Appendix

The Theory
Behind
AHP/ANP

1. Hierarchies, Paired Comparisons, Eigenvectors and

Consistency
The Analytic Hierarchy Process (AHP) is a theory of relative measurement
with absolute scales applied to both tangible and intangible criteria based on
the judgment of knowledgeable and expert people. The main concern of the
mathematics behind the AHP deals with the measurement of intangibles. In the
end we must fit our entire world experience into our system of priorities if we
want to understand it. The AHP reduces a multidimensional scaling problem
into a one-dimensional one. Decisions are determined by a vector of priorities
that gives an ordering of the different possible outcomes. The best outcome is
that with the highest priority. For a group the most basic way to combine
judgments for individuals is to multiply them together and take the geometric
mean, after suitable debate and discussion. If it is not possible to have the
group together for discussion the judgments can be solicited by questionnaire
then combined. A different approach is to find the final overall priority vector
obtained from a set of judgments for each individual, then combine them by
taking the geometric mean of the corresponding components of the priority
vectors, or if the judges have priorities, combine them by raising each
component to the appropriate power and multiplying to produce the final
answer. This latter process is generally used when the judges are experts who
do not wish to be influenced by others.

Paired Comparisons and the Fundamental Scale

To make tradeoffs among the many objectives and criteria, the judgments that
are usually made in qualitative terms are expressed numerically. To do this,
rather than simply assigning a score out of a persons memory that appears
reasonable (an ordinal number that can be justified), one uses a paired
comparison. In paired comparisons, the smaller or lesser element is used as the
unit and the larger or greater element is estimated as a multiple of that unit
with respect to the common property or criterion for which the comparisons
are made. In this sense, measurement with judgments can be more scientific
than assigning numbers that are more or less arbitrary. Because human beings

The Encyclicon: Volume 2

are limited to their intermediate size range and the firings of their neurons are
limited in intensity, it is clear that there is a limit on their ability to compare the
very small with the very large. Pairwise comparisons should be made on
elements or alternatives that are similar in size or homogeneous, that is, no
element in a set of elements to be pairwise compared is more than 9 times
another. The more separated they become, the more the need to put them in
different groups and link these groups with a common element from one group
to an adjacent group of slightly greater or slightly smaller elements.
From all the paired comparisons, one derives a scale of relative values as
priorities. Because of inconsistency among the judgments, it is mathematically
necessary to derive the priorities in the form of the principal right eigenvector
of the matrix of paired comparisons. The priorities must be invariant, which
means if the judgments are weighted by the priorities of the corresponding
elements then added one should get back these priorities. It has been proven
that only the principal eigenvector satisfies this invariance requirement. The
eigenvector (to which corresponds a left principal eigenvector that is its
reciprocal, particularly when the matrix is consistent) captures the dominance
of one element over another along paths of different lengths. Priorities from
path dominance at any power of the matrix are obtained by summing each row
and normalizing the resulting vector. The left eigenvector is obtained from the
normalized column sums and captures priorities of being dominated. The
upshot of all this is that we want to compute the vector of positive entries from
the limiting powers of the matrix to get the eigenvector. In 1907 Oskar Perron
proved that this vector is the principal eigenvector and it corresponds to the
largest eigenvalue of the matrix that is real, positive, and a simple root of its
characteristic equation and it dominates all other eigenvalues in modulus.
We learn from making paired comparisons in the AHP that if A is 5 times
larger than B and B is 3 times larger than C, then A is 15 times larger than C
and A dominates C 15 times. That is different from A having 5 dollars more
than B and B having 3 dollars more than C, which implies that A has 8 dollars
more than C. Defining intensity of dominance along the arcs of a graph and
raising the matrix to powers measures the first kind of dominance precisely
and never the second. It has definite meaning, and, as we shall see below, in
the limit it is measured uniquely by the principal eigenvector. The use of ratios
used to represent dominance leads to this idea of dominance and can be
verified by taking a matrix whose entries are written as ratios, raising it to
powers and checking that the resulting coefficients give the desired dominance
from any point to any other. There is a useful connection between what we do
with dominance priorities in the AHP and what is done with transition
probabilities both of which use matrix algebra to find their answers.
Transitions between states are multiplied and added. To compose the priorities
for the alternatives of a decision with respect to different criteria, it is also

