The College of Wooster Libraries

Open Works
Senior Independent Study Theses

2014

The Analytic Hierarchy Process: A Mathematical

Model for Decision Making Problems
Giang Huong Nguyen
The College of Wooster, [email protected]

Follow this and additional works at: https://openworks.wooster.edu/independentstudy

Part of the Mathematics Commons

Recommended Citation
Nguyen, Giang Huong, "The Analytic Hierarchy Process: A Mathematical Model for Decision Making Problems" (2014). Senior
Independent Study Theses. Paper 6054.
https://openworks.wooster.edu/independentstudy/6054

This Senior Independent Study Thesis Exemplar is brought to you by Open Works, a service of The College of Wooster Libraries. It has been accepted
for inclusion in Senior Independent Study Theses by an authorized administrator of Open Works. For more information, please contact
[email protected].

© Copyright 2014 Giang Huong Nguyen

 the analytic hierarchy
process: a mathematical
model for decision making
problems

Independent Study Thesis

Presented in Partial Fulfillment of the

Requirements for the Degree Bachelor of Arts in
the Department of Mathematics and Computer
Science at The College of Wooster

by
Giang Huong Nguyen
The College of Wooster
2014

Advised by:
Dr. John Ramsay
Abstract

The ability to make the right decision is an asset in many areas and lines of
profession including social work, business, national economics, and
international security. However, decision makers often have difficulty
choosing the best option since they might not have a full understanding of
their preferences, or lack a systematic approach to solve the decision making
problems at hand. The Analytic Hierarchy Process (AHP) provides a
mathematical model that helps the decision makers arrive at the most logical
choice, based on their preferences. We investigate the theory of positive,
reciprocal matrices, which provides the theoretical justification of the method
of the AHP. At its heart, the AHP relies on three principles: Decomposition,
Measurement of preferences, and Synthesis. Throughout the first five chapters
of this thesis, we use a simple example to illustrate these principles. The last
chapter presents a more sophisticated application of the AHP, which in turn
illustrates the Analytic Network Process, a generalization of the AHP to
systems with dependence and feedback.

iii
Acknowledgements

I would like to thank Professor John Ramsay, my advisor, for his tremendous
support with this thesis. His knowledge of the subject, passion for
mathematics, and relentless encouragement were the main source of
motivation that got me through a research project that at times seemed
impossible. Being his student has been the most extraordinary experience in
my sixteen years of schooling. I am most grateful to complete an Independent
Study project under his guidance.
My sincere thanks goes to all of the professors in the Department of
Mathematics and Computer Science. Each of them has offered advice that
helped shape the outcome of this thesis.
I would also like to thank my friends, Dung Nguyen’14 and Chris and
Christy Ranney, for their support. Without them, this thesis would not have
been possible.
Last but not least, thank you, Mom and Dad, for having faith in my
abilities and judgment.

v
Contents

Abstract iii

Acknowledgements v

1 Introduction 1

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 The Initial Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.3 Research Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 The Consistent Decision Maker 9

2.1 The Eigenvalue Problem . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2 Finding the Weight Vector for the Alice Example . . . . . . . . . . 16

3 The Inconsistent Decision Maker 19

3.1 From Consistent to Inconsistent . . . . . . . . . . . . . . . . . . . . 19

3.2 Positive Matrices and Their Eigenvalues . . . . . . . . . . . . . . . 22

3.3 Finding the Weight Vector for the Inconsistent Alice Example . . 41

vii
4 Finding the Score of an Alternative on an Objective 45
4.1 Finding the Score of Each Alternative in the Alice Example . . . . 46
4.2 The Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3 Extension of the Hierarchy in the Alice Example . . . . . . . . . . 53

5 Metrics 59
5.1 Measure of Consistency . . . . . . . . . . . . . . . . . . . . . . . . 60
5.1.1 Consistency of a Pairwise Reciprocal Matrix . . . . . . . . 60
5.1.2 Consistency Index . . . . . . . . . . . . . . . . . . . . . . . 68
5.1.3 Finding the Consistency Ratio for the Alice Example . . . 70
5.2 Measure of Pairwise Preferences . . . . . . . . . . . . . . . . . . . 72

6 An Application in Medical Diagnosis 75

6.1 The Supermatrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.1.1 The Supermatrix Approach for the Alice Example . . . . . 76
6.1.2 The Supermatrix Approach for a System with Feedback . 82
6.2 A Case Study in Medical Diagnosis . . . . . . . . . . . . . . . . . . 87
6.2.1 Preliminary Analysis of the Case Study . . . . . . . . . . . 87
6.2.2 Finding the Likelihood of the Diseases . . . . . . . . . . . . 98
6.2.3 Finding the Priorities of the Alternatives with Respect to
the Diseases . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.2.4 Finding the Priorities of the Alternatives . . . . . . . . . . 111

7 Conclusion 113
List of Figures

4.1 The Hierarchy for the Alice Example . . . . . . . . . . . . . . . . . 52

4.2 The Extended Hierarchy for the Alice Example . . . . . . . . . . . 54

6.1 Network for the Alice Example . . . . . . . . . . . . . . . . . . . . 77

6.2 Network for a System with Feedback . . . . . . . . . . . . . . . . . 84
6.3 Network for the Case Study in Medical Diagnosis . . . . . . . . . 89
6.4 Reduced Network for the Case Study in Medical Diagnosis . . . . 95
6.5 Linear Network for the Case Study in Medical Diagnosis . . . . . 96

ix
x LIST OF FIGURES
List of Tables

5.1 Values of the Random Index (R.I.) . . . . . . . . . . . . . . . . . . . 69

5.2 Interpretations of Entries in a Pairwise Comparison Matrix . . . . 73

6.1 Priorities of Symptoms with Respect to Diseases . . . . . . . . . . 100

6.2 Priorities of Diseases with Respect to Symptoms . . . . . . . . . . 103
6.3 Priorities of Symptoms with Respect to Symptoms . . . . . . . . . 105
6.4 Priorities of Alternatives with Respect to Diseases . . . . . . . . . 110

xi
Chapter 1

Introduction

1.1 Motivation

We make decisions every day. High school graduates decide which college to
attend. Moms make up their minds as to how much time they should let their
children watch TV each day. Chefs choose the spices to use in dishes. The
Federal Reserve makes judgments about its monetary policy. Decision making
is a part of everyday life, whether it be a personal choice, a decision with
global impact, or anything in between. Being able to make right decisions can
have a tremendously positive impact on individuals’ lives.

However, making the right decision is often easier said than done. For
starters, assuming that the decision makers are aware of their options, they
need to identify their goals in choosing among these options. (As a simple
example, consider a buyer, who is trying to decide between two kinds of
tomatoes. He or she needs to know what qualities in tomatoes he or she
values: the color, the shape, or the price.) Then, the decision makers need to

1
2 CHAPTER 1. INTRODUCTION

determine how well each option satisfies their goals. Even when the decision
makers have clear ideas of their goals and the qualities of their options, their
judgments can still exhibit inconsistency, or can vary according to time.
Variations can also occur due to the environment that surrounds the decision
making process.

The Analytic Hierarchy Process (AHP), developed by Thomas Saaty [10] is

a model that helps the decision makers arrive at the most logical choice. Using
the AHP, we first determine the options available to the decision makers and
their goals in making decisions. These goals and options will then be used to
construct an analytic hierarchy, which reflects the various factors in the
decision making process and their importance. The outcome of the AHP is a
priority vector, which gives us an insight into the best option for the decision
makers. In order to understand the strengths of the AHP, in the next section,
we consider a simple problem of choosing universities.

1.2 The Initial Problem

Alice is in her last year of high school. She has been accepted to three major
universities and is trying to decide which one to attend. Those universities are
City University, Suburb University, and Town University. In evaluating a
university, Alice considers three factors: the academic quality of the university,
the financial aid package that the university offers her, and the quality of the
town or city that surrounds the university. In our discussion of the AHP, we
will refer to the universities that extend their acceptance to Alice as
alternatives and the three factors as objectives.
CHAPTER 1. INTRODUCTION 3

One solution to Alice’s problem is to directly assess her opinion of each

university. For example, Alice can assign a score to each university, on a scale
of 1–10, based on each of her objectives. Then the total score of each university
is the sum of the scores of that university on the three objectives. The
university with the highest total score is the one that Alice should attend. The
solution to the problem seems easy enough.

However, what will happen if Alice has ten, instead of three, objectives?
Then scoring the universities will be a more complicated and error-prone task.
How about the possibility that each objective has a different level of
importance to Alice? Then we have to take into account not only the score of
each university on each objective, but also the weight of each objective to
Alice’s decision. The problem quickly escalates in terms of complexity if Alice
also has different potential majors in mind, and her choice of university affects
her choice of major. For example, one of the universities might not have a
Pre-Health program, but it has a strong Music department (assuming that
Alice is considering studying Pre-Health and Music). Another university has a
prestigious Pre-Health curriculum, but it does not have a Music major. The
third university has both Pre-Health and Music, allowing Alice more
opportunities to explore her interest, but both of the programs are only
average.

The AHP provides a systematic method of solving problems such as the

one that Alice is facing. We mentioned in the last section that in order to apply
the AHP, we first identify clearly the objectives and alternatives available to
the decision makers. Then there are two steps that we need to carry out. First,
we measure the importance of each objective to the decision maker, compared
4 CHAPTER 1. INTRODUCTION

to the importance of the other objectives. This relative importance is called the
weight of the objective. For ease of computing, we require that the measured
weights sum to 1 [4, p. 30]. For this example, suppose that the AHP has found
the weights for academic quality, financial aid, and quality of location to be
0.6, 0.3, and 0.1, respectively. According to these weights, Alice considers the
quality of a university’s education the most important factor in making her
decision. Following in importance are the financial aid package that she is
offered, and then the quality of the university’s location.

In the next step, the AHP measures how well each of the decision maker’s
alternatives satisfies each of the objectives. The extent to which each
alternative meets the decision maker’s expectation in each objective is
measured as a numerical value. This value is referred to as the score of the
alternative on that objective. In this example, suppose that the AHP has found
the scores for one of the universities, City University, on academic quality,
financial aid, and quality of location, to be 5, 7.5, and 10, respectively. These
scores indicate that Alice loves the setting of City University, while the
financial aid that she is offered is moderately good, and the academic quality is
only average. We noted earlier that a decision maker’s preferences can exhibit
inconsistency. The scores found by the AHP are not simply the decision
maker’s assessment of his or her own preferences. The AHP achieves these
scores using a systematic approach that will be discussed in Chapter 4, helping
the decision maker make the most logical choice given his or her preferences.

The reader might have guessed that in order for Alice to make a decision,
both the weights of her objectives and the score of each alternative on those
objectives must be taken into account. Indeed, we utilize both kinds of
CHAPTER 1. INTRODUCTION 5

information to compute the total score for each alternative. For the first
alternative, City University, we first take the products of each of City
University’s scores on an objective and the weight of that objective. Taking the
sum of those products will give us the total score of City University. Using the
values in this example, the total score of City University is computed as
follows:
5 × 0.6 + 7.5 × 0.3 + 10 × 0.1 = 6.25.

Given that we also know the score of the other alternatives on Alice’s
objectives, we can compute the total score of those alternatives using the same
method explained above. The alternative with the highest total score should
be chosen.

In subsequent chapters, it will be clear that the AHP involves more than
just finding the weights of the objectives and the scores of the alternatives. In
fact, the two steps outlined above are merely part of a hierarchy that is used to
model the decision makers’ preferences. The same idea of a hierarchy that is
used to solve Alice’s problem could be used to answer questions about
promotion and tenure in higher education [10, p. 162], the optimum choice of
coal plants [10, p. 156], and measuring the world influence of nations such as
the U.S., China, France, the U.K., Japan, . . . [10, p. 40]. For the purpose of
understanding the foundations of the AHP, the next three chapters will
investigate the two outlined tasks: finding the weights of the objectives and
the scores of the alternatives. The discussion of the hierarchy will soon follow.

At this stage, the AHP problem is summarized in mathematical notations

as follows: Suppose a decision maker has n objectives and m alternatives. In
6 CHAPTER 1. INTRODUCTION

the first step of the AHP, for the ith objective, the AHP generates a weight wi . In
the second step, for the ith objective and the kth alternative, the AHP obtains a
score sik of the kth alternative on the ith objective. The total score of the kth
alternative is then computed by the following formula:

n
X
wi sik . (1.1)
i=1

After the total scores of all alternatives have been calculated, the decision
maker should choose the alternative that has the highest total score.

1.3 Research Outline

This thesis focuses on the method and mathematical reasoning of the AHP.
The next two chapters cover the task of finding the weights of the objectives.
Specifically, in Chapter 2, we will develop Alice’s problem, introducing the
basic terms and concepts of the AHP. As will be clear in this chapter, an
essential concept of the AHP is the consistency of the decision maker’s
preferences. We will include the definition of consistency, and then focus on
the consistent decision maker. In Chapter 3, we investigate the case of an
inconsistent decision maker. We turn our attention to finding the scores of the
alternatives in Chapter 4. The discussion of the scores naturally leads to the
idea of a hierarchy. Having discussed the hierarchy, we apply it to extend our
Alice’s problem. In Chapter 5, we examine some measurement issues of the
AHP. Specifically, we will provide an explanation for the measure of the
decision maker’s consistency, and comment on the scale that is used to
CHAPTER 1. INTRODUCTION 7

measure the decision maker’s pairwise preferences.

In the second-to-last chapter, we present a more sophisticated application
of the AHP in medical diagnosis. The application in turn illustrates the
Analytic Network Process, a generalization of the AHP to systems with
dependence and feedback. Included in Chapter 7 are final remarks and ideas
for future research.
The terminologies of the AHP are defined as they are first used in the
subsequent chapters. Unless otherwise noted, these definitions apply
throughout the thesis.
8 CHAPTER 1. INTRODUCTION
Chapter 2

The Consistent Decision Maker

2.1 The Eigenvalue Problem

In this chapter, we tackle the first task of the AHP: finding the weight of each
objective of the decision maker. We continue with the Alice example presented
in the last chapter, introducing terms and concepts that are fundamental to the
AHP. One such concept is what is called the pairwise comparison matrix. This
concept will naturally lead to the idea of the decision maker’s consistency in
preferences. The AHP problem will then be divided into two sub-problems:
one for the consistent and the other for the inconsistent decision maker. In the
rest of the chapter, we present the method of finding the weight vector for the
consistent case. This method involves the eigenvalue problem, a concept
central to the AHP. Finally, we provide an illustration of the eigenvalue
problem for the Alice example.

It is the writer’s intention to start the discussion of the AHP with a simple,
instructive example such as the problem of choosing university for Alice.

9
10 CHAPTER 2. THE CONSISTENT DECISION MAKER

However, the AHP has a wide array of applications in more complicated

situations. The complexity of these situations is enhanced by the fact that

(i) the decision maker has a large number of objectives,

(ii) the objectives are divided into layers of importance,

(iii) there exists a relationship between objectives and alternatives,

or any combination of those mentioned above. For more sophisticated

applications of the AHP, the interested reader is referred to [10, p. 91–163],
which covers topics such as prediction of the number of children in a
household, designing the transport system for the Sudan, and the future of
higher education in the United States.

In decision making problems that involve a large number of objectives, it is

difficult to compare each objective to the rest of the objectives. The process
tends to result in error, much like the process of scoring each alternative on a
scale of 1–10, compared to the other alternatives, that we mentioned in the
introduction. A method that provides a more accurate assessment of the
available objectives is to compare the importance of each objective to that of
each other objective. This method results in the pairwise comparisons of
objectives.

Returning to the Alice example, the AHP assumes that Alice knows the
pairwise comparisons of her objectives. Suppose she values the quality of a
university’s education twice as much as the financial aid that the university
offers, and six times as much as the quality of the university’s location.
Similarly, financial aid is three times as important to Alice as the quality of
CHAPTER 2. THE CONSISTENT DECISION MAKER 11

location. These numerical values that represent the pairwise comparisons of

objectives are placed into a pairwise comparison matrix A:

 
 1 2 6 
 
 
A =  21 1 3  .
 
 
 1 1 

6 3
1 

The entry in the ith row and the jth column of matrix A is the importance of
objective i compared to that of objective j. For example, a13 = 6 means that the
first objective, academic quality, is six times as important as the third objective,
location. It follows that the entries in the diagonal of matrix A are equal to 1.
Moreover, if objective i is twice as important as objective j, then objective j
must be half as important as objective i. In other words, a ji = 1
aij
. We also note
that if Alice is consistent in her preferences, then ai j a jk = aik . That is, if she
prefers academic quality twice as much as financial aid, and financial aid three
times as much as location, then she must prefer academic quality six times as
much as location to be considered a consistent decision maker. According to
this definition, Alice in our example is consistent. However, a decision maker
in real life is rarely consistent in her preferences. In Chapter 3, we will discuss
the process of obtaining the weights of the objectives in the inconsistent case.
It is in this inconsistent case that the mathematical reasoning becomes more
complicated. For now, we turn our attention to the consistent case.

