Dual Simplex Method For Linear Programming PDF

Fuzzy Sets and Systems 158 (2007) 1961 1978
www.elsevier.com/locate/fss
Duality results and a dual simplex method for linear programming

problems with trapezoidal fuzzy variables
N. Mahdavi-Amiria, b, , S.H. Nasseria
a Department of Mathematical Sciences, Sharif University of Technology, Tehran, Iran
b Institute for Studies in Theoretical Physics and Mathematics (IPM), Tehran, Iran
Received 8 October 2005; received in revised form 2 May 2007; accepted 7 May 2007
Available online 25 May 2007
Abstract
Linear programming problems with trapezoidal fuzzy variables (FVLP) have recently attracted some interest. Some methods
have been developed for solving these problems by introducing and solving certain auxiliary problems. Here, we apply a linear
ranking function to order trapezoidal fuzzy numbers. Then, we establish the dual problem of the linear programming problem with
trapezoidal fuzzy variables and hence deduce some duality results. In particular, we prove that the auxiliary problem is indeed the
dual of the FVLP problem. Having established the dual problem, the results will then follow as natural extensions of duality results
for linear programming problems with crisp data. Finally, using the results, we develop a new dual algorithm for solving the FVLP
problem directly, making use of the primal simplex tableau. This algorithm will be useful for sensitivity (or post optimality) analysis
when using primal simplex tableaus.
2007 Elsevier B.V. All rights reserved.
Keywords: Fuzzy variable linear programming; Duality; Ranking function; Dual simplex; Trapezoidal fuzzy number
1. Introduction
Fuzzy set theory has been applied to many disciplines such as control theory and management sciences, mathemat-
ical modeling and industrial applications. The concept of fuzzy mathematical programming on general level was rst
proposed by Tanaka et al. [25] in the framework of the fuzzy decision of Bellman and Zadeh [4]. The rst formulation
of fuzzy linear programming (FLP) is proposed by Zimmermann [31]. Afterwards, many authors considered various
types of the FLP problems and proposed several approaches for solving these problems [58,1620,26]. Chanas [7]
showed an application of parametric programming techniques in FLD and obtained the set of solutions maximizing
the objective function, being analytically dependent on a parameter. Delgado et al. [8] studied a general model for FLP
problems which includes fuzziness both in the coefcients and in the accomplishment of the constraints. Buckley and
Feuring [5] considered the extreme case of fully fuzzied linear program with all the parameters and variables as fuzzy
numbers. They turned the problem into a multi-objective FLP one. They showed that the fuzzy exible programming
could be used to explore the undominated set to the multi-objective problem and proposed an evolutionary algorithm
Corresponding author. Tel.: +98 21 66165607.

E-mail addresses: nezamm@sina.sharif.edu, nezamm@sharif.edu (N. Mahdavi-Amiri).
0165-0114/$ - see front matter 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.fss.2007.05.005
1962 N. Mahdavi-Amiri, S.H. Nasseri / Fuzzy Sets and Systems 158 (2007) 1961 1978
to solve the fuzzy exible program. Some authors used the concept of comparison of fuzzy numbers for solving FLP
problems [6,16]. In effect, most convenient methods are based on the concept of comparison of fuzzy numbers by
use of ranking functions [9,17,18,20]. Of course, ranking functions have been proposed by researchers to suit their
requirements of the problem under consideration [6,18,20], and conceivably there are no generally accepted criteria for
application of ranking functions. Nevertheless, usually in such methods authors dene a crisp model which is equivalent
to the FLP problem and then use optimal solution of the model as the optimal solution of the FLP problem. A review of
some common methods for ranking fuzzy numbers can be seen in [27]. Moreover, a review of the literature concerning
fuzzy mathematical programming as well as comparison of fuzzy numbers can be seen in Klir and Yuan [12] and also
Lai and Hwang [13]. Some authors considered types of single and multi-objective linear programming problems in
which the variables and the right-hand sides of the constraints are fuzzy parameters [10,1719].
The study of duality theory for fuzzy parameter linear programming problems has attracted researchers in fuzzy
decision theory. The duality of fuzzy parameter linear programming was rst studied by Rodder and Zimmermann [23].
Verdegay [26] dened the fuzzy dual problem with the help of parametric linear programming and showed that the fuzzy
primal and dual problems both have the same fuzzy solution under some suitable conditions. The fuzzy primal and
dual linear programming problems with fuzzy coefcients were formulated by using the fuzzy scalar product proposed
in [28]. Liu et al. [15] proposed the fuzzy primal and dual problems by considering the fuzzy-max and fuzzy-min
in the objective functions as the crisp pattern of linear programming problems. Bector and Chandra [1] discussed
duality in FLP based on a modication of the dual formulation stated by Rodder and Zimmermann [23]. Afterwards
Bector et al. [2] considered a fuzzy matrix game and proved its equivalence to two crisp linear programming problems
which constitute a primaldual pair in the sense of duality for linear programming with fuzzy parameters (see also
[3]). Inuiguchi et al. [11] studied FLP duality in the setting of fuzzy relations. Ramik [22,21] discussed a class of FLP
problems based on fuzzy relations and a new concept of duality and deduced the weak and strong duality theorems.
Wu [29] offered a concept of fuzzy scalar product and proved the weak and strong duality theorems using a dual fuzzy
mathematical programming problem.
The fuzzy variable linear programming (FVLP) problems have been explored by Zimmermanns discussion [32] of
the so-called nonsymmetric exible linear programming (NFLP) problems, where the problem data are considered to
be crisp but certain constraints are considered to be fuzzy inequality" constraints (see also [13,14]). It was shown that
under certain conditions the NFLP problem is equivalent to an FVLP problem. Moreover, as we will show later, an
FVLP problem is the dual of a fuzzy number linear programming (FNLP) problem, in which the coefcients of the cost
function are fuzzy. Therefore, methods for solving FVLP problems can be used for solving both the NFLP and FNLP
problems. Maleki et al. [17,18] gave an auxiliary problem, having only fuzzy cost coefcients, for an FVLP problem.
In [18], they obtained some results leading to an algorithm for solving the auxiliary problem as a method for solving
the FVLP problem (we will see later that their algorithm is a dual algorithm for solving a primal FVLP problem). In
[17], they used the algorithm to solve the NFLP problems. Here, we rst show that the auxiliary problem given in [18]
is indeed the dual of the FVLP problem. This leads us to duality results for the two problems. Our main contributions
here are the establishment of duality and complementary slackness, upon the use of certain linear ranking functions to
order trapezoidal fuzzy numbers. Using the results, we develop and present a dual simplex algorithm directly using the
primal simplex tableau (as opposed to solving the dual problem considered by Maleki et al. [17]). The new algorithm
tenders the capability for sensitivity (or post optimality) analysis using primal simplex tableaus. In Section 2, we rst
give some necessary notations and denitions of fuzzy set theory. Then we provide a discussion of fuzzy numbers and
linear ranking functions for ordering them. In particular, a certain linear ranking function for ordering trapezoidal fuzzy
numbers is emphasized. The denition of the FVLP problem is given in Section 3. Section 4 explains the notion of
fuzzy basic feasible solution. We establish duality for the FVLP problem in Section 5 and deduce the duality results. In
Section 6, we develop and present the dual simplex algorithm, using the primal tableau, for solving the FVLP problems.
We conclude in Section 7.
2. Preliminaries
2.1. Denitions and notations
We review the fundamental notions of fuzzy set theory, initiated by Bellman and Zadeh [4].
N. Mahdavi-Amiri, S.H. Nasseri / Fuzzy Sets and Systems 158 (2007) 1961 1978 1963
Denition 2.1. A convex fuzzy set A on R is a fuzzy number if the following conditions hold:
(a) Its membership function is piecewise continuous.
(b) There exist three intervals [a, b], [b, c] and [c, d] such that A is increasing on [a, b], equal to 1 on [b, c],
decreasing on [c, d] and equal to 0 elsewhere.
Denition 2.2. Let A = (a L , a U , , ) denote the trapezoidal fuzzy number, where (a L , a U + ) is the support
of A and [a L , a U ] its core.
Remark 2.1. We denote the set of all trapezoidal fuzzy numbers by F(R).
We next dene arithmetic on trapezoidal fuzzy numbers. Let a = (a L , a U , , ) and b = (bL , bU , , ) be two
trapezoidal fuzzy numbers. Dene,
x > 0, x R; x a = (xa L , xa U , x, x),
x < 0, x R; x a = (xa U , xa L , x, x),
a + b = (a L + bL , a U + bU , + , + ).
We point out that the arithmetic on trapezoidal fuzzy numbers follows the Extension Principle (for a discussion
see [13]).
2.2. Ranking functions
One convenient approach for solving the FLP problems is based on the concept of comparison of fuzzy numbers
by use of ranking functions (see [27]). An effective approach for ordering the elements of F(R) is to dene a ranking
function R : F(R) R which maps each fuzzy number into the real line, where a natural order exists.
We dene orders on F(R) by
a b if and only if R(a) R(b), (2.1)
R
a > b if and only if R(a) > R(b), (2.2)

