Professional Documents
Culture Documents
Comment On Some Mathematical Properties of A DEA Model For The Joint Determination of Ef Ficiencies
Comment On Some Mathematical Properties of A DEA Model For The Joint Determination of Ef Ficiencies
Comment On Some Mathematical Properties of A DEA Model For The Joint Determination of Ef Ficiencies
1. Introduction (see model (1)) is convex, and thus the program is a convex
program, is not correct. It should be noted that the conclusion
The data envelopment analysis (DEA) is a mathematical
drawn in Molinero and Tsai (1997) should be applicable
programming technique that can be used to evaluate the relative
potentially not only to the model given by the authors, but also
efficiencies of a set of decision-making units (DMUs) involving
arguably to all the researches concerning a production entity
multiple inputs and outputs. The technique was originally
with shared inputs and/or outputs, if it were correct. Because of
introduced by Charnes et al (1978) (called CCR model). As is
its potential influence, we think it is worthwhile to show this
known, the CCR model captures both technical and scale
incorrectness and the significance of this incorrectness.
inefficiencies. Banker et al (1984) proposed a new approach
This paper unfolds as follows: in the next section, we revisit
(called BCC model), which extended the CCR model by
the model and point out the incorrect logic in establishing the
separating technical efficiency and scale efficiency. Since its
convexity; in Section 3, we present a counterexample and the
inception in 1978 under the name of DEA, it has been extended
significance of this incorrectness; Section 4 summarizes
from both a theoretical and a practical perspective.
this paper.
The evaluation of a production entity with shared inputs and/
or shared outputs attracts the attention of researchers (see, eg,
the recent works by Imanirad et al (2013) and Cherchye et al
(2013)). In this line of studies, Beasley (1995) first showed an 2. The model revisited
application and formulated a rigorous model characterized by
joint consumption of some inputs by multiple activities. In We now recapitulate the model considered in Molinero and
Molinero and Tsai (1997), the authors extended Molinero Tsai (1997) and Molinero (1996). The interested readers should
(1996) by looking into the estimation strategy through refer to the original works for more information. To study a
establishing the convexity and the first-order Kuhn-Tucker production with shared inputs and shared outputs, Molinero and
conditions. Tsai (1997) and Molinero (1996) presented model (1) below to
In the current work, without reducing the interestingness of evaluate the production system with two activities: the teaching
Molinero and Tsai (1997), as the establishment of Kuhn-Tucker activity and researching activity (see Figure 1).
condition is interesting by itself, we point out that the In Figure 1, inputT and inputR indicate the inputs consumed,
assertion that the feasible region of the mathematical program respectively, by Teaching and Researching activity. Similarly,
outputT and outputR indicate the outputs from, respectively,
*Correspondence: Chenpeng Feng, School of Management, University of
Teaching and Researching activity. Finally, sharedinput denotes
Science and Technology of China, Hefei, Anhui Province 230026, P.R. China. the inputs jointly consumed by Teaching and Research activity,
E-mail: dxx@mail.ustc.edu.cn and sharedoutput denotes the outputs produced together by
Jingjing Ding et al—Comment on some mathematical properties of a DEA model 1961
X
S Note that notations follow exactly Molinero and Tsai (1997).