The next definitions formalize our discussion of the pairwise comparison

matrix.

Definition 2.1 (Pairwise Comparison Matrix). The pairwise comparison

12 CHAPTER 2. THE CONSISTENT DECISION MAKER

h i
matrix for a decision maker with n objectives is an n × n matrix A = ai j such
that:

(i) ai j > 0 for i, j = 1, . . . , n, and

(ii) a ji = 1
aij
for i, j = 1, . . . , n.

A matrix A that satisfies condition (i) is defined to be a positive matrix. If

A satisfies condition (ii), then it is said to be a reciprocal matrix. In the next
chapters, when we refer to the pairwise comparison matrix of a decision
maker, the assumption is that the matrix is positive and reciprocal. Also note
that conditions (i) and (ii) in Definition 2.1 imply that aii = 1 for i = 1, . . . , n.

Definition 2.2 (Pairwise Comparison Matrix of a Consistent Decision Maker).

If a decision maker is consistent, then the pairwise comparison matrix A
satisfies the conditions in Definition 2.1, and

(iii) aik = ai j a jk for i, j, k = 1, . . . , n.

Recall that the ai j entry in A represents the importance of objective i,

compared to that of objective j. Let wi be the (unknown for the time being)
weight of objective i. We assume each of the weights is positive, and the
weights sum to 1. Then for a consistent decision maker, the ij entry of A can be
written as:
wi
ai j = .
wj

We note that the above equality is guaranteed to be true only if the

decision maker is consistent. Thus, we have an alternative definition of the
pairwise comparison matrix for a consistent decision maker:
CHAPTER 2. THE CONSISTENT DECISION MAKER 13

Definition 2.3. The pairwise comparison matrix A of a consistent decision

maker has the following form:

 
w1 w1 w1

 w1 w2
··· wn


 
 w2 w2 w2 
h i  w1 w2
··· wn 

A = ai j =  ..  ,

.. .. ..
. . . . 


 
 wn wn wn 


w1 w2
··· wn


Pn
where wi > 0 and i=1 wi = 1.

Definition 2.3 leads to the formal definition of the weight vector of the
decision maker:

Definition 2.4. The weight vector w of a decision maker is of the form:

 T
w = [wi ] = w1 w2 · · · wn ,

Pn
where wi > 0 and i=1 wi = 1. The weight vector w is also referred to as the
priority vector of the decision maker.

As mentioned in the introduction, the goal of the AHP is to find w. The

next theorem guarantees that for a consistent decision maker, we can always
obtain this weight vector from the pairwise comparison matrix A .

Theorem 2.1. Suppose a decision maker is consistent and has n objectives. Let A be
the corresponding pairwise comparison matrix, and w the weight vector. Then w is an
eigenvector of A with corresponding eigenvalue λ = n.
14 CHAPTER 2. THE CONSISTENT DECISION MAKER

Proof. By Definition 2.3 and Definition 2.4

    
w1 w1 w1 w1 w1 w1

 w1 w2
··· wn
  w  
  1   w
w1 1
+ w
w2 2
+ ··· + w
wn n


     
w2 w2 w2  w2 w2
+ + · · · + wwn2 wn
   w   

w1 w2
··· wn    2   w
w1 1
w
w2 2

Aw =  ..   .. = 
 
.. .. .. .. 
. . . .   .   .
 
 
     
     
 wn
w1
wn
w2
··· wn 
wn
  w  
n
wn
w
w1 1
+ wn
w
w2 2
+ · · · + wwnn wn 
   
 nw   w 
 1   1 
   
   
 nw2   w2 
=   = n  . = nw.
  
..  . 
.  . 
 
 
   
   
 nw   w 
n n

Thus, w is an eigenvector of A with corresponding eigenvalue n. 

From Theorem 2.1, we have a way to obtain the weight for each objective,
given that we know the number of objectives n and the pairwise comparison
matrix A. We know this by observing that:

Aw = nw

⇔ Aw − nw = 0

⇔ Aw − nIw = 0

⇔ (A − nI)w = 0.

Thus, w is in null(A − nI), where I is the identity matrix of an appropriate

dimension.
CHAPTER 2. THE CONSISTENT DECISION MAKER 15

The equation
Aw = nw (2.1)

proved in Theorem 2.1 is critical to the AHP. In Chapter 3, we will discuss

certain conditions that allow us to apply Equation 2.1 to solve the decision
making problem when the decision maker is inconsistent. In Chapter 5, this
equation is once again important to our discussion of the decision maker’s
consistency. Throughout the rest of this thesis, we shall refer to Equation 2.1 as
the eigenvalue problem.

The purpose of this section has been to present the eigenvalue problem in
Equation 2.1. We close the section by the next theorem, which provides a quick
and easy way to identify the weight vector when the decision maker is
consistent.
 
Theorem 2.2. The normalized form of any column of the matrix A = wwij is a
 T
solution to the eigenvalue problem Aw = nw, where w = w1 w2 · · · wn is the
weight vector solution and n is the number of objectives.

Proof. By Definition 2.3, the j column of A has the form

 T
w1
wj
w2
wj
··· wn
wj
,

for j = 1, 2, . . . , n. Therefore, each column of A is simply a constant multiple of

w. It follows that the normalized form of any column of A is a solution to the
eigenvalue problem. 

In the next section, we apply both Theorem 2.1 and Theorem 2.2 to find the
16 CHAPTER 2. THE CONSISTENT DECISION MAKER

weights of Alice’s objectives. Our expectation is that the methods from both
theorems give the same weight vectors.

2.2 Finding the Weight Vector for the Alice

Example

In this section, we illustrate the process of finding the weight vectors for the
Alice example, using both Theorem 2.1 and Theorem 2.2. By our definition of
consistent pairwise comparison matrices, the matrix A in the Alice example,
which was given in the last section, is consistent. This guarantees that we can
apply both theorems to obtain the weight vectors to solve Alice’s problem.

First, applying Theorem 2.1, we know that λ = 3 is an eigenvalue of A and

that w is in the null space of A − 3I. In order to find the solutions of the
homogeneous system (A − 3I)w = 0, we first find:

     
 1 2 6   3 0 0   −2 2 6 
     
     
A − nI = A − 3I =  21 1 3  =   .
     1
 −  0 3 0

     2 −2 3 
 1 1     1 1

1   0 0 3 −2
   
6 3 6 3

 
 1 0 −6 
 
 
The reduced row echelon form of A − 3I is:  0 1 −3 . Using
 

 
 
 0 0 0 
Gauss-Jordan elimination [8, p. 78], the eigenvectors of A with corresponding
 T
eigenvalue 3 have the form 6s 3s s (s ∈ R).

It can be easily seen that an eigenvector of A with corresponding

CHAPTER 2. THE CONSISTENT DECISION MAKER 17

 T
eigenvalue 3 is 6 3 1 . Since the weights of all of the objectives sum to 1,
the weight vector is
 T
w= 6
10
3
10
1
10
.

According to this weight vector, the weights of the three objectives—academic

quality, financial aid, and location—are 0.6, 0.3, and 0.1, respectively. This
weight vector is indeed a constant multiple of any of the columns in A. So the
methods of Theorem 2.1 and Theorem 2.2 yield the same result.
In this chapter, we have discussed the components essential to the
theoretical reasoning of the AHP. The discussion of the pairwise comparison
matrix at the beginning of the chapter brought about the bifurcation of the
AHP problem into the consistent and the inconsistent sub-problems. We also
justified the existence of a solution to the AHP in the consistent case by the
eigenvalue problem. In the next chapter, we shall see how this problem is
utilized in solving for the weight vector in the inconsistent case.
18 CHAPTER 2. THE CONSISTENT DECISION MAKER
Chapter 3

The Inconsistent Decision Maker

3.1 From Consistent to Inconsistent

In reality, decision makers are usually inconsistent. For example, if a

prospective college student prefers academic quality twice as much as
financial aid, and financial aid three times as much as location, it is unlikely
that she will prefer academic quality six times as much as location. As a result,
h i
for the n × n pairwise comparison matrix A = ai j , it no longer holds that
aik = ai j a jk for i, j, k = 1, . . . , n. Therefore, we cannot directly apply the
eigenvalue problem presented in the last chapter to the inconsistent case. (We
note that A is still a positive, reciprocal matrix. This means that ai j > 0 and
ai j = 1
a ji
for all i and j.)

Saaty contends that if the entries of a positive reciprocal matrix change by

small amounts, then the eigenvalues of that matrix also change by small
amounts [10, p. 51]. In addition, the corresponding eigenvectors do not vary
by much. Provided that our decision maker is not too inconsistent, the

19
20 CHAPTER 3. THE INCONSISTENT DECISION MAKER

inconsistent pairwise comparison matrix will not deviate much from the
consistent matrix. We can find the weight vector for an inconsistent decision
maker based on the weight vector in the consistent case. Thus, we want to
further explore the eigenvalues and corresponding eigenvectors of a consistent
pairwise matrix. The next theorem serves that purpose.

Theorem 3.1. Suppose a decision maker is consistent and has n objectives. Then the
pairwise comparison matrix has a unique largest eigenvalue n. All of the other
eigenvalues are zero.

Proof. In the last chapter, we proved that

Aw = nw.

So w is an eigenvector of A with corresponding eigenvalue λ = n.

Consider:
          
 w w1 w1 w1   w
 1
 
  w1 w2
··· wn   1



 w1 − w1  
  0 

 w
 1


           
   w2 w2 w2          

 0  
  w1 w2
··· wn 
 
  0 


 w2 − w2  
  0 


 0 

 ..  
 =  .. .. ..  
  .. 
 = 
 ..  
 =  .. 
 = 0 
 .. 
 .
A  .   . . ··· .   .   .   .   . 
           
wn−1 wn−1 wn−1 
           
 0   ···   0   wn−1 − wn−1   0   0 
   w1 w2 wn          
           
   wn wn wn    −w       
−wn w1 w2
··· wn n wn − wn 0 −wn
CHAPTER 3. THE INCONSISTENT DECISION MAKER 21

 
 w 
 1  
 


 0 

.

.

.  is an eigenvector of A with corresponding eigenvalue

Thus, the vector  

 
 

 0 
 
 
−wn 
λ = 0. Using the same approach, we can prove that the following vectors are
the eigenvectors corresponding to the eigenvalue λ = 0 for A:

       
 w   0   0   0 
 1       
       
       
 0   w2   0   0 
       
       
0 0   w3 0
       
     
 ,   ,   , . . . ,   .
.. .. .. ..

       

 .  
  .  
  . 


 . 

       
       
 0   0   0   wn−1 
       
       
−wn −wn −wn −wn
       
| {z }
n-1 vectors

It can be easily seen that these n − 1 vectors are linearly independent, since
none of them could be written as a linear combination of the others. So the
basis of the eigenspace associated with the eigenvalue λ = 0 has at least n − 1
vectors. In other words, the geometric multiplicity of the eigenvalue λ = 0,
which is the dimension of the eigenspace associated with that eigenvalue, is at
least n − 1. Since the algebraic multiplicity of the eigenvalue λ = 0, which is
the number of times its factor occurs in the characteristic polynomial, is
always bigger than or equal to its geometric multiplicity, the factor of the
eigenvalue 0 occurs at least n − 1 times in the characteristic polynomial.

The characteristic polynomial has a degree of n. In the proof of Theorem

22 CHAPTER 3. THE INCONSISTENT DECISION MAKER

2.1, we already proved that λ = n is an eigenvalue of A. Therefore, the factor of

the eigenvalue λ = 0 can only occur in the characteristic polynomial n − 1
times, since the factor of the eigenvalue λ = n has to occur at least once in the
characteristic polynomial with degree n. Therefore, the characteristic
polynomial of A is p(λ) = λn−1 (λ − n).

Thus, the matrix A has a unique largest eigenvalue λ = n and all of the
other eigenvalues equal zero. 

From the proof above, when the eigenvalue of A is 0, the corresponding

eigenvectors violate the assumption wi > 0. Therefore, the only useful
eigenvalue of A is λ = n. Provided that the decision maker is slightly
inconsistent, we expect that A has a unique largest eigenvalue that
approximates n. The corresponding eigenvector, denoted w0 , approximates w.

Our problem for the inconsistent case becomes: given a decision maker
that is inconsistent in her preferences, find the weight vector w0 that satisfies

Aw0 = λmax w0 , (3.1)

where λmax is the unique largest eigenvalue for A. In Section 3.2, we will
discuss Perron’s theorem, which guarantees that Equation 3.1 always has a
unique solution.

3.2 Positive Matrices and Their Eigenvalues

In this section, we discuss Perron’s theorem for positive matrices. The proof of
this theorem provides the theoretical foundation for the method of finding the
CHAPTER 3. THE INCONSISTENT DECISION MAKER 23

weight vector in the inconsistent case. We begin by introducing several

concepts and theorems about stochastic matrices. These results will be useful
for our proof of Perron’s theorem later. The statement and the proof of
Perron’s theorem are presented in the second half of the section. The idea of
the proof of Perron’s theorem in this section draws from an outline provided
by Saaty [10, p. 170–176].

In Chapter 2, we mentioned the condition that makes a matrix positive.

For the purpose of the theorems in this section, we provide a formal definition
of the terms non-negative matrix and positive matrix.

Definition 3.1. A real matrix A is non-negative (or positive) if all of the entries
of A are non-negative (or positive). We write A ≥ 0 (or A > 0).

Definition 3.2. A non-negative matrix M is a stochastic matrix if each of the

row sums equal 1 [16, p. 189].

In another common definition of stochastic matrix, the entries in each of the

columns of M sum to 1. We can also say that the columns of M are probability
 
 0.7 0.2 
vectors. An example of a column stochastic matrix is M =   .
 
 0.3 0.8 

Theorem 3.2. For a positive, n × n, row stochastic matrix M

lim Mk = ev,
k→∞

 
where v = v1 v2 · · · vn is a positive row vector, ni=1 vi = 1, and
P
 T
e = 1 1 ··· 1 .
24 CHAPTER 3. THE INCONSISTENT DECISION MAKER

Proof. Theorem 3.2 states that for large k, Mk approaches a matrix with
identical rows. In order to prove this theorem, we prove that each column of
Mk approaches a column vector with identical entries. Let
 T
y0 = y0 y0 · · · y0
1 2 n be an arbitrary column vector in Rn . Define
yk = Mk y0 , with k = 0, 1, 2, . . .. Let ak and bk be the maximum and minimum
components of yk , respectively. Further, let α be the minimum entry in M.

The outline of the proof is as follows: (i) first, we demonstrate that the
sequences (ak ) and (bk ) are monotone, and (ii) bounded. As a result, (ak ) and
(bk ) converge. (iii) Next, we prove that the difference between (ak ) and (bk )
approaches 0 as k approaches infinity. Therefore, (ak ) and (bk ) tend to a
common limit, and all of the components in yk also approach this limit. So yk
approaches a column vector with identical entries. (iv) Lastly, we choose y0 so
that for each y0 , yk represent a column of Mk . Putting everything together,
each column of Mk approaches a column vector whose entries are the same,
and Mk approaches a matrix whose rows are identical.

(i) The sequences (ak ) and (bk ) are monotone:

We observe that for k = 0, 1, 2, . . ., yk+1 = Mk+1 y0 = MMk y0 = Myk . Let

j
yik+1 be the ith component of yk+1 , yk be the jth component of yk , and mi,j be
the entry in the ith row and jth column of M, we have

n
X
j
yik+1 = mi, j yk = mi,1 y1k + mi,2 y2k + . . . + mi,n ynk . (3.2)
j=1

Without loss of generality, assume that y1k is the maximum component of

CHAPTER 3. THE INCONSISTENT DECISION MAKER 25

yk . Then y1k = ak . Equation 3.2 could be rewritten as

yik+1 = mi,1 ak + mi,2 y2k + . . . + mi,n ynk .

Since bk is the minimum component of yk ,

yik+1 = mi,1 ak + mi,2 y2k + . . . + mi,n ynk

≥ mi,1 ak + mi,2 bk + . . . + mi,n bk

= mi,1 ak + (1 − mi,1 )bk .

The last line was achieved because the entries in each row of M sum to
unity. Further, since α is the minimum entry in M, α ≤ mi,1 and
1 − α ≥ 1 − mi,1 . Therefore

yik+1 ≥ mi,1 ak + (1 − mi,1 )bk

≥ αak + (1 − α)bk .

Similarly, without loss of generality, assume that y2k is the minimum

component of yk . Then y2k = bk . Equation 3.2 could be rewritten as

yik+1 = mi,1 y1k + mi,2 bk + mi,3 y3k + . . . + mi,n ynk .

26 CHAPTER 3. THE INCONSISTENT DECISION MAKER

With the same reasoning as the case above, we could write

yik+1 = mi,1 y1k + mi,2 bk + . . . + mi,n ynk

≤ mi,1 ak + mi,2 bk + . . . + mi,n ak

= mi,2 bk + (1 − mi,2 )ak

≤ αbk + (1 − α)ak .