The Theory Behind AHP/ANP

necessary that the priorities of the alternatives with respect to each criterion be
multiplied by the priority of that criterion and then added over all the criteria.
The Fundamental Scale used for the judgments is given in Table 1. Judgments
are first given verbally as indicated in the scale and then a corresponding
number is associated with that judgment. The vector of priorities is the
principal eigenvector of the matrix. This vector gives the relative priority of
the criteria measured on a ratio scale. That is, these priorities are unique to
within multiplication by a positive constant. However, if one ensures that they
sum to one they are then unique and belong to a scale of absolute numbers.
Table 1: Fundamental Scale

1
Equal importance
3
Moderate importance of one over another
5
Strong or essential importance
7
Very strong or demonstrated importance
9
Extreme importance
2,4,6,8 Intermediate values
1.1-1.9 For comparing things that are very close
Use reciprocals for inverse comparisons
Use actual measurements from ratio scales to form
ratios or estimate this ratio by an absolute number
In the judgment matrix A, instead of assigning two numbers wi and wj and
forming the ratio wi/wj, we assign a single number drawn from a fundamental
scale of absolute (thus dimensionless) numbers to represent the ratio (wi/wj)/1.
It is a nearest integer approximation to the ratio wi/wj. The derived scale will
reveal the values of wi and wj. This is a central fact about the relative
measurement approach. It needs a fundamental scale to express numerically
the relative dominance relationship. The general eigenvalue formulation is
obtained by perturbation of the following consistent formulation:
A1 K
An
w1 K w1

w n
A 1 w1
M M
M
Aw =

w n K w n
w n
A n w1

w 1
M= n

wn

w1
M= nw.

w n

where A has been multiplied on the right by the transpose of the vector of
weights w ( w1 ,..., wn ) . The result of this multiplication is nw . Thus, to
recover the scale from the matrix of ratios, one must solve the problem
Aw nw or ( A nI ) w 0 . This is a system of homogeneous linear equations. It
has a nontrivial solution if and only if the determinant of A nI vanishes, that

The Encyclicon: Volume 2

is, n is an eigenvalue of A . Now A has unit rank since every row is a constant
multiple of the first row. Thus all its eigenvalues except one are zero. The sum
of the eigenvalues of a matrix is equal to its trace, that is, the sum of its
diagonal elements. In this case, the trace of A is equal to n . Thus n is an
eigenvalue of A , and one obtains a nontrivial solution. The solution consists of
positive entries and is unique to within a multiplicative constant.
Associated with the weights is an inconsistency. The consistency index of a
matrix is given by C.I . (max n) /( n 1) . The consistency ratio C.R. is
obtained by forming the ratio of C.I. and the R.I. that is associated with the
order of the matrix A from Table 2. The R.I. numbers are the average random
consistency indices computed for very large samples for matrices up to the
order of 10. A large number of reciprocal matrices are randomly generated
using the fundamental scale numbers 1/9, 1/8,,1/2, 1, 2,, 8, 9 and their
principal eigenvalues are averaged to form the Random Consistency Index R.I.
Table 2: Random Consistency Indices
Order
of
Matrix

10

R.I.

0.52

0.89

1.11

1.25

1.35

1.40

1.45

1.49

It is recommended that C.R. should be less than or equal to 0.05 when 3

elements are compared, .08 when 4 elements are compared and 0.10 when
more than 4 elements are compared. Inconsistency may be thought of as an
adjustment needed to improve the consistency of the comparisons. But the
adjustment should not be as large as the judgment itself, nor so small that it
would have no consequence. Thus inconsistency should be just one order of
magnitude smaller. On a scale from zero to one, the overall inconsistency
should be around 10 %. The requirement of 10% cannot be made smaller such
as 1% or .1% without trivializing the impact of inconsistency. But
inconsistency itself is important because without it, new knowledge that
changes preference cannot be admitted.

Validation Example and Homogeneity

Table 3 shows how an audience of about 30 people, using consensus to arrive
at each judgment, provided judgments to estimate the dominance of the
consumption of drinks in the United States (which drink is consumed more in
the US and how much more than another drink?). The derived vector of
relative consumption and the actual vector, obtained by normalizing the
consumption given in official statistical data sources, are at the bottom of the
table.

The Theory Behind AHP/ANP

Homogeneity, an important concept to ensure consistency in the paired

comparisons, requires the elements to be of the same order of magnitude which
means that our perceptions in comparing them should be of nearly the same
order of magnitude. It is a fact that people are unable to directly compare
widely disparate objects such as a ping-pong ball and a basketball according to
volume. To do that, we need a range greater than the 1-9 scale. To resolve this
problem, we can use a method in which we cluster different elements so we
can rate them within a cluster and then rate them across the clusters. We need
to add other objects to make the comparison possible and then form groups of
comparable elements. A common element, the pivot, could be the largest in one
cluster and the smallest in the next cluster of the next higher order of
magnitude. The weights of the elements in the second group are divided by the
priority of the pivot in that group and then multiplied by the priority of the
same pivot element from the first group, making them comparable with the
first group. The process is then continued.
Table 3: Relative Consumption of Drinks

Which Drink is Consumed More in the U.S.?

