國 立 交 通 大 學
應用數學系
碩
士
論
文
搜尋散佈謠言者的數學模型
A Mathematical Model for Finding the Culprit Who
Spreads Rumors
研 究 生:李姿慧
指導老師:傅恆霖 教授
搜尋散佈謠言者的數
學模型
A Mathematical Model for Finding the Culprit Who
Spreads Rumors
研究生
:李姿慧
Student: Zi-Hui Lee
指導教授:傅恆霖 教授 Advisor: Hung-Lin Fu
國 立 交 通 大 學
應 用
數
學 系
碩 士 論 文
A Thesis
Submitted to Department of Applied Mathematics
College of Science
National Chiao Tung University
in Partial Fulfillment of the Requirements
for the Degree of
Master
in
Applied Mathematics
June 2012
Hsinchu, Taiwan, Republic of China
中 華 民 國 一
○一 年 六 月
搜
搜
搜尋
尋
尋散
散
散佈
佈
佈謠
謠
謠言
言
言者
者
者的
的
的數
數
數
學
學
學模
模
模型
型
型
研究生
:李姿慧
指導教授:傅恆霖 教授
國 立 交 通 大 學
應 用
數
學 系
摘要
在這篇論文中,我們介紹謠言傳播模型,它的設計是根據一個在流
行
病學領域著名的易感–感染模型。我們描述在一個圖上散佈謠言
的源頭的最大概似估計值並計算預測到散佈謠言的源頭的機率。我
們發現:對於路徑的圖形,機率會隨著時間增加趨近到0,其關係
為t
−1/2;對於正則樹,機率有一個明確的範圍。當d = 3,其機率
值會隨著時間增加趨近到1/4,此結果已利用隨機圖模型得到 [6]。
中 華 民 國 一
○一 年 六 月
A Mathematical Model for Finding the
Culprit Who Spreads Rumors
Student: Zi-Hui Lee
Advisor: Hung-Lin Fu
Department of Applied Mathematics
National Chiao Tung University
June, 2012
Abstract
In this thesis, we introduce a rumor spreading model based
on the common susceptible-infected (SI) model which is a well
known epidemiological model. We describe the maximum
likeli-hood estimators of graphs and we evaluate the detection
proba-bilities of finding the rumor source in d-regular trees. We observe
that: For paths, the detection probability of finding the rumor
source scales as t
−1/2, which approaches 0 as t approaches
infin-ity. For regular trees, we find an explicit bound of the detection
probabilities of finding the source in d-regular trees. As a
conse-quence, for d = 3, the detection probability approaches 1/4, this
result has been obtained earlier by using a random graph model
[6].
誌
誌
誌謝
謝
謝
在交大的兩年,謝謝你們。
謝謝我的指導教授傅恆霖老師,謝謝老師讓我成為他的學生,
跟著老師的
一年半我學到了許多,尤其是態度,老師總是期許
我們要多讀點書,多運動等等,所有的一切我會謹記在心。謝
謝與我同師門的學長姐:惠蘭學姐、貓頭學長、明輝學長、惠娟
學姐、施智懷學長、敏筠學姐、軒軒學長、小泡學姐還有瑩晏學
姐。謝謝你們在這兩年給我的支持、建議與鼓勵。因為過程中有
你們的幫助我才能夠還算順利的完成我的碩士論文。
謝謝系
上的老師們指導我完成學業,讓我順利的走到這。特別
感謝的還有
:謝謝資訊技術服務中心的蔡錫鈞主任提供我解決問
題的方法與建議;謝謝系上的資訊人員詹宗智還有麻將學長教我
使用系上的機器做運算;謝謝游森朋教授抽空看我的問題;謝謝
陳秋媛老師的提點,讓我的論文內容更加嚴謹。
謝謝我的好朋友
們:米奇、小雨、憶妏、沙發、惠雯、吉米、
連
、DK、嗨派,有事情就幫助我,有玩的就找我。還要謝謝所
有我在這裡認識的人,因為有你們我的碩士生活才能如此多采多
姿。
最後一定要謝謝我的父母,因為有他們的期許我才一路走到。
在外地念書不能常常回家,謝謝家人的體諒以及時不時打電話來
關心問候,我愛我的爸爸媽媽。
Contents
Abstract (in Chinese) i
Abstract (in English) ii
Acknowledgement iii
Contents iv
List of Figures v
1 Introduction 1
1.1 Preliminaries . . . 2
1.2 Rumor spreading (RS) model . . . 4
2 Rumor Source Estimator 5
2.1 The ML estimator of Gn in a regular tree . . . 5
2.2 Rumor centrality . . . 8
3 Main Result 11
3.1 Detection probabilities of d-regular trees, d ≤ 3 . . . 11
3.2 Detection probabilities of d-regular trees, d ≥ 4 . . . 17
List of Figures
1 Illustration of subtree Tuv . . . 3
2 Network with 4 infected vertices. . . 6
3 Network of calculating rumor centrality. . . 9
1
Introduction
Social network, internet and electrical power grid network are the common
networks everywhere in our life. We are surrounded by all kinds of networks
and are very easily subjected to the influence of network risks. Although
every network has its different structure, the common phenomenon is: an
isolated risk is enlarged because it can be spread by the network.