We just showed that an arbitrary component of yk+1 has the following

bounds:
αak + (1 − α)bk ≤ yik+1 ≤ αbk + (1 − α)ak . (3.3)

Since the bounds hold for the largest component of yk+1

ak+1 ≤ αbk + (1 − α)ak , (3.4)

which is equivalent to

ak+1 − ak ≤ α(bk − ak ).

Because bk ≤ ak for k = 0, 1, 2, . . .

ak+1 − ak ≤ α(bk − ak ) ≤ 0

for k = 0, 1, 2, . . .. Thus, (ak ) is decreasing.

Similarly, the bounds in Equation 3.3 hold for the smallest component of
CHAPTER 3. THE INCONSISTENT DECISION MAKER 27

yk+1 . As a result
bk+1 ≥ αak + (1 − α)bk , (3.5)

which is equivalent to

bk+1 − bk ≥ α(ak − bk ) ≥ 0,

for all k = 0, 1, 2, . . .. Thus, (bk ) is increasing.

(ii) The sequences (ak ) and (bk ) are bounded and convergent:

We showed that (ak ) is decreasing and (bk ) is increasing. Thus, for

k = 0, 1, 2, . . ., ak ≤ a0 and bk ≥ b0 .

For k = 0, 1, 2, . . ., bk ≤ ak ≤ a0 . Thus, the increasing sequence (bk ) is

bounded above by the number a0 .

Similarly, for k = 0, 1, 2, . . ., ak ≥ bk ≥ b0 . Therefore, the decreasing

sequence (ak ) is bounded below by the number b0 .

By the Monotone Convergence Theorem [1, p. 51], since (ak ) is monotone

and bounded, it converges. Likewise, (bk ) converges.

(iii) The vector yk approaches a column vector with identical entries:

Combining Equation 3.4 and Equation 3.5, we have

ak+1 − bk+1 ≤ αbk + (1 − α)ak − (αak + (1 − α)bk ),

28 CHAPTER 3. THE INCONSISTENT DECISION MAKER

which is equivalent to

ak+1 − bk+1 ≤ (1 − 2α)(ak − bk ). (3.6)

Next, we prove that (ak − bk ) ≤ (1 − 2α)k (a0 − b0 ) by induction.

Choosing k = 0 and applying Equation 3.6, we have

a1 − b1 ≤ (1 − 2α)1 (a0 − b0 ),

so the base case is satisfied.

Assume an − bn ≤ (1 − 2α)n (a0 − b0 ). Then by Equation 3.6

an+1 − bn+1 ≤ (1 − 2α)(an − bn ).

By our assumption of the induction hypothesis, the quantity on the

left-hand side of the above inequality is less than or equal to
(1 − 2α)(1 − 2α)n (a0 − b0 ). Thus

an+1 − bn+1 ≤ (1 − 2α)n+1 (a0 − b0 ).

By induction,
ak − bk ≤ (1 − 2α)k (a0 − b0 ). (3.7)

We showed that (ak ) and (bk ) converge. Therefore, the Algebraic Limit
Theorem [1, p. 45] implies that (ak − bk ) also converges. Next, we
compute the limit of (ak − bk ).
CHAPTER 3. THE INCONSISTENT DECISION MAKER 29

Since M is a positive matrix whose row sums are ones, 0 < α < 1. Thus

0 < 2α < 2 ⇔ −1 < 1 − 2α < 1 ⇔ |1 − 2α| < 1,

which means limk→∞ (1 − 2α)k = 0 [1, p. 56].

From Equation 3.7 and the way we defined ak and bk , we have

0 ≤ ak − bk ≤ (1 − 2α)k (a0 − b0 ).

We know limk→∞ 0 = 0 and

limk→∞ ((1 − 2α)k (a0 − b0 )) = (a0 − b0 ) limk→∞ ((1 − 2α)k ) = (a0 − b0 )0 = 0.

Using the Squeeze Theorem [1, p. 49], limk→∞ (ak − bk ) = 0.

We showed that lim(ak ) and lim(bk ) exist. Again, by the Algebraic Limit
Theorem
lim (ak − bk ) = 0 = lim (ak ) − lim (bk ).
k→∞ k→∞ k→∞

Thus, (ak ) and (bk ) approach a common limit. Since ak and bk are the
maximum and minimum components of yk , respectively, all of the
components in yk also approach this limit. In other words, as k
approaches infinity, the vector yk approaches a column vector whose
entries are all the same. Since the decreasing sequence (ak ) is bounded
below by b0 and the increasing sequence (bk ) is bounded above by a0 ,

limk→∞ yk = Ce,
30 CHAPTER 3. THE INCONSISTENT DECISION MAKER

 T
where b0 ≤ C ≤ a0 and e = 1 1 · · · 1 .

(iv) Mk approaches a matrix with identical rows:

 T
j
Let y0 = y0 y0 · · · y0 , with yi0 = 1 and y0 = 0 for any j , i. Then
1 2 n

yk = Mk y0 is the ith column of Mk .

We showed that as k approaches infinity, the vector yk approaches a

column vector with identical entries. Therefore, the ith column of Mk also
approaches a column vector with identical entries, as k approaches
infinity.
 T
Let Ce = ci ci · · · ci . Since ci is an entry in Mk and M is positive, ci
is also positive. Let v be a row vector such that v j = ci , we have

     

 c1 c2 · · · cn   1   1 
 
     
       
 c1 c2 · · · cn   1    1  
lim Mk =  ..  = 
 c c · · · c =   v v · · · v = ev.

.. .. ..  1 2  ..  1 2
 
n n
k→∞ 
 . . .   .   . 
 
     
     
 c c ··· c  
1 2 n 1   1 

Since ci is positive, v is a positive row vector. The last item in the proof is
to show that the sum of all of the components of v is 1.
 T
Since the rows of M all sum to 1, Me = e, with e = 1 1 · · · 1 . We
prove that Mk is also a stochastic matrix by induction.

First, since M1 e = e as shown above, the base case is satisfied.

Assume that Mn e = e. Then Mn+1 e = M(Mn e) = Me = e.

Therefore, Mk e = e, by induction. So each of the rows of Mk sums to

CHAPTER 3. THE INCONSISTENT DECISION MAKER 31

Pn
unity, and therefore j=1 v j = 1.

We just proved that the powers of a positive, row stochastic matrix

eventually reach a stable stage. This result will be important to the proof of the
next lemma, which is in turn useful for the proof of Perron’s theorem.

Lemma 3.1. If A is a positive n × n matrix, then

Ak
lim = wv,
k→∞ λk

where λ is a positive constant, v a positive row vector, and w is a positive column

vector.

Proof. Let S be the set of all non-negative, column n-vectors such that the
entries of each of those vectors sum to 1. We denote

n
X
S = {x|x = [xi ]n×1 , xi ≥ 0, xi = 1, i = 1, 2, . . . , n}.
i=1

For any vector y, define the function

n
X
f (y) = yi ,
i=1

and the matrix transformation

1
T(y) = Ay,
f (Ay)

where A is a positive, n × n matrix.

32 CHAPTER 3. THE INCONSISTENT DECISION MAKER

The outline of the proof is as follows: (i) first, we demonstrate that the
transformation T is a continuous function, and T(S) ⊂ S. That is for an
arbitrary x ∈ S, T(x) ∈ S. (ii) Next, using the Brouwer Fixed Point Theorem
with the function T and the space S, we are able to find a positive fixed point
w. (iii) Finally, we apply Theorem 3.2 to prove the proposed limit.

(i) T is a continuous function which transforms S to S:

For an arbitrary x ∈ S, the matrix transformation Ax is continuous, since

each component of Ax is a linear function of x1 , x2 , . . . , xn and therefore
continuous. As a result, T(x) is continuous.

Let (Ax)i be the ith component of Ax. Consider

n n
X X 1
(T(x))i = ( (Ax))i
i=1 i=1
f (Ax)
1
= f( Ax)
f (Ax)
(Ax)1 (Ax)2 (Ax)n
= + + ... +
f (Ax) f (Ax) f (Ax)
n
1 X 1
= (Ax)i = f (Ax)
f (Ax) i=1 f (Ax)

= 1.

Since A is n × n and x is n × 1, T(x) is n × 1.

Pn
Further, since x ≥ 0 and xi = 1, x has at least one positive component.
i=1

Since A is positive, Ax > 0, which leads to f (Ax) = ni=1 (Ax)i > 0.

P

Therefore, T(x) > 0.

Pn
To sum up, for i = 1, 2, . . . , n, T(x) = [(T(x))i ]n×1 , T(x) > 0, i=1 (T(x))i = 1.
CHAPTER 3. THE INCONSISTENT DECISION MAKER 33

Therefore, T(x) ∈ S. In other words, T : S → S.

(ii) Find w and λ such that Aw = λw:

The space S is a closed bounded n − 1 dimensional convex set in Rn .

Thus S is topologically equivalent to a closed disk in Rn−1 .

By the Brouwer Fixed Point Theorem [14, p. 277], there exists a fixed
point w ∈ S such that
T(w) = w.

By our definition of T, the above equation could be written as

1
Aw = w, (3.8)
f (Aw)

which is equivalent to
Aw = f (Aw)w.

Like x, w is non-negative, with entries that sum to 1, so w has at least one

positive component. So Aw > 0. Therefore, the right-hand side of the
above equation has to be positive. Since we know that w is non-negative,
f (Aw) > 0 and therefore w > 0.

Set λ = f (Aw). Then we have

Aw = λw, (3.9)

where λ is a positive real number and w is a positive column n-vector.

(iii) Proof of the proposed limit:

34 CHAPTER 3. THE INCONSISTENT DECISION MAKER

Let D be a diagonal matrix, dii = wi and di j = 0 whenever i , j (wi is the ith

 T
component of w). Then it follows that w = De, with e = 1 1 · · · 1 .

For an arbitrary column n-vector z, Dz = 0 has only the trivial solution

z = 0. Therefore, D is invertible. D−1 is a diagonal matrix, with diagonal
entries 1
wi
. It follows that D−1 w = e.

We observe that

1 1 1
(D−1 ( )AD)e = (D−1 ( )A)De = D−1 ( )Aw = D−1 w = e.
λ λ λ

Thus, D−1 ( λ1 )AD is a row stochastic matrix. Further, since D−1 and D are
non-negative n × n matrices, A is a positive n × n matrix, and λ is
positive, D−1 ( λ1 )AD is a positive n × n matrix. Using Theorem 3.2,

1
lim (D−1 ( )AD)k = ev∗ ,
k→∞ λ
h i Pn
where v∗ = v∗i > 0, and i=1 v∗i = 1.
1×n

On the other hand,

1 1 1 1
limk→∞ (D−1 ( )AD)k = lim (D−1 ( )AD)(D−1 ( )AD) . . . (D−1 ( )AD))
λ k→∞ λ λ λ
| {z }
kterms
1
= lim D−1 ( A)k D
k→∞ λ
1
= D−1 (lim k Ak )D.
k→∞ λ
CHAPTER 3. THE INCONSISTENT DECISION MAKER 35

Therefore,

1 k
D−1 (lim A )D = ev∗
k→∞ λk
Ak
⇔ lim = Dev∗ D−1 = wv∗ D−1 .
k→∞ λk

Let v = v∗ D−1 . Since v∗ is a positive, row n-vector and D−1 is a

non-negative, n × n matrix, v is a positive row n-vector.

We have shown that for a positive n × n matrix A,

Ak
lim = wv,
k→∞ λk

where λ is a positive real number, v a positive row vector, and w is a

positive column vector.

We are now ready to state and prove Perron’s theorem, which justifies the
existence of a unique largest eigenvalue and a corresponding eigenvector w0
that satisfies Equation 3.1.

Theorem 3.3 (Perron’s Theorem). Let A be a positive matrix. Then

1. A has a real positive simple (i.e., not multiple) eigenvalue λmax , whose modulus
is larger than the modulus of any other eigenvalues.

2. Each of the right and left eigenvectors of A corresponding to the eigenvalue λmax
has positive components, and is essentially (to within multiplication by a
constant) unique.
36 CHAPTER 3. THE INCONSISTENT DECISION MAKER

3. The number λmax (also called the Perron root of A) is bounded above by the
maximum row (or column) sum of A, and bounded below by the minimum row
(or column) sum of A.

Proof. We claim that for a matrix A > 0, the real number λ and the vectors w
and v constructed in Lemma 3.1 satisfy Perron’s theorem. We present the
proof in the following steps: (i) We first prove that λ is a eigenvalue of A with
corresponding right eigenvector w and left eigenvector v, with λ, w, and v
constructed in Lemma 3.1. (ii) Next, we prove that λ is the unique, largest
eigenvalue of A, which is the first item in Perron’s theorem. (iii) The proof that
w and v are unique will be presented next. (iv) Lastly, we prove that λ is
bounded above by the maximum row (or column) sum, and bounded below
by the minimum row (or column) sum of A.

(i) The real number λ is an eigenvalue of A with corresponding right

eigenvector w and left eigenvector v:

From Equation 3.9, we know that λ is an eigenvalue of A with

corresponding right eigenvector w.

Using the result of Lemma 3.1

1 1 Ak Ak+1
wvA = lim k A = lim k+1 = wv.
λ λ k→∞ λ k→∞ λ

Therefore
wvA = λwv,

which is equivalent to
ewvA = λewv,
CHAPTER 3. THE INCONSISTENT DECISION MAKER 37

 
in which e = 1 1 · · · 1 is a row n-vector. Since ew is a constant, we
have
vA = λv, (3.10)

which means that v is a left eigenvector of A with eigenvalue λ.

We just showed that w and v are the right and left eigenvectors of A,
respectively, with eigenvalue λ. The vectors w and v, as well as λ, are
constructed in Lemma 3.1. In Lemma 3.1, we proved that λ is a positive
real number, w is a positive column vector and v is a positive row vector.

(ii) The eigenvalue λ is the unique, largest eigenvalue of A:

Suppose h is another eigenvalue of A with corresponding eigenvector u.

Then Au = hu. For any positive integer k, Ak u = hk u [8, p. 307]. This
means 1 k
λk
Au = ( λh )k u.

Taking the limit of both sides of the above equality and using the result
of Lemma 3.1, we have: wvu = limk→∞ ( λh )k u.

Since wvu is an column n-vector, the limit on the right-hand side of the
above equality has to exist, which means that limk→∞ ( λh )k has to exist.
Therefore, | λh | < 1, which is equivalent to

|h| < |λ|,

in which case limk→∞ ( λh )k = 0. Thus, λ is the unique, largest eigenvalue of

A.

(iii) The right and left eigenvectors w and v of A corresponding to λ are

38 CHAPTER 3. THE INCONSISTENT DECISION MAKER

unique to within multiplication by a constant:

We first prove that w is unique to within multiplication by a constant.

Suppose u is another right eigenvector of A corresponding to the

eigenvalue λ. Then Au = λu. For any positive integer k, Ak u = λk u [8,
p. 307]. Therefore 1 k
λk
Au = u.

Taking the limit of both sides of the above equality and using the result
of Lemma 3.1, we have: limk→∞ ( λ1k Ak u) = limk→∞ u ⇔ wvu = u. Since vu
is a constant, we can set a = vu. Thus

aw = u.

Similarly, to prove that any other left eigenvector of A corresponding to

λ is a constant multiple of v, suppose yA = λy. Then yAk = λk y (for any
k ∈ Z, k > 0)[8, p. 307]. Therefore 1
λk
yAk = y, which means y λ1k Ak = y.

Taking the limit of both sides of the above equation and using the result
Ak
of Lemma 3.1, we have y limk→∞ λk
= limk→∞ y ⇔ ywv = y. Since yw is a
constant, we can set c = yw. Thus

cv = y.

λ is also called the principal eigenvalue of A, with corresponding

principal eigenvectors w and v. Up to this point, we have proved the
first two items in Perron’s theorem. What remains to be shown is the
upper and lower bounds of λ.
CHAPTER 3. THE INCONSISTENT DECISION MAKER 39

(iv) The eigenvalue λ is bounded above by the maximum row (or column)
sum of A and bounded below by the minimum row (or column) sum
of A:
 T
Let e = 1 1 · · · 1 be a column n-vector. The row sums of A are
given by the components of Ae. Let M be the maximum row sum and m
be the minimum row sum of A. Then

me ≤ Ae ≤ Me. (3.11)

We proved that v is a left eigenvector of A with corresponding

eigenvalue λ. From Equation 3.10, we have vAe = λve.

In addition, from 3.11, we have

vme ≤ vAe ≤ vMe.

Therefore
vme ≤ λve ≤ vMe. (3.12)

Since ve is a positive real number (we proved that v > 0), we can divide
3.12 by ve, yielding:
m ≤ λ ≤ M.

Using similar techniques, we can prove that λ is bounded above by the

maximum column sum and bounded below by the minimum column
 
sum. This time, let e = 1 1 · · · 1 be a row n-vector. Then the column
40 CHAPTER 3. THE INCONSISTENT DECISION MAKER

sums of A are given by the components of eA. Let N be the maximum

column sum and n be the minimum column sum of A. Then
en ≤ eA ≤ eN.

