Nonlinear Analysis Forum 11(2), pp.

137–145, 2006

ON A DUALITY SCHEME IN LINEAR FRACTIONAL

PROGRAMMING

Ta Quang Son

Nhatrang College of Education

1 Nguyen Chanh Str.
Nhatrang, Vietnam
E-mail : [email protected]

Abstract. Some remarks on Seshan’s duality scheme in linear fractional programming

are proposed in this paper. Not assuming the boundedness of the feasible set of the
primal problem, we obtain the converse duality theorem by a new method of reasoning.
We also clarify the relationship between the feasible set of the dual problem and the
solution set of the primal problem.

1. Introduction
Duality theory for linear fractional programming is a special case of the duality
theory for fractional programming. It has attracted much attention in the last four
decades (see [1, 2, 6-8]). Several different duality schemes have been proposed. In
particular, the parametric method and the change variable method have been used
to linearize the primal problem before constructing the duality problem (see [10]). In
these ways, the objective function is changed. So, the role of the objective function
in duality theorems becomes unclear, or it is lost completely.
The ordinary Lagrangian duality scheme is not suitable for fractional programming
problems, because it may lead to a duality gap (see, for instance, [1], p. 682). The
main reason is that these problems are nonconvex in general. Since the Fenchel and
the Lagrangian duality schemes are equivalent [6], the same remark is valid for the
Fenchel duality scheme.
In building duality schemes for fractional programming, one naturally wants that
the objective function of the dual problem is also a fractional function. Another
approach is to use a symmetric duality scheme [12]. Although the ordinary Lagrange
duality theory is inadequate for dealing with fractional programming, Gol’shtein [3]
managed to introduce a ‘fractional Lagrangian’ as a fractional dual problem and
obtained a saddle-point duality correspondence. Bitran and Magnanti [1] has com-
plemented these results by presenting some theorems for fractional duality.

Received Jun. 2006.

2000 Mathematics Subject Classification: 49N15, 90C32.
Key words and phrases: Linear fractional programming, primal problem, dual problem, duality
theorem, optimal solution set.
This research has been supported partially by the project “Optimization and Scientific Comput-
ing”, Vietnam, and FUNDP, Namur University, Belgium. The author thanks Prof. N. D. Yen and
Dr. T. D. Phuong (Hanoi Institute of Mathematics) for valuable comments and suggestions.
138 Ta Quang Son

There is an interesting duality scheme for linear fractional programming suggested

by Seshan [9]. It has been quoted in detail in the book of Stancu-Minasian [10].
The scheme is very specific: the primal and dual problems have the same objective
function. Note that the weak duality and the strong duality theorems in [9] (see
also [10], pp. 164–168) were established under rather strict conditions (namely, the
positiveness of the denominator of the objective function on whole the feasible set of
the dual problem, and the boundedness of that feasible set).
Our aim in this paper is to propose some extensions for the above-mentioned
duality scheme of Seshan. Not assuming the boundedness of the feasible set of the
primal problem, we obtain strong duality theorems by a new method of reasoning.
We also clarify the relationship between the feasible set of the dual problem and the
solution set of the primal problem.
We now describe the contents of the next sections. Section 2 recalls the main
results of [9] together with our analysis and examples. Section 3 presents some new
duality theorems. Section 4 discusses the relationship between the dual and the
primal problems. Section 5 is devoted to some applications.

2. Seshan’s duality theorems

Consider the linear fractional programming problem:
 
cT x + c0
(P ) max F (x) := T Ax ≤ b, x ≥ 0 .
d x + d0

Here A is an (m × n)-matrix; c, d, x are column vectors with n components; b is a

column vector of m components; c0, d0 are arbitrary constants; the superscript T
denotes matrix transposition. Let

X = {x ∈ Rn | Ax ≤ b, x ≥ 0}

denote the feasible set of (P ).

The dual problem corresponding to the above primal problem (P ) is defined [9] as
follows:  
cT u + c0
(D) min I(u, v) := T (u, v) ∈ Y ,
d u + d0
where the feasible set Y consists of all (u, v) ∈ Rn × Rm satisfying the constraints

(dT u)c − (cT u)d − AT v ≤ c0 d − d0c, (1)

T T T
c0d u − d0c u + b v ≤ 0, (2)
u ≥ 0, v ≥ 0. (3)

In [10], the next three assumptions were imposed on the problem (P ):

(A1) dT x + d0 > 0 for every x ∈ X;
(A2) X is a bounded set;
(A3) F is not a constant function on X.
 On a duality scheme in linear fractional programming 139

