Duality in Linear Programming

Dual Problem:
Associated with every linear programming problem there always exist another Linear Programming problem
which is based upon the same data and having the same solution. The original problem is called the primal
problem while the associated one is called its dual problem. It is important to note that either of the two linear
programming problems can be treated as primal and the other as its dual. The two problems, thus, constitute a
primal-dual pair.

Importance and Uses:

Duality is an extremely important and interesting feature of linear programming. The various useful aspects of this
property are as follows.

a) If the primal contains a large number of rows (constraints) and a smaller number of columns (variables),
the computational procedure can be reduced by converting it into dual and then solving it. This is a great
advantage in may applications.

b) There is additional information that it describes how the optimal solution changes as a result of the
changes in the coefficients and the formulation of the problem. This is known as post optimality or
sensitivity analysis.

c) Duality in linear programming has reach consequences of economic nature.

d) Calculation of the dual checks the accuracy of the primal solution.

Dual Problem when Primal is in the Canonical Form:

The general linear programming problem in canonical form is written as
Maximize Z = c1 x1 + c 2 x 2 +  + c n x n
Subject to Constrants a11 x1 + a12 x 2 +  + a1n x n  b1 

a 21 x1 + a 22 x 2 +  + a 2n x n  b2 
    (1)
  
a m1 x1 + a m 2 x 2 +  + a mn x n  bm 
x1 , x 2 ,  , x n  0
If the above problem is referred to as primal, then its associated dual will be
Minimize W = b1 y1 + b2 y 2 +  + bm y m
Subject to Constrants a11 y1 + a 21 y 2 +  + a m1 y m  c1 

a12 y1 + a 22 y 2 +  + a m 2 y m  c 2 
    (2)
  
a1n y1 + a 2n y 2 +  + a mn y m  c n 
Where the associated dual var iables y1 , y 2 ,  , y m  0

Relations (1) and (2) are called symmetric primal-dual pairs.

 The above pair of programs can be written as

Primal Dual
n m
Maximize Z =  c j x j Minimize W =  bi y i
j =1 i =1
n m
Subject to  aij x j  bi ; i = 1, 2,  , m Subject to  aij y i  c j ; j = 1, 2,  , n
j =1 i =1
Where x j  0 ; j = 1, 2,  , n Where y i  0 ; i = 1, 2,  , m

From the above two programs, the following points are clear:
a) If the primal contains n variables and m constraints, the dual will contain m variables and n
constraints.
b) If the primal is maximization type then the dual will be minimization type and vice versa.
c) The maximization problem has less than ( ) type constraints while the minimization problem has greater

than ( ) type constraints.

d) The constants c1 , c 2 ,  , c n in the objective function of the primal appear in the constraints of the dual.

e) The constants b1 , b2 ,  , bm in the constraints of the primal appear in the objective function of the dual.

f) The variables in both problems are non-negative.

Theorem: The dual of the dual is the primal.

Proof:
The general linear programming problem involving n decision variables and m constraints in canonical form is
given by following relations.
n
Maximize Z = c jx j
j =1
n
Subject to  aij x j  bi ; i = 1, 2,  , m    (1)
j =1

xj 0 ; j = 1, 2,  , n

If we treat this as the primal then the dual of this problem may be written as follows.

Suppose the y1 , y 2 , , y m be the dual variable then the dual problem is

m
Minimize W =  bi y i
i =1
m
Subject to  aij y i  c j ; j = 1, 2,  , n    (2)
i =1
yi  0 ; i = 1, 2,  , m

Multiplying each of the relation (2 ) by −1 we obtain

m
Maximize (− W ) = − bi y i
i =1
m
Subject to −  a ij y i  −c j ; j = 1, 2,  , n    (3)
i =1
yi  0 ; i = 1, 2,  , m
 Now, if we consider the above problem as the primal and convert it as a dual we will get the following problem.
n
Minimize − Z = − c j x j
j =1
n
Subject to − aij x j  bi ; i = 1, 2, , m
j =1

xj  0 ; j = 1, 2, , n

Multiplying each of the above relation by −1 we will get

n
Maximize Z = cj xj
j =1
n
Subject to  aij x j  bi ; i = 1, 2, , m
j =1

xj  0 ; j = 1, 2, , n

This is the actual primal problem. Thus the dual of the dual is a primal. (Proved )

Theorem: The value of the objective function Z for any feasible solution at the primal is  the value of the objective
function W for any feasible solution of the dual.