Using Equation 3.9, we have eAw = eλw. Therefore, enw ≤ eλw ≤ eNw,
which is equivalent to
n ≤ λ ≤ N.

Perron’s theorem guarantees that in the inconsistent case, we can always

find the weight vector from a positive, reciprocal pairwise comparison matrix
A. (Equation 3.1 has a unique solution). The next theorem provides the
justification for the method of finding the weight vector.

Theorem 3.4.
Ak e
lim = w1 ,
k→∞ eT Ak e

where A > 0, w1 is its principal eigenvector corresponding to the maximum

eigenvalue λ1 , such that ni=1 (w1 )i = 1.
P

For the proof of this theorem, the reader is referred to [10, p. 176]. Theorem
3.4 states that in order to compute the weight vector of an inconsistent
pairwise comparison matrix, we raise the matrix to an arbitrarily large power,
and then divide the sum of each row by the sum of the entries in the matrix .

In the next section, we illustrate this method of finding the weight vector
for an inconsistent decision maker.
CHAPTER 3. THE INCONSISTENT DECISION MAKER 41

3.3 Finding the Weight Vector for the Inconsistent

Alice Example

In this section, we illustrate the use of Theorem 3.4 with the inconsistent
version of our Alice example. The consistent pairwise comparison matrix A is
modified so that Alice is inconsistent in her preferences:

 
 1 2 5 
 
 
A =  21 1 3  .
 
 
 1 1 
1
 
5 3

Since a12 a23 = 6 , 5 = a13 , A is not consistent. For k = 1:

    
 1 2 5   1   8 
     
     
    9 
A1 e = Ae =  12 1 3     2  ,
  1  = 


     
 1 1     23 

5 3
1 1 15

  
 1 2 5   1 
   
 
   421
   
eT A1 e = 1 1 1  21 1 3   1  = .

    30
 1 1   

5 3
1  1 

Applying Theorem 3.4 and approximating the result to five decimal places,
the eigenvector is

   
 8   0.57007 
   
1
A e 30
   
w = T 1 =
1 9  =  0.32066  .

   

e A e 421   2   
 23   
   0.10926 
15
42 CHAPTER 3. THE INCONSISTENT DECISION MAKER

Replicating this process for larger values of k yields:

 
 0.58176 
 
2
A e
 
w2 = T 2 =  0.30896 

e A e  
 
 0.10928 

 
 0.58157 
 
3
A e
 
w = T 3 =  0.30898 
3 
 
e A e  
 
 0.10945 

 
 0.58155 
 
4
A e
 
w4 = T 4 =  0.30900 

e A e  
 
 0.10945 

 
 0.58155 
 
5
Ae
 
w = T 5 =  0.30900  .
5  

e A e  
 
 0.10945 

The values of the eigenvector have stabilized after five iterations. In

general, we stop when
kwi − wi+1 k < ,

where  > 0 is predetermined. The weights of Alice’s objectives are the entries
of w5 : the weight of the first objective is the first entry of w5 , the weight of the
second objective is the second entry, and so on. From w5 , we can see that the
weights of academic quality, financial aid, and location are 0.58155, 0.30900,
and 0.10945, respectively. These weights are close to the weights in the
CHAPTER 3. THE INCONSISTENT DECISION MAKER 43

consistent case, which were found in Chapter 2 to be 0.6, 0.3 and 0.1. This
indicates that Alice is not too inconsistent in her preferences. The method to
measure the degree of inconsistency will be presented in Chapter 5.
Up until now, we have explained the methods used to find the weight
vectors in both the consistent and the inconsistent sub-problems of the AHP. If
a decision maker is consistent, Equation 2.1 states that the desired weight
vector w is an eigenvector of A with corresponding eigenvalue n. By Theorem
2.2, we are assured that w is simply the normalized form of any column in A.
If the decision maker is inconsistent, we find the largest eigenvector of A
corresponding to the principal eigenvalue λmax . (In other words, we find a
solution to Equation 3.1). Theorem 3.3 (Perron’s theorem) guarantees the
existence of a unique solution to Equation 3.1. In order to find this unique
principal eigenvector, we apply Theorem 3.4: raising A to an arbitrarily large
power, and then dividing each row sum by the sum of the entries in the
matrix. The iterations are stopped when the difference between two resulting
vectors is less than a prescribed value. Since the goal is for the resulting
vectors to converge, a quick way to obtain the weight vector is to raise A to the
2 · i power at the i iteration, for i = 1, . . . , n [10, p. 179].
In the next chapter, we turn our attention to the second task of the AHP:
finding how well each alternative satisfies each objective.
44 CHAPTER 3. THE INCONSISTENT DECISION MAKER
Chapter 4

Finding the Score of an Alternative

on an Objective

In this chapter, we present the method to determine the score of each

alternative on each objective. After the scores have been obtained, we will be
able to calculate the priority of each alternative. As it turns out, the method to
compute the scores of the alternatives is by nature similar to the process used
to obtain the weights of the objectives. This leads to the generalization of the
AHP, which at its heart involves the construction of a hierarchy with different
levels (or strata) in order to model the various elements in a decision maker’s
preferences.

45
 CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
46 OBJECTIVE
4.1 Finding the Score of Each Alternative in the

Alice Example

The computation of the scores starts with the construction of a pairwise

comparison matrix. In Chapter 2 and Chapter 3, the pairwise comparison
matrix has been used to compare each pair of the objectives in terms of their
importance to the goal. In this chapter, we construct a pairwise comparison
matrix for each objective, assessing how well each alternative satisfies that
objective, compared to each other alternative. In other words, we are now
interested in the pairwise comparisons of alternatives on each objective.

Not surprisingly, we aim to obtain the score vector s on an objective from

the pairwise comparison matrix for that objective. If the matrix is consistent,
we apply Theorem 2.2 to find s. If the matrix is inconsistent, we find s by using
Theorem 3.4. In a nutshell, finding the scores of the alternatives on an
objective uses the same process as finding the weights of the objectives on the
decision maker’s final choice. The only difference is that we have to repeat this
process for each objective. We illustrate this process with the Alice example.

Recall that Alice has been accepted into three universities: City University,
Suburb University, and Town University. Suppose further that we know how
well each university satisfies each objective, compared to how well each other
university satisfies the same objective. The extent to which alternative i
satisfies an objective, compared to the extent to which alternative j satisfies
that same objective, is measured on an integer-valued, 1–9 scale.1 These

1
This scale is the same scale that was used to measure the pairwise comparisons of objectives
in Chapter 2. This scale will be further discussed in Chapter 5.
CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
OBJECTIVE 47

pairwise comparative scores are placed into a pairwise comparison matrix for
each objective. For example, for the first objective, academic quality, the
pairwise comparison matrix comparing each pair of the three universities is

 
 1 4 8 
 
 
B1 =  41 1 3  .
 
 
 1 1 
1
 
8 3

The ij entry in B1 reflects the score of university i on academic quality,

compared to the score of university j on the same objective. In this example,
suppose we refer to City University, Suburb University, and Town University
as the first, second, and third university, respectively. Then the entry in the
second row and the first column of B1 , which is 14 , means that Suburb
University scores one fourth as well as City University on academic quality.

Though B1 is a positive, reciprocal matrix as required, it is inconsistent,

since b12 b23 = 4 · 3 = 12 , 8 = b13 . Applying Theorem 3.4, we find the score
vector s1 for B1 :  
 0.69488 
 
1
B e  
s11 = T 1 1 =  0.22717 
 
e B1 e  

0.07795 


 
 0.71788 
 
2
B e  
s21 = T 1 2 =  0.20459 
 
e B1 e  
 
 0.07753 
 CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
48 OBJECTIVE

 
 0.71677 
 
3
B 1
e  
s1 = T 3 =  0.20496 
3 
 
e B1 e  

0.07826 


 
 0.71664 
 
4
B 1
e  
s41 = T 4 =  0.20509 

e B1 e  

0.07826 


 
 0.71665 
 
5
B 1
e  
s51 = T 5 =  0.20509 

e B1 e  

0.07826 


 
 0.71665 
 
6
B1
e  
s1 = T 6 =  0.20509  .
6  

e B1 e  

0.07826 


The values of s1 are within our convergence tolerance  = 0.00001 after the
sixth iteration. The first entry of this vector is the score of the first alternative
on the objective academic quality, the second entry corresponds to the score of
the second alternative on that objective, and so on. Based on s61 , we know that
the scores of City University, Suburb University, and Town University on the
objective academic quality are 0.71665, 0.20509, and 0.07826, respectively. Note
that just like the weights of the objectives, these scores sum to 1.

We can use the same process to compute the score vectors of the three
universities for the other two objectives. Suppose the pairwise comparison
matrix comparing each pair of the three universities on the objective financial
CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
OBJECTIVE 49

aid is  
 1 1 1 

 7 4 

B2 =  7 1 2  .
 
 
 
 4 1 1 
2

Like B1 , B2 is also inconsistent. Applying Theorem 3.4, we find that the

values of the score vector s2 are within our convergence tolerance  = 0.00001
after the fifth iteration. The score vector for B2 is

 
 0.08234 
 
5
B 2
e  
s2 = T 5 =  0.60263  .
5 
 
e B2 e  

0.31503 


Based on this vector, the scores of City University, Suburb University, and
Town University on financial aid are 0.08234, 0.60263, and 0.31503, respectively.

Finally, the pairwise comparison matrix comparing each pair of the three
universities on location is  
 1 1 4 

 2 

B3 =  2 1 9  .
 
 
 1 1 
1
 
4 9

The score vector s3 , whose values are within our convergence tolerance
 = 0.00001 after five iterations of Theorem 3.4, is

 
 0.30116 
 
5
B e  
s53 = T 3 5 =  0.62644  .
 
e B3 e  

0.07239 

 CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
50 OBJECTIVE

The interpretation of this score vector is the same as the interpretations of

the two previous score vectors. For the objective location, the scores of the City,
Suburb, and Town University are 0.30116, 0.62644, and 0.07239, in that order.

Now that we have calculated the weights of the objectives, as well as the
scores of the alternatives on the objectives, we can compute the priorities of
the three alternatives. Applying Equation 1.1, we find the total score of City
University:

0.71665 × 0.58155 + 0.08234 × 0.30900 + 0.30116 × 0.10945 = 0.4751728295,

the total score of Suburb University:

0.20509 × 0.58155 + 0.60263 × 0.30900 + 0.62644 × 0.10945 = 0.3740466175,

and the total score of Town University:

0.07826 × 0.58155 + 0.31503 × 0.30900 + 0.07239 × 0.10945 = 0.1507794585.

We observe that these total scores can be computed by matrix

multiplication. We put the three column score vectors into a 3 × 3 score matrix
called S, and then right-multiply this matrix by the weight vector w, computed
in Chapter 3. The result is the column 3-vector, whose entries are the priorities
of the alternatives. This method provides a short computational way for
problems in which the decision makers have a large number of objectives.
CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
OBJECTIVE 51

    
 0.71665 0.08234 0.30116   0.58155   0.475172829500000 
     
     
Sw =  0.20509 0.60263 0.62644   0.30900  =  0.374046617500000  .
     
     
     
 0.07826 0.31503 0.07239  
0.10945   0.150779458500000
  

One way to think of the computation Sw is that we are weighing the scores
of each university on the objectives by the importance of those objectives.
From the obtained priority matrix, the priorities (or total scores) of City
University, Suburb University, and Town University, are 0.48, 0.37, and 0.15,
respectively. These scores are the same as the scores calculated separately for
each alternative. Based on these scores, Alice should choose to go to City
University, as it has the highest total score.

4.2 The Hierarchy

The similarity of the methods used to compute the weights of the objectives
and the scores of the alternatives might have led the reader to guess that there
is a connection between the roles of the objectives and the alternatives in the
AHP. Indeed, this similarity results from the fundamental idea of the AHP: the
construction of a hierarchy, which consists of different levels, in order to
reflect the various layers of factors that affect the decision making process.

There is no set of rules that prescribes the construction of a hierarchy.

However, it is ideal to choose layers that reflect the decision maker’s
preferences as well as the complexity of the decision making problem [10,
p. 14]. Figure 4.1 shows the hierarchy for the Alice example. The first level is
 CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
52 OBJECTIVE

Figure 4.1: The Hierarchy for the Alice Example

the goal of the decision making process, choosing a university. The second
level consists of the objectives used when choosing a university, academic
quality, financial aid, and location. The third level specifies the alternatives
available to Alice: City University, Suburb University, and Town University.

In Figure 4.1, each line segment represents the influence (or impact) that
an element of a higher level has on an element of a lower level. Through the
intermediary level (which is Objectives in this case), the goal has an impact on
the alternatives. In our Alice example, the influence of the second level on the
lowest level is how well each alternative satisfies each objective. This influence
CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
OBJECTIVE 53

is expressed mathematically by the score matrix S. The influence of the first

level on the second level is the weights of the objectives, represented by the
weight vector w. Finally, the overall influence of the first level on the lowest
level is the values of the alternatives to Alice. These values are computed by
the matrix multiplication Sw.

The hierarchy in Figure 4.1 is of the simplest form. It is a linear system,

extending from one level down to the next. A more complicated form of the
hierarchy is a system with feedback, in which case the levels in the hierarchy
interact with one another in a nonlinear manner. An example of this nonlinear
system will be presented in Chapter 6. In the next section, we perform a
simple extension of the linear hierarchy for the Alice example.

4.3 Extension of the Hierarchy in the Alice Example

The hierarchy for the Alice example can be easily extended to solve a slightly
different problem. Suppose Alice is trying to decide between studying
Engineering and studying Medicine in college. All of the three universities
that she considers offer these two programs, and her choice of major will affect
her choice of university. The two majors add a fourth layer to the hierarchy
called the Sub-alternatives. Figure 4.2 shows this extended hierarchy.

In Figure 4.2, each alternative has an influence on each of the

sub-alternatives. In this example, this influence is interpreted as the academic
strength and reputation of each major in each university. Each of the objectives
also has an influence on each major through the alternatives in the third level.
Likewise, the overall goal has an influence on the majors through the
 CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
54 OBJECTIVE

Figure 4.2: The Extended Hierarchy for the Alice Example

CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
OBJECTIVE 55

objectives in the second level and the alternatives in the third level.

We aim to find the score of the sub-alternatives on each alternative. The

computation of the score vectors introduced earlier in this chapter is easily
applicable to this new layer of sub-alternatives. We begin with a pairwise
comparison matrix for each university. The ij entry in each matrix represents
the strength of program i, compared to the strength of program j, at the
university. In this example, let Engineering be the first, and Medicine be the
second program. Suppose that the pairwise comparison matrices for the City,
Suburb and Town Universities are C1 , C2 , and C3 , respectively:

     
 1 2   1 4   1 1 
3
C1 =  C2 =  C3 =   .
     
 
 1 1   1 1   3 1 
2 4

We observe that all of the three matrices are consistent. This is in fact the
case for every 2 × 2 positive, reciprocal matrix, since it always holds that
aik = ai j a jk . The corresponding eigenvectors of these matrices are:

     
 2   4   1 
3 5 4
s1 =  s2 =  s3 =   .
     
 
 1   1   3 
3 5 4

In order to compute the priority vector of the two majors, we first put s1 ,
s2 , and s3 into a 2 × 3 matrix S0 . Right-multiplying S0 by Sw, we obtain the
desired priority vector.
 CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
56 OBJECTIVE

  
   0.71665 0.08234 0.30116   0.58155  
   
 2 4 1       0.65371 
3 5 4
S0 Sw =    0.30900  =   .
       
  0.20509 0.60263 0.62644
1 1 3       

3 5 4
    0.34628
0.07826 0.31503 0.07239 0.10945 
   

According to this result, the priority of studying Engineering is 0.65, while

the priority of studying Medicine is 0.35. Therefore, studying Engineering is
the most logical choice for Alice, given her objectives and alternatives.

This example serves its purpose to illustrate an extension of the AHP.

However, the example does not provide a good model for a real-world
problem. The sub-alternatives have been chosen as if Alice is "partially" going
to all of the three universities. Since the calculation of the priorities of the
sub-alternatives takes into account the priorities of the alternatives, the
priorities of Engineering and of Medicine have been weighed accordingly to
the weights of the three Universities. Specifically, the priority of Engineering
has been calculated as if Alice would divide her Engineering curriculum into
47% at City University, 37% at Suburb University, and 15% at Town University.
The same argument can be made for the calculation of the priority of Medicine.