Remark 2.1. In [9], it has been noted that if d ≥ 0 and d0 > 0 then (A1) is
fulfilled and, moreover, the objective functions of (P ) and (D) are well defined on X
and Y , respectively. But this condition is a very strict one.
Remark 2.2. Seshan [9] assumed that dT x + d0 > 0 for all x ≥ 01. (Then
(A1) is satisfied.) This strong condition guarantees that both the primal problem
and the dual problem are well defined. Stancu-Minasian [10] quoted the duality
scheme of Seshan with a little modification: the condition dT x + d0 > 0 is only
assumed to be satisfied for all x ∈ X. In this case, the objective function I(u, v) =
(cT u+c0 )/(dT u+d0 ) of the dual problem may not be well defined on Y . To overcome
this obstacle, the following condition has been introduced:
(A4) dT u + d0 > 0 for every (u, v) ∈ Y .
 
1
Example 2.1. Consider the problem max x ≥ 1 . According to the above
x
duality scheme, the dual problem is written as follows
 
1
min v ≤ 1, u − v ≤ 0, u ≥ 0, v ≥ 0 .
u

It is clear that the feasible set Y here contains the line segment {(0, v) | 0 ≤ v ≤ 1}
which lies outside the effective domain of I(u, v).
We now recall three duality theorems which are the main results in [9] (see also
[10], pp. 164–168). Detailed proofs of the first two theorems are demonstrated here
to make our development of Shesan’s duality scheme (see Section 3 below) easier for
understanding. The longer proof of the third theorem will be omitted. Actually, in
the sequel we will give a shorter proof for a sharpened version of that theorem. It is
important to stress that the proofs given in [9] and [10] are not rigorous (see Example
2.1 above) if one does not impose the assumption (A4) (which was not stated clearly
in [9] and [10]).
Theorem 2.1. (Weak duality) If (A1) and (A4) are satisfied, then

sup{F (x) | x ∈ X} ≤ inf{I(u, v) | (u, v) ∈ Y }. (4)

Proof. By the usual convention, sup ∅ = −∞ and inf ∅ = +∞. So (4) holds if
one of the sets X and Y is empty, or both of them are empty. Now suppose that
X 6= φ and Y 6= φ. Take any x ∈ X and (u, v) ∈ Y . Since x ≥ 0, from (1) we deduce
that
(dT u)(cT x) − (cT u)(dT x) − xT AT v ≤ c0dT x − d0 cT x. (5)

In addition, since v ≥ 0 by (3), from (2) and the inequality Ax ≤ b it follows that

c0 dT u − d0cT u ≤ −bT v ≤ −xT AT v. (6)

1 In fact, it was assumed that dT x + d > 0 for all x ∈ Rn . It is easy to see that this requirement
0
cannot be fulfilled if d 6= 0.
140 Ta Quang Son

Combining (5) and (6) gives

(dT u)(cT x) − (cT u)(dT x) + c0 dT u − d0cT u ≤ c0dT x − d0 cT x,

which implies
(cT x + c0 )(dT u + d0) ≤ (cT u + c0)(dT x + d0).
Then we have
cT x + c0 cT u + c0
≤ ,
dT x + d0 dT u + d0
because dT x + d0 > 0 and dT u + d0 > 0 by (A1) and (A4). The proof of (4) is
complete.

Theorem 2.2. (Direct duality) Suppose that (A1) and (A4) are satisfied. If x̄ is
a solution to (P ), then there exits a solution (ū, v̄) to (D) such that

F (x̄) = I(ū, v̄). (7)

cT x̄+c0
Proof. Given a solution x̄ of (P ), we set λ = dT x̄+d0 and consider the linear
programming problem

(LP ) max{(cT x + c0 ) − λ(dT x + d0 ) | Ax ≤ b, x ≥ 0}.

Using (A1) and the definition of λ, one can verify easily that x̄ is a solution to (LP )
whose optimal value is 0. According to the duality theorem in linear programming,
the optimal value of the dual problem

(LD) min{bT v + c0 − λd0 | AT v ≥ c − λd, v ≥ 0, v ∈ Rm }

of (LP ) is 0, and (LD) must have a solution v̄1 ∈ Rm . Then v̄1 ≥ 0 and

bT v̄1 + c0 − λd0 = 0. (8)

We choose ū = x̄ and v̄ = (dT x̄ + d0)v̄1 . This pair (ū, v̄) is a solution to (D). Indeed,
we have
(dT ū)c − (cT ū)d − AT v̄ ≤ (dT x̄)c − (cT x̄)d − (dT x̄ + d0)(c − λd)
= (dT x̄)c − (cT x̄)d − (dT x̄ + d0)c + (cT x̄ + c0 )d
= c0 d − d0 c.