An Example of Estimation Using Judgments

Drink
Consumption
Coffee Wine
in the U.S.

Tea

Beer

Sodas

Milk

Water

Coffee

1/2

Wine

1/9

1/3

1/9

1/9

1/9

1/9

Tea

1/5

1/3

1/4

1/3

1/9

Beer

1/2

1/2

1/3

Sodas

1/2

Milk

1/2

1/3

Water

The derived scale based on the judgments in the matrix is:

Coffee Wine Tea
Beer Sodas Milk Water
.177
.019
.042
.116
.190
.129
.327
with a consistency ratio of .022.
The actual consumption (from statistical sources) is:
.180
.010
.040
.120
.180
.140
.330

Since a ping-pong ball and a basketball are too far apart to link them through
comparisons, we need to add other balls that are each gradually larger and
larger than the ping-pong ball but smaller than the basketball. We add a golf
ball, a tennis ball, a baseball, a handball, volleyball and a soccer ball. We first
compare the ping-pong ball with the golf ball and the tennis ball on the 1-9
scale. The tennis ball is then used again in a second cluster with the baseball on
the 1-9 scale in a separate homogeneous group. We then divide the relative
weight of the baseball by the relative weight of the tennis ball and multiply by

The Encyclicon: Volume 2

the tennis balls relative weight in the first cluster. Then we use the baseball as
the pivot ball with the handball in the third cluster and so on. Finally, we obtain
the result that 284.13 ping-pong balls are equal to the basketball as shown in
Table 3. Table 4 gives the actual relative volumes of the balls. We have not
considered the amount of space between the balls as they are packed inside
another ball (generally near 74% efficiency).
Table 4: Estimated Volumes of Eight Different Balls
First
group

Ping pong ball

0.142

Golf ball

0.179

Tennis ball

0.678

Inconsistency
ratio: 0.05

Second
group

Tennis ball 0.2

0.2/0.2=1
0.687*1=0.687

Inconsistency
ratio: 0.00

Baseball 0.8
0.8/0.2=4
0.687*4=2.748

Table 4 (contd): Estimated Volumes of Eight Different Balls

Third
group
Inconsistency
ratio: 0.00

Fourth
group

Baseball 0.143
0.143/0.143=1
2.748*1=2.748

Handball 0.857
0.857/0.143=5.993
2.748*5.993=
16.468

Handball 0.15
0.15/0.15=1
16.468*1=
16.468

Volleyball 0.20
0.20/0.15=1.33
16.468*1.33=
21.902

Volleyball

Soccer ball 0.282

0.282/0.15=1.88
16.468*1.88=
30.959

Soccer ball

40.346 / 0.142 = 284.13 ping pong balls are equal to the basketball ball.

Basketball 0.368
0.368/0.15= 2.45
16.468*2.45=40.346
Inconsistency ratio:
0.03
Basketball

The Theory Behind AHP/ANP

Table 5: Actual Volume Weights of the Eight Different Balls in the Four Clusters
First Cluster

Second Cluster

Third Cluster

Fourth Cluster

Balls

Actual

Balls

Actual

Balls

Actual

Balls

Ping-pong
Golf
Tennis

0.135
0.183
0.680

Tennis
Baseball

0.253
0.746

Baseball
Handball

0.108
0.891

Handball
Volleyball
Soccer ball
Basketball

Actual

0.162
0.217
0.265
0.354
Total
1
1
1
1
3
Actual standard volumes of the balls (cm ): Ping pong ball: 28.72, Golf: 38.77, Tennis:
143.72, Baseball: 423.67, Handball: 3471.73, Volleyball: 4642.26, Soccer ball: 5674.71,
Basketball: 7567.87

2. Networks, Dependence and Feedback

In Figure 1, we exhibit a hierarchy and a network. A hierarchy is comprised
of a goal, levels of elements and connections between the elements. These
connections are oriented only to elements in lower levels. A network has
clusters of elements, with the elements in one cluster being connected to
elements in another cluster (outer dependence) or the same cluster (inner
dependence). A hierarchy is a special case of a network with connections
going only in one direction. The view of a hierarchy such as that shown in
Figure 1 the levels correspond to clusters in a network.