R
a = b if and only if R(a) = R(b), (2.3)

R
where a and b are in F(R). Also we write a b if and only if b a.

R R
We restrict our attention to linear ranking functions, that is, a ranking function R such that
R(k a + b) = kR(a) + R(b) (2.4)
for any a and b belonging to F(R) and any k R.
Remark 2.2. For any trapezoidal fuzzy number a, the relation a 0 holds, if there exist 0 and 0 such that
R
a (, , , ). We realize that R(, , , ) = 0 (we also consider a = 0 if and only if R(a) = 0). Thus, without
R R
loss of generality, throughout the paper we let 0 = (0, 0, 0, 0) as the zero trapezoidal fuzzy number.
The following lemma is now immediate.
Lemma 2.1. Let R be any linear ranking function. Then,

(i) a b if and only if a b 0 if and only if b a.
R R R
(ii) If a b and c d, then a + c b + d.
R R R
We consider the linear ranking functions on F(R) as
R(a) = cL a L + cU a U + c + c , (2.5)
where a = (a L , a U , , ), and cL , cU , c , c are constants, at least one of which is nonzero. A special version of the
above linear ranking function was rst proposed by Yager [30] (see also [9,24]) as follows:

1 1
R(a) = (inf a + sup a ) d, (2.6)
2 0
which reduces to
aL + aU 1
R(a) = + ( ). (2.7)
2 4
Then, for trapezoidal fuzzy numbers a = (a L , a U , , ) and b = (bL , bU , , ), we have
a b if and only if a L + a U + 21 ( )bL + bU + 21 ( ). (2.8)