Moreover, multiplying both sides of (8) by dT x̄ + d0 > 0 we can deduce that

0 = (dT x̄ + d0 )bT v̄1 + c0 (dT x̄ + d0) − d0 (cT x̄ + c0 ) = bT v̄ + c0dT ū − d0 cT ū.

From what it has been said, it is easy to see that (1)–(3) are valid, so (ū, v̄) ∈ Y .
Since F (x̄) = I(ū, v̄), we can invoke Theorem 2.1 to deduce that (ū, v̄) is a solution
to (D).

Theorem 2.2 states that, under suitable conditions, if (P ) is solvable then (D)
is also solvable. The next theorem shows that, under some tighter conditions, the
converse is true: if (D) is solvable then (P ) is also solvable.
 On a duality scheme in linear fractional programming 141

Theorem 2.3. (Converse duality) Suppose that (A1)–(A4) are satisfied. If (ū, v̄)
is a solution to (D), then there exists a solution x̄ to (P ) such that (7) holds.
The proof given in [9] and [10] (pp. 167–168) for this theorem is rather long. Apart
from the assumptions (A1) and (A4), it relies heavily on the assumptions (A2) and
(A3).
Remark 2.3. Assumption (A4) is vital for the proofs of Theorems 2.1–2.3. But
it is not satisfied even for the simple problem described in Example 2.1. The duality
scheme of Seshan allows us to assign to each linear fractional programming problem
of the form (P ) a dual problem (D). Since there is no explicit methods to decide
whether (A4) is valid for this (D) or not, we cannot know definitely whether the
above three duality theorems can be applied to the given pair (P ) and (D) or not.
So, it is not clear to which linear fractional programming problems (except for those
described in Remark 2.1), Seshan’s duality theory is applicable.
We are going to propose some duality theorems which do not suffer from the need
of the assumption (A4).

3. New duality theorems

Consider the following dual problem of (P ):
 
cT u + c0
(D1 ) min I(u, v) := T (u, v) ∈ Y1 ,
d u + d0

where Y1 is the set of all (u, v) ∈ Rn × Rm satisfying (1)–(3) and the inequality

Au ≤ b. (9)

It is obvious that if (A1) is satisfied then dT u + d0 > 0 for every (u, v) ∈ Y1 . This
means that the dual problem is well defined if the primal problem is well defined . The
next two theorems follow from this simple observation and the proofs of Theorems
2.1–2.2.
Theorem 3.1. (Weak duality) If (A1) is satisfied, then

sup{F (x) | x ∈ X} ≤ inf{I(u, v) | (u, v) ∈ Y1}. (10)

Theorem 3.2. (Direct duality) Suppose that (A1) is satisfied. If x̄ is a solution

to (P ), then there exits a solution (ū, v̄) to (D1 ) such that the equality (7) is valid.
Not assuming the fulfilment of (A2) and (A3), we can establish the following
converse duality theorem for the pair (P ), (D1 ).
Theorem 3.3. (Converse duality) Suppose that (A1) is satisfied. If (ū, v̄) is a
solution to (D1 ), then there exists a solution x̄ to (P ) such that (7) holds.
142 Ta Quang Son

Proof. Suppose that (ū, v̄) is a solution to (D1 ). Then (ū, v̄) ∈ Y1 and

I(ū, v̄) = min{I(u, v) | (u, v) ∈ Y1 }. (11)

Let PrRn (Y1 ) = {u ∈ Rn | (u, v) ∈ Y1 }. On one hand, from the definition of Y1 it

follows that
PrRn (Y1 ) ⊂ X.
On the other hand, (11) implies that

I(ū, v̄) ≤ inf{I(u, v) | u ∈ PrRn (Y1 ), v ∈ Rm } = inf{F (u) | u ∈ PrRn (Y1)}.

Therefore,

I(ū, v̄) ≤ sup{F (u) | u ∈ PrRn (Y1 )} ≤ sup{F (x) | x ∈ X}. (12)

By (10),
sup{F (x) | x ∈ X} ≤ inf{I(u, v) | (u, v) ∈ Y1 } = I(ū, v̄). (13)
From (12) and (13) it follows that

F (ū) = I(ū, v̄) = max{F (x) | x ∈ X)}. (14)

Since each pair (u, v) ∈ Y1 must satisfy (9), setting x̄ = ū we have x̄ ∈ X. Then (14)
shows that x̄ is a solution of (P ). The desired equality (7) is a consequence of (14).