Linear Hierarchy
Goal
Criteria

component,
cluster
(Level)

Subcriteria
element
Alternatives
A loop indicates that
each element depends
only on itself.

The Encyclicon: Volume 2

Feedback Network with Components having
Inner and Outer Dependence among Arc from component
Their Elements
C4 to C2 indicates the

C4

C1
Feedback

outer dependence of the

elements in C2 on the
elements in C4 with
Respect to a common
property.

C2

C3
Loop in a component indicates inner dependence of
the elements in that component with respect to a
common property.

Figure 1: How a Hierarchy Compares to a Network

There are two kinds of influence: outer and inner. In the first, one compares the
influence of elements in a cluster on elements in another cluster with respect to
a control criterion. With inner influence, one compares the influence of
elements in a group on each one. For example, if one takes a family of father
mother and child, and then takes them one at a time, with say the child first,
one asks who contributes more to the child's survival, its father or its mother,
itself or its father, itself or its mother. In this case the child is not so important
in contributing to its survival as its parents are. But if we take the mother and
ask the same question on who contributes to her survival more, herself or her
husband, herself would be higher, or herself and the child, again herself.
Another example of inner dependence is making electricity. To make
electricity steel is needed to make turbines, and fuel is also needed. So we have
the electric industry, the steel industry and the fuel industry. What does the
electric industry depend on more to make electricity, itself or the steel
industry? Does the electric industry depend more on the steel industry or on the
fuel industry; on itself or on the fuel industry? The electric industry does not
need its own electricity to make electricity. It needs fuel. Its electricity is only
used to light the rooms, which it may not even need. If we think about it
carefully, everything can be seen to influence everything including itself
according to many criteria. The world is far more interdependent than we
know how to deal with using our existing ways of thinking and acting. The
ANP is our logical way to deal with such dependence.
The priorities derived from pairwise comparison matrices are entered as parts
of the columns of a supermatrix. The supermatrix represents the influence
priority of an element on the left of the matrix on an element at the top of the
matrix with respect to a particular control criterion. A supermatrix along with
an example of one of its general entry matrices is shown in Figure 2. The
component C1 in the supermatrix includes all the priority vectors derived for
nodes that are parent nodes in the C1 cluster. Figure 3 gives the supermatrix

The Theory Behind AHP/ANP

of a hierarchy and Figure 4 shows the kth power of that supermatrix which is
the same as hierarchic composition in the (k,1) position.
The Supermatrix of a Network
C1
e11
e12

C1

e1n

e211
e22

W = C2

e2n

CN

C2

e1n

e21e22

CN

e2n

eN1eN2

eNn

W11

W12

W1N

W21

W22

W2N

WN1

WN2

WNN

eN1
eN2
e

e11e12

NuN

W ij Component of Supermatrix
(j1)

Wi1
Wij =

(j1)

Wi2

(j )

Wini1

(j2)

Wi1

(j2)

Wi2

(j )

Win2i

(jnj)

Wi1

(jnj)

Wi2

(jn )

Wini j

Figure 2: The Supermatrix of a Network and Detail of a Component in it

10

The Encyclicon: Volume 2

Supermatrix of a Hierarchy
C1

C2

e11
e11

C1

e1n

e21
e2n

e(N-2)1

e(N-2) n
N-2
e(N-1)1

CN
eN1
e(N-1) n

eNnN

N-1

W21 0

0 W32

0
0

eN1

CN

e2n2

CN-1

C2

W=

e1n1e21

CN-2

eNn

0
0

Wn-1, n-2 0
0
0 Wn, n-1 I

Figure 3 The Supermatrix of a Hierarchy

Wk

0
0
0
0
M
M
0
0
Wn ,n 1Wn 1,n 2 K W32W21 Wn ,n 1Wn 1,n 2 K W32

K
0
0
0
K
0
0
0

M
M
M
M
M M

K
0
0
0
K Wn ,n 1Wn 1,n 2 Wn, n1 I

Figure 4: The Limit Supermatrix of a Hierarchy (corresponds to Hierarchical

Composition)

The (n,1) entry of the limit supermatrix of a hierarchy as shown in Figure 4

above gives the hierarchic composition principle.
In the ANP we look for steady state priorities from a limit super matrix. To
obtain the limit we must raise the matrix to powers. Each power of the matrix
captures all transitivities of an order that is equal to that power. The limit of
these powers, according to Cesaro Summability, is equal to the limit of the sum
of all the powers of the matrix. All order transitivities are captured by this
series of powers of the matrix. The outcome of the ANP is nonlinear and
somewhat complex. The limit may not converge unless the matrix is column
stochastic that is each of its columns sums to one. If the columns sum to one
then from the fact that the principal eigenvalue of a matrix lies between its
largest and smallest column sums, we know that the principal eigenvalue of a
stochastic matrix is equal to one. To show that we have for a row stochastic
matrix (each of its rows sums to one):