For example, in an electrical power grid network, an isolated failure can
lead to a rolling blackout. Computer viruses utilize the internet to infect
millions of computers everyday. The malicious rumors or misinformation
can be spread in the social networks quickly and the person concerned will
be deeply hurt and offended. In all of these situations, power network
operator, internet service provider or victim of a malicious rumor would
like to infer the source of risks as quickly as possible and then blockade the
spread of risks. All of these situations can be modeled as rumors spreading
through networks, where the goal is to find the source in order to control
and prevent these network risks based on limited information about the
network structure and the rumor infected vertices.
Prior work on rumor spreading has primarily focused on infectious
dis-eases in populations. The standard model of infectious disdis-eases is known
as the susceptible-infected-recovered (SIR) model [1]. In this model, there
are three types of vertices: (i) susceptible vertices that are capable of being
infected; (ii) infected vertices that can spread the virus further; and (iii)
recovered vertices that are cured and can’t be infected anymore. Research
infec-tion/cure [2, 3, 4, 5]. However, there has no idea of identifying the source
of an epidemic. Now, a mathematical model has been developed to identify
the rumor source in a network based on rumor infected vertices [6]. But,
not much is known if the network is getting more complicate. In this thesis,
we set forth to study the networks defined on d-regular trees.
1.1
Preliminaries
A graph G is a triple consisting of a vertex set V (G), and edge set
E(G), and a relation that associates with each edge two vertices called its
endpoints. The order of a graph G is the number of vertices in G and the
size of a graph G is the number of edges in G. A loop is an edge whose
endpoints are equal, and multiple edges are edges having the same pair
of endpoints. A simple graph G is a graph having no loops or multiple
edges. We consider a simple graph with a countably infinite vertex set.
A subgraph of graph G is a graph H such that V (H) ⊆ V (G) and
E(H) ⊆ E(G) and the assignment of endpoints to edges in H is the same
as in G. We then write H ⊆ G and say that “G contains H”. A path is a
simple graph whose vertices can be ordered so that two vertices are adjacent
if and only if they are consecutive in the list. A graph G is connected if
each pair of vertices in G belongs to a path.
In a graph G, the contraction of edge e with endpoints u, v is the
replacement of u and v with a single vertex whose incident edges are the
edges other than e that were incident to u or v. The resulting graph G · e
has one less edge e than G, and the new vertex {u, v}(or {v, u}).
Figure 1: Illustration of subtree Tv u
e = vu), they are adjacent and are neighbors. We write u ↔ v for ”u
is adjacent to v”. The neighborhood of v, written NG(v) or N (v), is the
set of vertices adjacent to v. The degree of vertex v in a graph G, written
dG(v) or d(v), is the number of edges incident to v. And G is regular if the
degrees of all vertices are the same. It is k-regular if the common degree
is k.
A graph with no cycles is acyclic. A tree is a connected acyclic graph. In
this thesis, A d-regular tree is a tree where every vertex has d neighbors.
Therefore, it is an infinite graph. A rooted tree Tr is a tree with one
vertex r chosen as root. For each vertex v, Let p(v) be the unique path
from r to v. The parent of v is its neighbor on p(v); its children are its
other neighbors. The set child(v) is the set of all children of v.
A branch of a tree is a subtree Tr
v of Tr induced by v and all its
de-scendants. In addition, tr
v denotes the number of vertices in Tvr. A d-regular
trees are branches of Tv and Pu∈N (v)tvu = tvv − 1. A simple example T is
shown in Figure 1. T1
7 is a branch of T1; T76 is a branch of T6. Clearly,
T1
7 = T76. And T can be decomposed into 6 subtrees (T21, T31, T41, T51, T61),
such that these trees are branches of T1.