Further, since we have already determined that City University is the most
logical choice for Alice, among the three universities, and Engineering is twice
as strong as Medicine at City University (the entry in the first row and the
second column of C1 is 2), we can easily conclude that Alice should study
Engineering at City University, without using the pairwise comparison
matrices to find S0 . A reasonable argument would be that if the strength of
CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
OBJECTIVE 57

Medicine greatly exceeds the strength of Engineering at a university that has a

lower score than City University, that would tilt the odds in this lower-scored
university’s favor. We changed the pairwise comparison matrix C2 for the next
best university, Suburb University, so that Medicine is 9 times as strong as
Engineering at this university. The AHP recommends that Alice should study
Medicine (priority of 0.61, as opposed to 0.39 for Engineering). Again, this
recommendation assumes that Alice studies Medicine 47% at City University,
37% at Suburb University, and 15% at Town University, which is a condition
that cannot be satisfied in real life.
1
It is interesting that changing the relative strength of Medicine from 4
to 9
times of Engineering in C2 does not affect the AHP’s recommendation that
Alice goes to City University. This strange result is due to the fact that the
influence among levels in a hierarchy extends downward. In Figure 4.2, the
choice of universities affects the choice of majors, but not vice versa. A
problem in which the choice of majors affects the choice of universities
requires a system with feedback, a generalization of the hierarchy. In Chapter
6, after our understanding of the fundamental components of the AHP has
been complete, we will present an example that illustrates this kind of system.

In this section, we have discussed the hierarchy, an essential component of

the AHP. As mentioned in the introduction, the hierarchy has been applied to
solve complicated problems in the social sciences. For a more detailed analysis
of the advantages of hierarchies, we refer the reader to [10, p. 14]. The
construction of a hierarchy for each decision making problem requires an
in-depth understanding of the various factors of that problem. Saaty proposes
a few suggestions on the construction a hierarchy. These suggestions are not
 CHAPTER 4. FINDING THE SCORE OF AN ALTERNATIVE ON AN
58 OBJECTIVE

based on theories in mathematics and are therefore beyond the scope of this
thesis. For more information, the interested reader can refer to [10, p. 14–16].
At this stage, we have investigated most of the foundations of the AHP,
including the idea of hierarchy, and the justification and computation of the
weights of the elements in a hierarchy. However, we still need to fill some
remaining gaps in the theory of the AHP. These are two of the metrics utilized
in the AHP: measure of a decision maker’s consistency, and the integer-valued
1–9 scale that was used to quantify the pairwise comparisons. We will spend
more time developing both of these metrics in the next chapter.
Chapter 5

Metrics

In Chapter 3, we presented the argument made by Saaty [10, p. 179] that when
the entries of a positive reciprocal matrix change by small amounts, then the
eigenvalues change by small amounts. This argument provides the foundation
necessary for us to move from an eigenvalue problem in the consistent case to
the same problem in the inconsistent case. In this chapter, we shall provide the
theoretical justification for our argument. The central content of this
justification is the first metric utilized in the AHP to measure a decision
maker’s consistency.

As we shall see, a positive, reciprocal matrix A is consistent if and only if

the principal eigenvalue λmax equals the number of objectives n. In addition, if
the perturbation of the entries ai j is small, and the number of objectives n is
also small, then the principal eigenvalue λmax does not deviate much from its
original value n. This leads to a method to measure a decision maker’s
consistency. In the second half of the chapter, we discuss a second metric used
in the AHP: the integer-valued 1–9 scale that was used to measure the decision

59
60 CHAPTER 5. METRICS

maker’s pairwise preferences.

5.1 Measure of Consistency

5.1.1 Consistency of a Pairwise Reciprocal Matrix

 
In Chapter 2, we defined the pairwise comparison matrix A = ai j to be a
 
positive, reciprocal matrix. The weight vector w = wi is a positive, column
n-vector, where wi is the weight of objective i, and ni=1 wi = 1.
P
 
Recall that in the consistent case, A0 = wwij , and the weight vector w0 is the
solution to the following equation:

A0 w0 = nw0 ,

where n is the number of objectives.

For an inconsistent pairwise comparison matrix A, the priority vector w is

the solution to the following equation:

Aw = λmax w. (5.1)

We found that λmax is the principal eigenvalue of A corresponding to the

principal eigenvector w.

The objective of this subsection is to show that if the perturbation of the

entries in A from those in A0 is small, then λmax does not deviate much from n,
provided that n is small. (Saaty contends that n should be less than 10 [10,
p. 181]). For a discussion of the sensitivity of the eigenvector, see section 7-7 of
CHAPTER 5. METRICS 61

[10, p. 192].

We assume that all perturbations of the entries in the original pairwise

wi
comparison matrix A can be represented by ai j =  .
wj i j
Specifically,
wi wi wj
wj
+ αi j = wj
(1 + α ).
wi i j
Then A is consistent when i j = 1. We note that since A
is a positive matrix in both the consistent and the inconsistent cases, i j > 0.
Furthermore, since A is reciprocal,  ji = 1
ij
for all i and j.

From Equation 5.1, for each i = 1, 2, . . . , n:

n
X
ai j w j = λmax wi .
j=1

Therefore

n
X wj
λmax = ai j
j=1
wi
w1 w2 wi wn
= ai1 + ai2 + . . . + aii + . . . + ain
wi wi wi wi
X wj
= ai j + 1,
j,i
wi

for each i = 1, 2, . . . , n.

The last line was achieved because by construction, the entries on the
diagonal of the pairwise matrix A are ones. Taking the sum of λmax over n rows
yield:

X wj X wj X wj
nλmax = a1j +1+ a2 j + 1 + ... + an j + 1.
j,1
w1 j,2
w2 j,n
wn
62 CHAPTER 5. METRICS

We can rewrite the above equation as:

X wj X wj
nλmax = ai j + ai j + n. (5.2)
1≤i<j≤n
wi 1≤j<i≤n wi

The first quantity on the right-hand side of Equation 5.2 corresponds to the
entries above the diagonal of A, and the second quantity corresponds to the
entries below the diagonal. Combining these two quantities, we have

X wj wi
nλmax = (ai j + a ji ) + n,
1≤i<j≤n
wi wj

which is equivalent to

X wj wi
nλmax − n = (ai j + a ji ). (5.3)
1≤i<j≤n
wi wj

The eigenvalue of the consistent pairwise comparison matrix is n, and the

eigenvalue of the inconsistent pairwise comparison matrix is λmax . Therefore,
λmax − n gives an intuitively reasonable measure of consistency. As it turns out,
the measure of consistency depends on both λmax − n and the number of
objectives n.

Define
λmax − n
µ= . (5.4)
n−1

We shall prove later that µ is a measure of consistency.

We have
λmax − 1 + 1 − n n(λmax − 1) 1 − n
µ= = + ,
n−1 n(n − 1) n−1
CHAPTER 5. METRICS 63

and by Equation 5.3

1 X wj wi
µ = −1 + (ai j + a ji ). (5.5)
n(n − 1) 1≤i<j≤n wi wj

We want to express µ in terms of i j so that µ is tractable for our discussion

of the inconsistent case. Since the relation a ji = 1
aij
still holds when the decision
maker is inconsistent, and by assumption

wi
ai j = i j , (5.6)
wj

we have
wj 1
a ji = . (5.7)
wi i j

Substituting Equation 5.6 and Equation 5.7 into Equation 5.5, we have

1 X w wj wj 1 w
i i
µ = −1 + ( i j + ).
n(n − 1) 1≤i<j≤n w j wi wi i j w j

Thus
1 X 1
µ = −1 + (i j + ). (5.8)
n(n − 1) 1≤i<j≤n i j
64 CHAPTER 5. METRICS

We note that

1 X 1
lim µ = −1 + lim (i j + )
ij →1 n(n − 1) 1≤i<j≤n ij →1 i j
1 X
= −1 + 2
n(n − 1) 1≤i<j≤n
1 2n(n − 1)
= −1 +
n(n − 1) 2
= 0.

The second to last line was achieved because the number of entries that are
n(n−1)
above the diagonal of A is n−1
i=1 i =
P
2
.

The above limit states that as consistency is approached (i j approaches 1),
µ approaches 0. This suggests that we choose µ as a measure of the decision
maker’s inconsistency. We will show that the general form of µ in Equation 5.4
suffices as a measure of consistency. In order to accomplish this goal, we need
to introduce the next three theorems.

First, we define 

i j − 1 i j ≥ 1



δi j = 


 ji − 1 i j < 1.




Since A is reciprocal, if i j < 1, then  ji = 1

ij
> 1. Thus,

δi j ≥ 0,

for all i and j. We note that δi j = 0 when A is consistent (because then

CHAPTER 5. METRICS 65

i j =  ji = 1). Developing from Equation 5.8 we have:

1 X 1
µ= (i j + ) − 1
n(n − 1) 1≤i<j≤n i j
1 X 1 2n(n − 1)
= ( (i j + ) − ).
n(n − 1) 1≤i<j≤n i j 2

2n(n−1)
In computing the limit of µ, we found that =
P
2 1≤i<j≤n 2. Therefore we
can write µ as
1 X 1
µ= (i j + − 2).
n(n − 1) 1≤i<j≤n i j

If i j > 1, then i j + 1
ij
= 1 + δi j + 1
1+δij
.

If i j < 1, and consequently  ji > 1, then i j + 1

ij
= 1
 ji
+  ji = 1
1+δij
+ 1 + δi j .

Therefore
1 X 1
µ= (1 + δi j + − 2).
n(n − 1) 1≤i<j≤n 1 + δi j

Further simplification yields

1 X δ2i j
µ= . (5.9)
n(n − 1) 1≤i<j≤n 1 + δi j

With this new notion of µ, we can proceed to prove the next theorems,
which will be helpful to our discussion of the measure of consistency.

Theorem 5.1. For an n × n positive, reciprocal matrix A, the principal eigenvalue

λmax ≥ n.
66 CHAPTER 5. METRICS

Proof. From Equation 5.4 and Equation 5.9

λmax − n 1 X δ2i j
=µ= ,
n−1 n(n − 1) 1≤i<j≤n 1 + δi j

which is non-negative, since δi j ≥ 0 by its construction.

Thus,
λmax ≥ n. (5.10)

Theorem 5.2. An n × n positive, reciprocal matrix A is consistent if and only if

λmax = n.

Proof. If A is consistent, then δi j = 0 (by construction of δi j ). As a result, µ = 0,

λmax −n
by Equation 5.9. This means that n−1
= µ = 0, by Equation 5.4. Thus,
λmax = n.

Conversely, if λmax = n, by Equation 5.4, µ = λmax −n

n−1
= 0. Then, by Equation
δij
2

1≤i<j≤n 1+δij = µ = 0. This implies that δi j = 0 for each choice of i and

1
P
5.9, n(n−1)
j. As a result, A is consistent. 

Theorem 5.3. Let

δ = max δi j .
i, j

Then
n−1 2
λmax − n < δ.
2
CHAPTER 5. METRICS 67

Proof. From Equation 5.4 and Equation 5.9

1 X δ2i j
λmax − n =
n 1≤i<j≤n 1 + δi j
1 X 2
< δ
n 1≤i<j≤n i j
1 X 2
≤ δ
n 1≤i<j≤n
1 n(n − 1) 2
= δ
n 2
n−1 2
= δ.
2

To arrive at the conclusion above, we have used the fact that δi j > 0 for some i, j
(by construction of δi j ), assuming that A is inconsistent. Consequently,
δ2ij
1 + δi j ≥ 1 for all i,j and 1 + δi j > 1 for at least one i, j. Therefore, 1+δij
< δ2i j . This
justifies the second step of our argument. To arrive at the fourth step, we used
the fact that the number of the entries above the diagonal of a matrix is
Pn−1 n(n−1)
i=1 i = 2
. 

Theorem 5.3 states that the deviation of λmax from n depends on the
maximum perturbation δ of the entries in A and the number of objectives n.
Small perturbations in the entries of A would cause λmax to deviate from n by a
small amount. Therefore, we can find the weight vector of the inconsistent
case using the same eigenvalue problem in the consistent case, which is
A0 w0 = nw0 . Theorem 5.3 thus provides the theoretical reasoning for our
method in Chapter 3.

However, one issue remains: a decision maker can be extremely

68 CHAPTER 5. METRICS

inconsistent. In this case, the entries of A will deviate from the entries of a
consistent matrix by large amounts. The weight vector that results from our
approach will be meaningless. As such, it is necessary that we develop a
measure of consistency. This measure will provide some insight into the
accuracy of the weight vector obtained from our eigenvalue approach. In the
next subsection, we will present this measure of consistency.

5.1.2 Consistency Index

We have argued that an n × n positive, reciprocal matrix A is consistent if and

only if its principal eigenvalue λmax equals the number of objectives n.
Theorem 5.3 shows that the difference λmax − n depends on the magnitude of
the maximum perturbation of the entries in A and n − 1. For this reason,

λmax − n
µ=
n−1

is used as a measure of the closeness of A to consistency. The quantity µ is

called the consistency index (C.I.) [10, p. 21].

From Theorem 5.3, we have

λmax − n δ2
µ= < .
n−1 2
δ2
Therefore, 2
provides an upper bound for our measure of the consistency
δ2
index. However, we note that if n is large, then µ < 2
even if λmax is far away
from n. Therefore, for a large number of objectives, µ might not provide a
meaningful measure of consistency. For n small, µ provides a reasonable
CHAPTER 5. METRICS 69

measure of how far λmax is from n, and, consequently, how far the decision
maker is from consistency. Saaty suggests that n should be less than 10 [10,
p. 181].

In order to check for consistency, Saaty uses both the consistency index
and another measure called the random index. A random sample of 500
pairwise reciprocal matrices is constructed. Each matrix is generated
randomly, and its entries are subject to the constraints that

(i) ai j are values from the integer scale 1–9, and

(ii) ai j = 1 if i = j, and

(iii) ai j = 1
a ji
for i, j = 1, 2, . . . , n.

The average value of the consistency indexes of these 500 matrices is called
the random index (R.I.). Table 5.1 gives the random indexes and the
corresponding matrix orders [10, p. 21]. Since Saaty suggests using the AHP
when the number of objectives is less than 10, this table only lists the R.I. for
matrices up to order 10.

Table 5.1: Values of the Random Index (R.I.)

n 1 2 3 4 5 6 7 8 9 10

R.I. 0.00 0.00 0.58 0.90 1.12 1.24 1.32 1.41 1.45 1.49

The consistency ratio (C.R.) of a matrix is the ratio of the C.I. of that matrix
to the R.I. for the same matrix order. Thus, for a specific decision maker, the
consistency ratio takes into account both the measure of consistency for that
70 CHAPTER 5. METRICS

specific decision maker (through the consistency index), and the measure of
consistency of a random sample of 500 other decision makers (through the
random index). Therefore, the consistency ratio provides a reasonable
measure of the consistency of the decision maker in question.

If the consistency ratio is 0.10 or less, the decision maker is not too
inconsistent and the result obtained by the AHP is acceptable. However, if the
C.R. is larger than 0.10, more serious inconsistency exists and the priority
vector might not provide an accurate solution to the decision making process
[17, p. 788]. We illustrate the use of the consistency ratio with our Alice
example.

5.1.3 Finding the Consistency Ratio for the Alice Example

In Chapter 3, we discussed the eigenvalue problem of the Alice example:

Aw0 = λmax w0 , (5.11)

where  
 1 2 5 
 
 
A =  21 1 3  .
 
 
 1 1 

5 3
1 

We found the weight vector to be

 
 0.58155 
 
A5 e
 
w0 = T 5 =  0.30900  .
 
e A e  
 
 0.10945 
CHAPTER 5. METRICS 71

From Equation 5.11 and the rule of matrix multiplication, we know that

3
X wj
λmax = ai j ,
j=1
wi

h i
for i = 1, 2, 3, A = ai j , and w0 = [wi ]. Thus, λmax can be computed as the ratio
of any component in Aw0 to the corresponding component in w0 . Computing
these ratios for all pairs of components in Aw0 and w0 yield

3
X wj 1(0.58155) + 2(0.30900) + 5(0.10945)
a1j = = 3.003694708,
j=1
w1 0.58155

3
X wj 1
(0.58155) + 1(0.30900) + 3(0.10945)
a2j = 2 = 3.003694450,
j=1
w2 0.30900

3
X wj 1
5
(0.58155) + 31 (0.30900) + 1(0.10945)
a3j = = 3.003694636.
j=1
w3 0.10945

Taking λmax = 3.003694, we have the consistency index

λmax − n 3.003694 − 3
= = 0.001847.
n−1 3−1

For a matrix of order 3, the random index (as in Table 5.1) is 0.58. Therefore,
the consistency ratio is

0.001847
= 0.003184 < 0.1.
0.58

So the inconsistency is acceptable and the weight vector w0 provides an

accurate solution to the Alice example.
72 CHAPTER 5. METRICS

In this section, we have investigated the theories that allow us to apply the
same eigenvalue approach in the consistent case to the inconsistent case. In
λmax −n
the process, we also derived the consistency index µ = n−1
, which is used
together with the random index to measure the decision maker’s inconsistency.
In the next section, we will discuss the integer-valued scale 1–9 that was used
to represent the decision maker’s preferences for each pair of objectives.

5.2 Measure of Pairwise Preferences

In Chapter 2, we introduced the pairwise comparison matrix A. Recall that the

ai j entry in A represents the importance of objective i for the decision maker,
compared to that of objective j. In this section, we explain the metric that is
used to measure the decision maker’s pairwise preferences. We will not aim to
investigate the derivation of this metric, since this derivation is based on
theories in the social sciences.

