Convex sets and simplex method

Things left to be proved:

The correspondence between corner point solutions and basic solutions
Why it is enough just consider the corner points?

The matrix form of LP model

A general LP model in the standard form is
Max z = c 1 x 1 + c 2 x 2 + ` + c n x n
Subject to
a 11 x 1 + a 12 x 2 + ` + a 1n x n b 1
a 21 x 1 + a 22 x 2 + ` + a 2n x n b 2
_
a m1 x 1 + a m2 x 2 + ` + a mn x n b m
and
x 1 , x 2 , ..., x n 0
After we added the slack variables, it becomes
Max z = c 1 x 1 + c 2 x 2 + ` + c n x n
Subject to
a 11 x 1 + a 12 x 2 + ` + a 1n x n + x n+1 = b 1
a 21 x 1 + a 22 x 2 + ` + a 2n x n + x n+2 = b 2
_
a m1 x 1 + a m2 x 2 + ` + a mn x n + x n+m = b m
and
x 1 , x 2 , ..., x n , x n+1 , ..., x n+m 0
Let

c=

A=

c1 c2 ` cn 0 ` 0
a 11 a 12 ` a 1n 1

a 21 a 22 ` a 2n 0

_
a m1 a m2

a mn 0

x1
x=

x2
_
x n+m
b1

b=

b2
_

bm
The system can be expressed as
Max z = cx
s.t.

Ax = b
x0

Definition: A set S is convex if, for any two points, x 1 , x 2 5 S, and J 5 0, 1 imply that
Jx 1 + 1 ? Jx 2 5 S.
An equivalent definition would be:
Definition: A set S is convex if, for any positive integer p and p points, x 1 , x 2 , ..., x p 5 S and
J 1 , J 2 , ...J p 5 0, 1, J 1 + J 2 + ` + J p = 1 imply that J 1 x 1 + J 2 x 2 + ` + J p x p 5 S.
J 1 x 1 + J 2 x 2 + ` + J p x p is called a convex combination of x 1 , x 2 , ..., x p . If none of J 1 , J 2 , ...J p is
0 or 1, then it is a strict convex combination of x 1 , x 2 , ..., x p .
Definition: x is a basic feasible solution if it satisfies the constraints, and the columns of A that
corresponding to the positive components of x are linearly independent.
Theorem: The set S = x : Ax = b, x 0 is a convex set. (4.5-3)
Definition: A corner point or extreme point of a convex set S is a point x such that x is not a strict
convex combination of any two other points in S.
Theorem: A point x of the set S = x : Ax = b, x 0 is a corner point of S if and only if it is a
basic feasible solution.
Reminder: Ax = b can be understood as > x i A i = b. Here A i is the i-th column of A.
Proof of the theorem: Without loss of generality, we may assume that the components of x are
zero except for the first p components, namely

x1

x=

where x =

_
xp

We also denote the first p columns of A by A, namely A =

A1 ` Ap

. We know that

Ax = Ax = b.
Suppose that A 1 , u, A p are linearly dependent, then there are w 1 , u, w p not all zero such
w1
that > w i A i = 0 or if we let w =
, we have Aw = 0. Since x 1 > 0, ux p > 0, we may
_
wp
choose a N small enough so that x i Nw i > 0 for all i = 1, up. (We can pick N < min |wx ii | .)
Let y = x + Nw, z = x ? Nw, y =

,z =

then we have y 0, z 0. Also we

0
0
have Ay = Ay = Ax + Nw = b, Az = Az = Ax + Nw = b. Therefore y, z 5 S. Also we have
1
y + 12 z = x, a contradiction to the fact that x is a corner point.
2
Suppose that x is not a corner point. Then there are distinct points y, z in S such that
x = Jy + 1 ? Jz for some 0 < J < 1. Since y, z 0, the last n ? p components of y, z must be
zero. Then w = x ? y has last n ? p components being zero as well. Therefore Aw = Aw = 0.
w 0 w 0. That implies that the first p columns of A are linearly dependent.
Theorem: Let S = x : Ax = b, x 0. If S is bounded, then every point in S is a convex
combination of some corner points x 1 , u, x n .
[Proof] Let x be a point in S. We show that x is a convex combination of some corner points of
S by using induction and an idea similar to the one used in the previous proof.
Induction on the number of positive components in x. Obviously that if that number is zero then
x is a BF solution therefore a corner point itself. Suppose that x is not a corner point, then the
columns of A corresponding to the positive components of x are linearly dependent. To simplify
the notations, we may assume that these are the first r columns of A:
x1
_
x=

xr
0

_
0
There are p 1 , p 2 , u, p r not all of them zero such that p 1 A 1 + p 2 A 2 + ` + p r A r = 0. Let

p1
_
p=

pr
0

, then

_
0
Ap = p 1 A 1 + p 2 A 2 + ` + p r A r + 0 + `0 = 0.
For all P > 0, we have
Ax + Pp = Ax + PAp = b + 0 = b
So x + Pp is also in S if it satisfise the nonnegative constraints. It is the same for x ? Pp.
Since x i > 0 for i = 1, u, r. So we can choose P 1 small enough such that x i + P 1 p i 0 for all i
and we can choose P 1 > 0 such that for one or more i 1 i r, x i + P 1 p i = 0 (otherwise S
would be unbounded). Simlarly we can choose P 2 > 0 such that x i ? P 2 p i 0 for all i and
x i ? P 2 p i = 0 for at least one i. (Remember that both x + P 1 p and x ? P 2 p are in S.) Let
y = x + P 1 p and z = x ? P 2 p. Easy to see that
x = P P+2 P y + P P+1 P z
1

so x is a convex combination of y and z. Both y and z has at least one less positive components
than x. By the inductive hypothesis, y and z are convex combinations of the corner points of S:

> Jiyi
z = > Kizi

y=

where > J i = > K i = 1, J i 0, K i 0 for all i and y 1 , y 2 , u, z 1 , z 2 , u are corner points of

S. Let L 1 = P 1P+P2 2 , and L 2 = P 1P+P1 2 . We have
x = L1y + L2z
=

> L1Jiyi + > L2Kizi

Since > L 1 J i + > L 2 K i = 1 and L 1 J i 0, L 2 K i 0 for all i, x is a convex combination of the

corner points.
Theorem: If a linear programming problem has an optimal solution, then it has an optimal solution
on a corner point of the feasible region.
Proof: Suppose that z D = cx is an optimal solution. If x is a corner point, then we are done.
Suppose that x is not a corner point, then there are corner points of the feasible region such that
there are x i and J i that > J i x i = x and > J i = 1. Then
cx = c> J i x i = > cJ i x i = > J i cx i > J i z D z D . Since z D = cx, every step here is an
equality. So cx i = z D for all i. Also this shows that x is only optimal if it is a convex
combination of optimal corner point solutions. That is a part of 4-5.4.
Theorem: The feasible region of a linear programming problem has a finite number of corner points.
Proof: Suppose that there are n equations and n + m variables. Then there are at most mn+n
ways to choose that n-linearly independent columns of A.
Theorem: If a corner point has no adjacent corner point that has better objective function value, then

there is no other corner point with better objective function value.

Proof: Clear from the row zero of the simplex tableau..