R
3. Linear programming problem with trapezoidal fuzzy variables
A linear programming problem with trapezoidal fuzzy variables (FVLP) is dened as [17,18]
min z = cx,
R
s.t. Ax b,
R
x 0 = (0, 0, 0, 0), (3.1)
R
where b (F(R))m , A Rmn , cT Rn are given and x (F(R))n is to be determined, and R is a linear ranking
function as dened by (2.5).
Denition 3.1. We say that fuzzy vector x (F(R))n is a fuzzy feasible solution to (3.1) if x satises the constraints
of the problem.
Remark 3.1. It should be emphasized that in solving problem (3.1) , we are concerned with nding a minimizer x so
that Ax b, x 0 (implying that R(Ax)R(b), R(x) 0). In fact, the operator R plays an implicit role here and
R R
the concern is nding x to satisfy an equivalent fuzzy equality relation Ax + y = b, as described by (5.11) in proof of
R
Theorem 5.3. We will see that in solving problem (3.1), fuzzy basic solutions (xB = B 1 b, xN = 0), as explained in the
next section, of a corresponding equality constrained problem play the decisive roles. Every basic solution is, of course,
unique (see Theorem 5.3 in Section 5), in the absence of degeneracy. Thus, while the ranking function has no role in
computing the basic solutions but the constraint relations are satised by the ranking function at every basic solution.
Any other fuzzy vector satisfying R(Ax) R(b) is not of interest here because it is not a solution of the constraints
in the rst place. However, the ranking function will play its crucial role in ordering the trapezoidal numbers being
used for testing the optimality (inequality) conditions (see Theorem 4.1) and making the decision for pivoting (see
Algorithm 6.1). These have been used in computing the dual simplex tableaus in Example 6.1.
Denition 3.2. A fuzzy feasible solution x is a fuzzy optimal solution for (3.1), if for all fuzzy feasible solution x for
(3.1), we have cx cx.
R
Remark 3.2. In practice, there are several linear programming models that are related to linear programming problems
with fuzzy variables. One such problem is the FNLP problem. In Section 5, we will see that the dual of any FVLP
problem is a linear programming problem with fuzzy cost coefcients. Knowing this, one may use the solutions of the
FVLP problem to obtain the solutions of the linear programming problem with fuzzy cost coefcients (see Theorem
5.3). Another problem is the NFLP considered by Zimmermann [32] (see also [13,14]).
A method for solving the FVLP problem based on solving an auxiliary problem has been given by Maleki et al.
[17,18]. We will show that the auxiliary problem is indeed the dual of the FVLP problem. This leads us to the concept
of duality in linear programming with trapezoidal fuzzy variables. We state and prove duality results obtained by a
natural extension of the results in crisp linear programming. We nally introduce a dual simplex algorithm for solving
the FVLP problem directly, using the simplex tableau corresponding to the original primal problem (3.1). The new
algorithm will then be useful for post optimality analysis based on the use of primal simplex tableaus.
4. Fuzzy basic feasible solution
Here, we explore the concept of fuzzy basic feasible solution for FVLP problems. Consider the FVLP problem,
min z = cx,
R
s.t. Ax = b,
R
x 0, (4.1)
R
where the parameters of the problem are as dened in (3.1). Let A = [aij ]mn . Assume rank(A) = m. Partition A as
[B N] where B, m m, is nonsingular. It is obvious that rank(B) = m. Let yj be the solution to By = aj , where
aj is the jth column of the coefcient matrix A. It is apparent that the basic solution
xB = (xB1 , . . . , xBm )T = B 1 b, xN = 0 (4.2)
is a solution of Ax = b. We call x, accordingly partitioned as (xBT T )T , a fuzzy basic solution corresponding to basis
xN
R
B. If xB 0, then the fuzzy basic solution is feasible and the corresponding fuzzy objective value is z = cB xB , where
R R
cB = (cB1 , . . . , cBm ). Now, corresponding to every index j, 1 j n, dene
zj = cB yj = cB B 1 aj . (4.3)
Observe that for any basic index j = Bi , 1 i m, we have B 1 aj = ei where ei = (0, . . . , 0, 1, 0, . . . , 0)T is the ith
unit vector, since Bei = [aB1 , . . . , aBi , . . . , aBm ]ei = aBi = aj , and so we have:
zj cj = cB B 1 aj cj = cB ei cj = cBi cj = cj cj = 0. (4.4)
The following theorem characterizes optimal solutions. The converse part of the result needs the nondegeneracy
assumption of the problem, where all fuzzy basic variables corresponding to every basis B are nonzero (and hence
positive).
Theorem 4.1 (Optimality conditions). Assume the linear programming problem with trapezoidal fuzzy variables (4.1)
is nondegenerate and B is a feasible basis. A fuzzy basic feasible solution xB = B 1 b 0, xN = 0 is optimal to (4.1)
R
if and only if zj = cB B 1 aj cj for all j, 1 j n.
T )T is a fuzzy basic feasible solution to (4.1), where x = B 1 b, x

Proof. Suppose x =(xBT xN B N = 0. Then the
R
corresponding fuzzy objective value is:
z = cx = cB xB = cB B 1 b. (4.5)
R R R
On the other hand, for any fuzzy basic feasible solution x to (4.1), we have
b = Ax = B xB + N xN . (4.6)
R R
Hence, we can rewrite (4.6) as follows:

xB = B 1 b B 1 N xN . (4.7)
R
Then, for any fuzzy basic feasible solution to (4.1), we have
z = cx = cB xB + cN xN = cB B 1 b (cB B 1 N cN )xN
R R R

n
n
1 1 1
= cB B b (cB B aj cj )xj = cB B b (zj cj )xj .
R R
j =1 j =1
Hence, using (4.4) and (4.5), we have

z = z (zj cj )xj . (4.8)
R
j =Bi
Now, if for all j, 1j n we have zj cj , then from feasibility of x we have (zj cj )xj 0, and then we obtain
R
(z
j =Bi j c )x
j j 0. Therefore, it follows from (4.8) that z z , and so x is optimal. For only ifpart, let x be
R R
a fuzzy optimal basic feasible solution to (4.1). For j = Bi , 1 i m, from (4.4) we know that zj cj = 0. From
(4.8) it is obvious that if for any nonbasic variable xj we have zj > cj , then we can enter xj into the basis and obtain
z > z (because the problem is nondegenerate and xj > 0 in the new basis). This is a contradiction to z being optimal.
R R
Hence we must have zj cj , 1 j n.
Remark 4.1. A basic solution of the problem, xB = y0 = B 1 b = (xB1 , . . . , xBm )T , where xBi = (xBLi , xBUi , Bi , Bi ),
xN = 0 is a unique fuzzy vector due to b being fuzzy. If one is interested in a crisp solution, a natural defuzzication
of the trapezoidal fuzzy solution is obtained by setting x = R(x ), that is, xB = R(xB ), xN = 0.
In the next section, we develop the duality results.
5. Duality
5.1. Formulation of the dual problem
Consider the FVLP problem below:

n
min z = cx = cj xj ,
R R
j =1

n
s.t. Ax b or aij xj bi , i = 1, . . . , m,
R j =1 R
x 0 or xj 0, j = 1, . . . , n, (5.1)
R R
where R is a linear ranking function, x = (x1 , . . . , xn )T (F(R))n , b = (b1 , . . . , bm )T (F(R))m , A = [aij ]mn ,

c = (c1 , . . . , cn )T Rn , bi = (biL , biU , bi , bi ) F(R), xj = (xjL , xjU , xj , xj ) F(R).
It is obvious that problem (5.1) is equivalent to problem (5.2) below:
min z = R(cx),
s.t. R(Ax) R(b),
R(x) 0, (5.2)
where, R(x) = (R(x1 ), . . . , R(xn R , R(b) = (R(b1 ), . . . , R(bm
))T n
R . ))T m

In the above notations, R(xj ) = cL xj + c U xj + c xj + c xj , R(bi ) = cL biL + c U biU + c bi + c bi , where
L U
cL , cU , c , c are constants, at least one of which is nonzero.
Theorem 5.1. Problems (5.1) and (5.2) are equivalent to the following problem:
min z = cR(x),
s.t. AR(x)R(b),
R(x)0. (5.3)
Proof. Since, R is a linear ranking function, we have

n
n
n
R(cx) = R cj xj = R(cj xj ) = cj R(xj ) = cR(x),
j =1 j =1 j =1
where, R(x) = (R(x1 ), . . . , R(xn ))T Rn .