The values of these pairwise comparisons are drawn from a scale of

integers ranging from 1 to 9. The values of this scale and their interpretations
are presented in Table 5.2 [17, p. 787]. The interpretations of the values in this
table are modified in wording so that they fit the context of our decision
making problem. For the original table by Saaty, see [10, p. 54].
CHAPTER 5. METRICS 73

Table 5.2: Interpretations of Entries in a Pairwise Comparison Matrix

Value of ai j Interpretations

1 Objectives i and j are of equal importance.

3 Objective i is weakly more important than objective j.
5 Experience and judgment indicate that objective i is
strongly more important than objective j.
7 Objective i is very strongly or demonstrably more
important than objective j.
9 Objective i is absolutely more important than objective j.
2, 4, 6, 8 Intermediate values. For example, a value of 8 means
that objective i is midway between strongly and absolutely
more important than objective j.

Since A is reciprocal, if ai j is assigned one of the values in the above table,

a ji is simply the reciprocal of that value. The derivation of Table 5.2 is based on
the theories of stimulus and response in psychology. For a discussion of this
derivation, we refer the interested reader to the section on scale comparison
written by Saaty [10, p. 53–64].
The discussion in this chapter has completed our understanding of the
fundamental components of the AHP. At this stage, we have presented the
method and theories of the AHP. The next chapter will focus on an application
of the AHP in medical diagnosis. As mentioned in Chapter 4, this application
illustrates a system with feedback, a generalization of the hierarchy to include
feedback among levels of the hierarchy.
74 CHAPTER 5. METRICS
Chapter 6

An Application in Medical
Diagnosis

In the previous chapters, we have used the simple Alice example to illustrate
the method of the AHP. The objective of this chapter is to present a more
sophisticated application of the AHP in medical diagnosis. We first introduce a
few concepts that are useful to the execution of this application. Such concepts,
e.g. the supermatrix, are not fundamental elements of the AHP and therefore
are presented in this chapter as extensions of the AHP. The second half of the
chapter is dedicated to the application of the AHP in medical diagnosis.

75
76 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

6.1 The Supermatrix

6.1.1 The Supermatrix Approach for the Alice Example

We make use of this section to develop several key ideas for the medical
application of the AHP. In Chapter 4, we mentioned that the hierarchy for the
Alice example was a linear system: we start with the highest level, and then
extend downward from one level to the next. In terms of the relationship
between two consecutive levels, the higher has an influence on the lower, as
was pointed out in the extended Alice example (where the sub-alternatives
were added). However, in a real-world problem, it is possible that the lower
level of a hierarchy has an influence on a higher level as well, or the elements
in a level have dependent relationships. In fact, the hierarchy for the medical
application that will be presented in the next section exhibits both of those
characteristics. A hierarchic structure with such nonlinear relationship
between layers, or between elements in a layer, is called a system with
feedback. We will soon demonstrate how a linear hierarchy, such as that for
the Alice example, is modified to display the feedback relationship. In order to
do that, we first represent the hierarchy in the form of a network, which shows
clearly the relationship between levels in the hierarchy.

Figure 6.1 shows the network for the Alice example. The network has
three nodes, which correspond to the three levels in the hierarchy given in
Figure 4.1. Each node in turn includes one or more elements. For example, the
node Goal consists of only one element: the goal of choosing a university,
while the node Objectives has three elements: academic quality, financial aid,
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 77

Figure 6.1: Network for the Alice Example

and location. Each arrow in the network represents an influence that the
elements in the higher node has on the elements in a lower node. In this
example, the influence of the elements in the node Objectives on the elements
in the node Alternatives is represented by the scores of the alternatives on the
objectives. Similarly, the influence of the element in the node Goal on the
elements in the node Objectives corresponds to the weights of the objectives
on the goal of choosing a university. In Chapter 4, we expressed these scores
and weights as the score matrix S and the weight vector w, respectively.

In order to obtain the priorities of the three alternatives on Alice’s goal of

choosing a university, we enter S and w into a matrix W that displays the
78 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

interactions among elements in the nodes of a network.

Goal Objectives Alternatives

 
Goal  W W12 W13 
 11 
 
W = Objectives  W21 .
 
W22 W23  (6.1)
 

Alternatives  W31 W32 W33


Equation 6.1 shows the matrix W for a network with three nodes. The
construction of W is as follows. Let the first, second, and third node of the
network correspond to the first, second, and third row (column) of W,
respectively. Then the ij component in W reflects the influence of the elements
in the j node on the elements in the i node of the network. By influence, we
mean the priorities of the elements in the i node, with respect to the elements in
the j node of the network.

Each ij component in W is itself a matrix. For this reason, W is referred to

as a supermatrix. (In the rest of this thesis, whenever we refer to the
components of W, we mean the matrices that reflect the relationship between
elements in the nodes in a network, as opposed to the entries in W.) If there is
no dependent relationship between the i and the j nodes of the network, then
the i j component in W is the zero matrix. If dependent relationship exists, Wi j
is nonzero. In the Alice example, the weight vector w reflects the influence of
the goal on the objectives. Therefore, w is entered in the 2,1 position of W.
Using the same reasoning, the score matrix S is entered into the 3,2 position of
W. The form of the supermatrix for the Alice example is as follows:
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 79

 
 0 0 0 
 
 
W =  w 0 0  .
 
 (6.2)
 

 0 S I 

Since w is a weight vector, the entries in w are positive and sum to 1. Each
of the columns in S is itself a weight vector, so S is a column stochastic matrix.
By putting the identity matrix in the 3,3 position of W, we have made W a
column stochastic matrix. Saaty contends that the powers of W will eventually
reach a stable stage, denoted W k , and that by raising W to powers, we will
obtain the desired priorities [13, p. 494]. Since the purpose of this chapter is to
illustrate a sophisticated application of the AHP, we will not investigate the
theoretical justification of this claim. However, for more theoretical discussion,
we refer the reader to [10, p. 206–214] and [11].
The target priority vector can be found as a column in W k or as a
component in W k , depending on the form of W. Thus, we need several
definitions about certain types of matrix that will help us categorize W.

Definition 6.1. A square matrix is irreducible (by permutations) if it cannot be

 
 A1 0 
decomposed into the form  , where A and A are square matrices
 1 3
 A A 
2 3
and 0 is the zero matrix. Otherwise the matrix is said to be reducible [10, p. 168].

Definition 6.2. A non-negative, irreducible matrix A is primitive if and only if

there is an integer m ≥ 1 such that Am > 0. Otherwise A is called imprimitive
[10, p. 176].

Thus, W can be reducible, irreducible and primitive, or irreducible and

imprimitive. If W is reducible, the priorities of the elements in the n node of
80 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

the network, with respect to the elements in the j node, is given by the
component at the n, j position of W k , where n is the number of nodes in the
hierarchy. Thus, the priorities of the elements in the lowest node, with respect
to the elements in the remaining nodes, can always be read from the last row
of components in W. We note that by component of W k , we are also referring
to a matrix, since W k is the result of raising the supermatrix W to powers.

If W is primitive, W k has identical columns. Each of these columns gives

the desired priorities. For the theoretical justification of the cases when W is
reducible, or primitive, the reader is referred to [13, p. 494] and [10,
p. 208–214]. If W is imprimitive, Saaty shows that W can always be made
primitive by substituting arbitrarily small positive numbers for the zero
entries in W, subject to the condition that W remains column stochastic. More
detailed discussion on this topic is available at [11], [2], and [12]. In the
remaining of this chapter, we will apply the supermatrix approach to obtain
the target priorities both when W is reducible (the Alice example) and when it
is primitive (the application in medical diagnosis).

Substituting the entries of S and w into W in Equation 6.2, we have the

supermatrix for the Alice example, where the entries in w are in bold, and the
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 81

entries in S are italicized:

 
 0 0 0 0 0 0 0

 
 
0.58155
 0 0 0 0 0 0
 

0.30900 0 0 0 0 0 0

 
W = 0.10945 0 .
 
 0 0 0 0 0
 

 0
 0.71665 0.08234 0.30116 1 0 0
 
 
 0
 0.20509 0.60263 0.62644 0 1 0
 

0 0.07826 0.31503 0.07239 0 0 1


It is obvious that W is reducible. W can be decomposed into the form

 

 0 0 0 0 
   
 A1 0   
 0.58155 0 0 0 
, where A1 =   is a square matrix, and A3 is the
 

 
 
A2 A3   0.30900 0 0 0 

 
 
 0.10945 0 0 0 
3 × 3 identity matrix.

Raising W to powers, we find that the entries in W stabilize after three

iterations. In other words, W 3 is the stable form of W:

 

 0 0 0 0 0 0 0 

 
 

 0 0 0 0 0 0 0 

 
0 0 0 0 0 0 0
 
 
 
W 3 =   .
 
0 0 0 0 0 0 0

 
 
 0.47517282950 0.71665 0.08234 0.30116 1 0 
 0 

 

 0.37404661750 0.20509 0.60263 0.62644 0 1 0


 
 
0.15077945850 0.07826 0.31503 0.07239 0 0 1
 
82 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

The component in the 3,1 position of W 3 (in bold) gives the priorities of the
elements in the last node (Alternatives) of the network, with respect to the
element in the first node (Goal). These priorities are precisely what we found
by multiplying Sw in Chapter 4. The component in the 3,2 position of W 3
gives the priorities of the alternatives with respective to the objectives. This
component is exactly the score matrix S that we found in Chapter 4. For the
Alice example, raising the supermatrix W to powers yields the same result as
successively weighing the elements in a hierarchy.

The purpose of our discussion thus far has been to introduce the network
and the supermatrix, so that we can utilize them to solve a decision making
problem in a more complex system. The next subsection will present the
supermatrix approach in light of a system with feedback.

6.1.2 The Supermatrix Approach for a System with Feedback

We open this section with a brief discussion of the usefulness of systems with
feedback, which itself necessitates the supermatrix approach to obtain the
desired priority vector. Saaty contends that many problems in the social
sciences have such complex situations that the linear form of a hierarchy fails
to capture. An example is the various forms of organizations that cannot be
put in a hierarchic structure [10, p. 199].

In both the initial and the extended versions of the Alice example, the
desired priority vector was obtained by matrix multiplication. In other words,
the priorities of the elements in the lowest level, with respect to the element in
the highest level of the hierarchy, were obtained by successively weighing the
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 83

priorities of the elements in each level with respect to the elements in the level
immediately above. In the initial Alice example, the priorities of the three
alternatives on the goal were obtained by weighing the priorities of the
alternatives on the objectives, and then the priorities of the objectives on the
goal. In terms of matrix multiplication, this successive weighing process is
represented by Sw. However, this approach is impossible to execute in a
system with feedback, since we no longer have a hierarchy extending from
one level down to the next. We now have a network, in which it is no longer
clear which node is lower and which is higher. Furthermore, there might be
more than one interaction between the elements in two nodes. The
supermatrix comes to our rescue. A supermatrix enables us to express and
compute two-way interactions between elements in different nodes, as well as
the relationship among elements in the same node.

Figure 6.2 illustrates the network for a system with feedback. As in the
network for the Alice example, each arrow represents the influence that the
elements in the node at the starting point of the arrow has on the elements in
the node at the ending point of the arrow. In this network, the elements in the
node Objectives exhibit dependent relationship. The matrix multiplication
approach used in Chapter 4 has no procedure to take into account this
relationship while deriving the priorities of the alternatives. The supermatrix
approach, however, offers a solution to this problem. Let us take a look at the
corresponding supermatrix for this system with feedback:
84 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

Figure 6.2: Network for a System with Feedback

CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 85

Goal Objectives Alternatives

 
Goal  0 0 0
 

 
 
W0 = Objectives  W21 W22 0
.

 
 
 
Alternatives  0 W32 I

The construction of W0 follows the same rule that we used to construct W

in Equation 6.1: the component at the i,j position of W0 reflects the priorities of
the elements in the i node of the network, with respect to the elements in the j
node. To express the dependent relationship among elements in the i node, we
use the component Wii . A zero matrix in the ij component of W0 signifies that
there is no relationship between the elements in the i node and those in the j
node. We note that the 3,3 position of W is the identity matrix, instead of the
zero matrix, even though the alternatives are independent of one another. The
purpose is that the transformed form of W will be column stochastic. From
here on, any recurrence of the placement of I in the component in the last row
and the last column of the supermatrix will serve the same purpose.

This rule for construction of the supermatrix extends to the general case, as
given in the next definition.

Definition 6.3. Let N be a network with n nodes, denoted N1 , N2 , . . ., Nn . The

86 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

supermatrix W for the network N is:

N1 N2 ··· Nn
 
N1  W11 W12 ··· W1n 
 
 
N2  W21 W22 ··· W2n 
W = .  . .. .. .. ,

..  .. . . . 
 
 
N W Wn2 ··· W 

n n1 nn

where each Wi j is a matrix [10, p. 207].

We recall that the network for a hierarchy is of a linear form, in which

there are only one-way interactions, extending from the highest node down to
the lowest node. Thus, the supermatrix for a hierarchy always has the form

N1 N2 N3 ··· Nn−2 Nn−1 Nn

 
N1  0 0 0 ··· 0 0 0
 

 
 

N2  W21 0 0 ··· 0 0 0 

 
 
N 3
 0 W32 0 ··· 0 0 0 
W =
H
,
 
..  .. .. .. .. .. .. ..

.  . . . . . . .


 
 
 
Nn−1  0 0 0 ··· Wn−1,n−2 0 0 

 
 
N  0 0 0 ··· 0 Wn,n−1 I
 
n

where each WiHj is a matrix [10, p. 209].

As in the case with W, each component of W H is a matrix. We can obtain

the target priorities by looking at the last row of components in the stable form
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 87

of W H , as illustrated in the Alice example. We observe that W H is simply a

special case of W defined in Definition 6.3. More generally, our method of
using the hierarchy to find the desired priorities is a special case of the
supermatrix approach.

In this section, we have introduced the network, as well as the important

supermatrix approach, which allows us to solve problems in systems with
feedback. We also noted that the matrix multiplication method to obtain target
priorities is a special case of the supermatrix approach. Before ending this
section, we make an important observation that W might not be column
stochastic, even though each of its components is a column stochastic matrix.
An example is the matrix W0 introduced in this section. In order to apply the
method of raising W to powers to obtain the target priorities, we need to
transform W into a column stochastic matrix. This is achieved by deriving the
pairwise priorities of the appropriate nodes, and then weighing each
component in those nodes by the priorities. The execution of this idea will be
illustrated in the case study in medical diagnosis in the next section.

6.2 A Case Study in Medical Diagnosis

6.2.1 Preliminary Analysis of the Case Study

In this section, we apply the ideas developed in the first half of the chapter to
solve a problem in medical diagnosis. The case study in this section, as well as
the data on the pairwise comparison matrices and the supermatrix, is drawn
from an article by Saaty [13].
88 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

The medical case is described as follows.

Case Study: A woman in her second trimester of pregnancy was admitted

to a local hospital. The tests at the hospital revealed the following seven
symptoms:

• Anemia (An),

• Low Platelets (LP),

• Abnormal Liver (ABL),

• Blood Clotting (BC),

• High Activated Partial ThromboPlastin Time (APTT-H),

• High AntiNuclear Antibody (ANA-H), and

• High AntiCardiolipin Antibody (ACA-H).

At this point the physicians considered four possible diseases that could
cause the symptoms:

• Lupus,

• Thrombotic Thrombocytopenic Purpura (TTP),

• Hemolysis, Elevated Liver function, and Low Platelets (HELLP), and

• AntiCardiolipin Antibody Syndrome (ACA Syn).

For more information on these medical terms, the reader can consult the
encyclopedic reference [3].
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 89

Figure 6.3: Network for the Case Study in Medical Diagnosis

Given the condition of the patient, the physicians need to make a decision
between two alternatives: to terminate the pregnancy, or to treat the patient
for her symptoms and let her proceed with the pregnancy. The network for
this case study is given in Figure 6.3.

As in the previous networks, the arrow reflects the influence of the

elements in one node on the elements in another, or the same, node. As we can
see in Figure 6.3, the diseases and the symptoms have two-way interactions,
which would be impossible to depict a linear hierarchic structure.
Furthermore, each symptom also has an influence on other symptoms. This
dependent relationship among symptoms is in fact the reason for the
desirability of the AHP in solving this problem. An alternative approach,
Bayes’ theorem, has been particularly popular in medical diagnostics. (See, for
90 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

example, [5], [15], and [9].) However, the use of Bayes’ theorem in diagnosing
diseases requires an assumption that is seldom satisfied in real life: the
symptoms are independent of one another [6].

Saaty suggests that the reason for this assumption is the lack of
information on the relationship of symptoms, as well as the daunting task of
conducting a sufficient number of experiments to obtain the necessary
statistical data to apply Bayes’ theorem [10, p. 492]. Using the AHP, on the
other hand, allows the incorporation of physicians’ judgment in order to take
into account the relationship among symptoms.

From Figure 6.3, the interactions among the elements in the nodes of the
network for the case study are described as follows.