These definitions about graph theory are cited from the book
”Intro-duction to Graph Theory” written by Douglas B. West. If this part is not
sufficient, please refer to the reference [8] for details.
1.2
Rumor spreading (RS) model
We consider a discrete time susceptible-infected (SI) model. Let S(t) be a
set of people who don’t know rumors yet at time t and let I(t) be a set of
people who have known rumors at time t in which the number of people of
S(t) and I(t) are denoted St and It, respectively. Using a fixed population,
St+ It= N . For convenience of study, we shall assume
St+1 = St− 1, It+1 = It+ 1, S0 = N, I0 = 0.
Let each person be a vertex and the relationship between two people be an
edge, and these form the vertex set and edge set of a graph G. Let Gt be a
subgraph of order t of G. This graph is compose of t infected vertices which
are people who have known rumors at time t. The graph G1 is a vertex
which is called rumor source and in each discrete time-step t + 1, t > 0,
Gt+1 develops from Gt by adding a vertex z with an edge. We assume that
every vertex is chosen with the following probability distribution:
Pt+1(z) =
1 P
v∈V (Gt)d(v) − 2(t − 1)
2
Rumor Source Estimator
Consider a network G and a subgraph Gn of G that is a graph with n
infected vertices. The Maximum Likelihood(ML) estimator is the vertex v
which has the maximum of P (Gn|v). where P (Gn|v) is the probability of
observing Gn under the RS model if v is the source.
2.1
The ML estimator of G
nin a regular tree
Continue from the paragraph above, if we want to know that what’s the
ML estimator of Gn, we would like to evaluate P (Gn|v) for all v ∈ Gn
and then choose v so that P (Gn|v) is maximum. In general, evaluating
every vertex’s P (Gn|v) is complicated. Let us consider a simple example
as shown in Figure 2 with n = 4. First, we suppose that the vertex 1 is the
source, and we would like to calculate P (G4|1). Then there are six ways
or vertex infection orders that a rumor can be spread to every vertex in
G4 with vertex 1 as the source: (1, 2, 3, 4), (1, 2, 4, 3), (1, 3, 2, 4), (1, 3, 4, 2),
(1, 4, 2, 3) and (1, 4, 3, 2). However, if we suppose that the vertex 2 is the
source, infection order (2, 3, 1, 4) is not possible, since 2 = 3 (2 and 3 are not adjacent in G4). Therefore, in general to evaluate P (Gn|v), we need
to find all such possible n-permutations of V (Gn) and their corresponding
probabilities.
Let S (v, Gn) be the set of all possible n-permutations of V (Gn) starting
with vertex v where v ∈ V (Gn), and σ = (v1 = v, v2, · · · , vn) for each σ ∈
S (v, Gn). The probability P (σ|v) is the probability of observing Gn under
the RS model if v is the source in the infected order σ. Let Gk(σ) be the set
Figure 2: Network with 4 infected vertices.
for 1 ≤ k ≤ n. Then we have
P (σ|v) = n−1 Y k=1 1 P vi∈Gk(σ)d(vi) − 2(k − 1) . (2)
For d-regular trees,
P (σ|v) = n−1 Y k=1 1 dk − 2(k − 1). (3)
From equation (3), we can see that every permutation σ has the same
probability and is independent of the source. Specifically, for any source v
and permutation σ, P (σ|v) is a constant. Thus
P (Gn|v) = X σ∈S(v,Gn) P (σ|v) = |S (v, Gn) | · n−1 Y k=1 1 dk − 2(k − 1) ∝ |S (v, Gn) |.
