Shihui Wu et al.: Marginal optimization method to improve the inconsistent comparison matrix in the analytic hierarchy process 1143
denoted as a = {a
12
,...,a
1n
,a
23
,...,a
2n
,...,a
(n−1)n
},
and b = {b
12
,...,b
1n
,b
23
,...,b
2n
,...,b
(n−1)n
} is called
the superior decision vector, which is defined as
b
ij
=
⎧
⎨
⎩
a
ij
,a
ij
1
1
a
ij
, otherwise
.
Obviously, b
ij
1,forallb
ij
∈ b.
Then, the original PCM A =(a
ij
)
n×n
can be uniquely
determined by vector a or b.
Let c = {c
12
,...,c
1n
,c
23
,...,c
2n
,...,c
(n−1)n
} be the
superior decision vector of the revised PCM, and the max-
imum revision σ is defined as the maximum deviation be-
tween the revised PCM and original PCM [13]:
σ =max{|c
ij
− b
ij
|},c
ij
∈ c,b
ij
∈ b. (3)
Obviously, the dimension of vectors a, b and c are
(n − 1) + (n − 2) + ···+1 = n(n − 1)/2.Forexam-
ple, if A is a 4×4 PCM, it can be uniquely determined by
a six dimension decision vector.
3. Marginal optimization method
3.1 Tolerance deviations interval
The maximum revision in (3) that is less than 1 is called
within the tolerance deviations interval [13]. Accordingly,
tolerance deviations intervals are listed in Table 2. For ex-
ample, when i is strongly important to j, the scale of a
ij
is 5 , and the to lerance deviations interval is [4,6), which
means a
ij
can b e substituted by a ny real number on the in -
terval [4,6). Since we only consider a
ij
1 here, when i is
equally important to j (a
ij
=1), the tolerance deviations
interval is set to [1,2). For a
ij
< 1, the tolerance deviations
interval can be obtained by [a
L
ij
,a
U
ij
]=[1/a
U
ji
, 1/a
L
ji
].
Table 2 Linguistic description of scale and its corresponding toler-
ance deviations interval
Linguistic description Scale
Tolerance
de viations
interval
Extreme importance 9 [8,10]
Demonstrated importance 7 [6,8)
Strong importance 5 [4,6)
Moderate importance 3 [2,4)
Equal importance 1 [1,2)
From Table 2, we can see that Saaty’s one to nine scales
is using integer number ranging from 1 to 9 to quantify the
linguistic description of importance, while the tolerance
deviations interval actually extends the integer number to
an acceptable interval. Also, Fig. 2 gives the correspond-
ing relationship between tolera nce deviations intervals and
linguistic descriptions of importance scales.
Fig. 2 Relationship between tolerance deviations interval and lin-
guistic description
3.2 Marginal effect analysis
Marginal effect analysis is a b asic mathematic method
widely used in economy issues. Because many of the vari-
ables of interest are interaction terms, to see the over-
all effect of a condition on growth, one must look at
the marginal effect of that variable. Besides economy
problems, marginal optimization based on marginal effect
analysis has found many applications, such as the optimal
spare parts inventory problems [14,15].
The main id e a of the marginal optimizatio n method to
improve the inconsistent PCM is to increase (or decrease)
one element a
ij
in superior decision vector by a fixed value
d for all i, j (d is usually set to be 0.1, 0.01, 0.001, and
we will have more discussion on this in Section 3.3 and
Section 4), and calculate the marginal effect of each mo-
dification a
ij
± d, that is, the ratios of the reduction of CR
caused by modifying one element to a
ij
±d, i.e., we denote
as CR(a
ij
)−CR(a
ij
±d), to maximum revision σ derived
by (3) for the modified matrix. The modification with the
biggest ratio means that it is of greatest marginal effect,
or benefit/cost ratio, in other words, it is sensible to revise
this element of the superior decision vector to improve CR.
Therefore, this element should increase (or decrease) by d,
and whether it increases or decreases depends on if it is
able to reduce CR to a lower value or not.
Theorem 1 Let A =[a
ij
]
n×n
be a continuous PCM
that is inconsistent with CR(A) > 0.1, then there must be
one element a
ij
that can make CR(A) decrease when a
ij
is modified.
Proof According to Saaty [ 7], let r
ij
=
∂λ
max
∂a
ij
:
r
ij
=
v
i
u
j
− u
i
v
j
a
−2
ij
,a
ij
1
v
i
u
j
a
2
ij
− u
i
v
j
,a
ij
< 1
(4)
where u =(u
1
,u
2
,...,u
n
)
T
and ν =(ν
1
,ν
2
,...,ν
n
)
T
are the principal right eigenvectors of A and A
T
respec-
tively. Obviously, when A is in consistent with CR(A) >
0.1,matrixR =(r
ij
)
n×n
holds:
r
ij
=
0,i= j
−r
ji
,i= j
. (5)