On the other hand,

n n
n
R aij xj = R(aij xj ) = aij R(xj ).
j =1 j =1 j =1
So, if we denote the ith row of matrix A by ai , we have

T
n n
R(Ax) = ((R(a1 x), . . . , R(am x))T = R a1j xj , . . . , R amj xj
j =1 j =1
T

n
n
= a1j R(xj ), . . . , amj R(xj )
j =1 j =1
= (a1 R(x), . . . , am R(x))T = AR(x).
We note that for any feasible x (F(R))n for (5.1) we have

n
aij R(xj )R(bi ), i = 1, . . . , m,
j =1
R(xj )0, j = 1, . . . , n,
or, using a linear ranking function,

n

aij (cL xjL + c U xjU + c xj + c xj )cL biL + c U biU + c bi + c bi , i = 1, . . . , m,
j =1

cL xjL + c U xjU + c xj + c xj 0, j = 1, . . . , n. (5.4)
Now, for any optimal x for (5.1) we have
cx cx
R
for any feasible solution x for (5.1), or

R(cx ) = cR(x ) R(cx) = cR(x),
or

n
n
cj R(xj ) cj R(xj ),
j =1 j =1
or, using a linear ranking function,

n
n

cj (cL xj L + c U xj U + c xj + c xj ) cj (cL xjL + cU xjU + c xj + c xj ). (5.5)
j =1 j =1
Therefore, x = (x1 , . . . , xn )T , where xj = (xj L , xj U , xj , xj )T F(R), is optimal for (5.1), if and only if
x = (x1 L , x1 U , x1 , x1 , . . . , xn L , xn U , xn , xn )T R4n is an optimal solution for (5.2).
We are now in a position to naturally extend the usual denition of a dual problem in linear programming to the
problem with trapezoidal fuzzy variables.
Denition 5.1. The dual of (5.1) is dened to be
max u = y b,
R
s.t. yAc,
y 0. (5.6)
We name the above problem as the DFVLP problem.
Now, we shall show that the dual of (5.1) is indeed (5.6), using the usual denition of the dual in linear programming.
Using Theorem (5.1) and applying a linear ranking function, we can write (5.1) as follows:

n

min z= cj (cL xjL + c U xjU + c xj + c xj ),
j =1

n

s.t. aij (cL xjL + c U xjU + c xj + c xj )R(bi ), i = 1, . . . , m,
j =1

cL xjL + c U xjU + c xj + c xj 0, j = 1, . . . , n. (5.7)
The dual of (5.7) is

m

m
max u= yi (cL biL + c U biU + c bi + c bi ) = yi R(bi ),
i=1 i=1

m
s.t. aij cL yi + cL ym+1 = cL cj ,
i=1

m
aij cU yi + cU ym+1 = cU cj ,
i=1

m
aij c yi + c ym+1 = c cj ,
i=1

m
aij c yi + c ym+1 = c cj , j = 1, . . . , n,
i=1
y1 , . . . , ym 0, ym+1 0. (5.8)

Note that if any of cL , cU , c , c is equal to zero then the corresponding constraint is

not present, and since at least
one of the values cL , c U , c , c is nonzero, the corresponding constraints reduce to m i=1 aij yi + ym+1 = cj , for
j = 1, . . . , n. Then we have
m
max u = yi R(bi ),
i=1

m
s.t. yi aij cj , j = 1, . . . , n,
i=1
yi 0, i = 1, . . . , m, (5.9)
or, using matrix notations,
max u=y R(b),
s.t. yA c,
y 0. (5.10)
Since problems (5.6) and (5.10) are equivalent, then problem (5.6) is the dual of (5.1) in the usual denition of the dual
in linear programming.
Lemma 5.1. The dual of the problem (5.6) is the problem (5.1).
Proof. Problems (5.6) and (5.1) are, respectively, equivalent to problems (5.9) and (5.7). Knowing from linear pro-
gramming that the dual of (5.9) is (5.7), it follows that the dual of (5.6) is (5.1).
5.2. Relations between FVLP and DFVLP problems
Theorem 5.2 (The weak duality property). If x0 and w0 are feasible solutions to FVLP and DFVLP problems, respec-
tively, then cx0 w0 b.
R
Proof. Multiplying Ax0 b on the left by w0 0 and w0 A c on the right by x0 0 and using Lemma 2.1, we get
R R
cx0 w0 Ax0 w0 b.
R R
Corollary 5.1. If x0 and w0 are feasible solutions to FVLP and DFVLP problems, respectively, and cx0 = w0 b, then
R
x0 and w0 are optimal solutions to their respective problems.
Proof. It is straightforward, using Theorem 5.2.
The following corollary relates unboundedness of one problem to infeasibility of the other. We use the denition
below.
Denition 5.2. We say FVLP problem (or DFVLP problem) is unbounded if feasible solutions exist with arbitrary
small (or large) ranking function for the fuzzy objective value.
Corollary 5.2. If any one of the FVLP or DFVLP problem is unbounded, then the other problem has no feasible
solution.
Proof. It is straightforward, using Theorem 5.2.
We are now ready to present the strong duality result. The converse part of the result was given in [18] before. Here,
we state and prove a stronger result.
Theorem 5.3 (Strong duality). If any one of the FVLP or DFVLP problem has an optimal solution, then both problems
have optimal solutions and the two optimal value of ranking functions for the fuzzy objective values are equal. (In fact,
if x with xB = B 1 b, and xN
= 0 is optimal solution of the primal problem then the crisp vector w = c B 1 ,
B
where B is the optimal basis, is optimal solution of the dual problem.)
Proof. First, assume that the FVLP problem has a fuzzy optimal solution, and rank(A) = m. Let y 0 be the fuzzy
R
slack variables for the constraints Ax b. The new equivalent problem to the FVLP problem is
R
min z = cx + 0 y,
R
s.t. Ax + y = b,
R
x 0,
R
y 0. (5.11)
R
Assume B is the optimal basis matrix and x =(xBT 0T )T =(bT B T 0T )T is the fuzzy basic optimal solution corre-
R R
sponding to the FVLP problem. From Theorem 4.1 we have
cB B 1 aj cj 0, j = 1, . . . , n, n + 1, . . . , n + m,
or equivalently,
cB B 1 aj cj , j = 1, . . . , n,
cB B 1 ei 0, i = 1, . . . , m.
Hence, we must have
cB B 1 Ac
cB B 1 0.
Now, let w = cB B 1 . Using the above inequalities, we can write,
w Ac,
w 0.
Thus, w is a feasible solution to the DFVLP problem and
w b = cB B 1 b = cB xB = cx
R R R
and hence
w b = cx .
R
Therefore, the result follows immediately from Theorem 5.2. The proof for the converse is given in [18].
Remark 5.1. We emphasize that the basic solution to the FVLP problem is a fuzzy vector that is uniquely determined
by the optimal basis B. The solution to the dual problem is a unique crisp solution, however, depending solely on the
basis matrix B, with the objective value of the dual having the same rank as the primal (see also Remark 4.1). We realize
that the duality results obtained here are independent of the choice of the linear ranking function. It is clear that we
can use any other linear ranking function, and although the solution obtained may be different but the duality results
are still valid for the new solution. In particular, the dual simplex algorithm that we present later will nd the fuzzy
solution of the primal as well as the crisp solution of the dual problem along with the induced basis corresponding to
the ranking function being used (we point out that the optimal basis corresponding to a linear ranking function does
not change if the ranking function is multiplied by a positive constant). As for the types of the fuzzy data in the model
and the assumption of fuzziness in the variables, the choice and compatibility of the ranking function for the linear
programming model should be the decision makers main concerns. For trapezoidal fuzzy numbers and variables, the
linear ranking function (2.5) is deemed to be appropriate (see [9,24]).
For an illustration of the above theorem consider the following FVLP problem and its dual.
Example 5.1.
min z = 6x1 + 10x2 ,