The following theorem shows that there is a close connection between the dual
problems (D1 ) and (D).
Theorem 3.4. Suppose that (A1) and (A4) are satisfied. Then, if (P ) has a
solution then each one of the problems (D) and (D1 ) also has a solution, and

max{F (x) | x ∈ X} = min{I(u, v) | (u, v) ∈ Y } = min{I(u, v)|(u, v) ∈ Y1 }. (15)

Proof. Since Y1 ⊂ Y , it holds

inf{I(u, v) | (u, v) ∈ Y } ≤ inf{I(u, v) | (u, v) ∈ Y1 }. (16)

Let x̄ be a solution to (P ). Since (A1) and (A4) are satisfied, by Theorem 2.2 there
is a solution (ū, v̄) of (D) such that (7) holds. According to Theorem 3.2, (D1 ) also
has a solution (ū1, v̄1) ∈ Y1 satisfying F (x̄) = I(ū1 , v̄1). Combining this equality
with (7) and (16) yields

F (x̄) = I(ū, v̄) = min{I(u, v) | (u, v) ∈ Y }

≤ min{I(u, v)|(u, v) ∈ Y1} = I(ū1 , v̄1 ) = F (x̄).

Since F (x̄) = max{F (x) | x ∈ X}, it follows that (15) is valid.

 On a duality scheme in linear fractional programming 143

4. Feasible set of the dual problem

In this section we show that there is a special connection between the feasible set
of the dual problem (D1 ) and the solution set of (P ), which we denote by Sol(P ).
Theorem 4.1. Under the assumption (A1), it holds

Sol(P ) = PrRn (Y1 ). (17)

Proof. On one hand, the inclusion PrRn (Y1 ) ⊂ X yields

sup{F (x) | x ∈ X} ≥ sup{F (x) | x ∈ PrRn (Y1 )},

so
sup{F (x) | x ∈ X} ≥ F (u) (∀ u ∈ PrRn (Y1 )).
On the other hand, Theorem 3.1 gives

sup{F (x) | x ∈ X} ≤ inf{I(u, v) | u ∈ PrRn (Y1 ), v ∈ Rm },

which implies that

sup{F (x) | x ∈ X} ≤ F (u) (∀ u ∈ PrRn (Y1 )).

Consequently,
sup{F (x) | x ∈ X} = F (u) (∀ u ∈ PrRn (Y1 )). (18)
Hence PrRn (Y1) ⊂ Sol(P ). To prove the reverse inclusion, we fix any x̄ ∈ Sol(P )
and apply the arguments in the first part of the proof of Theorem 2.2 to find a pair
(ū, v̄) ∈ Y which belongs to the solution set of (D). Since ū = x̄ by the construction,
we deduce that (ū, v̄) ∈ Y1 . The desired inclusion x̄ ∈ PrRn (Y1) now follows from the
equality x̄ = ū.

Corollary 4.2. Under the assumption (A1), the solution set of (P ) is nonempty
if and only if the feasible set of (D1 ) is nonempty.
Proof. The desired conclusion follows directly from (17).

Corollary 4.3. Under the assumptions (A1) and (A4), if the feasible set of (D1 )
is nonempty then the dual problem (D) has a solution, and it holds

Sol(P ) = PrRn (Y1 ) ⊂ PrRn (Sol(D)), (19)

where Sol(D) denotes the solution set of (D).

Proof. The first conclusion follows from Corollary 4.2 and Theorem 2.2. The sec-
ond conclusion follows from Theorem 4.1 and the arguments in the proof of Theorem
2.2.

It is still not clear to us whether the inclusion in (19) can be replaced by an

equality provided that (A1) and (A4) are satisfied.
144 Ta Quang Son

5. A dual algorithm and illustrative examples

Equality (17) in Theorem 4.1 leads to the following dual algorithm for solving the
linear fractional programming problem (P ):
Step 1. Construct the duality problem in form (D).
Step 2. Reduce the feasible set Y of (D) to Y1 by adding the constraint Au ≤ b.
Step 3. Use the simplex method to find a point (ū, v̄) ∈ Y1. Then x̄ := ū =
PrRn (ū, v̄) is a solution of (P ).
In Step 3, if whole the set Y1 could be found, then one obtains whole the solution
set Sol(P ) by formula (17).
We now consider several examples.