11

The Theory Behind AHP/ANP

max

wj

j 1

j 1

wi

aij aij
n

min

a a
j 1

ij

ij

j 1

wj
wi

max for max wi

max for min wi

Thus for a row stochastic matrix we have

n

aij max max

1 min

j 1

a
j 1

ij

The same kind of argument applies to a matrix that is column stochastic.

Now we know, for example, from a theorem due to J.J. Sylvester that when the
eigenvalues of a matrix W are distinct that an entire function f(x) (power series
expansion of a function f(x) of a single variable x converges for all finite
values of x) with x replaced by W, is given by

( I A)
f (W ) f ( )Z ( ),Z ( )
, Z ( ) I , Z ( ) Z ( ) 0, Z
( )
n

i 1

j i

j i

i 1

( i ) Z ( i )

where I and 0 are the identity and the null matrices respectively.
A similar expression is also available when some or all of the eigenvalues have
multiplicities. We can see that if, as we need in our case, f (W ) W k , then
f (i ) ik and as k the only terms that give a finite nonzero values are
those for which the modulus of i is equal to one. The fact that W is stochastic
ensures this because its largest eigenvalue is equal to one. The priorities of the
alternatives (or any set of elements in a component) are obtained by
normalizing the corresponding values in the appropriate columns of the limit
matrix.
When W has zeros and is reducible (its graph is not strongly
connected so there is no path from some point to some other point) the limit
can cycle and a Cesaro average over the different limits of the cycle is taken.
For complete treatment, see the 2001 book by Saaty on the ANP, and also the
manual for the ANP software.

ANP Formulation of the Classic AHP School Example

Figure 5 shows the hierarchy developed in a school selection decision.
Following the hierarchy, Figure 6 shows the corresponding supermatrix and its
limit supermatrix used to obtain the priorities of three schools involved. They
are precisely what one obtains by hierarchic composition using the AHP. The
priorities of the criteria with respect to the goal and those of the alternatives
with respect to each criterion are clearly discernible in the supermatrix itself.

Figure 5: The School Choice Hierarchy

12

The Encyclicon: Volume 2

Goal
Satisfaction with School

Learning

Friends

School
A

School
Life

Vocational
Training

School
B

College
Prep.

Music
Classes

School
C

Figure 5: The School Selection Hierarchy

Note that there is an identity submatrix for the alternatives with respect to the
alternatives in the lower right hand part of the matrix in Figure 6. The level of
alternatives in a hierarchy is a sink cluster of nodes that absorbs priorities but
does not pass them on. This calls for using an identity submatrix for them in
the supermatrix.

The School Hierarchy as Supermatrix

Goal
Learning
Friends
School life
Vocational training
Collegepreparation
Music classes
AlternativeA
AlternativeB
AlternativeC

Goal
0
0.32
0.14
0.03
0.13
0.24
0.14
0
0
0

Learning
0
0
0
0
0
0
0
0.16
0.59
0.25

Friends
0
0
0
0
0
0
0
0.33
0.33
0.34

School life Vocational trainingCollegepreparation Music classes

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.45
0.77
0.25
0.69
0.09
0.06
0.5
0.09
0.46
0.17
0.25
0.22

A
0
0
0
0
0
0
0
1
0
0

B
0
0
0
0
0
0
0
0
1
0

C
0
0
0
0
0
0
0
0
0
1

Limiting Supermatrix & Hierarchic Composition

Goal
Goal
0
Learning
0
Friends
0
School life
0
Vocational training
0
Collegepreparation
0
Musicclasses
0
AlternativeA
0.3676
AlternativeB
0.3781
AlternativeC
0.2543

Learning
0
0
0
0
0
0
0
0.16
0.59
0.25

Friends School life Vocational trainingCollegepreparationMusicclasses

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.33
0.45
0.77
0.25
0.69
0.33
0.09
0.06
0.5
0.09
0.34
0.46
0.17
0.25
0.22

A
0
0
0
0
0
0
0
1
0
0

B
0
0
0
0
0
0
0
0
1
0

C
0
0
0
0
0
0
0
0
0
1

Figure 6: The Limit Supermatrix of the School Choice Hierarchy showing the same
Result as Hierarchic Composition.