Fuzzy Optimization and Decision Making

https://doi.org/10.1007/s10700-020-09323-y

A direct consistency test and improvement method for the

analytic hierarchy process

Kang Xu1,2 · Jiuping Xu1

© Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract
The consistency test is a vital component of pairwise comparison matrices if mean-
ingful results are to be guaranteed, and it has been studied extensively since the
analytic hierarchy process was developed by Saaty. However, when using the exist-
ing methods, it is imperative to carry out matrix operations, which are usually
not intuitional. In this paper, a direct method, from the perspective of 3 tuples,
is proposed to test and improve the consistency of a pairwise comparison matrix,
which is far more intuitional and easier to understand. In the proposed 3 tuples
iterative method, only simple mathematical calculations are needed, without the
need for matrix operations. Furthermore, the calculation of Saaty’s consistency
ratio is only conducted to verify the proposed method. Some related theorems and
propositions are proved mathematically or verified by random simulations. The
proposed method is applied to some published examples to verify its effective-
ness and practicability. Finally, we provide an improved iterative method, which
is based on the 3 tuples iterative method. The improved iterative method and
the 3 tuples iterative method are closely linked and each has its own different
perspective. Moreover, they are applied on different occasions and reinforce one
another.

Keywords Multiple criteria decision making · Analytical hierarchy process ·

Pairwise comparison matrix · Consistency test · 3 tuples

B Jiuping Xu
[email protected]
Kang Xu
[email protected]

1 Business School, Sichuan University, Chengdu 610065, People’s Republic of China

2 School of Economics and Management, Hubei University of Automotive Technology, Shiyan
442002, Hubei, People’s Republic of China

123
 K. Xu, J. Xu

1 Introduction

The analytical hierarchy process (AHP), developed by Saaty (1977, 1978, 1979,
1980), is a powerful tool in multiple criteria decision making (MCDM) when resolving
problems that combine qualitative and quantitative analysis. In recent years, the AHP
has become a highly important research topic (Goyal and Kaushal 2018; Krejčí et al.
2017), and the scope of its application is very extensive.
The AHP is a powerful tool for dealing with multiple criteria decision problems in
which the criteria (or objectives) are compared pairwise with respect to their impor-
tance. Moreover, the judgements regarding the paired comparisons of objectives are
collected using pairwise comparison matrices.
Consistency is a basic requirement for pairwise comparison matrices if meaningful
results are to be guaranteed, and checking the consistency of pairwise comparison
matrices is a crucial step to avoid misleading solutions (Zhu et al. 2016). Since an
inconsistency has a serious impact on the results of a priority vector, many studies
have focused on the consistency test, and how to improve it, for a number of decades
(Chiclana et al. 2009; Ramík 2018; Saaty 1986, 1990; Wu et al. 2019; Yager and
Alajlan 2015).
However, some of the existing methods are complicated and difficult to use when
revising the inconsistent comparison matrix, whilst others make it difficult to preserve
most of the original comparison information, as a new matrix has to be constructed
to replace the original one (Ergu et al. 2011). Ergu et al’s method has overcome
the aforementioned problems. They proposed a method, which is simpler and more
efficient and accurate than the previous approaches, to improve the consistency ratio
of the pairwise comparison matrix (Ergu et al. 2011). The latest research, such as
Zhang et al. (2018), proposed there were two paradoxes in Saaty’s consistency test
in the pairwise comparison method using Saaty’s scale. They then investigated the
consistency by dividing an n × n pairwise comparison matrix into 3 × 3 submatrices.
However, in Ergu et al’s method, it is still imperative to carry out matrix oper-
ations, which are usually not intuitional. Furthermore, the existing methods can
not respond to certain problems directly. For example, assume a 3 × 3 positive
pairwise comparison matrix A = (ai j )3×3 ; according to our current understand-
ing, it can be found that a13 = a12 a23 or a13 /(a12 a23 ) = 1 means that
A is of perfect consistency. Moreover, consider the following questions. What
values should a13 /(a12 a23 ) take when it means that A is of acceptable consis-
tency? Does it matter whether or not a13 /(a12 a23 ) = 0.99 or a13 /(a12 a23 ) =
1.01 means that A is of acceptable consistency? What about a13 /(a12 a23 ) =
0.9 or a13 /(a12 a23 ) = 1.1? Or even a13 /(a12 a23 ) = 0.5 or a13 /(a12 a23 ) =
2?
The objective of this paper is to provide a direct consistency test and improvement
method for the pairwise comparison matrix in AHP, and subsequently reply to the
questions above directly. As it does not carry out matrix operations, the proposed
method is only for testing and improving the consistency according to the 3 tuples
(aik , ak j , ai j ) in the pairwise comparison matrix.
The contributions of this paper are as follows. First, a direct method, from the
perspective of 3 tuples (aik , ak j , ai j ), rather than carrying out matrix operations,

123
A direct consistency test and improvement method for the…

is provided to test the consistency of a pairwise comparison matrix. We divide an

n × n pairwise comparison matrix into 3 tuples (aik , ak j , ai j ). Moreover, the inter-
nal quantitative relationships between the consistency ratio of the comparison matrix
and the 3 tuples are ascertained. Second, this paper presents a simpler algorithm for
improving the consistency of a pairwise comparison matrix. The proposed method
is far more intuitional and easy to understand, only simple mathematical calcula-
tions are needed, and there are no matrix operations. Furthermore, the calculation
of Saaty’s consistency ratio is only conducted to verify the proposed method. Third,
the proposed method can also retain most of the information provided by the orig-
inal comparison matrix, indicate the modification direction and provide the optimal
values; using the proposed method, the inconsistent elements in the matrix can be
found accurately and rapidly. Finally, we provide an improved iterative method,
which is based on the 3 tuples iterative method. The improved iterative method
and the 3 tuples iterative method are closely linked and each has its own differ-
ent perspective. Moreover, they are applied on different occasions and reinforce one
another.
The rest of this paper is organized as follows. In Sect. 2, some concepts and prop-
erties associated with the topic are briefly reviewed. In Sect. 3, a direct acceptable
consistency test is proposed, sufficient and approximately equivalent conditions for
an acceptable consistency are developed and some valuable theorems and propo-
sitions are proved mathematically or verified by random simulations. In Sect. 4,
a 3 tuples iterative method for improving the consistency process is proposed. In
order to compare the proposed method with those suggested by other researchers,
it is applied to some published examples, and the results are compared. In Sect. 5,
an improved iterative method, which is based on the 3 tuples iterative method,
and is convenient for preference relations with high orders. Some examples are
calculated by employing the improved iterative method, and an actual case study
is used in order to demonstrate the effectiveness of the proposed approach. In
the final section, conclusions are drawn and future research directions are dis-
cussed.

2 Preliminaries

In this section, some concepts and properties associated with the topic are briefly
reviewed.

2.1 The 1–9 ratio scale and pairwise comparison matrix

Saaty (1980) proposed a 1–9 ratio scale (shown in Table 1) as a basis for decision
makers (DMs) to provide judgements over the paired comparisons of objectives.
Then the judgements are collected by multiplicative preference relations, where reci-
procity is required, which are called pairwise comparison matrices, or simply called
comparison matrices. A pairwise comparison matrix A = (ai j )n×n can be shown
as:

123
 K. Xu, J. Xu

Table 1 The 1–9 ratio scale

Scale Meaning

1 Equally preferred
3 Moderately preferred
5 Strongly preferred
7 Very strongly preferred
9 Extremely preferred
Other values between 1 and 9 Intermediate values used to represent compromise

⎡ ⎤
1 a12 a13 ··· a1n
⎢ 1 a23 ··· a2n ⎥
⎢ ⎥
⎢ .. .. ⎥
⎢ . 1 . ⎥
A=⎢ ⎥
⎢ .. .. .. .. ⎥
⎢. 1/ai j . . . ⎥
⎣ ⎦
..
. 1
1
where A = (ai j )n×n (n ≥ 2), ai j a ji = 1, ai j ∈ 9 , 8 , 7 , . . . , 2 , 1, 2, . . . 7, 8, 9
1 1 1
, i,
j = 1, 2, . . . , n.

2.2 The consistency of a pairwise comparison matrix

Saaty (1980) presented a consistency ratio in the AHP, and also developed the concepts
of perfect consistency and acceptable consistency. Saaty (1977) also developed an
eigenvector method to derive priorities from comparison matrices, and then defined a
consistency index to measure their consistency degrees.
For a set of objectives X = {x1 , x2 , . . . , xn } and a constructed comparison matrix
A = (ai j )n×n , the eigenvector method is based on solving the following equation:

n
Aω = λmax ω, ωi = 1 (1)
i=1

where λmax is the maximum eigenvalue of A, and ω is the priority vector of the
objectives.
Let

λmax − n
CI = (2)
n−1

then the consistency index is defined as

CI
CR = (3)
RI

123
A direct consistency test and improvement method for the…

Table 2 The random index (R I )

by Saaty (2001) n 1 2 3 4 5 6 7 8 9 10
RI 0 0 0.52 0.89 1.11 1.25 1.35 1.40 1.45 1.49

where R I is a random consistency index, which is the average value of C I obtained

from 10,000 random pairwise comparison matrices, whose entries were randomly
generated using the 1 to 9 scale. Saaty considers that a value of C R under 0.10
indicates the pairwise comparison matrix is of acceptable consistency. Table 2 gives
values of R I for different matrix orders (n). Some cases require the consistency ratio
to be less than 5% for n = 3 and less than 8% for n = 4 (Saaty 2001).

