Topic 4:
Economic interpretation and applications of duality:
Picking up from where we left last, this lecture can be viewed with two fold aims. The first is
to provide a formal interpretation of the dual variables and second is to establish the celebrated
result called Farkas Lemma using the principle of duality.
Consider the following LPP
max z = cT x
subject to Ax = b
x 0.
Suppose B is an optimal basis and (xB , 0) is an optimal solution. Then xB = B 1 b, and
the optimal solution of the associated dual problem is wT = cTB B 1 (recall the strong duality
theorem).
Suppose the resource vector b is perturbed slightly to b + b such that this perturbation does
not alter the optimal basis. In other words it means that B remains optimal basis for the LPP.
max z = cT x
subject to Ax = b + b
x 0.
However, the optimal solution will definitely see a change. The new optimal solution of the
perturbed LPP is say (xB + xB , 0), where,
xB + xB = B 1 (b + b)
giving
xB = B 1 b.
Now we figure out the effect of this perturbation on the objective value, say the new value be
z + z. Then,
z = cTB (xB + xB ) cTB xB
= (cTB B 1 )b
= wT b.
This implies
z
= wi , i = 1, 2, . . . , m.
bi
The rate of change of optimal objective function value with respect to the ith resource availability
bi is the ith optimal dual variable wi . In general, the ith optimal dual variable wi is an answer
to the question, If the amount of the constraint bi on the right of the ith constraint
c
Copyright
Reserved IIT Delhi
2
changes, how does the value of objective function change? In this sense, the dual variables
are to be treated as prices. Economists call them as shadow prices or implicit prices. In a more
general context, they are refereed to as Lagrange multipliers. In nutshell, shadow price is the
maximum price that the management/person is willing to pay for an extra unit of a limited
resource.
To illustrates the above discussion, we present the following example.
Example:
(P ) min 3x1 + 2x2
subject to 7x1 + 2x2 30
5x1 + 4x2 20
2x1 + 8x2 16
x1 , x2 0.
Its dual problem is
(D) max 30w1 + 20w2 + 16w3
subject to 7w1 + 5w2 + 2w3 3
2w1 + 4w2 + 8w3 2
w1 , w2 , w3 0.
The optimal table of (D) is as follows
cB vB
30 w1
16 w3
cj
wB
5/13
2/13
30 20 16
0
0
y1
y2
y3
y4
y5
1 8/13 0
2/13 1/26
0 9/26 1 1/26 7/52
zj cj
30
31
1
Now suppose we change the resource vector 20 to 20 , i.e, b = 0 .
16
16
0
We can verify (!) that the current optimal basis remains optimal even on change in b. (Do not
worry even if you are not able to verify it right now. We shall be elaborating on this aspect in
Module 6. For the time being you can take it on its face value). Observe that
z = wT b =
5
13
c
Copyright
Reserved IIT Delhi
3
5
and thus the optimal value gets change from 14 to 14 + 13
= 14.385. Thus, one unit re5
We next move towards our second aim. Herein we illustrate the application of duality principle
studied so far to the celebrated Farkas Lemma. This lemma was originally proved by Von Hern
Julius ( Gyula) Farkas in 1902 and falls to a broad class of theorems of alternatives. The latter
denotes the class of theorems which construct two systems of linear inequalities or equations in
which one or the other has a solution but not both.
Farkas lemma states that a vector is either in a given convex cone or there exists a hyperplane
seperating the vector from the cone but not the both. More precisely , let A be an m n matrix
and b be an m 1 vector. Then, exactly one of the following two statements is true.
1. x Rn such that Ax = b, x 0.
2. y Rm such that AT y 0, bT y < 0.
The Farkas lemma has witnessed several variants off late but we have restricted ourselves to the
statement made above. Proof of Farkas lemma can be found in almost any optimization text
book. Though early proofs of this lemma do not make it clear as to why this lemma works; the
recent proofs are much simple to work out. Before pondering on the detailed proof, we try to
see the insight of the lemma.
For convenience, let A be a 2 2 matrix and let a1 , a2 be its two rows (in order) and A1 , A2 be
its two columns (in order).
c
Copyright
Reserved IIT Delhi
a2
Ax
a1
A2
A y>0
A1
1 3
4 2
4
5
"
and b =
3
4
satisfy
Ax = b, x 0
i.e, the system 1 has a solution. In this case system 2 is
y1 + 4y2 0
3y1 + 2y2 0
3y1 + 4y2 < 0
which is inconsistent (left to see it graphically).
c
Copyright
Reserved IIT Delhi
5
"
While, if we take same A but take b =
3
4
#
then system 1 is inconsistent as there is no
x1 , x2 0 for which
x1 + 3x2 = 3
On the other hand system 2 translates to the following system of inequalities.
y1 + 4y2 0
3y1 + 2y2 0
3y1 + 4y2 < 0
which is consistent (for instance choose y1 = 2, y2 = 1).
Note: System 1 and system 2 can not hold simultaneously. Since if x Rn and y Rm
such that
Ax = b, x 0,
AT y 0, bT y < 0
then,
(AT y)T x 0
giving y T Ax 0 or y T b 0. But this is not true as bT y = y T b < 0.
c
Copyright
Reserved IIT Delhi
6
But the same is not true. Thus (P) has to be unbounded.
Consequently (D) is infeasible, thereby, system 1 fails to have a solution.
c
Copyright
Reserved IIT Delhi