R

2x1 + 5x2 (5, 8, 2, 5),

R

s.t. 3x1 + 4x2 (6, 10, 2, 6),

R

x1 , x2 0.
R
max u =(5, 8, 2, 5)w1 + (6, 10, 2, 6)w2 ,

R

2w + 3w2 6,
1
s.t. 5w1 + 4w2 10,

w1 , w2 0.
We see that the fuzzy basic optimal solution for the FVLP problem is x1 =( 5 2 30 30 38
7 , 7 , 7 , 7 ), x2 =( 7 , 7 , 7 , 7 )
12 18 19
R R
62 300 360 418
with R(x1 ) = 15
28 , R(x2 ) = 7 , and also the optimal objective value is z = ( 7 , 7 , 7 , 7 ) with R(z) = 19.0714,
16
with the optimal basis matrix

5 2
B= .
4 3
Hence, if we let w = cB B 1 , we obtain w = ( 67 , 10

7 ), which is equal to the optimal solution of the DFVLP problem.
On the other hand, we have R(u) = 19.0714. Therefore, both problems have optimal solutions and the two optimal
value of ranking functions for the fuzzy objective values are equal.
Remark 5.2. Note that in Example 5.1, we considered trapezoidal fuzzy number and used linear ranking function as
given in (2.5).
We can now state the fundamental duality result.
Theorem 5.4 (Fundamental theorem of duality). For any FVLP problem and its corresponding DFVLP problem, ex-
actly one of the following statements is true.
(1) Both have optimal solutions x and w with cx = w b.
R
(2) One problem is unbounded and the other is infeasible.
(3) Both problems are infeasible.
Proof. Using the above results, the rst and second statements are obviously correct. We give Example 5.2 to show
the third.
Example 5.2. Consider the FVLP problem and its corresponding DFVLP problem as follows:
min z = x1 2x2 ,
R
s.t. x1 x2 (1, 1, 4, 2),
R
x1 + x2 (2, 1, 3, 7),
R
x1 , x2 0.
R
max u =(1, 1, 4, 2)w1 + (2, 1, 3, 7)w2 ,
R
s.t. w1 w2 1,
w1 + w2 2,
w1 , w2 0.
We see that both problems are infeasible.

We now state and prove an important result of duality theory, commonly named as complementary slackness.
Theorem 5.5 (Complementary slackness). Let x and w be any feasible solutions to FVLP problem and its corre-
sponding DFVLP problem. Then x and w are respectively optimal if and only if
(w A c)x = 0, w (b Ax ) = 0. (5.12)
R R
Proof. Suppose that x and w are feasible solutions to FVLP and DFVLP problems, respectively. Therefore,
Ax b (5.13)
R
and
w Ac. (5.14)
Multiplying Ax b on the left by w 0 yields
R
w Ax w b. (5.15)
R
Multiplying w Ac on the right by x 0 yields

R
w Ax cx . (5.16)
R
Therefore, we will have

w b w Ax cx . (5.17)
R R
On the other hand, since x and w are optimal solutions to the primal and dual problems, respectively, then by Theorem
5.2, we have w b = cx , and (5.17) is written as
R
w b = w Ax = cx . (5.18)
R R
From, (5.18) we immediately have

(w A c)x = 0, w (b Ax ) = 0. (5.19)
R R
The converse of the theorem follows from the fact that (w A c)x = 0 and w (b Ax ) = 0 imply that cx = w b.
R R R
Therefore, optimality of x and w follows from Corollary 5.1.
Remark 5.3. The complimentary slackness conditions (5.12) is equivalent to (below, ai refers to the ith row and aj
refers to the jth column of A)
w aj < cj xj = 0, or xj > 0 w aj = cj , j = 1, . . . , n,
R R
ai x < bi wi = 0, or wi < 0 ai x = bi , i = 1, . . . , m.
R R
Maleki et al. [17,18] consider the FVLP problem and introduce a new method for solving it. The method proposed,
solves a related problem called the auxiliary problem. In the next section, we introduce a new dual method for solving
the FVLP problem directly, working on the primal problem and its associated simplex tableaus.
6. Dual simplex method
6.1. Primal optimality and dual feasibility
Consider the following FVLP problem,

min z = cx,
R
s.t. Ax b,
R
x 0, (6.1)
R
where x = (x1 , . . . , xn )T , b = (b1 , . . . , bm )T , c = (c1 , . . . , cn ), A = [aij ]mn , bi , xj F(R), cj , aij R, for all
i, j , and R is a linear ranking function.
Now, we may rewrite (6.1) as follows:
min z = cx + 0 y,
R
s.t. Ax + y = b,
R
x 0, y 0, (6.2)
R R
where y = (y1 , . . . , ym )T .
Dene x (F(R))n+m and c Rn+m as