s:t: λTs xTas ⩽ xTak The peculiarities of model (1) are that two radial expansion
s=1 variables (wT and wR) are utilized to measure the equipropor-
tionate expansion potentials of the Teaching activity and
X
S
λRs xRbs ⩽ xRbk Researching activity. In addition, user provided weights are
s=1 applied to incorporate a decision maker’s preference informa-
tion. By using matrices A, DT, DR, b, xT and xR, Molinero and
X
S X
S
μc λTs xTR Tsai (1997) were able to give a matrix form for the constraints
cs + ð1 - μc ÞλRs xTR
cs ⩽ xck
TR
X
S
λTs yTds ⩾ wT yTdk DT AxT + DR AxR ⩽ b; (2)
s=1
X
S where
λRs yRes ⩾ wR yRek
s=1
0 1
xTa1 xTa2 xTak xTaS 0
B C
X
S X
S B C
μf λTs yTR + 1 - μf λRs yTR B R C
fs ⩾ w μf yfk
T TR
fs B xb1 xRb2 xRbk xRbS 0 C
s=1 s=1 B C
B C
B C
+ wR 1 - μf yTR B xTR xTR xTR xTR 0 C
fk B c1 c2 ck cS C
B C
A=B C;
wT ⩾ 1; wR ⩾ 1; 0 ⩽ μc ⩽ 1; 0 ⩽ μf ⩽ 1; B T C
B - yd1 - yTd2 - yTdk - yTdS yTdk C
B C
B C
λTs ⩾ 0; λRs ⩾ 0: ð1Þ B C
B - yR - yRe2 - yRek - yReS yRek C
B e1 C
Here, B C
@ A
- yTR
f1 - yTR
f2 - yTR
fk - yTR
fS yTR
fk
a is an index assigned to inputs which relate solely to the
Teaching activity 0 1
b is an index assigned to inputs which relate solely to the 1 0 0 0 0 0
B C
Research activity B C
B C
c is an index assigned to shared inputs B0 0 0 0 0 0 C
B C
d is an index assigned to outputs which relate solely to the B C
B C
Teaching activity B0 0 μ 0 0 0 C
B c C
e is an index assigned to outputs which relate solely to the B C
D =B
T
C;
Research activity B C
B0 0 0 1 0 0 C
B C
f is an index assigned to joint outputs B C
B C
S is the total number of DMUs in the data set B0 0 0 C
B 0 0 0 C
k is the DMU under investigation B C
xTas is the amount of input a associated only with the @ A
Teaching activity of DMUs 0 0 0 0 0 μf
1962 Journal of the Operational Research Society Vol. 66, No. 12
0 1
P
10
0 1
inputT11
The feasibility status, True or False, is presented in the last
B C
B inputT21 C row of the table. Clearly, if LHS is less or equal to RHS for all
B C
B C
B inputT31 C constraints, then the corresponding solution is feasible, and we
B C
B C report True in the last row.
B inputR11 C
B C As we can see from Table 3, the convex combinations of two
B C
B inputR21 C
B C feasible solutions are not necessarily feasible solutions of
B C
B sharedoutput11 C model (1), since cvx_sol1 and cvx_sol2 are infeasible. We
B C
B C report the violations in boldface. Note also that these violations
B sharedoutput21 C
B C
B C are too large to be attributed to the rounding error as floating
B
RHS = B 0 C
C point numbers are handled less precisely than the operations on
B C
B0 C integers by a computer. Now we are in position to conclude that
B C
B C
B0 C the feasible region of model (1) is not a convex set, and the
B C
B C convexity argument in Molinero and Tsai (1997) is not correct.
B0 C
B C To interpret the significance of the incorrectness, a concrete
B C
B0 C
B C example is warranted. To facilitate, it is convenient to separate
B C
B0 C the constraints of model (1) into linear constraints and nonlinear
B C
B C ones, each of which has the form LHS ⩽ 0(See model (6)).
B0 C
@ A
The order of the constraints in model (6) is important in that
0 the Lagrangian multipliers provided in Tables 4 and 5 are
1964 Journal of the Operational Research Society Vol. 66, No. 12
associated with the constraints in this order. superscript T of boldfaced vectors and matrices indicates
transpose, while the superscript T of scalars such as λ and w
min - θT wT - θR wR
indicates they are relating to the teaching activity.
8 S Let η1, η2, η3 ⩾ 0 denote, respectively, Lagrangian multi-
> P T T
>
> λs xas - xTak ⩽ 0
>
> pliers at the solution d for the linear inequalities, nonlinear
>
> s=1
>
> inequalities associated with shared inputs and shared outputs.
>
>
>
> PS The entries of these vectors are ηij, where the first index
>
> λRs xRbs - xRbk ⩽ 0
>
> i = 1, …, 3 indicates the types of the constraints. Tables 4 and
>
> s = 1
>
> 5 report a specific set of values for Lagrangian multipliers.
>
>
>
> PS Then, d*, η*1, η*2 and η*3 satisfy the Kuhn-Tucker conditions
>
> - λTs yTds + wT yTdk ⩽ 0
>
> provided the following conditions hold.