It follows that P (Gn|v) is proportional to |S (v, Gn) |. Let R(v, Gn) be the
number of distinct ways to spread a rumor to every vertex in Gn with v as
the source. Clearly, |S (v, Gn) | = R(v, Gn). Then
In summary, the ML estimator of Gn in a regular tree can be obtained by
finding the maximum of R (v, Gn) for all v. For the above example as shown
in Figure 2 the estimation of vertices are
P (G4|1) = X σ∈S(1,G4) P (σ|1) = P ((1, 2, 3, 4)|1) + P ((1, 2, 4, 3)|1) + P ((1, 3, 2, 4)|1) + P ((1, 3, 4, 2)|1) + P ((1, 4, 2, 3)|1) + P ((1, 4, 3, 2)|1) = 6 · 1 3· 1 4 · 1 5 = 1 10. P (G4|2) = X σ∈S(2,G4) P (σ|2) = P ((2, 1, 3, 4)|2) + P ((2, 1, 4, 3)|2) = 2 · 1 3· 1 4 · 1 5 = 1 30. P (G4|3) = X σ∈S(3,G4) P (σ|3) = P ((3, 1, 2, 4)|3) + P ((3, 1, 4, 2)|3) = 2 · 1 3· 1 4 · 1 5 = 1 30. P (G4|4) = X σ∈S(4,G4) P (σ|4) = P ((4, 1, 2, 3)|4) + P ((4, 1, 3, 2)|4) = 2 · 1 3· 1 4 · 1 5 = 1 30.
Clearly, we can now evaluate R (v, G4) for all v. R (1, G4) = 6, R (2, G4) =
R (3, G4) = R (4, G4) = 2. Thus vertex 1 is the probable rumor source in
2.2
Rumor centrality
In this paragraph, we shall explain how to calculate the rumor centrality
for each vertex in a given graph Gn. Moreover, the rumor center of Gn is a
vertex with the maximum of rumor centrality.
Let Gn be a rooted tree with root v and assume v has a rumor. The
next infected vertex must be one of the children of v. For all u ∈ child(v),
there are R(u, Tv
u) ways to spread a rumor in the branch Tuv with u as the
source. Thus R(v, Gn) = (n − 1)! Y u∈child(v) R(u, Tuv) tv u! . (5)
To understand the above expression, the number of ways to permute n − 1
steps from different subtrees is a permutation of the multiset {tv v1 · 1, t
v v2 · 2, · · · tv
vd · d}. If we continue this recursion (5) until we reach the leaves of the tree, we obtain
R(v, Gn) = (n − 1)! Y u∈child(v) R(u, Tuv) tv u! = (n − 1)! Y u∈child(v) (tv u− 1)! tv u! · Y w∈child(u) R(w, Tv w) tv w! = (n − 1)! Y u∈child(v) 1 tv u · Y w∈child(u) R(w, Tv w) tv w! = (n − 1)! Y u∈Gn−v 1 tv u .
Since a leaf vertex l has one vertex so R(l, Tv
l ) = 1. By the fact that tvv = n,
then R(v, Gn) = n! Y u∈Gn 1 tv u . (6)
Figure 3: Network of calculating rumor centrality.
Let us consider an example as shown in Figure 3.
R(1, G5) =
5!
5 · 3 · 1 · 1 · 1 = 8.
Indeed, there are 8 possible n-permutations of the network in Figure 4 with
vertex 1 as the source. They are
(1, 2, 4, 5, 3), (1, 2, 4, 3, 5), (1, 2, 3, 4, 5), (1, 3, 2, 4, 5)
(1, 2, 5, 4, 3), (1, 2, 5, 3, 4), (1, 2, 3, 5, 4), (1, 3, 2, 5, 4)
In order to find the rumor center of a given graph Gn, we have to find
the rumor centrality of every vertex in Gn respectively. In fact, the rumor
centrality of v can be deduced from rumor centrality of its neighbors. To
this end, consider two adjacent vertices u and v in Gn, we have
tvu = n − tuv. (7)
And tvw = tuw for each w ∈ Gn− {u, v}. Thus,
R(u, Gn) R(v, Gn) = t v u n − tv u . (8)
For example in Figure 4, we have
R(2, G5) R(1, G5) = 12 8 = t1 2 5 − t1 2 = 3 2.
It shows that for each vertex u, we can calculate its rumor centrality by
using its neighbor’s rumor centrality and tv u.
The following is an important property of the rumor center:
Theorem 2.2.1. [6] Given an n vertices tree, vertex v is the rumor center
if and only if
tvu ≤ n 2
for all u 6= v. Furthermore, a tree can have at most 2 rumor centers.
3
Main Result
In this section, we shall explain the behavior of the detection probabilities
of finding the rumor source in different graphs. Let Et(G) be the event of
correct rumor source detection under the ML rumor source estimator after
time t on a graph G. If the graph G considered is prescribed, then we use Et
to denote Et(G). Then P(Et) is the correct detection probability of finding
the rumor source in a given graph G.
