Plan

Directions of Fea(LPP)

Extreme directions of Fea(LPP)

Representation Theorem for Fea(LPP)

Necessary and sufficient conditions for the existence of

optimal solutions

Optimal solutions in atleast one corner point

Fea(LPP) = { x ∈ Rn : Am×n x ≤ b,x ≥ 0}
Observation 5: Suppose if a LPP has an unbounded
feasible region, then there exists a vector d /= 0 such that
starting from any x ∈ Fea(LPP) and moving in the positive
direction of d, always gives elements of Fea(LPP).
This d /= 0 is called a direction of S = Fea(LPP)
Definition: Given a non empty convex set S, S ⊂ Rn,
d /= 0 is called a direction of S if for all x ∈ S, x + αd ∈ S
for all α ≥ 0
If an LPP has an unbounded feasible region, then there
exists a d /= 0 such that A(x + αd) ≤ b and x + αd ≥ 0 for
all α > 0, for all x ∈ Fea(LPP)
If d is a direction of a convex set S, then for all γ > 0, γd
is a direction of S
since x + α(γd) = x + (αγ)d ∈ S for all α > 0
Directions d1, d2 of S are said to be distinct if d1 /= γd2
for any γ > 0
( or equivalently d2 /= βd1 for any β > 0)
Example 2: (revisited) Consider the problem
Min −x + 2y
subject to
x + 2y ≥ 1
−x + y ≤ 1
x ≥ 0,y ≥ 0
d1 = [1,1]T , dJ1 = [2,2]T ,..
are all equal as directions of Fea(LPP)
Similarly d2 = [1,0]T , dJ2 = [2,0]T ,..
are all equal as directions of Fea(LPP)
Whereas d1 = [1, 1]T , d2 = [1, 0]T give two distinct
directions
Result: The set of all directions of a non empty feasible
region S = {x ∈ Rn : Am×nx ≤ b, x ≥ 0} is given by
D = { d ∈ Rn : d /= 0, Am×n d ≤ 0, d ≥ 0}
If d1, d2 are directions of S, then d = αd1 + βd2 is also a
direction of S, for any α, β non negativesuch that both α, β
are not equal to zero, or α + β /= 0
The set of all directions of S = Fea(LPP) is a convex set
If D denotes the set of all directions of S
( D = φif S is bounded ), then
Σ
DJ = { d ∈ Rn : d ≥ 0,Ad ≤ 0, i di = 1} is a set of all
distinct directions of S
Also each d ∈ D is of the form d = αdJ for some dJ ∈ DJ
Σ
where α = i di (> 0)
A 0
DJ = d ∈ Rn : d ≥ 0, 1 1,.., 1 d≤ 1
−1 −1,.., −1 −1
The set DJ now looks exactly like the feasible region of a
LPP
Definition: A direction d of S is called an extreme
direction of S, if it cannot be written as a positive linear
combination of two distinct directions of S
that is, if d an extreme direction of S and
d = αd1 + βd2 , for α,β > 0 and d1 ,d2 ∈ D
then d1 = γd2 for some γ > 0
Since DJ is a Fea(LPP) for some LPP, if DJ /= φ, DJ
must atleast one extreme point
have
Result: d is an extreme direction of S if and only if
dJ = Σ d d is an extreme point of DJ
i i