Definition 1 Saaty (1980) A reciprocal judgement matrix A = (ai j )n×n is of perfect

consistency if for any positive integers i, j, k (≤ n), the following result holds:

ai j = aik ak j (4)

3 A direct acceptable consistency test

To check the consistency of pairwise comparison matrices using the existing meth-
ods, it is imperative to carry out matrix operations, which are usually not intuitional.
Moreover, as outlined in the introduction, the existing methods cannot resolve certain
problems directly. A direct acceptable consistency test method is provided to test the
consistency of a pairwise comparison matrix in this paper. 3 × 3 pairwise comparison
matrices are investigated first.

3.1 A direct acceptable consistency test for 3 × 3 pairwise comparison matrices

In this section a correspondence between the consistency ratio C R and the ratio
a13 /(a12 a23 ) is proposed.

Theorem 1 For a 3 × 3 pairwise comparison matrix A = (ai j )3×3 (a ji =


1/ai j , i, j = 1, 2, 3), ai j ∈ 19 , 18 , 17 , . . . , 21 , 1, 2, . . . 7, 8, 9 , there is a correspon-
dence between Saaty’s consistency ratio and the ratio a13 /(a12 a23 ).
(1) When the threshold value of the consistency ratio is 0.05,

a13
C R < 0.05 ⇔ 0.505 < < 1.979
a12 a23

where ‘⇔’ means being equivalent, namely, C R < 0.05 is equivalent to 0.505 <
a13 /(a12 a23 ) < 1.979.
(2) When the threshold value of the consistency ratio is 0.1,

a13
C R < 0.1 ⇔ 0.382 < < 2.620
a12 a23

123
 K. Xu, J. Xu

where ‘⇔’ means being equivalent, namely, C R < 0.1 is equivalent to 0.382 <
a13 /(a12 a23 ) < 2.620.
Hereby 0.505, 1.979, 0.382, 2.620 all are approximate values, which are numbers to
three decimal places.

Proof Since ai j ∈ 19 , 18 , 17 , . . . , 21 , 1, 2, . . . 7, 8, 9 , there are only 17 possible values
from 19 to 9 for a12 , a23 or a13 . By using the multiplication principle of combinatorial
mathematics, there are 17 × 17 × 17 = 4, 913 different combinations of a12 , a23 , a13 .
Furthermore, due to a ji = 1/ai j (i, j = 1, 2, 3), there are only 4913 combinations of
ai j . Therefore, Theorem 1 can be proved by enumerating all 4913 cases one by one.
It is completed by using MATLAB (The MATLAB codes are listed in “Appendix 1”).
The proof of Theorem 1 can be included with that of Theorem 2. In other words, if
Theorem 2 holds, then Theorem 1 also holds.
Theorem 2 When ai j (a ji = 1/ai j , i, j = 1, 2, 3) is continuous, i.e., for arbitrary
ai j ∈ [1/9, 9], the conclusions in Theorem 1 also hold.
Proof Suppose that a pairwise comparison matrix
⎡ ⎤
1 a c
Ȧ = ⎣ 1/a 1 b⎦
1/c 1/b 1

where a, b, c ∈ [1/9, 9] and

c
=k (5)
ab
for brevity, hereby we let a12 = a, a23 = b, a13 = c, then we calculate the character-
istic polynomial of Ȧ as follows:

λ−1 −a −c
λE − Ȧ = −(1/a) λ−1 −b
−(1/c) −(1/b) λ−1
ab c
= (λ − 1)3 − 3(λ − 1) − +
c ab

where E denotes a 3 × 3 identity matrix and λ denotes the eigenvalues; i.e., the roots
of the characteristic equation λE − Ȧ = 0, namely,

ab c
(λ − 1)3 − 3(λ − 1) − + =0 (6)
c ab

According to Eq. (5), Eq. (6) is equivalent to

1
(λ − 1)3 − 3(λ − 1) − +k =0 (7)
k

123
A direct consistency test and improvement method for the…

Fig. 1 The correspondence between C R and k

Fig. 2 h(x) = x 3 − 3x − t

Let
1
x = λ − 1, t = +k
k
then Eq. (6) is equivalent to

x 3 − 3x − t = 0 (8)

According to Eq. (7), it can be found that the roots of Eq. (7), or equivalent Eqs.
(6) and (8) are only dependent on the parameter k, and thus the consistency ratio C R
is only dependent on the parameter k. Figure 1 shows a corresponding line chart of
C R and k. The data for Fig. 1 is omitted for the purposes of brevity.
In order to find the roots of Eq. (8), suppose that
h(x) = x 3 − 3x − t

where t is a parameter.
As shown in Fig. 2, when t increases, the largest root of h(x) = 0 also increases.
h is continuous and differentiable everywhere, the derivative of h is

123
 K. Xu, J. Xu

dh
= 3x 2 − 3 = 3(x 2 − 1)
dx

When x > 1, ddhx > 0, and since it is continuous everywhere, h is strictly increasing
on the interval [1, +∞) and due to

1
h(1) = −2 − t < 0 t =k+ ≥2
k
lim h(x) = +∞
x→+∞

therefore, function h has just one zero point on the interval [1, +∞), i.e., Eq. (8) has
just one root on the interval [1, +∞); accordingly the root on the interval [1, +∞)
is the largest root of Eq. (8), so in the following discussion we only consider the
interval [1, +∞). Suppose the largest root of Eq. (8) is xmax , since we have supposed
that x = λ − 1, the corresponding λ, namely, λmax = xmax + 1, is also the largest
eigenvalue according to Eq. (6) or (7). According to Eq. (8), we have

t = x 3 − 3x

It is obvious that t = x 3 −3x is also strictly increasing on the interval [1, +∞). Notice
that we now only consider the interval [1, +∞); according to Eqs. (2) and (3) and
Table 2, we have

CI
C R < 0.1 ⇔ < 0.1
RI
λmax − n λmax − 3
⇔ = < 0.1
0.52(n − 1) 0.52 × 2
(xmax + 1) − 3
⇔ < 0.1
0.52 × 2
⇔ xmax < 2.104
⇔ t = x 3 − 3x < 2.1043 − 3 × 2.104
1 1
⇔ k + < 2.1043 − 3 × 2.104 t =k+
k k
⇔ 0.382 < k < 2.620

where ‘⇔’ means being equivalent. t = k + k1 is decreasing on the interval (0, 1], and
increasing on the interval [1, +∞), solving the equation

1
k+ = 2.1043 − 3 × 2.104 (9)
k

the approximations of the two roots are obtained:

k1 = 0.382,

123
A direct consistency test and improvement method for the…

Table 3 Values of C R (the

threshold value is 0.05) for k < 0.505 0.505 (0.505, 1.979) 1.979 > 1.979
different k CR > 0.05 0.05 < 0.05 0.05 > 0.05

Table 4 Values of C R (the

threshold value is 0.1) for k < 0.382 0.382 (0.382,2.620) 2.620 > 2.620
different k CR > 0.1 0.1 < 0.1 0.1 > 0.1

k2 = 2.620

which are numbers to three decimal places. The domain of k is k1 < k < k2 . This
completes the proof of case (2); the proof of the other case is similar.
Notice that in this paper 0.505, 1.979, 0.382, 2.620 are all approximate values,
which are numbers to three decimal places. If it is necessary, more accurate approx-
imations or even the exact values can be found by solving Eq. (9), but are omitted
here for brevity; this is the same for other parts of the paper. Using the exact roots to
simulate, from 1,000,000 random pairwise comparison matrices, Theorem 2 is found
to be satisfied. The exact roots of Eq. (9) and the Matlab codes are listed in “Appendix
2”. The random simulation is also further proof of Theorem 2.
As shown in Tables 3 and 4, if the threshold value of C R is 0.05, then k ∈
(0.505, 1.979) means that Ȧ is of acceptable consistency; if the threshold value of
C R is 0.1, then k ∈ (0.382, 2.620) means that Ȧ is of acceptable consistency.

Example 1 If a12 = 3, a23 = 4, take 0.1 as the threshold value of the consistency ratio,
then

0.382 × 3 × 4 < a13 < 2.620 × 3 × 4

where a13 ∈ [1/9, 9], namely,

4.58 < a13 ≤ 9

1
means that A is of acceptable consistency. If ai j ∈ 9 , 8 , . . . , 2 , 1, 2, . . . , 8, 9
1 1
is
needed, then a13 = 5, 6, . . . , 9.

3.2 A direct consistency test method for higher order pairwise comparison
matrices

In this section, we investigate the approximate condition for an acceptable consistency

in the environment of a 4 × 4 and higher order pairwise comparison matrix.

Proposition 1 Suppose a pairwise comparison matrix A,

 
1 1 1 1
A = (ai j )n×n (n ≥ 3), ai j a ji = 1, ai j ∈ , , , . . . , , 1, 2, . . . 7, 8, 9
9 8 7 2

123
 K. Xu, J. Xu

ai j
θikj = (1 ≤ i < k < j ≤ n),
aik ak j

considering that if θikj > 1 (i = 1, 2, 3, 4), there is an upward deviation from a

perfect consistency, while if θikj < 1, there is a downward deviation. In order to avoid
a situation that the upward and downward deviations cancel each other out, suppose
that the adjusted θikj , i.e.,


1/θikj , i f θikj < 1;
†θikj =
θikj , else

then we have

n−2 n j−1
6 approximately
†θikj < 2.620 ⇔ C R(A) < 0.1 (10)
n(n − 1)(n − 2)
i=1 j=i+2 k=i+1

approximately
where ⇔ means that in formula (10), the former is approximately equivalent
to the latter with a very small percentage error. 2.620 is an approximate value, which
is a number to three decimal places; the total number of all †θikj is Cn3 = n(n−1)(n−2)
6
6 n−2 n  j−1 k is the average of all †θ k .
and n(n−1)(n−2) i=1 j=i+2 k=i+1 †θij ij

Proof Proposition 1 involves some approximately equivalent conditions, which can be

verified by random simulations. The MATLAB codes for 4 × 4 and 5 × 5 comparison
matrices are listed in “Appendix 3”, the rest are similar in nature.
 