3.1
Detection probabilities of d-regular trees, d ≤ 3
In this paragraph, we shall present shorter proofs by using combinatorial
methods than the original proofs in reference [6]. We first consider the
detection probability of finding the source in a 2-regular tree. By (4), rumor
center is the ML estimator of Gn where Gn is a subtree of a d-regular tree.
Theorem 3.1.1. Suppose a rumor has spread in a 2-regular tree. Then we
have that P(Et) = O 1 √ t . Proof:
Consider a path which compose of 2n (respectively 2n + 1) infected vertices.
By Theorem (2.2.1) and (6), v is a unique rumor center in G2n+1 and we
have R (v, G2n+1) = 2n n ,
and there are two vertices v, v0 both are rumor centers in G2n then we have
R (v, G2n) = R (v0, G2n) =
2n − 1 n − 1
.
Every graph Gn in a 2-regular tree is a path. The correct detection
proba-bility of finding the rumor source in a path is
P(source = vGn) = P(Gn source = v)P(source = v) P i∈GnP(Gn source = i)P(source = i) .
The probabilities of vertices which are rumor sources in a d-regular tree are
equal possible. Thus
P(source = vGn) = P(Gn source = v) P i∈GnP(Gn source = i) . By (4), P(source = vGn) = R (v, Gn) P i∈GnR (i, Gn) (9) Hence P(source = vG2n+1) = 2n n 22n 0 +2n 1 + · · · + 2n n − 1 +2n n =2n n · 1 22n, and P(source = vG2n) = 2n − 1 n − 1 22n − 1 0 +2n − 1 1 + · · · +2n − 1 n − 1 =2n − 1 n − 1 · 1 22n−1 = 2n n · 1 22n. By Stirling’s formula, 2n n · 1 22n ∼ √ 4πn 2ne 2n 22n√2πn n e n√ 2πn n e n = 1 √ πn.
Thus P(Et) = P(source = vGt) = r 2 πt if t is even, 1 2P(source = v Gt) + 1 2P(source = v 0 Gt) = s 2 π(t − 1) if t is odd. (10)
It follows that the path detection probability scales as t−1/2, which
ap-proaches 0 as t apap-proaches infinity.
Now, we consider the detection probability of finding the source in a
d-regular tree, d ≥ 3.
Theorem 3.1.2. Suppose a rumor has spread in a regular tree. Then we
have that
0 < P (Et) ≤
1 2.
Proof:
Consider a graph Gn in a regular tree. We can regard the number of ways
to spread a rumor to every vertex in Gn with v as the source as the sum of
ways that rumor can be spread from v through u, where u is the neighbor
of v. Hence
R (v, Gn) =
X
u∈N (v)
R ({v, u}, Gn· vu) . (11)
Given any two adjacent vertices u and v, we have
R (u, Gn) =
X
w∈N (u)
R ({u, w}, Gn· uw) and
R (v, Gn) =
X
u∈N (v)
Since u is adjacent to v, there is at least one common term in above two
sums on the RHS. Thus we have
R (v, Gn) ≤ X i∈Gn,i6=v R (i, Gn) ⇒ R (v, Gn) ≤ 1 2 X i∈Gn R (i, Gn) . (12)
Since R (v, Gn) is positive, we have
0 < PR (v, Gn)
i∈Gn
R (i, Gn)
≤ 1
2.
This concludes that for every graph which composes of infected vertices,
the detection probability is greater than 0 and less than 1/2.
For d-regular trees with d > 2, Theorem (3.1.2) states that the event
that positive detection probability happens is independent of the order of the
graph. In what follows, we shall evaluate the explicit detection probability
of finding the source in a d-regular tree.
Our goal is to calculate the detection probability of finding the rumor
source. Recall the Theorem (2.2.1), v is the unique rumor center if and only
if tv u <
n
2 for all u 6= v. So we consider the number of vertices of branches
of Tv, and these branches are rooted trees with root u, u ∈ N (v). Without
loss of generality, we assume tv
u ≥ 1 for all u ∈ N (v). Let Ad and Bdbe two
sets such that
Ad= {(a1, a2, · · · , ad)| 1 ≤ ai < n 2, d X i=1 ai = n − 1}, and Bd= {(b1, b2, · · · , bd)| bi ∈ N, d X i=1 bi = n − 1}.
Clearly, Ad⊆ Bd. Moreover, |Bd| = n − 1 − d + d − 1 d − 1 =n − 2 d − 1 .