If D /= φ, then S(/= φ) must have atleast one extreme

direction
If S =
/ φis unbounded then D /= φand S must have
atleast one extreme direction
The number of distinct extreme directions of S is finite
(why?)
The extreme directions of S which are extreme points of DJ
(after suitable normalization) will lie on n LI hyperplanes
defining DJ
Since d ∈ Rn, d /= 0 cannot be orthogonal to n LI vectors,
so the extreme points d of DJ cannot lie on n LI
hyperplanes of the (m + n) hyperplanes given by
{ d ∈ Rn : aTi d = 0} for i = 1,2 ,...,m, and
{ d ∈ Rn : −ejT d = 0} for j = 1,2,...,n
If d ∈ DJ, is an extreme direction of S then it should lie on
(n − 1) LI hyperplanes of the above mentioned (m + n)
hyperplanes, and on the hyperplane
{ d ∈ Rn : [1,1,...,1]d = 1}
which together gives a collection of n LI hyperplanes
Any d ∈ DJ, which lies on (n − 1) LI hyperplanes out of the
(m + n) hyperplanes given by
{ d ∈ Rn : aTi d = 0} for i = 1,2 ,...,m, and
{ d ∈ Rn : −ejT d = 0} for j = 1,2,...,n
is an extreme point of DJ
Any d ∈ D, which lies on (n − 1) LI hyperplanes out of the
(m + n) hyperplanes given by
{ d ∈ Rn : aTi d = 0} for i = 1,2 ,...,m, and
{ d ∈ Rn : −ejT d = 0} for j = 1,2,...,n
is an extreme direction of S
Example 2: (revisited) Consider the problem,
Min −x + 2y
subject to
x + 2y ≥ 1
−x + y ≤ 1
x ≥ 0,y ≥ 0
S = Fea(LPP) = { d ∈ R2 : −x − 2y ≤ −1,−x + y ≤ 1,x ≥
0,y ≥ 0}
The set of all directions of S is given by
D = {d ∈ R2 : d /= 0,[−1, −2]d ≤ 0, [−1,1]d ≤ 0,d ≥ 0}
If d ∈ D is an extreme direction of S then it has to lie on
exactly one of the hyperplanes given by
(i) { d ∈ R2 : [−1,−2]d = 0}
(ii) { d ∈ R2 : [−1,1]d = 0}
(iii) { d ∈ R2 : d1 = 0}
(iv) { d ∈ R2 : d2 = 0}
There exists no d ≥ 0, d /= 0 such that [−1, −2]d = 0
If d ≥ 0, d /= 0, satisfies the condition d1 = 0, then
[−1,1]d ≤ 0 cannot be satisfied, hence such a d does not
belong to D
If d ∈ D, is an extreme direction of S then it lies on one of
the following Hyperplanes:
{ d ∈ R2 : [−1,1]d = 0} , { d ∈ R2 : d2 = 0}
dJ = [1,1]T and any positive scalar multiple of dJ, are
extreme directions of S
dJJ = [1,0]T and any positive scalar multiple of dJJ, are
extreme directions of S
These are the only extreme directions of S
Theorem:
If S = { x ∈ Rn : Am×n x ≤ b,x ≥ 0} is nonempty, then S
has atleast one extreme point
Remark: The above result is not necessarily true for all
polyhedral sets
A single half space, or a straight line in Rn, are polyhedral
sets, but does not have any extreme points
The theorem works for Fea(LPP) because of the non
negativity constraints, and Fea(LPP) has n LI , defining
hyperplanes
Exercise: Can you find a nonempty polyhedral set S,
S ⊂ R2 which has two defining hyperplanes but does not
have any extreme point?
Exercise: Can you find a nonempty polyhedral set S,
S ⊂ R3 which has two LI defining hyperplanes but does
not have any extreme point?
Exercise: Is it possible to find a nonempty polyhedral set
S, S ⊂ R3 which has three LI defining hyperplanes ( not
necessarily the nonnegativity constraints) but does not
have any extreme point
Definition: Given S, a nonempty subset of Rn, and
x1 ,x2 ,...,xk ∈ S
Σ
i=1 λi xi , is called a convex combination of Σ
x1, x2, ...,xk
k