xj , j = 1, . . . , n, cj , j = 1, . . . , n,
xj = cj = (6.3)
yj n , j = n + 1, . . . , n + m, 0, j = n + 1, . . . , n + m.
Suppose that a basic solution for (6.2) is given by xB = B 1 b, with the basis matrix B. Now, let zj = cB B 1 aj , y0 =
R R
B 1 b, where cB = (cB1 , . . . , cBm ), and aj is the j th column of the coefcient matrix [A I ]. Consider Table 1, where
(xB )r is the rth fuzzy basic variable, yj = B 1 aj , and R(yr0 ) is the real number corresponding to the fuzzy number
yr0 obtained by a linear ranking function. Suppose that for j = 1, . . . , n + m, we have
zj cj 0. (6.4)
Dene w = cB B 1 , where w = (w1 , . . . , wm ). For j = 1, . . . , n, we have
y0j = zj cj = cB B 1 aj cj = w aj cj .
Therefore, for j = 1, . . . , n, from zj cj 0, it follows that waj cj 0. Then,
wAc. (6.5)
Table 1
A dual feasible simplex tableau
Basis x1 xl xn+m R.H.S. R(R.H.S.)
z z1 c1 zl cl zn+m cn+m cB y0 R(cB y0 )
(xB )1 y11 y1l y1,n+m y1 0 R(y1 0 )

. . . . . .
. . . . . .
. . . . . .
(xB )r yr1 yrl yr,n+m yr 0 R(yr 0 )
. . . . . .
. . . . . .
. . . . . .
(xB )m ym1 yml ym,n+m ym 0 R(ym 0 )
On the other hand, using (6.4), we have,

0 zn+i cn+i = cB B 1 ei 0 = wei = wi , i = 1, . . . , m,
and hence,
w 0, (6.6)
yielding the dual feasibility. If R(yr0 ) 0, for all r = 1, . . . , m, then a fuzzy feasible solution for the FVLP problem
is at hand. Moreover, we will have,
cx = cB y0 = cB B 1 b = w b,
R R R
and thus, by Corollary 5.1, establishing the optimality of x and w for the FVLP and DFVLP problems, respectively.
Therefore, we have the following result.
Corollary 6.1. The optimality criteria zj cj 0 for all j, for the FVLP problem is equivalent to the feasibility
condition for the DFVLP problem. If, in addition, x corresponding to a basis B is primal feasible then x is optimal for
the FVLP problem and w = cB B 1 is optimal to the DFVLP problem.
Now, lets assume that the DFVLP problem is feasible and x, corresponding to a basis B, is dual feasible but primal
infeasible (and hence not optimal). That is, we have,
y0j = zj cj 0, j = 1, . . . , n + m
and there exists at least one r such that yr 0 < 0. We can immediately deduce that the primal problem must be nite.
R
Thus, the FVLP problem can be either infeasible (in which case, the DFVLP problem is unbounded), or it has an
optimal solution. In what follows we will show how to work on row r of the simplex tableau corresponding to B, as the
pivoting row, and either (1) detect infeasibility of the FVLP problem (or unboundedness of the DFVLP problem), or
(2) nd a column l, as the pivoting column, to pivot on yrl and obtain a new dual feasible tableau with a nondecreasing
primal objective value. We explain these cases as Lemmas 6.1 and 6.2 below.
Lemma 6.1. If in a dual feasible simplex tableau an r exists such that yr0 < 0 and yrj 0, for all j, then the FVLP
R
problem is infeasible.
Proof. Suppose that a dual simplex tableau is feasible, and an r exists such that yr0 < 0 and yrj 0, for all j. Corre-
R
sponding to the rth row of the tableau, we have j yrj xj = yr0 . Since yrj 0 for all j and xj is required to be nonneg-
R
ative, then j yrj xj 0 for any fuzzy basic feasible solution. However, yr0 < 0. This shows that FVLP problem is
R R
infeasible.
Lemma 6.2. If in a dual feasible simplex tableau, an r exists such that yr0 < 0 and there exists a nonbasic index j such
R
that yrj < 0, then a pivoting column l can be found so that pivoting on yrl will yield a dual feasible tableau with a
corresponding nondecreasing objective value.
Proof. We need a criterion for choosing a nonbasic fuzzy variable to enter the basis so that the new simplex tableau
will remain dual feasible and the new objective value is nondecreasing. Assume column l is the pivot column. Pivoting
on the pivot yrl will result in the new 0th row as follows:
y0l
y0j = y0j yrj , j = 1, . . . , n + m, j = Bi . (6.7)
yrl
For the new tableau to be dual feasible we need to have
y0j 0, j = Bi , (6.8)
which, using (6.7), results in
y0l y0j
, j = Bi . (6.9)
yrl yrj
To satisfy (6.9), it is sufcient to let

y0l y0j
= min | yrj < 0 . (6.10)
yrl j =Bi yrj
We note that the new objective value is nondecreasing, since

y0l y0l
y00 = R y00 yr0 = R(y00 ) R yr0 R(y00 ),
yrl yrl
using the fact that

y0l
R yr0 0.
yrl
Now, using the above results, we introduce a new dual algorithm to solve the FVLP problem directly, making use of
the dual feasible simplex tableau. Thus, we refer to the new algorithm as a dual simplex method.
Algorithm 6.1. A dual simplex method
(1) {Dual feasibility} Given a basis B for the FVLP problem such that
y0j = zj cj 0 for all j. Compute the simplex tableau.
(2) If y0 0 then Stop (the current solution is optimal)
R
else select the pivot row r with yr0 < 0 (that is, r so that R(yr0 ) < 0).
R
(3) If yrj 0 for all j then Stop (the FVLP problem is infeasible)
else select the pivot column l by the following minimum ratio test:

y0l y0j
= min | yrj < 0 .
yrl j =Bi yrj
(4) Pivot on yrl and go to (2).
Remark 6.1. One suggestion for choice of r in step (2) may be r so that
R(yr0 ) = min1 i m {R(yi0 )}.
Remark 6.2. Pivoting on yrl in step (4) is the usual Gaussian elimination process that, using yrl , converts column l to
the unit vector er yielding the new simplex tableau corresponding to the new basis (the one that is obtained by replacing
column r of B with al ).
6.2. A numerical example
For an illustration of the above method we consider an example used by Maleki [17], which was solved by use of an
auxiliary problem. Note that here we use from linear ranking function (2.5) as a pattern of linear ranking function on
F(R). Moreover, we suppose the decision making variables be trapezoidal fuzzy number. In practice we consider the
known coefcients and the decision making variables are the same type of fuzzy number.
Example 6.1.
min z = 6x1 + 10x2 ,
R