Now, let Si = {(x1, x2, · · · , xd) ∈ Bd : xi ≥ n2}. By principle of Inclusion
and Exclusion, |Ad| = |Bd| − d [ i=1 Si = |Bd| − d X i=1 |Si|. (13)
Let xj = yj+ 1, j 6= i and xi = dn2e + yi. Pdi=1yi = (n − 1) − (d − 1) − dn2e =
bn 2c − d. Then we have |Si| = bn2c−d+d−1 d−1 = bn2c−1 d−1 . Hence |Ad| = n − 2 d − 1 − d · bn 2c − 1 d − 1 .
For any vertex in a d-regular tree, say v, let (tv v1, t v v2, · · · , t v vd) denote the orders of branches (Tv v1, T v v2, · · · , T v
vd) which is the decomposition of Gn in a
d-regular tree in which P
vi∈N (v)vi = n − 1. Consider (a1, a2, · · · , ad) ∈ Ad, v is a unique rumor center of any graph that the orders of branches of it is
satisfied tvv1 = a1, tvv2 = a2, · · · , t
v
vd = ad.
We want to calculate the total number of ways to spread a rumor to n
vertices from v. In addition, the order of breaches of this form graph Gn is
(tv v1, t
v
v2, · · · , t
v
vd). First, assume that the vertex v had spread a rumor to u where u ∈ N (v), and then it can be spread to u’s descendants only. There
are d − 1 choice of the next infected vertices since u has d − 1 children.
Now, there are two vertices (u and a child of u) have rumors, then there are
2d − 3 choice of the next infected vertices. Therefore, the number of ways
to spread a rumor to m vertices in that graph Tu is m
Y
i=1
Hence, given (tv v1, t
v
v2, · · · , t
v
vd), the total number of ways to spread a rumor is (n − 1)! d Y k=1 Qtvvk i=1((d − 2) (i − 1) + 1) tv vk! . (15)
Note that the number of ways to permute n − 1 steps from different subtrees
is a permutation of the multiset {tv v1 · 1, t
v
v2· 2, · · · t
v vd· d}.
Let Pd(n) be the ratio of the number of ways to spread a rumor to n
vertices such that (tvv1, tvv2, · · · , tvvd) ∈ An to the number of ways to spread a
rumor to n vertices such that (tv v1, t v v2, · · · , t v vd) ∈ Bn. Thus Pd(t) ≈ P (Et). We have Pd(n) = X (tv v1,tvv2,··· ,tvvd)∈Ad d Y k=1 Qt v vk k=1((d − 2) (i − 1) + 1) tv vk! ! X (tv v1,tvv2,··· ,tvvd)∈Bd d Y k=1 Qtvu k=1((d − 2) (i − 1) + 1) tv vk! ! . (16)
Theorem 3.1.3. If G is a 3-regular graph, then we have that
lim
t→∞P(Et) =
1 4.
Proof:
Let the source be v. By (16),
P3(n) = X (tv v1,tvv2,tvv3)∈A3 d Y k=1 Qt v vk i=1 i tv vk! ! X (tv v1,tvv2,tvv3)∈B3 d Y k=1 Qt v vk i=1 i tv vi! ! = X (tv v1,tvv2,tvv3)∈A3 1 X (tv v1,tvv2,tvv3)∈B3 1 = |A3| |B3| = n − 2 2 − 3 · bn 2c − 1 2 n − 2 2 = 1 − 3 (bn 2c − 1)(b n 2c − 2) (n − 2)(n − 3) .
This implies that P3(n) = 1 − 3 4 (n − 4) (n − 3) if n is even, 1 − 3 4 (n − 5) (n − 2) if n is odd. (17)
Hence, the proof follows.
3.2
Detection probabilities of d-regular trees, d ≥ 4
In what follows, we use (16) to calculate the detection probability of finding
the source in a d-regular tree, d ≥ 4. Let v be the source and let wdn =
Qn
i=1((d − 2) (i − 1) + 1) (14). We can rewrite it by (13).
Pd(n) = 1 − d X i=1 X (x1,x2,··· ,xd)∈Si wdx1 x1! wdx2 x2! · · ·wdxd xd! X (tv v1,tvv2,··· ,tvvd)∈Bd w dtv v1 tv v1! wdtv v2 tv v2! · · ·wdt v vd tv vd! . Let xd ≥ n2, Pd(n) = 1 − d X (x1,x2,··· ,xd)∈Sd wdx1 x1! wdx2 x2! · · ·wdxd xd! X (tv v1,tvv2,··· ,tvvd)∈Bd w dtv v1 tv v1! wdtv v2 tv v2! · · ·wdt v vd tv vd! .