where 0 ≤ λi ≤ 1 for all i = 1 ,2,...,k, and ki=1 λi = 1

All possible convex combinations of two distinct points
gives a straight line segment with those two points as
boundary points.
All possible convex combinations of three non colinear
points gives a triangle with those points as corner points.
All possible convex combinations of four points no three of
which are colinear gives a quadrilateral.
Result: Given φ/= S ⊂ Rn , S is a convex set ⇔for all
k ∈ N,
the convex combination of any k elements of S is in S.
Theorem: (Representation Theorem)
If S = {x ∈ Rn : Am×nx ≤ b, x ≥ 0} is nonempty and if
x1 ,x2 ,...,xk are the extreme points of S and d1 ,d2 ,..,dr
are the distinct extreme directions of S (the set of
directions is empty if S is bounded) then x ∈ S ⇔
Σ Σ
x = ki=1 λi xi + rj=1 µj dj Σ
where 0 ≤ λi ≤ 1 for all i = 1,2,...,k, i λi = 1
and µj ≥ 0, for all j = 1,2,...,r
x ∈ S ⇔x can be written as a convex combination of the
extreme points of S plus a non negative linear
combination of the extreme directions of S
If S = {x ∈ Rn : Am×nx ≤ b, x ≥ 0} is nonempty and
bounded and if x1, x2, ...,xk are the extreme points of S
then x ∈ S ⇔
Σ
x = ki=1 λi xi Σ
where 0 ≤ λi ≤ 1 for all i = 1,2,...,k, i λi = 1
Example 2: (revisited) Consider the problem with
Fea(LPP),
x + 2y ≥ 1
−x + y ≤ 1
x ≥ 0,y ≥ 0
The extreme points are [1,0]T ,[0,1]T and [0, 12]T
Distinct extreme directions are [1, 0]T ,[1, 1]T
[0, 43]T ∈ Fea(LPP)
[0, 43 ]T = 21 [0, 12 ]T + 12 [0,1]T
[ 32 ,0 ]T ∈ Fea(LPP)
[ 23 ,0]T = 1[1,0]T + 12 [1,0]T
[2,3]T ∈ Fea(LPP)
[2,3]T = 1[0,1]T + 2[1,1]T
[3,3]T ∈ Fea(LPP)
[3,3]T = 1[0,1]T + 2[1,1]T + 1[1,0]T
[3,3]T = 1 [0,1]T + 1 [1,0]T + 5 [1,1]T
2 2 2
If S = Fea(LPP) is nonempty and bounded then any
x ∈ S is a convex combination of the extreme points of S
Example 1:(revisited) Given the Fea(LPP)
3x + 2y ≤ 6
x + 2y ≤ 4
x ≥ 0,y ≥ 0
The extreme points are [2, 0]T ,[0, 2]T ,[0, 0]T and [1, 32]T
[1, 12]T ∈S
[1, 21 ]T = 21 [2,0]T + 41 [0,2]T + 14 [0,0]T
[1, 21 ]T = 31 [2,0]T + 13 [1, 32 ]T + 13 [0,0]T
[1, 21 ]T = 12
5 [2,0]T + 3 [0,2]T + 1 [1, 3 ]T +
24 6 2
7 [0,0]T
24
Since DJ, the set of distinct directions
Σ of S (if it is
nonempty) is bounded ( d ≥ 0 and ni=1 di = 1)
so any d ∈ DJ is a convex combination of the extreme
points of DJ
Any direction d ∈ D is a nonnegative linear combination of
the extreme directions of S
Observation : If there exists a d ∈ D such that cT d < 0
then the LPP(*)
(*) Min cT x
subject to Ax ≤ b, x ≥ 0
does NOT have an optimal solution
If LPP(*) has an optimal solution then cT d ≥ 0 for all d ∈ D
If cT dj ≥ 0 for all extreme directions dj of the nonempty
and unbounded feasible region S of an LPP, then cT d ≥ 0
for all directions d ∈ D, of S
Observation : Converse:
If S /= φand cT dj ≥ 0 for all j = 1,2,..,r, then LPP(*) has
an optimal solution
Observation : From the representation theorem it follows
that if S = Fea(LPP) is nonempty and bounded then the
LPP(*) has an optimal solution
If LPP(*) has an optimal solution then atleast one optimal
solution is attained at an extreme point of S
Let S be the feasible region of LPP(*)
Conclusion 1: If S /= φ, then LPP (*) has an optimal
solution ⇔one of the following is true:
(i)S is bounded (also seen before by using extreme value
theorem)
(ii) S is unbounded and cT dj ≥ 0 for all extreme directions
dj of the feasible region S
Conclusion 2: If LPP (*) has an optimal solution then there
exists an extreme point of S, which is an optimal solution
Exercise: Give an example of a nonlinear function
f : S → R, where S ⊂ R is a closed and bounded
polyhedral subset of R, (what are these sets?) such that f
has a minimum value in S but the minimum value is not
attained at any extreme point of S
Conclusion 3: If S is nonempty
and there exists an M ∈ R such that for all x ∈ S, cT x ≥ M
then LPP (*) has an optimal solution
Example 2: (revisited) Consider the problem
Min −x + 2y
subject to
x + 2y ≥ 1
−x + y ≤ 1
x ≥ 0,y ≥ 0
The distinct extreme directions are d‘ = [1, 0]T and
d“ = [1,1]T
cT d‘ = −1, hence this problem does NOT have an optimal
solution
If we change the above c to [1,−1]T then cT d‘ = 1 and
cT d“ = 0, and the new problem has an optimal solution
Consider a linear programming problem, LPP(**) as
(**) Max cT x
subject to Ax ≤ b, x ≥ 0
Conclusion 1a: If S = Fea(LPP) /= φ, then LPP (**) has
an optimal solution ⇔one of the following is true:
(i) S is bounded
(ii) S is unbounded and cT dj ≤ 0 for all extreme directions
dj of the feasible region S
Conclusion 2a: If LPP (**) has an optimal solution then
there exists an extreme point of the feasible region S,
which is an optimal solution
Conclusion 3a: If S = Fea(LPP) is nonempty, and there
exists an M ∈ R such that for all x ∈ S, cT x ≤ M
then LPP (**) has an optimal solution