(i) Each symptom observed in the patient has an influence on the possible
diseases. This influence is interpreted as the likelihood of the diseases,
given the symptom;

(ii) Each possible disease has an influence on the observed symptoms. This
influence means the extent to which the symptoms are characteristic of
the disease;

(iii) Each observed symptom has an influence on other observed symptoms.

This influence means the likelihood that a given symptom is associated
with or occurs jointly with other symptoms;

(iv) Each possible disease has an influence on the alternative treatments. This
influence is precisely the priority of the alternative treatments, given the
possible diseases.
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 91

We proceed to construct a supermatrix, whose components represent these

interactions among the elements in the nodes of the network. Let the nodes
Diseases, Symptoms and Alternative Treatments correspond to the first,
second, and third row (column) of W. Then using Definition 6.3, the
supermatrix for this case study is:

Diseases Symptoms Alternatives

 
Diseases 0 W12 0
 
 
 
 
W = Symptoms  W21 W22 0
.

 
 
 
Alternatives  W31 0 I

We would like to obtain the priorities of the alternative treatments from W.

Since W is not column stochastic, we cannot directly apply the method of
raising W to powers introduced in the last section. However, we can transform
W into a column stochastic matrix by the following method.

Consider the first column of components in W, which consists of the zero

matrix, W21 , and W31 . Let α1 be the weight of the node Symptoms with respect
to the node Diseases, and α2 the weight of Alternatives with respect to
Diseases, such that α1 and α2 are positive real numbers, and α1 + α2 = 1. We
obtain α1 and α2 , then multiply each entry in W21 by α1 , and each entry in W31
by α2 . Recall that each of W21 and W31 is a matrix of weight vectors and is
therefore column stochastic. By multiplying the entries in W21 and W31 by α1
92 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

 
 0 
 
 
and α2 , we will make the block W21  in W column stochastic.
 
 
 
W31 

To obtain α1 and α2 , we construct a pairwise comparison matrix M for

Diseases:

Diseases Symptoms Alternatives

 
Symptoms  1 m12
 

M= 

,


Alternatives  m21 1

where the ij entry in M is the weight of the i node with respect to Diseases,
compared to the weight of the j node with respect to Diseases. In particular,
the entry m21 reflects the importance of knowing about the alternatives,
compared to the importance of knowing about the symptoms, in diagnosing the
 T
disease(s). Then the eigenvector u = u1 u2 of M such that u1 + u2 = 1
gives the weights of the nodes Symptoms and Alternatives on the node
Diseases. By our construction of M, we have that u1 = α1 and u2 = α2 .

The reader must have realized that this process of finding the weight
vector for the nodes Symptoms and Alternatives, with respect to Diseases,
resembles the process of finding the priority vector that we discussed in
Chapter 2. Indeed, these two procedures are the same. In Chapter 2 and
Chapter 3, we discussed the method to find the weights of the objectives on
the goal, and of the alternatives on an objective. In light of our recent
discussion of network, what we did in those chapters was finding the weights
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 93

of a group of elements (in a node) on another element (in another node). We

now apply the same principles to find the weights of a group of nodes on
another node. This approach yields the desired priorities in this case as long as
we keep in mind that we are finding the weights of the nodes, not of the
elements in the nodes. As long as this condition is satisfied, the discussion in
Chapter 2 through Chapter 5 applies.

Let β1 and β2 be the weights of Diseases and Symptoms, respectively, with

respect to Symptoms. Using the same process outlined above, we can obtain
β1 ad β2 . By multiplying each entry in W12 by β1 , and each entry in W22 by β2 ,
 
W 
 12 
 
we will make the block W22  in W column stochastic. Thus, the transformed

 

 
 
 0 
form of W:

 
 0
 β1 W12 0
 
T = α1 W21 β2 W22 0
 

 
α2 W31
 
0 I

is column stochastic. Each of W12 , W21 , W22 , and W31 in T reflects the weights
of the elements in a node, with respect to the elements in another node. Each
of the coefficients α1 , α2 , β1 , and β2 reflects the weight of a node, with respect to
another node. Therefore, T represents all of the interactions of the factors in
this case study. Raising T to powers would give us the target priorities of the
alternatives.

Before proceeding to find T, we make two important observations:

94 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

First, in finding the weights of Symptoms and Alternatives with respect to

Diseases, we used the pairwise comparison matrix approach. Recall that in the
initial Alice example, the merit of using a pairwise comparison matrix lies in
the fact that this matrix allows utilization of the information on all the
objectives. Further, comparing the weight of each objective to the weights of
all other objectives is an error-prone task. However, when we have to find the
weights of only two nodes on another node, as in the supermatrix for this case
study, it is quicker to directly compare the importance of the two nodes,
instead of using the pairwise comparison matrix approach.

In particular, to find α1 and α2 , we ask the question: Is the diagnosis of a

disease more a direct result of knowing about the symptoms, or of knowing
about the alternatives, and how much more so? If the answer is that knowing
about the symptoms contributes 80% to the diagnosis of a disease, then
α1 = 0.8, and α2 = 0.2. To estimate β1 and β2 , we ask: Is knowledge of a
symptom and its usefulness in making the diagnosis more a direct result of
knowing about the diseases, or of knowing about the other symptoms, and
how much more so? If the answer is that knowing about the disease
contributes 40% to the knowledge of the symptom, then β1 = 0.4, and β2 = 0.6.
These two questions are answered by physicians, taking into account their
knowledge of the diseases, the symptoms and the medical history of the
patient.

Second, we note that the question asked in finding α1 and α2 does not
make much sense. Certainly, knowing about the alternatives, which can only
be "terminate" or "not terminate the pregnancy", does not have any
contribution whatsoever to the diagnosis of a disease. Thus, α2 must be 0,
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 95

Figure 6.4: Reduced Network for the Case Study in Medical Diagnosis

which makes α2 W31 the zero matrix. T is reduced to

Diseases Symptoms
 
Diseases  0 β1 W12
 

Q= 

,
 (6.3)
β2 W22

Symptoms  W21

which corresponds to the network in Figure 6.4.

By raising Q to powers, we will obtain the priorities of the possible

diseases, given the symptoms. However, the ultimate goal is to find the
priorities of the alternatives, given the diseases. We overcome this issue by
observing that the problem in this case study can be presented in the form of a
96 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

Figure 6.5: Linear Network for the Case Study in Medical Diagnosis

linear network as in Figure 6.5, provided that we have obtained the priorities of the
possible diseases.

The case study becomes a problem that can be solved in three steps:

1. Finding the likelihood of the diseases. In Figure 6.5, this is represented

as the influence of the element in the node Goal on the elements in the
node Diseases. We achieve this likelihood by applying the supermatrix
approach to the network in Figure 6.4. The supermatrix approach in turn
has the following five steps:

(a) Finding the priorities of the observed symptoms with respect to the
possible diseases. This is W21 in Q.
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 97

(b) Finding the priorities of the possible diseases with respect to the
observed symptoms. This is W12 in Q.

(c) Finding the priorities of the observed symptoms with respect to one
another, which is W22 in Q.

(d) Finding the weights β1 and β2 of Diseases on Symptoms and of

Symptoms on Symptoms, respectively.

(e) Putting W21 , W12 , W22 , β1 , and β2 into Q. Raising Q to powers to

obtain the priorities of the possible diseases, given the observed
symptoms.

2. Finding the priorities of the alternative treatments with respect to the

possible diseases. This is represented in Figure 6.5 as the influence of
the elements in the node Diseases on the elements in the node
Alternative Treatments. The priorities are obtained as principal
eigenvectors of the pairwise comparison matrices that compare the
alternatives for each disease.

3. Finding the priorities of the alternative treatments for the patient,

which is the influence of the element in the node Goal on the elements in
the node Alternative Treatments in Figure 6.5. We achieve this by the
matrix multiplication approach presented in Chapter 4. Specifically, we
right-multiply the matrix of priorities obtained in step 2 by the priority
matrix obtained in step 1.

The information on the relationship among diseases, symptoms, and

alternatives for the patient is obtained based on the physicians’ answers. The
98 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

rest of this chapter will be devoted to finding the priorities of the alternatives
using the presented approach.

6.2.2 Finding the Likelihood of the Diseases

In this subsection, we explain how Saaty found the likelihood of the possible
diseases. The method is to apply the supermatrix approach to the network in
Figure 6.4. Recall that the supermatrix associated with this network is Q given
in Equation 6.3. Each of the nonzero components in Q is a matrix of priorities,
comprised of principal eigenvectors of pairwise comparison matrices. We
illustrate the process to find these components below.

Finding the priorities of the observed symptoms with respect to the

possible diseases (W21 )

For each of the diseases, a pairwise comparison matrix is constructed. The

entry in the ij position of this matrix reflects the extent to which symptom i is
characteristic of the disease, compared to the extent to which symptom j is
characteristic of the disease, as judged by the physicians. In order to help
obtain the entries in the pairwise comparison matrix, the physicians answered
the following question:

For (the given) disease and for two symptoms, which symptom is more
characteristic of the disease, and how much more is it?

The judgments are provided verbally as equally, weakly, strongly, very

strongly, or absolutely (more characteristic of the disease). The judgments are
interpreted into numerical values according to Table 5.2, and these numerical
values are placed into the pairwise comparison matrix for the disease. The
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 99

method presented in Chapter 2 and Chapter 3 are used to obtain the principal
eigenvector of the pairwise comparison matrix, which gives the priorities of
symptoms with respect to the disease.

The process outlined above follows closely what we have described in the
first five chapters. In the rest of this chapter, we will encounter more
opportunities in which our knowledge of the pairwise comparison matrix and
the procedure to find the priority vector is utilized. In such situations, our
discussion in Chapter 2 through Chapter 5 applies.

For the disease Lupus, the pairwise comparison matrix A that provides
comparisons on the pairs of symptoms is

Lupus An LP BC APTT-H ANA-H ACA-H

 
1
An  1 5 4 4 2
 
9

 
 
 1 1 
LP  5
 1 1 1 9
1 

 
 1 1 1

BC 
 4 1 1 1 9 2

A= .


APTT-H  41 1 1
 
1 1 1 9 2


 
 
ANA-H  9 9 9 9 1 9
 
 
 

ACA-H  1 1 2 2 1
1
 
2 9

For example, the entry in the fifth row and the sixth column in A is 9, so
ANA-H is absolutely more characteristic of Lupus, compared to ACA-H. The
symptom ABL was not included in this matrix because ABL is not
characteristic of the disease Lupus. The priority of ABL with respect to Lupus
is therefore 0. The weight vector w, which is the principal eigenvector of A,
100 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

gives the priorities of all of the symptoms, except for ABL, with respect to
Lupus.

 T
w1 = 0.156 0.050 0.046 0.046 0.630 0.072 .

For the other diseases, the original article does not provide the pairwise
comparison matrices, but information about the corresponding eigenvectors is
available. Next, the priority vectors are placed into Table 6.1. The entries of w1
are placed in the first column of the table.

Table 6.1: Priorities of Symptoms with Respect to Diseases

Lupus TTP HELLP ACA Syn

An 0.156 0.133 0.313 0.053

LP 0.050 0.789 0.313 0.158
ABL 0.000 0.000 0.313 0.000
BC 0.046 0.026 0.000 0.263
APTT-H 0.046 0.026 0.061 0.263
ANA-H 0.630 0.000 0.000 0.000
ACA-H 0.072 0.026 0.000 0.263

In Table 6.1, the entries in the j column give the priorities of the symptoms,
with respect to the j disease. In terms of our case study, these entries represent
the relative probabilities that the symptoms are observed, given the j disease.
By relative probabilities, we mean, for example, the probability that the
symptom An, out of all of the other symptoms, is observed in the patient if she
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 101

has Lupus. According to the first column of the table, the probability that An
is observed in the patient if she has Lupus is 0.156. Thus, by looking at all of
the entries in a column, we can identify the symptom that is most prevalent for
a given disease. From the first column of the table, given that the disease is
Lupus, ANA-H is the most likely symptom to be exhibited, with a probability
of 0.63.

We note that Table 6.1 gives the entries in W21 . Since the table is obtained
from the eigenvectors of the pairwise comparison matrices, such that the
entries of each vector sum to 1, each column sum of the table is 1, and W21 is
column stochastic.

Before moving on to the next step, we note that the probabilities in Table
6.1 are not for a general, hypothetical pair of disease-symptom, but for the pair
of disease-symptom pertinent to the patient in this case study. This means that
the physicians whose judgments allowed the derivation of these probabilities
gave their answers based on their knowledge both of the disease-symptom
pair, and of the patient’s medical history and current condition. This subtle
difference also applies to the probabilities found in the next two steps.

Finding the priorities of the possible diseases with respect to the

observed symptoms (W12 )

For each of the symptoms, a pairwise comparison matrix is constructed.

The ij entry in this matrix represents the likelihood that the i disease exhibits
the given symptom, compared to the likelihood that the j disease exhibits the
symptom. By likelihood, we mean the chance that the disease causes the
symptom, as judged by the physicians. The following question was answered
by the physicians:
102 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

For (the given) symptom and two diseases, which disease is more likely to exhibit
this symptom, and how much more likely is it?

For the symptom An, the pairwise comparison matrix B that provides
paired comparisons of all the diseases, with respect to An, is

An Lupus TTP HELLP ACA Syn

 
 1 1 1 
Lupus 
 1 5 9 5


 
 
TTP  5 1 1 1 
B= .
 

 
HELLP  9 1 1 1 

 
 
ACA Syn  5 1 1 1
 

For example, the entry in the 2,1 position of B is 5. This means that TTP is
strongly more likely than Lupus to exhibit An as a symptom. The principal
eigenvector of B gives the priorities of all of the diseases, with respect to An:

 T
w2 = 0.052 0.299 0.350 0.299 .

The pairwise comparison matrices for the other symptoms are not
available in the original article. The priority vectors are available and were
placed into Table 6.2. The entries of w2 are placed in the first column of the
table.
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 103

Table 6.2: Priorities of Diseases with Respect to Symptoms

An LP ABL BC APTT-H ANA-H ACA-H

Lupus 0.052 0.231 0.000 0.279 0.222 0.706 0.119

TTP 0.299 0.461 0.000 0.070 0.056 0.088 0.030
HELLP 0.350 0.231 1.000 0.093 0.056 0.088 0.020
ACA Syn 0.299 0.077 0.000 0.558 0.666 0.118 0.831

In Table 6.2, the j column gives the priorities of the possible diseases with
respect to the j symptom. In other words, the entries in the j column
represents the relative probabilities that the possible diseases cause the j
symptom. Relative probabilities mean, for example, the probability that the
disease HELLP, among all of the other possible diseases, causes the symptom
LP. From the table, the probabilities that HELLP causes LP is 0.231, which
corresponds to the entry in the 3,2 position in the table. Therefore, by looking
at all of the entries in the j column, we can identify the disease that is most
likely to cause the symptom j. For example, by looking at the second column
in the table, we know that the disease that is most likely to cause LP is TTP,
with a probability of 0.461.

Table 6.2 gives the entries in W12 . Since each column in Table 6.2 is
obtained from a priority vector, the sum of the entries in each column is 1, and
W12 is column stochastic.

Finding the priorities of the observed symptoms with respect to one

another (W22 )

For each of the symptoms, a pairwise comparison matrix is constructed.

104 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

The entry in the ij position of this matrix reflects the likelihood that the i
symptom occurs at the same time as the given symptom, compared to the
likelihood that the j symptom occurs jointly with the given symptom.
Likelihood means the chance that a symptom occurs at the same time as the
given symptom, as judged by the physicians. The physicians answered the
following question in order to help derive the pairwise comparison matrix:

Given a symptom, e.g., ANA-H, and two other symptoms that may be related to
it, e.g., An and LP, which of the two latter symptoms is more likely to be associated
with, or occur jointly with, the given symptom? How much more likely is it?

The pairwise comparison matrix C that compare pairs of symptoms, with

respect to the symptom ANA-H is

ANA-H An LP BC ACA-H
 
 
An  1
 1 4 4 

 
 
LP  1 1 1 1 
C= .
 

 1 
BC 
 4 1 1 1 

 
 
ACA-H  14 1 1 1 

For example, the entry in the first row and the second column in C is 1,
which means that An is as equally likely as LP to occur jointly with ANA-H.
Note that ABL and APTT-H were omitted from C because these symptoms are
not related to ANA-H, as judged by the physicians. ANA-H is also skipped,
since the question asks about a symptom and two other symptoms that are
different from the original one. The priorities of ABL, APTT-H, and ANA-H,
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 105

with respect to ANA-H, are 0. The principal eigenvector w3 of C gives the

priorities of all of the other symptoms with respect to ANA-H:

 T
w3 = 0.455 0.235 0.155 0.155 .

The pairwise comparison matrices for the other symptoms are not
available from the original articles. The priority vectors are available and
placed into Table 6.3. The entries of w3 are placed in the sixth column of this
table.