Let f (x) be the exponential generating function for the sequence {wdn}∞n=1.
And we have (1 − ax)−1a = ∞ X n=0 −1 a n (−ax)n = 1 + ∞ X n=1 Qn i=1(a(i − 1) + 1) n! x n. (18)
From (18), we immediately know f (x) = (1 − ax)−1a − 1 where a = d − 2. Let Fk(x) = (f (x))k. We have Fk(x) = wd1 x 1!+ wd2 x2 2! + · · · k =(1 − ax)−1a − 1 k = k X l=0 k l (1 − ax)−al(−1)k−l = (−1)k+ k X l=1 (−1)k−lk l ∞ X n=0 −l a n (−ax)n = (−1)k+ k X l=1 (−1)k−lk l 1 + ∞ X n=1 −l a n (−ax)n ! = k X l=0 (−1)k−lk l + k X l=1 (−1)k−lk l ∞ X n=1 Qn i=1(l + a(i − 1)) n! x n = ∞ X n=1 k X l=1 (−1)k−lk l Qn i=1(l + a(i − 1)) n! ! xn. Let [xn]F (x) be the coefficient of xn in F (x). Then,
[xn]Fk(x) = k X l=1 (−1)k−lk l Qn i=1(l + a(i − 1)) n! . (19) The coefficient of xn in F (x) is wdn1 n1! wdn2 n2! · · · wdnk nk! whose degree is n = n1+ n2+ · · · + nk. Therefore, Let a = d − 2, the detection probability of finding
the source in a d-regular tree is
Pd(n) = 1 − d n−d X m≥n2 Qm i=1(a (i − 1) + 1) m! [x n−1−m]F d−1(x) xn−1 F d(x) . (20)
Clearly, this is too complicate to simplify the right side of (20). In Figure
5, we use computer to obtain the general behavior of this term for several
d’s and n ≤ 100, 000. As a matter of fact, we have 12 ≥ Pd(n) ≥ Pd0(n) ≥ 1
4
if d ≥ d0 ≥ 3
4
Conclusion
In this thesis, we have obtained a mathematical model for finding the
cul-prit who spreads rumors in a network defined on a d-regular tree (countably
infinite graph). We are concerned with the detection probabilities of finding
the culprit. By using this model, we are able to give a shorter and more
explicit proof for the cases when d ≤ 3. See [6] for a comparison.
Further-more, we can estimate the detection probabilities of finding the source in
d-regular trees for d > 3 by an explicit formula, though it is quite
com-plicate. It will be better if we can simplify the formula by using certain
combinatorial identities. Moreover, if we can reduce the estimation error
for general graphs, and generalize the estimator to networks with different
rumor spreading rate, then we have a much better result than the known
References
[1] N. T. J. Bailey, The Mathematical Theory of Infectious Diseases and
its Applications, second edition, Griffin, London, 1975.
[2] A. Ganesh, L. Massoulie, and D. Towsley, “The effect of network
topol-ogy on the spread of epidemics”, Proc. 24th Annual Joint Conference
of the IEEE Computer and Communications Societies (INFOCOM),
vol. 2, pp. 1455-1466, March 2005.
[3] C. Moore and M. E. J. Newman, “Epidemics and percolation in
small-world networks”, Phys. Rev. E, vol. 61, pp. 5678-5682, May 2000.
[4] M. E. J. Newman, “The spread of epidemic disease on networks”, Phys.
Rev. E, vol. 66, pp. 016128, July 2002.
[5] R. Pastor-Satorras and A. Vespignani, “Epidemic spreading in
scale-free networks”, Phys. Rev. Lett., vol. 86, pp. 3200-3203, April 2001.
[6] D. Shah and T. Zaman, “Rumors in a Network: Who’s the Culprit?”,
IEEE Transactions on Information Theory, vol. 57, pp. 5163-5181,
Au-gust 2011.
[7] J. H. Van Lint and R. M. Wilson, A Course in Combinatorics, second
edition, Cambridge University Press, Cambridge, 2001.
[8] D.B. West, Introduction to Graph Theory, second edition,