Proposition 2 When ai j ∈ 19 , 9 , i, j = 1, 2, . . . , n, i.e., when ai j is continuous, the
conclusions in Proposition 1 including formula (10) also hold.

Proof Similar to the verification of Theorem 2; the MATLAB codes are omitted.

According to Eq. (9), the two roots k1 , k2 are the inverse of each other, namely,
0.382 and 2, 620 are almost the inverse of each other, hence †θikj < 2.620 is equivalent
to 0.382 < θikj < 2.620. Thus a simple and direct consistency test method can be
obtained:

Definition 2 Let A be as before, and

n−2 n j−1
6
†θikj < 2.620
n(n − 1)(n − 2)
i=1 j=i+2 k=i+1

then A is considered to be of an acceptable consistency.

123
A direct consistency test and improvement method for the…

4 A 3 tuples iterative method for improving the consistency process

In this section, we investigate the consistency improvement process. A 3 tuples iterative

method is proposed for improving the consistency process.

4.1 A 3 tuples iterative method

According to the aforementioned theory, we propose a 3 tuples iterative method to

adjust a pairwise comparison matrix so that it is of acceptable consistency.
For convenience, (aik , ak j , ai j ) (1 ≤ i < k < j ≤ n) is called a 3 tuple. Further,
if
ai j
0.382 < < 2.620
aik ak j

then (aik , ak j , ai j ) is called a 3 tuple satisfying an acceptable consistency, otherwise

it is called an inconsistent 3 tuple.
Using the 3 tuples iterative method to improve the consistency of the comparison
matrix A, the general steps are as follows.
Step 1 Calculating. As mentioned in Proposition 1, calculate the ratios
ai j
θikj = (1 ≤ i < k < j ≤ n),
aik ak j

1/θikj , i f θikj < 1;
†θikj =
θikj , else

When calculating Saaty’s consistency ratio, if C R(A) < 0.1 is satisfied, then the
matrix A is of an acceptable consistency. This completes the calculation. Otherwise, if
n−2 n  j−1
C R(A) < 0.1 is not satisfied, or n(n−1)(n−2)
6
i=1 j=i+2 k=i+1 †θi j >> 2.620
k

is satisfied (by Definition 2), then the matrix A is not of an acceptable consistency;
therefore go to Step 2.
Step 2 Identifying. Find all the inconsistent 3 tuples, which satisfy †θikj ≥ 2.620.
Particularly for some fairly large values in †θikj , their corresponding 3 tuples are seri-
ously inconsistent, and should be adjusted first.
Step 3 Adjusting. Adjust the elements in the inconsistent 3 tuples. Then go to Step
1. When the elements in the inconsistent 3 tuples are adjusted, the following principles
should be obeyed as far as possible.
Principle 1 Adjust as few elements in the comparison matrix as possible. In this way
the original information from decision makers can be preserved as much as possible.
Principle 2 If an element appears in a tuple with the maximum of all ratios †θikj (1 ≤
i < k < j ≤ n), and this element simultaneously appears in several other seriously
inconsistent 3 tuples, then this element should be adjusted first.
If there are 3 tuple judgments with more than one common element that are out of
the acceptable consistency scope, then one must adjust the element with the maximum
deviation (ratios) and the highest frequency of occurrence in these inconsistent 3 tuples.
Step 4 Finish.

123
 K. Xu, J. Xu

Note 1 According to the reciprocity (ai j a ji = 1), in the pairwise comparison

matrix A, only the elements in the upper triangulation matrix of A are considered
by the authors. Similar processing can be found in Petra and Lidija (2012). In the
consistency improvement process, when we adjust one element ai j (i < j) in the
upper triangulation matrix of A without an explanation, the reciprocity is always
satisfied; i.e., the element a ji is adjusted simultaneously to let ai j a ji = 1 still hold.
This applies elsewhere in the paper.
In some special cases, if we adjust individual elements in the submatrix of a com-
parison matrix A, then the submatrix has a perfect consistency; at this moment, it is
necessary to firstly adjust the elements in the submatrix of the comparison matrix A.
Note 2 Due to the convergence of the monotone bounded series of numbers, the
iterative algorithm above is obviously convergent.

4.2 Some illustrative examples and comparisons

In order to compare our method with those of others, the proposed method in this
paper is applied to some published examples.

Example 2 The 4 × 4 pairwise comparison matrix A is inconsistent with C R =

1.0242 > 0.1.
⎡ ⎤
1 2 4 1/8
⎢ 1 2 4 ⎥
A=⎢
⎣
⎥
1 2 ⎦
1

Step 1 Calculating the ratios,

a13 4
θ13
2
= = = 1,
a12 a23 2×2
a24 4
θ24
3
= = = 1,
a23 a34 2×2
1
a14 1 1
θ14
2
= = 8 = , †θ14
2
= 2 = 64 > 2.620,
a12 a24 2×4 64 θ14
1
a14 1 1
θ14
3
= = 8 = , †θ14
3
= 3 = 64 > 2.620,
a13 a34 4×2 64 θ14

Step 2 Therefore the 3 tuples (a12 , a23 , a13 ) and (a23 , a34 , a24 ) are of perfect consis-
tency, while the 3 tuples (a12 , a24 , a14 ) and (a13 , a34 , a14 ) are seriously inconsistent.

2 + †θ 3 + †θ 2 + †θ 3
†θ13 1 + 1 + 64 + 64
24 14 14
= = 32.5  2.620
4 4

means that the comparison matrix A is seriously inconsistent.

123
A direct consistency test and improvement method for the…

Step 3 Adjust the 3 tuples (a12 , a24 , a14 ) and (a13 , a34 , a14 ), out of which a14 is a
common element. Therefore, a14 should be adjusted first. Let
a14 = a12 a24 = 2 × 4 = 8,

or
a14 = a13 a34 = 4 × 2 = 8

the adjusted comparison matrix A is obtained.

⎡ ⎤
1 2 4 8
⎢ 1 2 4⎥
Aad justed = ⎢
⎣
⎥
1 2⎦
1
C R(Aad justed ) = 0,

which completes Example 2.

Notice that in Example 2, when we let 0.382a12 a24 < a14 < 2.620a12 a24 or
0.382a13 a34 < a14 < 2.620a13 a34 , i.e., 3.056 < a14 < 20.96, and due to 19 ≤ a14 ≤
9, we have a14 ∈ (3.056, 9], or even a14 varies in a larger domain, A is still of an
acceptable consistency. By using formula (10), we can get an approximate interval
(1.887, 9] for a14 . The calculation process is as follows.
2 + †θ 3 + †θ 2 + †θ 3
†θ13 24 14 14
< 2.620
4
(1) When a14 < 8,
a14 a14 a14 a14
θ14
2
= = < 1, θ14
3
= = < 1,
a12 a24 8 a13 a34 8
1 8 1 8
2
†θ14 = 2 = , †θ14
3
= 3 = ,
θ14 a14 θ14 a14
2
†θ13 = θ13
2
= 1, †θ24
3
= θ24
3
= 1,
2 + †θ 3 + †θ 2 + †θ 3
†θ13 1 + 1 + a814 + a814
24 14 14
= < 2.620,
4 4
16
a14 > ≈ 1.887,
8.48
thus 1.887 < a14 ≤ 8 is obtained.
(2) When a14 ≥ 8,
a14 a14 a14 a14
θ14
2
= = ≥ 1, θ14
3
= = ≥ 1,
a12 a24 8 a13 a34 8
2 + †θ 3 + †θ 2 + †θ 3
†θ13 1 + 1 + a814 + a814
24 14 14
= < 2.620,
4 4
then a14 < 33.92 is obtained.

123
 K. Xu, J. Xu


Since a14 ∈ 19 , 18 , 17 , . . . , 21 , 1, 2, . . . 7, 8, 9 , finally the domain making the com-
parison matrix A be of an acceptable consistency is

a14 = 2, 3, 4, . . . , 9

(when a14 = 1, C R(A) = 0.2184; when a14 = 2, C R(A) = 0.0933; . . . ; when

a14 = 9, C R(A) ≈ 0).
Example 2 has appeared in Ergu et al. (2011); compared with their methods, the
results are the same, but the proposed method is much more intuitional and easy
to understand, as only simple mathematical calculations are needed, without matrix
operations. The calculation of the consistency ratio is only conducted to verify the
proposed method.
In the proposed method, the identification process of the inconsistent elements is
intuitional, rapid and efficient. The elements in the comparison matrix are adjusted as
little as possible, that is only a pair of elements are adjusted, and after that the pairwise
comparison matrix A is of an acceptable consistency, or even of perfect consistency.
In the proposed method, the elements needing to be adjusted are identified accu-
rately, and furthermore the specific domain of the adjusted element is also ascertained.
Example 3 The 8 order pairwise comparison matrix A was first introduced in
Kwiesielewicz and Uden (2002) as an example, then Ergu et al. (2011) also used
this matrix as an example, and compared their method with that of Kwiesielewicz and
Uden (2002). Now we will use this matrix as an example to facilitate a comparative
analysis.
⎡ ⎤
1 2 1/2 2 1/2 2 1/2 2
⎢ 1 4 1 1/4 1 1/4 1 ⎥
⎢ ⎥
⎢ 1 4 1 4 1 4⎥
⎢ ⎥
⎢ 1 1/4 1 1/4 1 ⎥
A=⎢ ⎢ ⎥
⎢ 1 4 1 4⎥⎥
⎢ 1 1/4 1 ⎥
⎢ ⎥
⎣ 1 4⎦
1
C R(A) = 0.1055