Proof:
The general linear programming problem involving n decision variables and m constraints in canonical form is
given by following relations.
n
Maximize Z = c jx j
j =1
n
Subject to  aij x j  bi ; i = 1, 2,  , m    (1)
j =1

xj 0 ; j = 1, 2,  , n

If we treat this as the primal then the dual of this problem will be m decision variables and n constraints.

Suppose the y1 , y 2 , , y m be the dual variable then the dual problem is

m
Minimize W =  bi y i
i =1
m
Subject to  aij y i  c j ; j = 1, 2,  , n    (2)
i =1
yi  0 ; i = 1, 2,  , m

Now, multiplying the first inequality of (1) by y i and then adding them all, we get
m n m
 aij x j y i   bi y i    (3)
i =1 j =1 i =1

Similarly, multiplying the first inequality of (2 ) by x j and then adding them all, we get

n m n
 aij x j y i   c j x j    (4)
j =1 i =1 j =1
 Now, the sums on the left hand side of inequalities of (3) and (4 ) are equal. Hence,
n m n m
 c j x j   aij x j yi   bi yi
j =1 i =1 j =1 i =1
n m
  c j x j   bi yi
j =1 i =1

That is, any feasible solution to the primal is less than or equal to any feasible solution to the dual. (Proved )

Dual Problem when Primal is in the Standard Form:

Consider the problem
Maximize Z = c1 x1 + c 2 x 2
Subject to Constrants a11 x1 + a12 x 2 = b1
a 21 x1 + a 22 x 2  b2
x1 , x 2  0

The first constraint of equality form, when expressed in canonical form is equivalent to
a11 x1 + a12 x 2  b1 and a11 x1 + a12 x 2  b1
 − a11 x1 − a12 x 2  −b1

Let, y1 , y1, and y 2 be the dual variables associated with the first second and third constrains. Then the dual

problem is
Minimize W = b1 ( y1 − y1 ) + b2 y 2
Subject to Constrants a11 ( y1 − y1 ) + a 21 y 2  c1
a12 ( y1 − y1 ) + a 22 y 2  c 2
y1 , y1, y 2  0

If we put y1 − y1 = y1 , then the dual problem becomes-

Minimize W = b1 y1 + b2 y 2
Subject to Constrants a11 y1 + a 21 y 2  c1
a12 y1 + a 22 y 2  c 2
y 2  0 and y1 is unrestrict ed in sign

This shows that the variable which represents equality constraint of the primal will be unrestricted in sign in the
primal. Conversely, when a primal variable is unrestricted in sign, its dual constraint must be in equality form.

In general, if the primal problem in standard form is

n
Maximize Z = cjx j
j =1
n
Subject to  aij x j = bi ; i = 1, 2,  , m
j =1

Where x j  0 ; j = 1, 2,  , n

then the dual problem is

m
Minimize W =  bi y i
i =1
m
Subject to  aij y i  c j ; j = 1, 2,  , n
i =1
Where y i (i = 1, 2,  , m ) are unrestrict ed in sign
 On the other hand, if the primal problem is
n
Maximize Z = cjxj
j =1
n
Subject to  aij x j  bi ; i = 1, 2,  , m
j =1

Where xj ( j = 1, 2,  , n ) are unrestrict ed in sign

then the dual problem is

m
Minimize W =  bi y i
i =1
m
Subject to  aij y i = c j ; j = 1, 2,  , n
i =1
Where yi  0 ; (i = 1, 2,  , m)

Properties of Primal and Dual Optimal Solutions:

The primal and the dual optimal solutions have the following properties:

a) The optimal value of the objective function of the primal is equal to the optimal value of the objective
function of the dual i.e. Z max = W min .

b) If the dual variable corresponds to a slack starting variable in the primal problem, the optimal value of
this dual variable is given directly by the coefficient of the slack variable, with changed sign, if
necessary.

c) If the dual variable corresponds to an artificial starting variable in the primal problem, the optimal value
of this dual variable is given by the coefficient of the artificial variable with changed sign, if necessary.

d) If the primal has an unbounded solution, the dual will not have a feasible solution and vice versa.