2x1 + 5x2 (5, 8, 2, 5),

R
s.t. 3x1 + 4x2 (6, 10, 2, 6),

R

x1 , x2 0.
R
We may write the rst dual feasible simplex tableau as follows:

2x1 5x2 + x3 =(8, 5, 5, 2),

R
3x1 4x2 + x4 =(10, 6, 6, 2),
R

x , . . . , x 0,
1 4
R
or equivalently,
Basis x1 x2 x3 x4 R.H.S. R(R.H.S.)
z 6 10 0 0 (0,0,0,0) 0
x3 2 5 1 0 (8, 5, 5, 2) 7.25
x4 3 4 0 1 (10, 6, 6, 2) 9
x4 is a leaving variable and x1 is an entering variable. The new tableau is
z 0 2 0 2 (12,20,4,12) 18
x3 0 73 1 23 (4, 53 , 19
3 , 6) 12
15
x1 1 4
3 0 13 (2, 10 2
3 , 3 , 2) 3
x3 is a leaving variable and x2 is an entering variable. The next tableau is
z 0 0 67 10
7 ( 62 300 360 418
7 , 7 , 7 , 7 )
267
14
x2 0 1 37 2
7 ( 57 , 12 18 19
7 , 7 , 7 )
15
28
x1 1 0 4
7 57 ( 27 , 30 30 38
7 , 7 , 7 )
16
7
2
Therefore, the optimal solution of the FVLP problem obtained by the dual method is x1 = 7 , 7 , 7 , x2 =
, 30 30 38
62 300 360 418 7
5 12 18 19
7 , 7 7 7 and z =
, , 7 , 7 , 7 , 7 .
Now if we consider the optimal solution, showing four decimal places for the fractional parts, we have
x1 = (0.2857, 4.2857, 4.2857, 5.4286),
x2 = (0.7143, 1.7143, 2.5714, 2.7143),
z = (8.8571, 42.8571, 51.4286, 59.7143) and R(z) = 19.0714.
The solution is matched with the solution obtained by Maleki [17].
7. Conclusions
We established the dual of linear programming problem with trapezoidal fuzzy variables and hence developed some
duality results, based on certain linear ranking functions, for the fuzzy primal and fuzzy dual problems. The duality
results have been established using certain general linear ranking functions on trapezoidal fuzzy numbers and appeared
to be natural extensions of the results for linear programming problems with crisp data. Using these results and the dual
feasible primal simplex tableau, we introduced a dual simplex algorithm for solving the primal and the dual problems
directly. The capabilities offered here will be useful for sensitivity (or post optimality) analysis using the primal simplex
tableau.
Acknowledgments
This research was supported in part by the Research Council of Sharif University of Technology, and in part by
a grant from IPM (Grant no. 85900040). We are grateful to the three anonymous referees and the editors for their
constructive comments, raising questions, and providing references. Consideration of all improved the presentation.
References
[1] C.R. Bector, S. Chandra, On duality in linear programming under fuzzy environment, Fuzzy Sets and Systems 125 (2002) 317325.
[2] C.R. Bector, S. Chandra, V. Vijay, Duality in linear programming with fuzzy parameters and matrix games with fuzzy pay-offs, Fuzzy Sets and
Systems 146 (2004) 253269.
[3] C.R. Bector, S. Chandra, Fuzzy Mathematical Programming and Fuzzy Matrix Games, Studies in Fuzziness and Soft Computing, Vol. 169,
Springer, Berlin, 2005.
[4] R.E. Bellman, L.A. Zadeh, Decision making in a fuzzy environment, Management Sci. 17 (1970) 141164.
[5] J. Buckley, T. Feuring, Evolutionary algorithm solution to fuzzy problems: fuzzy linear programming, Fuzzy Sets and Systems 109 (2000)
3553.
[6] L. Campos, J.L. Verdegay, Linear programming problems and ranking of fuzzy numbers, Fuzzy Sets and Systems 32 (1989) 111.
[7] S. Chanas, The use of parametric programming in fuzzy linear programming, Fuzzy Sets and Systems 11 (1983) 243251.
[8] M. Delgado, J.L. Verdegay, M.A. Vila, A general model for fuzzy linear programming, Fuzzy Sets and Systems 29 (1989) 2129.
[9] P. Fortemps, M. Roubens, Ranking and defuzzication methods based on area compensation, Fuzzy Sets and Systems 82 (1996) 319330.
[10] K. Ganesan, P. Veeramani, Fuzzy linear programming with trapezoidal fuzzy numbers, Ann. Oper. Res. 143 (2006) 305315.
[11] M. Inuiguchi, J. Ramik, T. Tanino, M. Vlach, Satiscing solutions and duality in interval and fuzzy linear programming, Fuzzy Sets and Systems
135 (2003) 151177.
[12] G.J. Klir, B. Yuan, Fuzzy Sets and Fuzzy Logic: Theory and Applications, Prentice-Hall, Englewood Cliffs, NJ, 1995.
[13] Y.J. Lai, C.L. Hwang, Fuzzy Mathematical Programming Methods and Applications, Springer, Berlin, 1992.
[14] Y.J. Lai, C.L. Hwang, A new approach to some possibilistic linear programming problem, Fuzzy Sets and Systems 49 (1992) 121133.
[15] Y.J. Liu, Y. Shi, Y.H. Liu, Duality of fuzzy MC2 linear programming: a constructive approach, J. Math. Anal. Appl. 194 (1995) 389413.
[16] N. Mahdavi-Amiri, S.H. Nasseri, Duality in fuzzy number linear programming by use of a certain linear ranking function, Appl. Math. Comput.
180 (2006) 206216.
[17] H.R. Maleki, Ranking functions and their applications to fuzzy linear programming, Far East J. Math. Sci. (FJMS) 4 (2002) 283301.
[18] H.R. Maleki, M. Tata, M. Mashinchi, Linear programming with fuzzy variables, Fuzzy Sets and Systems 109 (2000) 2133.
[19] H. Mishmast Nehi, H.R. Maleki, M. Mashinchi, Multiobjective linear programming with fuzzy variables, Far East J. Math. Sci. (FJMS) 5
(2002) 155172.
[20] H. Mishmast Nehi, H.R. Maleki, M. Mashinchi, Solving fuzzy number linear programming problem by Lexicographic ranking function, Italian
J. Pure Appl. Math. 15 (2004) 920.
[21] J. Ramik, Duality in fuzzy linear programming based on fuzzy relations, in: Proc. of 10th IFSA World Congress, June 29July 2, 2003, Istanbul,
Turkey.
[22] J. Ramik, Duality in fuzzy linear programming: some new concepts and results, Fuzzy Optim. and Decision Making 4 (2005) 2539.
[23] W. Rodder, H.J. Zimmermann, Duality in fuzzy linear programming, in: A.V. Fiacco, K.O. Kortanek (Eds.), Proc. Int. Symp. on Extremal
Methods and Systems Analysis, Berlin, Heidelberg, New York, 1980, University of Texas at Austin, 1977, pp. 415427.
[24] M. Roubens, Inequality constraints between fuzzy numbers and their use in mathematical programming, in: R. Slowinski, J. Teghem
(Eds.), Stochastic Versus Fuzzy Approaches To Multiobjective Mathematical Programming Under Uncertainty, Kluwer Academic Publishers,
Dordrecht, 1991, pp. 321330.
[25] H. Tanaka, T. Okuda, K. Asai, On fuzzy mathematical programming, J. Cybernet. 3 (1974) 3746.
[26] J.L. Verdegay, A dual approach to solve the fuzzy linear programming problems, Fuzzy Sets and Systems 14 (1984) 131141.
[27] X. Wang, E. Kerre, Reasonable properties for the ordering of fuzzy quantities (2 parts), Fuzzy Sets and Systems 118 (2001) 375405.
[28] H.C. Wu, Duality theory in fuzzy linear programming problems with fuzzy coefcients, Fuzzy Optim. and Decision Making 2 (2003) 6173.
[29] H.C. Wu, Duality theorems in fuzzy mathematical programming problems based on the concept of necessity, Fuzzy Sets and Systems 139
(2003) 363377.
[30] R.R. Yager, A procedure for ordering fuzzy subsets of the unit interval, Inform. Sci. 24 (1981) 143161.
[31] H.J. Zimmermann, Fuzzy programming and linear programming with several objective functions, Fuzzy Sets and Systems 1 (1978) 4555.
[32] H.J. Zimmermann, Applications of fuzzy sets theory to mathematical programming, Inform. Sci. 36 (1985) 2958.