Step 1 Calculating the ratios. This is a rather special kind of matrix, so first we consider
the 4 × 4 submatrix of A,
⎡ ⎤
1 2 1/2 2
⎢ 1 4 1⎥
⎢ ⎥ , C R(Asub ) = 0.4071 > 0.1,
4×4 = ⎣
Asub
1 4⎦ 4×4
1
1
a13 1 1
θ13
2
= = 2 = , †θ13 2
= 2 = 8 > 2.620,
a12 a23 2×4 8 θ13
a24 1 1 1
θ24
3
= = = , †θ24
3
= 3 = 16 > 2.620,
a23 a34 4×4 16 θ24

123
A direct consistency test and improvement method for the…

a14 2
θ14
2
= = = 1, †θ14
2
= θ14
2
= 1 < 2.620,
a12 a24 2×1
a14 2
θ14
3
= = 1 = 1, †θ14
3
= θ14
3
= 1 < 2.620
a13 a34 2 × 4

Step 2 Thus the 3 tuples (a12 , a24 , a14 ), (a13 , a34 , a14 ) are of an acceptable consistency,
or even of a perfect consistency, while the 3 tuples (a12 , a23 , a13 ), (a23 , a34 , a24 ) are
seriously inconsistent; out of which a23 is a common element. Thus a23 should be
adjusted first.
Step 3 Let

1
a13 1
a23 = = 2 = ,
a12 2 4

or

a24 1 1
a23 = = = ,
a34 4 4

then

⎡ ⎤
1 2 1/2 2
⎢ 1 1/4 1⎥
=⎢ ⎥ , C R(Asub-ad justed ) = 0 < 0.1
sub-ad justed
A4×4 ⎣ 1 4⎦ 4×4
1

Meanwhile,

⎡ ⎤
1 2 1/2 2 1/2 2 1/2 2
⎢ 1 1/4 1 1/4 1 1/4 1⎥
⎢ ⎥
⎢ 1 4 1 4 1 4⎥
⎢ ⎥
⎢ 1 1/4 1 1/4 1⎥
Aad justed =⎢
⎢
⎥ , C R(Aad justed ) = 0 < 0.1
⎢ 1 4 1 4⎥⎥
⎢ 1 1/4 1⎥
⎢ ⎥
⎣ 1 4⎦
1

Example 4 The following pairwise comparison matrix is obtained by adding one row
and one column with a random value to the comparison matrix in Example 3, and it
is also from Ergu et al. (2011).

123
 K. Xu, J. Xu

⎡ ⎤
1 2 1/2 2 1/2 2 1/2 2 1/3
⎢ 1 4 1 1/4 1 1/4 1 1/4 ⎥
⎢ ⎥
⎢ 1 4 1 4 1 4 1/7 ⎥
⎢ ⎥
⎢ 1 1/4 1 1/4 1 1/6 ⎥
⎢ ⎥
A=⎢
⎢ 1 4 1 4 6 ⎥⎥
⎢ 1 1/4 1 1/3 ⎥
⎢ ⎥
⎢ 1 4 7 ⎥
⎢ ⎥
⎣ 1 1/2 ⎦
1

As with Example 3, adjust a23 from 4 to 41 , then the submatrix from the first 8 rows
and the first 8 columns in the matrix A, i.e.,

⎡ ⎤
1 2 1/2 2 1/2 2 1/2 2
⎢ 1 1/4 1 1/4 1 1/4 1⎥
⎢ ⎥
⎢ 1 4 1 4 1 4⎥
⎢ ⎥
⎢ 1 1/4 1 1/4 1⎥
=⎢ ⎥
sub-ad justed
A8×8 ⎢
⎢ 1 4 1 4⎥⎥
⎢ 1 1/4 1⎥
⎢ ⎥
⎣ 1 4⎦
1

sub-ad justed
is of perfect consistency, C R(A8×8 ) = 0 < 0.1.
C R(Aad justed ) = 0.1335 > 0.1, at this moment, it is no longer necessary to adjust
sub-ad justed
the elements in the submatrix A8×8 , only the elements from the ninth row and
the ninth column in Aad justed need to be adjusted.
sub-ad justed
Step 1 Calculating the ratios. Since A8×8 has been of perfect consistency, it
sub-ad justed
is not necessary to consider the 3 tuples in A8×8 . The rest of the 3 tuples are
calculated as follows. Using the formulas

ai j
θikj = (1 ≤ i < k < j ≤ n),
aik ak j

1/θikj , i f θikj < 1;
†θikj =
θikj , else

the calculation results are listed in Table 5.

7 = 49 is the maximum,
Step 2 After observation, it can be found that out of all †θikj , †θ39
and †θ39 = 42, †θ39 = 14, †θ19 = 10.5, †θ49 = 10.5, †θ39
5 8 7 7 6 = 9.34, †θ 5 = 9,
19
5 = 9 are also quite large ratios; therefore the corresponding 3 tuples need to be
†θ49
adjusted, where the element a39 appears with the highest frequency. Thus a39 should
be adjusted first.

123
A direct consistency test and improvement method for the…

Table 5 Values of †θikj

i, k, j 129 139 149 159 169 179 189 239 249 259
†θikj 1.5 4.67 1 9 2 10.5 3 7 1.5 6
i, k, j 269 279 289 349 359 369 379 389 459 469
†θikj 1.33 7 2 4.67 42 9.34 49 14 9 2
i, k, j 479 489 569 579 589 679 689 789
†θikj 10.5 3 4.5 1.17 3 5.25 1.5 3.5

Table 6 Adjusted a39 and the corresponding consistency ratio

adjusted a39 ... 1/5 1/4 1/3 1/2 1 2

C R(Aad justed ) ... 0.1103 0.0980 0.0852 0.0721 0.0597 0.0566
adjusted a39 3 4 5 6 7 8 9
C R(Aad justed ) 0.0586 0.0618 0.0654 0.0654 0.0732 0.0771 0.0810

Step 3 Since the adjusted element a39 , i.e.,

 
ad justed 1 1 1
a39 ∈ , , . . . , , 1, 2, . . . , 8, 9
9 8 2

which can be enumerated one by one; the calculation results are listed in Table 6.
Out of which, a39 = 2 (a93 = 21 ) can maximize the consistent degree of A, at this
moment,
⎡ ⎤
1 2 1/2 2 1/2 2 1/2 2 1/3
⎢ 1 1/4 1 1/4 1 1/4 1 1/4 ⎥
⎢ ⎥
⎢ 1 4 1 4 1 4 2 ⎥
⎢ ⎥
⎢ 1 1/4 1 1/4 1 1/6 ⎥
⎢ ⎥
Aad justed =⎢
⎢ 1 4 1 4 6 ⎥⎥
⎢ 1 1/4 1 1/3 ⎥
⎢ ⎥
⎢ 1 4 7 ⎥
⎢ ⎥
⎣ 1 1/2 ⎦
1
C R(Aad justed ) = 0.0566 < 0.1

The adjusted matrix A has been of an acceptable consistency, and in the whole pro-
cess only two pairs of elements in the matrix A are adjusted; therefore compared to
the published methods, the calculation process of the proposed method is far more
intuitional and simple, without the need for matrix operations.

Example 5 The following pairwise comparison matrix was first introduced in Xu and
Wei (1999) as an example, then was also used by Saaty (2003) as an example to
describe the method embedded in the Expert Choice Software, which was used to

123
 K. Xu, J. Xu

Table 7 Values of †θikj

i, k, j 128 138 148 158 168 178 127 137 147 157
†θikj 2.86 2.4 3.5 4 4 1.5 3 54 3 3.6
i, k, j 167 126 136 146 156 125 135 145 124 134
†θikj 3.6 2.5 2 3.43 2 2.5 1.5 2.57 3.57 6
i, k, j 123 238 248 258 268 278 237 247 257 267
†θikj 1.8 2.14 4.37 3.5 3.5 1.43 10 3.57 3 3
i, k, j 236 246 256 235 245 234 348 358 368 378
†θikj 2.25 2.4 2 3 1.8 6 3.75 2.5 3.33 15
i, k, j 347 357 367 346 356 345 458 468 478 457
†θikj 7 10 7.50 2.67 2.67 1.5 2.25 3 1.75 2.14
i, k, j 467 456 568 578 567 678
†θikj 2.86 1.5 2 1.67 2 1.67

detect the inconsistencies, and also by Cao et al. (2008) as an inconsistent pairwise
comparison matrix to test their proposed heuristic method (Ergu et al. 2011). Later,
Ergu et al. (2011) also used this matrix to illustrate their method. Now we will also
apply this matrix to test the proposed method.
⎡ ⎤
1 5 3 7 6 6 1/3 1/4
⎢ 1 1/3 5 3 3 1/5 1/7 ⎥
⎢ ⎥
⎢ 1 6 3 4 6 1/5 ⎥
⎢ ⎥
⎢ 1 1/3 1/4 1/7 1/8 ⎥
A=⎢
⎢
⎥
⎢ 1 1/2 1/5 1/6 ⎥
⎥
⎢ 1 1/5 1/6 ⎥
⎢ ⎥
⎣ 1 1/2 ⎦
1
C R(A) = 0.1703 > 0.1

Step 1 Calculating the ratios. Using the formulas

ai j
θikj = (1 ≤ i < k < j ≤ n),
aik ak j

1/θikj , i f θikj < 1;
†θi j =
k
θikj , else

the calculation results are listed in Table 7.

