 2006.

89
90 BIBLIOGRAPHY

[6] R. Diestel, Graph Theory, Springer-Verlag, New York,

2005.

[7] D. Donato, L. Laura, S. Leonardi, S. Millozi, Large scale

properties of the web graph, Eurapean Physical Journal

(2004) 38 239-243.

[8] R. Durrett, Random, Graph Dynamics, Cambridge Uni-

versity Press, New York, 2006.

[9] S. Fortunato, A. Flammini, F. Menczer, A. Vespignani,

The egalitarian effect of search engines, Technical Re-

port 0604203, arXiv/Physics, 2006.

[10] M.L. Goldstein, S.A. Morris, G.G. Yen, Problems with

fitting to the power-law distribution, Eurapean Physical

Journal 4 1 (2004) 255-258.

[11] G.R. Grimmett, D.R. Stirzaker, Probability and Ran-

dom Processes, Oxford University Press, Oxford, 2001.

[12] O. Hernandez-Lerma, J.B. Lassere, Markov chains and

Invariant Probabilities, Birkhauser, Berlin, 2000.

[13] A.N. Langville, C.D. Meyer, Google's PageRank and

Beyond: The Science of Search Engine Rankings,

Princeton University Press, Princeton, 2006.

 BIBLIOGRAPHY 91

[14] N. Litvak, Y. Volkovich, W. Scheinhardt, In-degree

and PageRank of web pages: why do they follow sim-

ilar power laws?, In: Proceedings of the 4th Workshop

on Algorithms and Models for the Web-Graph, 2006.

[15] C D . Meyer, Matrix Analysis and Applied Linear Alge-

bra, SIAM, Philadelphia, 2000.

[16] M. Mitzenmacher, E. Upfal, Probability and Comput-

ing, Cambridge University Press, New York, 2005.

[17] C. Moler, The world's largest matrix computation,

Matlab News and Notes, October 2002, pp. 12-13.

[18] M.E.J. Newman, Power laws, Pareto distributions and

Zipf's law, Contemporary Physics 46 (2005) 323-351.

[19] E. Seneta, Non-negative Matrices and Markov chains,

Springer, New York, 2006.