Dual Simplex Method For Linear Programming PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Dual Simplex Method For Linear Programming PDF

Uploaded by

Copyright:

Available Formats

Fuzzy Sets and Systems 158 (2007) 1961 1978

Duality results and a dual simplex method for linear programming

Corresponding author. Tel.: +98 21 66165607.

2.1. Denitions and notations

2.2. Ranking functions

a > b if and only if R(a) > R(b), (2.2)

a = b if and only if R(a) = R(b), (2.3)

where a and b are in F(R). Also we write a b if and only if b a.

The following lemma is now immediate.

Lemma 2.1. Let R be any linear ranking function. Then,

We consider the linear ranking functions on F(R) as

a b if and only if a L + a U + 21 ( )bL + bU + 21 ( ). (2.8)

3. Linear programming problem with trapezoidal fuzzy variables

4. Fuzzy basic feasible solution

xB = (xB1 , . . . , xBm )T = B 1 b, xN = 0 (4.2)

T )T is a fuzzy basic feasible solution to (4.1), where x = B 1 b, x

Hence, we can rewrite (4.6) as follows:

In the next section, we develop the duality results.

5.1. Formulation of the dual problem

Consider the FVLP problem below:

Proof. Since, R is a linear ranking function, we have

where, R(x) = (R(x1 ), . . . , R(xn ))T Rn .

So, if we denote the ith row of matrix A by ai , we have

= (a1 R(x), . . . , am R(x))T = AR(x).

We note that for any feasible x (F(R))n for (5.1) we have

for any feasible solution x for (5.1), or

or, using a linear ranking function,

Denition 5.1. The dual of (5.1) is dened to be

We name the above problem as the DFVLP problem.

The dual of (5.7) is

y1 , . . . , ym 0, ym+1 0. (5.8)

Note that if any of cL , cU , c , c is equal to zero then the corresponding constraint is

5.2. Relations between FVLP and DFVLP problems

Proof. It is straightforward, using Theorem 5.2.

Proof. It is straightforward, using Theorem 5.2.

Hence, we must have

Now, let w = cB B 1 . Using the above inequalities, we can write,

Thus, w is a feasible solution to the DFVLP problem and

min z = 6x1 + 10x2 ,

max u =(5, 8, 2, 5)w1 + (6, 10, 2, 6)w2 ,

with the optimal basis matrix

Hence, if we let w = cB B 1 , we obtain w = ( 67 , 10

We can now state the fundamental duality result.

We see that both problems are infeasible.

Multiplying w Ac on the right by x 0 yields

Therefore, we will have

From, (5.18) we immediately have

6. Dual simplex method

6.1. Primal optimality and dual feasibility

Consider the following FVLP problem,

Basis x1 xl xn+m R.H.S. R(R.H.S.)

z z1 c1 zl cl zn+m cn+m cB y0 R(cB y0 )

(xB )1 y11 y1l y1,n+m y1 0 R(y1 0 )

On the other hand, using (6.4), we have,

We note that the new objective value is nondecreasing, since

Algorithm 6.1. A dual simplex method

(4) Pivot on yrl and go to (2).

6.2. A numerical example

Basis x1 x2 x3 x4 R.H.S. R(R.H.S.)

x4 is a leaving variable and x1 is an entering variable. The new tableau is

Basis x1 x2 x3 x4 R.H.S. R(R.H.S.)

x3 is a leaving variable and x2 is an entering variable. The next tableau is

Basis x1 x2 x3 x4 R.H.S. R(R.H.S.)

a b if and only if a L + a U + 21 ( )bL + bU + 21 ( ). (2.8)