The average of all †θikj is 4.32 > 2.620, by Proposition 3, the comparison matrix
is inconsistent (C R(A) = 0.1703 > 0.1). This is also a verification of Proposition 3.
3 = 54 is the maximum,
Step 2 After observation, it can be found that out of all †θikj , †θ17
7 = 15, †θ 3 = 10, †θ 5 = 10, †θ 6 = 7.5, †θ 4 = 7 are also quite large ratios,
and †θ38 27 37 37 37
therefore the corresponding 3 tuples

123
A direct consistency test and improvement method for the…

Table 8 Adjusted a37 and the corresponding consistency ratio

adjusted a37 1/9 1/8 1/7 1/6 1/5 1/4

C R(Aad justed ) 0.1044 0.1007 0.0969 0.0933 0.0897 0.0864
adjusted a37 1/3 1/2 1 2 3 ...
C R(Aad justed ) 0.0838 0.0828 0.0886 0.1058 0.1231 ...

(a13 , a37 , a17 ), (a23 , a37 , a27 ), (a37 , a78 , a38 ), (a35 , a57 , a37 ), (a36 , a67 , a37 ),
(a34 , a47 , a37 ),

are seriously inconsistent; out of which a37 is a common element, so a37 should be
adjusted first.
Step 3 Adjust the element a37 . Since the adjusted element a37 , i.e.,
 
ad justed 1 1 1
a37 ∈ , , . . . , , 1, 2, . . . , 8, 9
9 8 2

which can be enumerated one by one; the calculation results are listed in Table 8.
Therefore, a37 = 17 , 16 , . . . , 21 , 1 can make A be of an acceptable consistency, out of
which a37 = 21 (a73 = 2) can maximize the consistent degree of A, at this moment,
C R(Aad justed ) = 0.0828.

In order to facilitate a comparative analysis, we used several of the same examples

as in Ergu et al. (2011). Ergu et al. compared their method with some other meth-
ods and their comparison is quite sufficient. They pointed out that in Xu and Wei’s
method (1999), many elements in the original comparison matrix have been changed.
Based on the original inconsistent matrix, a consistent matrix is generated by an auto-
adaptive process instead of revising single elements. A similar problem occurs in
Cao et al. (2008). Compared with the other two methods, Saaty’s method is easier to
use because it is based on the ratio of priorities and designed for the Perron Eigen-
value Method (EM) (1977) and AHP. However, the ‘precise’ number recommended
by Saaty’s method is really an approximated value. Ergu et al. (2011) concluded that
all aforementioned three methods are based on the priority vector ratios, which are
calculated by the inconsistent comparison matrix. Apart from the EM, different meth-
ods, have been proposed to derive a priority vector with a given pairwise comparison
matrix, including the methods in Barzilai (1997), Saaty (1980), Chu et al. (1979),
Crawford and Williams (1985), Bryson (1995).
Ergu et al. (2011) also argued that in their method, all values provided by experts in
the original comparison matrix have been retained, except the inconsistent elements,
a37 and a73 . Furthermore, Ergu et al’s method does not violate the scale [1, 9], needs
fewer computations than the methods of Xu and Wei and Cao et al., and also preserves
more original comparison information than these two methods. Ergu et al’s method
can also show the modification direction and provide the optimal values. All of the
aforementioned advantages are also the advantages of our method. The computations in

123
 K. Xu, J. Xu

our method are simple, intuitional and easy to understand. Only simple mathematical
calculations are needed, without matrix operations. Furthermore, the calculation of
Saaty’s consistency ratio is only conducted to verify the proposed method. Our method
can also show the modification direction and provide the optimal values; moreover,
most of the information provided by the original comparison matrix can be retained
in our method.
In the current method, it is still imperative to carry out matrix operations, which
are usually not intuitional. The proposed method in this paper has preserved the afore-
mentioned advantages, while overcoming the disadvantages mentioned above.
The latest literature, such as Zhang et al. (2018) has some similar ideas to our
method, as they proposed two paradoxes regarding Saaty’s consistency test in the
pairwise comparison method with Saaty’s scale. They also investigated the consis-
tency by dividing an n × n pairwise comparison matrix into 3 × 3 submatrices, which
is similar to our method. However, they investigated the consistency test in the pair-
wise comparison method with a fixed numerical scale; thus there is a large difference
between their research direction and ours. Furthermore, they divided an n ×n pairwise
comparison matrix into 3 × 3 submatrices, while we divide an n × n pairwise com-
parison matrix into 3 tuples (aik , ak j , ai j ). Moreover, we have ascertained the internal
quantitative relationships between the consistency ratio of the comparison matrix and
the 3 tuples, which are intuitional and easy to understand.

5 An improved iterative method for preference relations with high

orders

The proposed 3 tuples method is intuitional, easy to understand, and more suitable
for manual calculations, especially for preference relations with low orders and some
specific preference relations with high orders. In this section, we provide an improved
iterative method, which is based on the 3 tuples iterative method, and is convenient
for preference relations with high orders.

5.1 An improved iterative method to increase the computational efficiency for

preference relations with high orders

We have the following proposition

Proposition 3 Suppose the comparison matrix A = (ai j )n×n is of a perfect consis-
tency, then the following Eqs. (11), (12), (13), and (14) all hold.
n
nai j = aik ak j (11)
k=1
nA = A 2
(12)

n
aik ak j
k=1
=1 (13)
nai j

123
A direct consistency test and improvement method for the…

bi j
=1 (14)
nai j

Proof According to Definition 1 and Eq. (4), it is known that if a comparison matrix
A = (ai j )n×n is of a perfect consistency, then for any positive integers i, j, k (≤ n),
ai j = aik ak j holds. Therefore, let k = 1, 2, . . . , n, then we have

ai j = ai1 a1 j
ai j = ai2 a2 j
······
ai j = ain an j

Add these formulas

 together, and Eq. (11) is obtained.
Notice that nk=1 aik ak j is the element in the ith row and jth column of the matrix
A2 , in other words, suppose that

A2 = (bi j )n×n

then we have
n
bi j = aik ak j
k=1

According to the arbitrariness of i, j, in fact Eq. (11) is equivalent to Eq. (12).

Since all the elements are positive, i.e., for any positive integers i, j = 1, 2, . . . , n,
ai j > 0, hence Eq. (11) or Eq. (12) is equivalent to Eq. (13) or Eq. (14), which
completes the proof process.

Let

bi j nai j
ci j = max , , i, j = 1, 2, . . . n (15)
nai j bi j

then in all the elements ci j , the largest one (suppose it is ci0 j0 ) corresponds to the
most serious inconsistency. According to the aforementioned Principle 1, adjust as
few elements in the comparison matrix as possible; at this moment, adjusting ai0 j0
has the highest efficiency. Therefore, the corresponding element in the matrix A, i.e.,
ai0 j0 should be adjusted first. Thus we have the following improved iterative method,
which is convenient for preference relations with high orders.
Step 1 Calculating. When calculating Saaty’s consistency ratio (C R), if it is found
that the consistency ratio of the matrix A satisfies C R(A) < 0.1, then the matrix A
is of an acceptable consistency and the computations are completed; therefore go to
Step 4. Otherwise, go to Step 2.
 2 Identifying. Calculate B = A = (bi j )n×n and n A, let C = (c i j )n×n  =
2
 Step
bi j bi j nai j
nai j ; C = (ci j )n×n , ci j = max c i j , c 1i j , , namely, ci j = max nai j , bi j ,
n×n
calculate the matrix C, ascertain the largest out of all the elements in the matrix C.

123
 K. Xu, J. Xu

Suppose the largest one is ci1 j1 , i.e., the element in the i 1 th row and j1 th column of
the matrix C, then go to Step 3.
Step 3 Adjusting. Adjust the corresponding element ai1 j1 in the matrix A. Because

ai j ∈ 19 , 18 , 17 , . . . , 21 , 1, 2, . . . 7, 8, 9 , all the values of ai j , which can make A be
of an acceptable consistency, can be ascertained by enumerating one by one; out of
which the optimal value of ai j , which can maximize the consistent degree of A, can
also be ascertained. Then go to Step 1.
Step 4 Finish.
Note 3 In theory, Eqs. (11), (12), (13), (14) are just necessary conditions, which can
help in identifying the most inconsistent element, but are not sufficient conditions for
A being of a perfect consistency. When Eqs. (11), (12), (13), (14) are satisfied, whether
or not A is of a perfect consistency or even an acceptable consistency, this needs to
be checked further. If the answer is ‘not’, further improvement can be achieved by
employing the proposed 3 tuples method.

Example 6 By reusing A in Example 5, i.e., let

⎡ ⎤
1 5 3 7 6 6 1/3 1/4
⎢ 1/5 1 1/3 5 3 3 1/5 1/7 ⎥
⎢ ⎥
⎢ 1/3 3 1 6 3 4 6 1/5 ⎥
⎢ ⎥
⎢ 1/7 1/5 1/6 1 1/3 1/4 1/7 1/8 ⎥
A = (ai j )8×8 =⎢
⎢ 1/6
⎥
⎢ 1/3 1/3 3 1 1/2 1/5 1/6 ⎥
⎥
⎢ 1/6 1/3 1/4 4 2 1 1/5 1/6 ⎥
⎢ ⎥
⎣ 3 5 1/6 7 5 5 1 1/2 ⎦
4 7 5 8 6 6 2 1

Step 1 Calculate Saaty’s consistency ratio. C R(A) = 0.1703 > 0.1, so A is not of
an acceptable consistency, go to Step 2.
Step
 2 Calculate
 B = A2 = (bi j )n×n and n A, let C = (ci j )n×n , ci j =
b na
max nai ij j , bi ij j , calculate the matrix C.
By using MATLAB, it is found that