Example 5.1. For the problem max x1 | x ≥ 1 , which has been considered in
Section 2, the dual problem in the form (D1 ) is the following one:
 
1
min | v ≤ 1, u − v ≤ 0, u ≥ 1, v ≥ 0 .
u
Then one has Y1 = {(1, 1)} and Sol(P ) = PrR (1, 1) = {1}. It is clear that (15) is
valid and the inclusion in (19) becomes an equality.
Example 5.2. The dual problem for the problem
 
x
max F (x) = x ≥ 0, x ≤ 1 ,
x+1
in the form (D1 ) is
 
u
min I(u, v) = v ≥ 1, v − u ≤ 0, u ≤ 1, v ≥ 0, u ≥ 0 .
u+1
Applying Theorem 4.1, we have
Sol(P ) = PrR (Y1) = {u | v ≥ 1, v − u ≤ 0, u ≤ 1, v ≥ 0, u ≥ 0} = {1}
and sup F (x) = F (1) = 1/2.
X
Example 5.3. For the problem
 
x1 + 2x2
max F (x) = 0 ≤ x1 ≤ 1, 0 ≤ x2 ≤ 1
x1 + x2 + 1
we have
( )
2 2
u = (u1 , u2) ≥ 0, v = (v1 , v2) ≥ 0, u2 + v1 ≥ 1, v2 − u1 ≥ 2
Y1 = (u, v) ∈ R × R .
u1 + 2u2 ≥ v1 + v2 , 0 ≤ u1 , u2 ≤ 1
By Theorem 4.1,
Sol(P ) = PrR2 (Y1 )
= {(u1, u2) | u2 + v1 ≥ 1, v2 − u1 ≥ 2, u1 + 2u2 ≥ v1 + v2, 0 ≤ u1, u2 ≤ 1}.
The optimal value of (P ) is 1 and Sol(P ) = {(1, 1), (0, 1)}.
When the denominator of objective function is a nonzero constant, (P ) is a linear
programming problem. The results in the previous section can be applied to this
special case.
 On a duality scheme in linear fractional programming 145

Example 5.4. For the problem

 

 3x1 − 2x2 ≤ 0, 

max F (x) = 6x1 + 5x2 3x1 + 2x2 ≤ 15

 
x ≥ 0, x ≥ 0 
1 2

it holds  

 v1 + v2 ≥ 2, 


 


 5 


 v1 − v2 ≤ − 


 2 

 
Y1 = (u, v) ∈ R2 × R2 6u1 + 5u2 ≥ 15v2 .

 3u1 − 2u2 ≤ 0 


 


 


 3u1 + 2u2 ≤ 15 


 

 u1 ≥ 0, u2 ≥ 0, v1 ≥ 0, v2 ≥ 0 

Since (0, 15/2) ∈ PrRn (Y1 ), we have max F (x) = F ((0, 15/2)) = 37.5.
X

References

[1] G. R. Bitran and T. L. Magnanti, Duality and sensitivity analysis for fractional programs,
Oper. Research 24 (1976), 675–699.
[2] B. D. Craven, Mathematical Programming and Control Theory, Chapman and Hall, London,
1978.
[3] E. G. Gol’stein, Dual problems of convex and fractionally-convex programming in functional
spaces, Soviet Math. Dokl. 8 (1967), 212–216.
[4] R. Jagannathan and S. Schaible, Duality in generalized fractional programming via Farkas’
lemma, J. Optim. Theory Appl. 41 (1983), 417–424.
[5] R. N. Kaul, S. Kaur and V. Lyall, Duality in inexact fractional programming with set-inclusive
constraints, J. Optim. Theory Appl. 50 (1986), 279–288.
[6] T. L. Magnanti, Fenchel and Lagrange duality are equivalent, Math. Program. 7 (1974), 253–
258.
[7] S. Schaible, Duality in fractional programming: A unified approach, Oper. Research 24 (1976),
452–461.
[8] , Fractional programming. I, Duality, Management Science 22 (1976), 858–867.
[9] C. R. Seshan, On duality in linear fractional programming, Proc. Indian Acad. Sci. (Math.
Sci.) 89 (1980), 35–42.
[10] I. M. Stancu-Minasian, Fractional Programming, Kluwer, Dordrecht, 1997.
[11] K. Swarup, Linear fractional functionals programming, Oper. Research 13 (1965), 1029–1036.
[12] T. Weir, Symmetric dual multiobective fractional programming, J. Austral. Math. Soc. (Series
A) 50 (1991), 67–74.