Table 6.3: Priorities of Symptoms with Respect to Symptoms

An LP ABL BC APTT-H ANA-H ACA-H

An 1.000 0.000 0.250 0.000 0.000 0.455 0.095

LP 0.000 0.000 0.750 0.105 0.106 0.235 0.048
ABL 0.000 0.500 0.000 0.000 0.000 0.000 0.000
BC 0.000 0.000 0.000 0.000 0.429 0.155 0.381
APTT-H 0.000 0.000 0.000 0.421 0.000 0.000 0.381
ANA-H 0.000 0.500 0.000 0.053 0.036 0.000 0.095
ACA-H 0.000 0.000 0.000 0.421 0.429 0.155 0.000

The j column in Table 6.3 gives the priorities of the observed symptoms
with respect to the j symptom. These priorities mean the relative probabilities
that given symptom j, the other symptoms that are observed in the patient are
also present. In particular, the entry in the i row and the j column of the table
represents the relative probability that the i symptom occurs jointly with the j
106 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

symptom. Relative probability means the probability that symptom i, among

all of the other symptoms, occurs jointly with symptom j. For example, among
all of the symptoms, the probability that An occurs jointly with ABL is 0.25,
and the probability that LP occurs jointly with ABL is 0.75. As a special case,
the ii entry of the table, which technically represents the probability that
symptom i occurs jointly with itself, is 0 if at least one of the other entries in the
i column is nonzero. The ii entry is 1 if all of the other entries in the i column is
0. The purpose is that each of the column vectors in W22 , which are precisely
the columns from Table 6.3, sums to 1. As a result, W22 is column stochastic.

Finding the weights of Diseases and Symptoms on Symptoms

Recall that β1 is the weight of the node Diseases on the node Symptoms,
and β2 is the weight of the node Symptoms on the node Symptoms. To
estimate β1 and β2 , we ask: Is knowledge of a symptom and its usefulness in making
the diagnosis more a direct result of knowing about the diseases, or of knowing about
the other symptoms, and how much more so?

The original article solves the case study problem with β1 = β2 = 0.5, which
implies that knowledge of the diseases and knowledge of the observed
symptoms both contribute equally to knowledge of a symptom and its
usefulness in making the diagnosis. The article then solves the problem with
β1 = 0.99, and β2 = 0.01, which means that knowledge of the dependent
relationships among symptoms contributes only 1 percent to the final
diagnosis. The results in both cases vary by an insignificant amount. (The
priority for one of the alternative treatment decreases by 0.03 if β2 decreases
from 0.5 to 0.01.) Since our purpose is solely to illustrate an application of the
extensions of the AHP, we choose to pursue only the case when β1 = β2 = 0.5.
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 107

Obtaining the priorities of the possible diseases

The idea of this subsection has been to apply the supermatrix approach to
the network in Figure 6.4. The supermatrix associated with this network is Q
in Equation 6.3. In this step, the entries of W21 , W12 , W22 , β1 , and β2 are placed
into Q. Note that each of the entries in W12 and in W22 is multiplied by
β1 = β2 = 0.5. We have the following matrix:
 108

Lupus TTP HELLP ACA Syn An LP ABL BC APTT-H ANA-H ACA-H

 
 
Lupus 
 0 0 0 0 0.025 0.115 0 0.140 0.111 0.353 0.059 

 
 
TTP 
 0 0 0 0 0.150 0.231 0 0.035 0.028 0.044 0.015 

 
 
HELLP  0 0 0 0 0.175 0.115 0.500 0.047 0.028 0.044 0.010 

 

 
ACA Syn  0 0 0 0 0.150 0.038 0 0.279 0.333 0.059 0.416 

 
 

An  0.156 0.133 0.313 0.053 0.500 0 0.125 0 0 0.227 0.048 

 
 
Q = LP 

 0.050 0.789 0.313 0.158 0 0 0.375 0.053 0.053 0.117 0.024 .

 
 
ABL 
 0 0 0.313 0 0 0.250 0 0 0 0 0 

 
 
BC  0.046 0.026 0 0.263 0 0 0 0 0.214 0.078 0.190 
 
 
 

APTT-H  0.046 0.026 0.061 0.263 0 0 0 0.211 0 0 0.190 
 
 
 
ANA-H  0.630 0 0 0 0 0.250 0 0.026 0.018 0 0.048 

 
 
ACA-H  0.072 0.026 0 0.263 0 0 0 0.211 0.214 0.078 0
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 109

When Q is raised to powers, the entries in Q reach a stable stage, in which

Q has identical columns. A column w of Q is:

 T
w= 0.073 0.068 0.087 0.106 0.169 0.144 0.063 0.067 0.066 0.088 0.070 .

Since w is positive, the stable form of Q is positive, and we know that Q is

primitive. The first four entries in w give the priorities of the diseases, with
respect to the symptoms. The last seven entries in w give the priorities of the
symptoms, with respect to the diseases. In order to obtain the relative
likelihood of the diseases, the first four entries are normalized. The
normalized form of these four entries is given in w0

 T
w0 = 0.218 0.203 0.262 0.317 .

The first, second, third, and fourth entry in w0 represents the relative
likelihood of Lupus, TTP, HELLP, and ACA Syn, respectively. This is the target
result for the first step in the three-step solution to the case study. The second
step is explained in the next subsection.

6.2.3 Finding the Priorities of the Alternatives with Respect to

the Diseases

In this subsection, we illustrate the process to find the priorities of the

alternative treatments with respect to the possible diseases. Recall that these
110 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

priorities correspond to the influence of the elements in the node Diseases on

the elements in the node Alternative Treatments in Figure 6.5. The method
used to find these priorities is the same as the method that allowed us to find
the priorities (or scores) of the alternatives on the objectives in Chapter 4.
In general, i.e. regardless of the number of alternatives, for each disease, a
pairwise comparison matrix is constructed. This matrix provides the paired
comparisons of the alternatives. The principal eigenvector of this matrix gives
the priorities of the alternatives, with respect to the disease.
There are two alternative treatments for the patient in this case study: to
terminate the pregnancy (denoted alternative T), and to proceed with the
pregnancy (denoted alternative NT). Observe that the goal in this step is to
find the weights of these two alternatives on each disease. As in the case of
finding β1 and β2 , it is quicker to directly compare the importance of the two
alternatives, instead of using the pairwise comparison matrix approach.
Specifically, the physicians assess the priorities of the two alternative
treatments with respect to each of the diseases. Those priorities are given in
Table 6.4.
Table 6.4: Priorities of Alternatives with Respect to Diseases

Lupus TTP HELLP ACA Syn

T 0.200 0.800 0.800 0.833

NT 0.800 0.200 0.200 0.167

The j column of Table 6.4 gives the priorities of the alternatives with
respect to the j disease. These priorities are interpreted as the extent to which
the alternative is appropriate for the given disease. For example, from the first
CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS 111

column of the table, the appropriate treatment for Lupus is to not terminate
the pregnancy (NT), with priority 0.8.

We note that the entries in Table 6.4 give the entries in W31 . Even though
W31 was not entered into the supermatrix Q, it is the priority matrix desired in
step 2 of the solution to the case study. In the next subsection, we illustrate the
final step of the solution.

6.2.4 Finding the Priorities of the Alternatives

This is the last step in the solution to this case study. It uses the results from
the first two steps. The method is to apply the matrix multiplication approach
presented in Chapter 4 to the linear hierarchy in Figure 6.5.

Recall that to obtain the priorities of the elements in the lowest level of a
hierarchy, with respect to the element in the highest level, we successively
weigh the priorities of the elements in each level with respect to the elements
in the level immediately above. In the initial Alice example in Chapter 4, we
weighed the priorities of the alternatives on the objectives, and then the
priorities of the objectives on the goal. This successive weighing process is
represented by right-multiplying the score matrix S by the weight vector w.
The result was the priorities of the alternatives with respect to the goal.

Using the same approach, we weigh the priorities of the alternative

treatments with respect to the diseases, and then the priorities of the diseases
with respect to the goal. The corresponding matrices are W31 (found in step 2),
and w0 (found in step 1). In terms of matrix multiplication, we right-multiply
W31 by w0 . This multiplication is shown below:
112 CHAPTER 6. AN APPLICATION IN MEDICAL DIAGNOSIS

 
0.218
 
     
0.200 0.800 0.800 0.833 0.203 0.68
W31 w0 =  =  .
     
  
 
0.800 0.200 0.200 0.167 0.262 0.32
 
 
0.317

The priority of T is 0.68, and the priority of NT is 0.32. Thus, the result of
the AHP recommends that the pregnancy be terminated. Saaty noted that this
recommendation was in agreement with the decision made by a doctor of the
patient [13, p. 500]. Even though this fact does not guarantee the infallibility of
the AHP, it gives some insight into its validity as a mathematical model for
solving decision making problems.
Chapter 7

Conclusion

The case study in the last chapter is in fact an illustration of the Analytic
Network Process (ANP). The ANP is a generalization of the AHP to the case
in which there exists dependence and feedback among factors in decision
making problems. In particular, the supermatrix approach is the
generalization of the hierarchy approach. Regardless, both the AHP and the
ANP rely on these three steps: Decomposition, Measurement of preferences,
and Synthesis [p. 492][10].

Decomposition is the process of breaking the problems into elements,

grouping these elements into levels, and representing those levels in such a
way that it reflects various factors in the decision making problems. The result
of this process is either a hierarchy (in which case we have the AHP), or a
network (in which case we have the ANP). We briefly discussed the
construction of a hierarchy in Chapter 4. Saaty’s suggestions for building a
hierarchy more or less stem from the social sciences [10, p. 14–16]. The
construction of a network, however, is based on graph theory, which is

113
114 CHAPTER 7. CONCLUSION

explained in [10, p. 200–204].

In the second step, Measurement of preferences, pairwise comparisons are

made about elements, which allows the derivation of pairwise comparison
matrices. For the AHP, each of the principal eigenvectors of these matrices
gives the priorities of the elements in a lower level with respect to an element
in a higher level. By putting these eigenvectors into a matrix, we have a
priority matrix, which represents the priorities of the elements in a lower level
with respect to the elements in a higher level. For the ANP, each principal
eigenvector gives the priorities of the elements in a node with respect to an
element in a different or the same node. Each of the priority matrices
represents the priorities of the elements in a node with respect to the elements
in a different or the same node.

The last step, Synthesis, occurs after we have obtained the priority
matrices for all of the interactions in the hierarchy (or network). The
supermatrix approach is the synthesizing step for the ANP, while for the AHP,
we successively weigh the priorities of the levels, from the top to the bottom of
the hierarchy.

Recall that the AHP yields meaningful results only when the number of
objectives is less than 10. Thus, the AHP is not particularly advantageous
when the decision making problem involves a very large number of objectives.
Moreover, the priority vector provides a meaningful result only when the
decision maker is not too inconsistent. Specifically, we found in Chapter 5 that
the consistency ratio needs to be 0.10 or less in order for the result of the AHP
to be acceptable.

The strength of the AHP, as well as the ANP, lies precisely in the three
CHAPTER 7. CONCLUSION 115

steps that we outlined above. By decomposing and grouping the elements of

the decision making problems, the decision makers gain a better
understanding of the problems and their preferences. The hierarchy and the
network allow the decision makers to look at the problems at hand in an
analytical manner. Finally, the synthesis of the priority matrices offers a
systematic approach to arrive at the best solution for the decision makers.

Due to time constraints, this thesis has not investigated in depth the
theoretical justification of the ANP. Researchers who wish to pursue further
work on the AHP might find it worthwhile to further explore the theory
behind the ANP. So far, two issues have emerged as worthy of future research.

The first issue is the explanation of the stable stage of the supermatrix. An
interesting fact, which is perhaps also useful for future research, is that the
method of the ANP parallels that of Markov chains, as Saaty himself notes in
[10, p. 206]. Specifically, the nodes in a network correspond to the states in
Markov chains, and the influence of the i node on the j node at time k
corresponds to the transition from state i to state j at time k. Further, the
concepts "priority" in the ANP and "probability" in Markov Chains coincide.
For a more comprehensive list of the correspondence of the terminologies
between the two systems, the reader is referred to the cited source.

Second, recall that the case study was solved by replacing the original
network (Figure 6.3) with a linear hierarchic structure (Figure 6.5), through the
reduction of the original network to a simpler one (Figure 6.4). The author of
this thesis based this approach on the discussion of the AHP in the thesis. This
can be viewed as an alternative explanation to the approach used by Saaty,
which is the same as the three-step solution presented in the last chapter.
116 CHAPTER 7. CONCLUSION

However, Saaty attributes the raising of the reduced supermatrix to powers to

the claim that the components in this supermatrix are the essential
components of the network. These components are called the sources or
impact-priority-diffusing components. By raising the reduced supermatrix Q
to powers, we obtain w, which is the limiting impact priorities of the
network. Further, by right-multiplying W31 by w0 , we obtain the limiting
absolute priorities of the network. For discussion of these topics by Saaty, the
reader is referred to [10, p.213] and [13, p. 498–500].
The AHP, as well its generalization, the ANP, is at heart a mathematical
model for solving decision making problems. The mathematics in this thesis,
especially in Chapter 2, Chapter 3, and Chapter 5 provide the theories
underpinning the method of these models. The case study in the last chapter
gives only a glimpse of the application of the ANP. In fact, these models have a
vast array of applications in diverse fields, including national security (e.g., an
analysis of terrorism for the Arms Control and Disarmament Agency [7]),
international peace (e.g., a study of conflict in Northern Ireland [4,
p. 225–241]), business (e.g., applications of the models in a consulting
environment [4, p. 192–212]), and development (selection of research projects
about surface water resources for South Africa [4, p. 122–137]). This list by no
means exhausts the possibilities of applications of the models. It will also be
beneficial for future research projects to devise creative applications of the
models to solve real-world problems.
Bibliography

[1] Stephen Abbott. Understanding analysis. Undergraduate Texts in

Mathematics. Springer-Verlag, New York, 2001.

[2] Andrew R. Blair, Robert Nachtmann, and Thomas L. Saaty. Incorporating

expert judgment in economic forecasts: The case of the us economy in
1992. Fundamentals of the Analytic Hierarchy Process, 4:175–192, 1993.

[3] Robert W. Colman. Hemostasis and thrombosis: basic principles and clinical
practice. Lippincott Williams & Wilkins, Philadelphia, PA, 2006.

[4] Bruce L. Golden, Edward A. Wasil, and Patrick T. (Eds.) Harker. The
Analytic hierarchy process: applications and studies. Springer-Verlag, Berlin,
1989.

[5] Glen W. Hamilton, Gene B. Trobaugh, James L. Ritchie, K. Lance Gould,

Timothy A. DeRouen, and David L. Williams. Myocardial imaging with
201thallium: An analysis of clinical usefulness based on bayes’ theorem.
Seminars in Nuclear Medicine, 8(4):358 – 364, 1978.

117
118 BIBLIOGRAPHY

[6] N.E.G. Jonson. Everyday diagnostics—a critique of the bayesian model.

Medical Hypotheses, 34(4):289–295, 1991.

[7] Robert H. Kupperman and Darrell M. Trent. Terrorism: threat, reality,

response. Publication Series. Hoover Institution Press, Stanford University,
1979.

[8] David Poole. Linear algebra: a modern introduction. Available 2011 Titles
Enhanced Web Assign Series. Cengage Learning, Boston, MA, 2010.

[9] Martin J. Prince. Predicting the onset of alzheimer’s disease using bayes’
theorem. American Journal of Epidemiology, 143(3):301–308, 1996.

[10] Thomas L. Saaty. The analytic hierarchy process: planing, priority setting,
resource allocation. Analytic Hierarchy Process Series. RWS Publ.,
Pittsburgh, PA, 1990.

[11] Thomas L Saaty. The analytic network process: decision making with
dependence and feedback. RWS Publ., Pittsburgh, PA, 2001.

[12] Thomas L. Saaty and David S. Turner. Prediction of the 1996 super bowl:
An application of the ahp with feedback (the supermatrix approach). In
Fourth International Symposium on The Analytic Hierarchy Process, pages
12–15, 1996.

[13] Thomas L. Saaty and Luis G. Vargas. Diagnosis with dependent

symptoms: Bayes theorem and the analytic hierarchy process. Operations
Research, 46(4):491–502, 1998.
BIBLIOGRAPHY 119

[14] Allan J. Sieradski. An introduction to topology and homotopy. The Prindle,

Weber & Schmidt Series in Advanced Mathematics. PWS-KENT
Publishing Co., Boston, MA, 1992.

[15] Steven J. Skates, Feng-Ji Xu, Yin-Hua Yu, Kerstin Sjövall, Nina Einhorn,
Yuchiao Chang, Robert C. Bast, and Robert C. Knapp. Toward an optimal
algorithm for ovarian cancer screening with longitudinal tumor markers.
Cancer, 76(S10):2004–2010, 1995.

[16] Shlomo Sternberg. Dynamical systems. Dover Publications Inc., Mineola,

NY, 2010.

[17] Wayne L. Winston. Operations research: applications and algorithms.

Thomson/Brooks/Cole, 2004.