B = A2 = (bi j )n×n
⎡ ⎤
8 27.8167 13.6389 103.3333 53.5000 46.9167 23.5667 4.8560
⎢ 3.3968 8 4.5976 36.9429 17.7238 13.1405 4.6667 2.1274 ⎥
⎢ ⎥
⎢ 22.0905 42.6000 8 97.9333 58.2000 53.2000 15.3683 5.8286 ⎥
⎢ ⎥
⎢ 1.4071 3.3980 1.6510 8 4.5881 4.2548 1.7400 0.5163 ⎥
=⎢⎢ 2.2897
⎥
⎢ 5.4333 2.7694 15.5667 8 7.0833 3.3841 1.0476 ⎥
⎥
⎢ 2.6548 5.8833 3.3111 21.0667 10.0833 8 3.3270 1.4060 ⎥
⎢ ⎥
⎣ 11.7222 33.7333 17.5833 100.0000 58.8333 50.9167 8 5.0393 ⎦
20.2095 64.6000 29.5000 165.0000 96.6667 92.0000 40.2762 8

123
A direct consistency test and improvement method for the…

bi j
C = (c i j )n×n =
nai j n×n
⎡ ⎤
1 0.6954 0.5683 1.8452 1.1146 0.9774 8.8375 2.4280
⎢ 2.1230 1 1.7241 0.9236 0.7385 0.5475 2.9167 1.8615 ⎥
⎢ ⎥
⎢ 8.2839 1.7750 1 2.0403 2.4250 1.6625 0.3202 3.6429 ⎥
⎢ ⎥
⎢ 1.2312 2.1238 1.2382 1 1.7205 2.1274 1.5225 0.5163 ⎥
=⎢
⎢ 1.7173
⎥
⎢ 2.0375 1.0385 0.6486 1 1.7708 2.1151 0.7857 ⎥
⎥
⎢ 1.9911 2.2062 1.6556 0.6583 0.6302 1 2.0794 1.0545 ⎥
⎢ ⎥
⎣ 0.4884 0.8433 13.1875 1.7857 1.4708 1.2729 1 1.2598 ⎦
0.6315 1.1536 0.7375 2.5781 2.0139 1.9167 2.5173 1
   
b na
Let ci j = max c i j , c 1i j , , namely, C = (ci j )n×n , ci j = max nai ij j , bi ij j , then we
have
⎡ ⎤
1 1.4380 1.7597 1.8452 1.1146 1.0231 8.8375 2.4280
⎢ 2.1230 1 1.7241 1.0828 1.3541 1.8264 2.9167 1.8615 ⎥
⎢ ⎥
⎢ 8.2839 1.7750 1 2.0403 2.4250 1.6625 3.1233 3.6429 ⎥
⎢ ⎥
⎢ 1.2312 2.1238 1.2382 1 1.7205 2.1274 1.5225 1.9370 ⎥
C =⎢
⎢ 1.7173
⎥
⎢ 2.0375 1.0385 1.5418 1 1.7708 2.1151 1.2727 ⎥
⎥
⎢ 1.9911 2.2062 1.6556 1.5190 1.5868 1 2.0794 1.0545 ⎥
⎢ ⎥
⎣ 2.0474 1.1858 13.1875 1.7857 1.4708 1.2729 1 1.2598 ⎦
1.5834 1.1536 1.3559 2.5781 2.0139 1.9167 2.5173 1

It can be ascertained that the largest element in the matrix C is c73 = 13.1875, which
means that there is a most serious inconsistency in the corresponding position of the
matrix A; therefore, the corresponding element a73 in the matrix A should be adjusted
first. The remaining steps are similar to Example 5; please refer to Step 3 and Table 8
in Example 5.

It is obvious that adjusting a73 and a37 lead to the same results, and thus the two
methods, i.e., the 3 tuples iterative method and the improved iterative method, lead to
the same results.

Example 7 Consider the following 16 × 16 matrix,

A = (ai j )16×16
⎡ ⎤
1 1 2 4 1/4 3 1/8 1/9 1/9 5 1/9 1/7 1/3 1/2 1/6 1/5
⎢ 1 1 2 3 1/5 3 1/9 1/9 1/9 4 1/2 1/8 1/4 1/3 1/7 1/6 ⎥
⎢ ⎥
⎢ 1/2 1/2 1 2 1/6 1 1/9 1/9 1/9 3 1/3 1/9 1/5 1/4 1/8 1/7 ⎥
⎢ ⎥
⎢ 1/4 1/3 1/2 1 1/8 1 1/9 1/9 1/9 2 1/5 1/9 1/7 1/6 1/9 1/9 ⎥
⎢ ⎥
⎢ 4 5 6 8 1 7 1/4 1/5 1/6 9 3 1/3 1 2 1/2 1 ⎥
⎢ ⎥
⎢ 1/3 1/3 1/8 ⎥
⎢ 1 1 1/7 1 1/9 8 1/9 2 1/4 1/9 1/6 1/5 1/9 ⎥
⎢ 8 3 ⎥
⎢ 9 9 9 4 9 1 1 1/2 9 7 1 5 6 2 ⎥
⎢ ⎥
⎢ 9 9 9 9 5 1/8 1 1 1 9 8 2 6 7 3 4 ⎥
= ⎢ ⎥
⎢ 9 9 9 9 6 9 2 1 1 1/7 9 3 7 8 4 5 ⎥
⎢ ⎥
⎢ 1/5 1/4 1/3 1/2 1/9 1/2 1/9 1/9 7 1 1/6 1/9 1/8 1/7 1/9 1/9 ⎥
⎢ ⎥
⎢ 9 2 3 5 1/3 4 1/7 1/8 1/9 6 1 1/6 1/2 1 1/5 1/4 ⎥
⎢ ⎥
⎢ 7 8 9 9 3 9 1 1/2 1/3 9 6 1 4 5 1 2 ⎥
⎢ ⎥
⎢ 3 4 5 7 1 6 1/5 1/6 1/7 8 2 1/4 1 1 1/3 1/2 ⎥
⎢ ⎥
⎢ 2 3 4 6 1/2 5 1/6 1/7 1/8 7 1 1/5 1 1 1/4 1/3 ⎥
⎢ ⎥
⎣ 6 7 8 9 2 9 1/2 1/3 1/4 9 5 1 3 4 1 1 ⎦
5 6 7 9 1 8 1/3 1/4 1/5 9 4 1/2 2 3 1 1

123
 K. Xu, J. Xu

Step 1 Calculate Saaty’s consistency ratio. C R(A) = 0.4473 > 0.1, so A is not of
an acceptable consistency, go to Step 2.
Step 2 Calculate
 B= A2 = (bi j )n×n and n A, let C = (ci j )n×n ,
b na
ci j = max nai ij j , bi ij j , calculate the matrix C.
By using MATLAB, it is found that

⎡
1.0000 1.0952 1.2643 1.7729 1.4555 1.5625 1.3856 14.7147 20.7002
⎢ 1.1076 1.0000 1.3587 1.4218 1.6590 1.6946 1.4022 14.5683 16.6871
⎢
⎢ 1.6039 1.4901 1.0000 1.4343 1.5426 1.1707 1.0262 5.2808 12.4058
⎢
⎢ 2.4419 1.7681 1.5289 1.0000 1.7269 1.1836 1.3341 5.0653 8.3203
⎢
⎢ 1.1214
⎢ 1.3801 1.1222 1.0115 1.0000 1.0489 1.8252 19.3072 25.5577
⎢ 15.3431
⎢ 15.2308 5.2815 5.5761 18.8524 1.0000 5.2391 7.5379 12.8049
⎢ 1.4848
⎢ 1.1406 1.5417 2.1806 1.4208 1.8273 1.0000 5.2919 9.2789
⎢
⎢ 1.5874 1.4013 1.8168 2.5460 1.4322 154.8750 1.1707 1.0000 4.6991
C =⎢
⎢ 1.9012 1.7121 2.2121 3.0074 1.3918 2.5778 1.4333 5.5459 1.0000
⎢
⎢ 22.0632 17.5877 13.5650 9.3964 24.9921 9.1244 8.3466 6.5703 7.7111
⎢
⎢ 3.8918 1.0410 1.1096 1.0766 1.8174 1.0927 2.0476 17.8936 25.5272
⎢
⎢ 1.3665 1.0098 1.2431 1.7847 1.3613 1.5039 1.1996 10.3205 13.5650
⎢
⎢ 1.0935 1.4628 1.2276 1.1626 1.2720 1.1800 1.8270 19.7363 26.2984
⎢
⎢ 1.1212 1.4468 1.2961 1.3249 1.2383 1.3038 1.7467 19.1101 26.1158
⎢
⎣ 1.2635 1.1200 1.1113 1.4323 1.1646 1.2104 1.3830 15.1118 17.6919
1.1804 1.2554 1.0086 1.1323 1.3126 1.0768 1.6868 17.7696 21.6856
⎤
1.6061 6.1556 1.3994 1.3799 1.1445 1.4301 1.4581
1.3895 1.2532 1.4491 1.6768 1.5971 1.5189 1.5972 ⎥
⎥
1.6680 1.4629 1.1796 1.6274 1.6407 1.3023 1.4278 ⎥
⎥
1.6216 2.0329 1.0672 1.9078 2.0399 1.1933 1.5272 ⎥
⎥
1.1568 1.4482 1.5931 1.3230 1.1394 1.2914 1.2578 ⎥
⎥
2.9269 17.5081 9.8690 19.4811 19.0666 14.5763 17.2130 ⎥
⎥
2.6394 1.1614 1.2636 1.3417 1.2177 1.2162 1.4010 ⎥
⎥
⎥
2.9472 1.0593 1.3155 1.2889 1.1364 1.5106 1.5262 ⎥
⎥
218.4375 1.0457 1.6680 1.2127 1.0486 1.6816 1.5642 ⎥
⎥
1.0000 25.4536 12.4058 26.0801 26.2620 16.5308 20.7245 ⎥
⎥
1.1873 1.0000 2.0086 1.5716 1.0361 1.9789 1.9296 ⎥
⎥
2.2121 1.2916 1.0000 1.3800 1.3079 1.2844 1.1862 ⎥
⎥
1.0084 1.2589 1.6823 1.0000 1.3241 1.5127 1.2557 ⎥
⎥
1.1387 1.2119 1.6664 1.2735 1.0000 1.5851 1.4632 ⎥
⎥
1.8162 1.4085 1.2225 1.3329 1.3544 1.0000 1.3001 ⎦
1.4637 1.4773 1.3330 1.1479 1.3156 1.2415 1.0000

It can be ascertained that the largest element in the matrix C is c9,10 = 218.4375,
which means that there is a most serious inconsistency in the corresponding position
of the matrix A; therefore, the corresponding element a9,10 in the matrix A should be
adjusted first.
Step 3 Adjust a9,10 to a9,10 −ad justed = 9 (a10,9 −ad justed = 1/9). At this point,
C R(A) = 0.2598. Furthermore, in a same way, the second iteration is carried out, a8,6
is adjusted to a8,6−ad justed = 9 (a6,8−ad justed = 1/9) with C R(A) = 0.0623 < 0.1;
at this moment, the adjusted matrix A is of an acceptable consistency.
In Example 7, the required computing time is very short and the calculation effi-
ciency is quite high.