>
> s=1
>
>
>
>
>
> PS Ad* + q⩽0
>-
> λRs yRes + wR yRek ⩽ 0
>
>
>
> s = 1 1 *T SI * SI *
>
> d Pc d + qc d + rc ⩽ 0
< 2
s:t: - wT + 1 ⩽ 0
>
> 1 *T SO * SO *
>
> - wR + 1 ⩽ 0 d Pf d + qf d ⩽ 0
>
>
>
> 2
>
>
>
> μc - 1 ⩽ 0
>
> ηT1 Ad* + q = 0
>
>
>
>
>
> μf - 1 ⩽ 0
>
>
1 *T SI * SI *
η2c d Pc d + qc d + rc = 0
>
>
>
> - λTs ⩽ 0
2
>
>
>
>
>
> R 1
>
> - λs ⩽ 0 η3f d*T PSO d*
+ qSO *
d =0
>
> f f
>
>
2
>
> 2
>
> - μc ⩽ 0
> P 1 T SI 3
>
>
: f T0 d + ηT1 ðAd + qÞ + η2c 2 d Pc d + qSI d
6 c 7
- μf ⩽ 0 ∂6
c 7
4 X 1 T SO 5
8 " # + η d Pf d + q f d
SO
>
> 1 X S XS f 3f 2
>
> 2μ λ x -
T TR
2μ λ x
R TR =0
>
> ∂d
> 2 s = 1 c s cs s = 1 c s cs
>
>
>
> X
>
>
S
>
< + λRs xTR
cs - xck ⩽ 0
TR
s=1 ð6Þ d = d* ; ηTi = η*T
>
i
>
> " #
>
> X S XS
η*1 ; η*2 ; η*3 ⩾ 0
>
>
1
2w μf yfk - 2w μf yfk - 2
T TR R TR
μf λs yfs + 2
T TR
μf λs yfs
R TR
>
>
>
>
2 s=1 s=1
>
> P
>
: -
S
λRs yTR + wR yTR ⩽0
fs fk
s=1 The last but one condition can be reduced to the following
equation:
By using proper matrices and vectors, we can rewrite
h iT
model (6) into (7) below: SI T *
f 0 + AT η*1 + qSO
f η *
3 + qc η2
min f T0 d
" #
s:t: Ad + q ⩽ 0 ðlinear contraintsÞ X X
+ η2c PSI
c + η3f PSO
f d* = 0
c f
1 T SI
d Pc d + qSI
c d + rc ⩽ 0
2 Now let θT = 0.7, θR = 0.3 and consider we measure the
ðnonlinear contraints associated with shared inputsÞ performance of DMU7. Both Sol6 and Sol7 in Table 2 are
feasible solutions of this particular case. Note that the objective
1 T SO function at Sol6, that is, − 1, is greater than that at Sol5, that is,
d Pf d + qSO
f d⩽0
2 − 1.26667. Obviously, Sol6 is not an optimal solution of
ðnonlinear contraints associated with shared outputsÞ ð7Þ model (6). Given the Lagrangian multipliers in Tables 4 and 5,
it is not difficult to show that the Kuhn-Tucker conditions are
where d = [λT1 , …, λTS , λR1 , …, λRS , wT, wR, μTc , μTf ]T is a column satisfied at Sol6. Therefore, we cannot claim that a solution
vector of the decision variables. It should be noted that the satisfying Kuhn-Tucker condition is a global optimum.
Jingjing Ding et al—Comment on some mathematical properties of a DEA model 1965
DMU1 5 6 10 2 5 7 6 14 4 12 4 2 2 4
DMU2 3 5 5 4 2 4 8 2 2 2 2 4 4 1
DMU3 10 1 7 4 1 7 2 6 5 4 12 10 6 7
DMU4 1 10 6 1 5 7 5 11 11 3 17 3 7 4
DMU5 1 5 8 5 2 7 6 3 7 15 20 4 5 7
DMU6 1 4 1 4 2 4 3 1 4 9 10 10 5 6
DMU7 3 9 10 2 4 8 1 11 12 9 1 7 6 1
DMU8 4 5 1 1 5 1 7 7 3 17 14 10 5 5
DMU9 10 10 8 2 4 3 2 15 12 17 4 7 5 5
DMU10 2 7 7 5 1 3 1 11 3 17 9 9 4 7
Table 2 Solutions
Table 4 Lagrangian multipliers at the solution Sol6 for the linear Table 5 Lagrangian multipliers at the solution Sol6 for the non-
inequalities linear inequalities
Lagrangian multipler (linear inequalities) Values Lagrangian multiplers (non-linear inequalities) Values