123
A direct consistency test and improvement method for the…

5.2 An actual case study

In this section, an actual case study is used for demonstrating the effectiveness of the
proposed approach.
Example 8 Many companies have recently tried to select suitable suppliers in order
to improve product quality. Therefore, supplier selection has become a key issue that
both enterprises and scholars are paying attention to. Supplier selection is particularly
important for seafood companies. Accordingly, in this case study a seafood company
needs to select a supplier. There are ten companies, x1 , x2 , ..., x9 and x10 , which can
be selected (the companies’ names will not be released due to confidentiality).
To select the best supplier, the seafood company employs a consultancy firm to
evaluate the ten competing suppliers. The experts provide their preference information
regarding the alternatives as follows:
⎡ ⎤
1 4 5 8 5 6 1/3 1/5 3 1/2
⎢ 1/4 1 6 5 6 3 1/5 1/8 4 1/4 ⎥
⎢ ⎥
⎢ 1/5 1/6 1 1/3 1/3 1/4 1/6 1/6 1 1/8 ⎥
⎢ ⎥
⎢ 1/8 1/5 3 1 1/3 1/4 1/7 1/7 1 1/5 ⎥
⎢ ⎥
⎢ 1/5 1/6 3 3 1 1/2 1/5 1/6 6 1/6 ⎥
A = (ai j )10×10 =⎢
⎢ 1/6
⎥
⎢ 1/3 4 4 2 1 1/5 1/7 1 1/3 ⎥
⎥
⎢ 3 5 6 7 5 5 1 1/2 8 1/3 ⎥
⎢ ⎥
⎢ 5 8 6 7 6 7 2 1 6 6 ⎥
⎢ ⎥
⎣ 1/3 1/4 1 1 1/6 1 1/8 1/6 1 1/7 ⎦
2 4 8 5 6 3 3 1/6 7 1

Step 1 It is found that C R(A) = 0.1336, so A is not of an acceptable consistency

and should be adjusted.
Step
 2 Calculate
 B = A2 = (bi j )n×n and n A, let C = (ci j )n×n , ci j =
b na
max nai ij j , bi ij j , calculate the matrix C.
By using MATLAB, it is found that the largest element in the matrix C is c95 =
5.1893, hence there is a most serious inconsistency, and a95 should be adjusted first.
Step 3 Adjust a95 to a95-ad justed = 1 (a59-ad justed = 1). At this point, C R(A) =
0.1157. Furthermore, in a same way, the second iteration is carried out, a10,8 is adjusted
to a10,8-ad justed = 1 (a8,10-ad justed = 1) with C R(A) = 0.0993 < 0.1; at this
moment, the adjusted matrix A is of an acceptable consistency.
Step 4 By using the row geometric mean method (RGMM) developed by Crawford
and Williams (1985), the weight vector for x1 , x2 , . . . , x9 and x10 is obtained:

(1.85, 1.13, 0.28, 0.35, 0.49, 0.67, 2.71, 3.92, 0.44, 3.22)

therefore these companies can be sorted as follows:

x8 x10 x7 x1 x2 x6 x5 x9 x4 x3

which indicates that x8 is the most desirable according to the consultancy firm.

123
 K. Xu, J. Xu

5.3 Comparisons and discussion

When comparing the two methods, i.e., the 3 tuples iterative method and the improved
iterative method, one can find that the 3 tuples iterative method can help in ascertaining
the internal quantitative relationships between the consistency ratio of the comparison
matrix and the 3 tuples. Furthermore, the 3 tuples iterative method is intuitional, easy
to understand, and more suitable for manual calculations, especially for preference
relations with low orders and some specific preference relations with high orders;
e.g. Examples 2, 3 and 4. The improved iterative method is also intuitional, easy to
understand, and easy to use for preference relations with high orders, e.g. Example 7.
However, the matrix operation utilizes the power of the algebraic
 operation.

ai j a a
The 3 tuples iterative method utilizes the ratios max aik ak j , ikai jk j (i, j, k =
1, 2, . . . , n, i < k < j); while the improved iterative method utilizes the ratios
n
nai j k=1 aik ak j
max n , nai j (i, j, k = 1, 2, . . . , n). In summary, the two methods
k=1 aik ak j
are closely linked and each has its own different perspective; moreover, they can be
applied on different occasions and reinforce one another.
It should be noted that since in Ergu et al’s method (2011) they reached the conclu-
sion that A2 − n A = 0, which is equivalent to our conclusion A2 = n A, the proposed
improved iterative method and Ergu et al’s method have a similar theoretical basis.
However, the latter’s method is based on the principle that the elements in the matrix
A2 − n A are near to 0 if A is close to being of a perfect consistency; while our pro-
posed improved iterative method is based on the principle that the ratios between the
corresponding elements in the matrices A2 and n A should be near to 1 if A is close to
being of a perfect consistency. Moreover, the proposed improved iterative method is
also more intuitional and easy to understand than their method.

6 Conclusions

In this paper, an intuitional method has been proposed to illustrate the consistency
problem from the perspective of 3 tuples (aik , ak j , ai j ), which does not require matrix
operations. We have found that in a pairwise comparison matrix, if all the 3 tuples
a
(aik , ak j , ai j ) satisfy 0.382 < aikiajk j < 2.620, then the pairwise comparison matrix
is of an acceptable consistency (Saaty’s consistency ratio is less than 0.1). Some
related theorems and propositions have been proved mathematically or verified by
random simulations. The effectiveness of the method has also been demonstrated
through a comparison with several published examples. Furthermore, an improved
iterative method has been proposed, which is based on the proposed 3 tuples method.
The proposed improved iterative method is also intuitional, easy to understand, and
convenient for preference relations with high orders.
Compared to the existing methods, the proposed method in this paper has the
following advantages and features:
(1) The internal quantitative relationships between the consistency ratio of the pair-
wise comparison matrix and the 3 tuples have been ascertained. A new method,

123
A direct consistency test and improvement method for the…

from the perspective of 3 tuples, is provided to test the consistency of a pairwise

comparison matrix. We divide an n × n pairwise comparison matrix into 3 tuples
(aik , ak j , ai j ).
(2) This paper presents a simpler algorithm for the consistency test. The proposed 3
tuples iterative method is more intuitional and easy to understand, as only simple
mathematical calculations are needed, without the need for matrix operations. The
calculation of Saaty’s consistency ratio is only conducted to verify the proposed
method.
(3) The proposed method can also retain most of the information provided by the
original comparison matrix, show the modification direction and provide the opti-
mal values. At the same time, by using the proposed method, the inconsistent
elements in the matrix can be found accurately and rapidly.
(4) The proposed improved iterative method is also intuitional, easy to understand,
and convenient for preference relations with high orders.

The two methods are closely linked and each has its own different perspective;
moreover, they are applied on different occasions and reinforce one another.
In future research, some similar methods may be applied to other related issues
regarding the pairwise comparison matrix; for example,the consensus of the pairwise
comparison matrix.

Appendix 1

S=0,v=[1/9 ,1/8 ,1/7 ,1/6 ,1/5 ,1/4 ,1/3 ,1/2 ,1 ,2 ,3 ,4 ,5 ,6 ,7 ,8 ,9];

for i =1:length(v) for j =1:length(v) for m=1:length(v)

x12=v( i ) ,x13=v( j ) ,x23=v(m) ,x21=1/x12,x31=1/x13,x32=1/x23;

B=[1,x12,x13;x21,1 ,x23;x31,x32, 1 ] ; [x,y]=eig (B) , eigenvalue=diag(y) ;
Lamda=eigenvalue(1) ,Lamda1=Lamda−3,CI=Lamda1/2 , CR=CI/0.52 ,

i f 0.505<x13/ (x12∗x23) & x13/ (x12∗x23)<1.979

k=1;
else
k=0
end

i f CR<0.05,
CR1=1;
else
CR1=0,
end

i f k==CR1,
S=S+1;
else
end
end
end
end

>> S
S=4913

123
 K. Xu, J. Xu

Appendix 2

>> syms x
x=solve ( 1 . /x+x==power(2.104,3)−3∗2.104)

x =

3379975011117215/2251799813685248 − (3∗705958741651544450761739614969^(1/2))/2251799813685248
(3∗705958741651544450761739614969^(1/2))/2251799813685248 + 3379975011117215/2251799813685248

>> 3379975011117215/2251799813685248 − (3∗705958741651544450761739614969^(1/2))/2251799813685248

ans = 0.3816

>> (3∗705958741651544450761739614969^(1/2))/2251799813685248 + 3379975011117215/2251799813685248

ans = 2.6204

>> S=0,N=1000000;for i =1:N;

x12=(9−(1/9))∗rand(1 ,1) +(1/9) , x13=(9−(1/9))∗rand(1 ,1) +(1/9) ,

x23=(9−(1/9))∗rand(1 ,1) +(1/9) ,

x21=1/x12,x31=1/x13, x32=1/x23,

B=[1,x12,x13;x21,1 ,x23;x31,x32, 1 ] ; [x,y]=eig (B) , eigenvalue=diag(y) ;

Lamda=eigenvalue(1) ,Lamda1=Lamda−3,CI=Lamda1/2 , CR=CI/0.52 ,

m=3379975011117215/2251799813685248−(3∗705958741651544450761739614969^(1/2))/2251799813685248
n=(3∗705958741651544450761739614969^(1/2))/2251799813685248+3379975011117215/2251799813685248

i f m<x13/ (x12∗x23) & x13/ (x12∗x23)<n

k=1;
else
k=0
end

i f CR<0.1,
CR1=1;
else
CR1=0,
end

i f k==CR1,
S=S+1;
else
end
end

>> S
S=1000000

Appendix 3
S=0,N=10000000;for i =1:N;
a=[1/9 ,1/8 ,1/7 ,1/6 ,1/5 ,1/4 ,1/3 ,1/2 ,1 ,2 ,3 ,4 ,5 ,6 ,7 ,8 ,9];
c=randperm(numel(a) ) ;b=a(c(1:10) ) ;
x12=b(1 ,1) ,x13=b(1 ,2) ,x14=b(1 ,3) ,x23=b(1 ,5) ,x24=b(1 ,6) ,x34=b(1 ,8)

x21=1/x12,x31=1/x13,x41=1/x14,x32=1/x23,x42=1/x24, x43=1/x34,

B=[1,x12,x13,x14;x21,1 ,x23,x24;x31,x32,1 ,x34;x41,x42,x43,1]

[x,y]=eig (B) , eigenvalue=diag(y) ; lamda=eigenvalue(1) ,
CI=(lamda−4)/3 ,CR=CI/0.89 ,

m=3379975011117215/2251799813685248−(3∗705958741651544450761739614969^(1/2))/2251799813685248
n=(3∗705958741651544450761739614969^(1/2))/2251799813685248+3379975011117215/2251799813685248

123
A direct consistency test and improvement method for the…

i f (max(x13/ (x12∗x23) ,(x12∗x23) /x13)

+max(x24/ (x23∗x34) ,(x23∗x34) /x24)+max(x14/ (x12∗x24) ,(x12∗x24) /x14)
+max(x14/ (x13∗x34) ,(x13∗x34) /x14) )<(4∗n) ,
k=1;
else
k=0
end

i f CR<0.1,
CR1=1;
else
CR1=0,
end

i f k==CR1,
S=S+1;
else
end
end
>> S

S=0,N=10000000;for i =1:N;
a=[1/9 ,1/8 ,1/7 ,1/6 ,1/5 ,1/4 ,1/3 ,1/2 ,1 ,2 ,3 ,4 ,5 ,6 ,7 ,8 ,9];
c=randperm(numel(a) ) ;b=a(c(1:10) ) ;
x12=b(1 ,1) ,x13=b(1 ,2) ,x14=b(1 ,3) ,x15=b(1 ,4) ,x23=b(1 ,5) ,x24=b(1 ,6) ,
x25=b(1 ,7) ,x34=b(1 ,8) ,x35=b(1 ,9) ,x45=b(1,10) ,

x21=1/x12,x31=1/x13,x41=1/x14,x51=1/x15,x32=1/x23,x42=1/x24,
x52=1/x25, x43=1/x34, x53=1/x35, x54=1/x45

B=[1,x12,x13,x14,x15;x21,1 ,x23,x24,x25;x31,x32,1 ,x34,x35;x41,x42,x43,1 ,x45;x51,x52,x53,x54,1]

[x,y]=eig (B) , eigenvalue=diag(y) ; lamda=eigenvalue(1) ,
CI=(lamda−5)/4 ,CR=CI/1.11 ,

m=3379975011117215/2251799813685248−(3∗705958741651544450761739614969^(1/2))/2251799813685248
n=(3∗705958741651544450761739614969^(1/2))/2251799813685248+3379975011117215/2251799813685248

i f (max(x13/ (x12∗x23) ,(x12∗x23) /x13)

+max(x24/ (x23∗x34) ,(x23∗x34) /x24)+max(x14/ (x12∗x24) ,(x12∗x24) /x14)
+max(x14/ (x13∗x34) ,(x13∗x34) /x14)+max(x15/ (x12∗x25) ,(x12∗x25) /x15)
+max(x15/ (x13∗x35) ,(x13∗x35) /x15)+max(x15/ (x14∗x45) ,(x14∗x45) /x15)
+max(x25/ (x23∗x35) ,(x23∗x35) /x25)+max(x25/ (x24∗x45) ,(x24∗x45) /x25)
+max(x35/ (x34∗x45) ,(x34∗x45) /x35) )<(10∗n) ,
k=1;
else
k=0
end

i f CR<0.1,
CR1=1;
else
CR1=0,
end

i f k==CR1,
S=S+1;
else
end
end
>> S

References
Barzilai, J. (1997). Deriving weights from pairwise comparison matrices. Journal of the Operational
Research Society, 48, 1226–1232.

123
 K. Xu, J. Xu

Bryson, N. (1995). A goal programming method for generating priority vectors. Journal of the Operational
Research Society, 46, 641–648.
Cao, D., Leung, L. C., & Law, J. S. (2008). Modifying inconsistent comparison matrix in analytic hierarchy
process: A heuristic approach. Decision Support Systems, 44, 944–953.
Chiclana, F., Herrera-Viedma, E., Alonso, S., & Herrera, F. (2009). Cardinal consistency of reciprocal
preference relations: A characterization of multiplicative transitivity. IEEE Transactions on Fuzzy
Systems, 17, 14–23.
Chu, A., Kalaba, R., & Springam, K. (1979). A comparison of two methods for determining the weights of
belonging to fuzzy sets. Journal of Operator Theory and Applications, 27, 531–541.
Crawford, G., & Williams, C. (1985). A note on the analysis of subjective judgment matrices. Journal of
Mathematical Psychology, 29, 387–405.
Ergu, D., Kou, G., Peng, Y., & Shi, Y. (2011). A simple method to improve the consistency ratio of the
pair-wise comparison matrix in ANP. European Journal of Operational Research, 213(1), 246–259.
Goyal, R. K., & Kaushal, S. (2018). Deriving crisp and consistent priorities for fuzzy AHP-based multicri-
teria systems using non-linear constrained optimization. Fuzzy Optimization and Decision Making,
17, 195–209.
Krejčí, J., Pavlačka, O., & Talašová, J. (2017). A fuzzy extension of analytic hierarchy process based on
the constrained fuzzy arithmetic. Fuzzy Optimization and Decision Making, 16, 89–110.
Kwiesielewicz, M., & Uden, E. (2002). Problem of inconsistent and contradictory judgements in pairwise
comparison method in sense of AHP. In P. M. A. Sloot, et al. (Eds.), Computational science-ICCS
2002: Lecture notes in computer science (Vol. 2329, pp. 468–473). Berlin: Springer.
Petra, G., & Lidija, Z. S. (2012). Acceptable consistency of aggregated comparison matrices in analytic
hierarchy process. European Journal of Operational Research, 223, 417–420.
Ramík, J. (2018). Strong reciprocity and strong consistency in pairwise comparison matrix with fuzzy
elements. Fuzzy Optimization and Decision Making, 17, 337–355.
Saaty, T. L. (1977). A scaling method for priorities in hierarchical structures. Journal of Mathematical
Psychology, 15(3), 234–281.
Saaty, T. L. (1978). Modeling unstructured decision problems—The theory of analytical hierarchies. Math-
ematics and Computers in Simulation, 20, 147–158.
Saaty, T. L. (1979). Applications of analytical hierarchies. Mathematics and Computers in Simulation, 21,
1–20.
Saaty, T. L. (1980). The analytical hierarchy process. New York: McGraw-Hill.
Saaty, T. L. (1986). Axiomatic foundation of the analytic hierarchy process. Management Science, 32(7),
841–855.
Saaty, T. L. (1990). How to make a decision: The analytic hierarchy process. European Journal of Opera-
tional Research, 48(1), 9–26.
Saaty, T. L. (2001). Deriving the AHP 1-9 scale from first principles. In ISAHP 2001 Proceedings, Bern,
Switzerland.
Saaty, T. L. (2003). Decision-making with the AHP: Why is the principal eigenvector necessary. European
Journal of Operational Research, 145(1), 85–91.
Wu, Z., Huang, S., & Xu, J. (2019). Multi-stage optimization models for individual consistency and group
consensus with preference relations. European Journal of Operational Research, 275(1), 182–194.
Xu, Z., & Wei, C. (1999). A consistency improving method in the analytic hierarchy process. European
Journal of Operational Research, 116, 443–449.
Yager, R. R., & Alajlan, N. (2015). On the consistency of fuzzy measures in multi-criteria aggregation.
Fuzzy Optimization and Decision Making, 14, 121–137.
Zhang, H., Chen, X., Dong, Y., Xu, W., & Wang, S. (2018). Analyzing Saatys consistency test in pairwise
comparison method: A perspective based on linguistic and numerical scale. Soft Computing, 22, 1933–
1943.
Zhu, B., Xu, Z., Zhang, R., & Hong, M. (2016). Hesitant analytic hierarchy process. European Journal of
Operational Research, 250, 602–614.

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

123