0% found this document useful (0 votes)
65 views

1 Extreme Points

This document summarizes key points from a lecture on convex optimization: 1) It discusses extreme points of convex sets and polyhedra. An extreme point cannot be written as a convex combination of other points in the set. 2) It presents two theorems showing that a point is an extreme point of a polyhedron if and only if it is a minimal face, where minimal means the face does not contain any other faces. 3) A test is given to check if a point is an extreme point by checking the rank of a submatrix formed from the active constraint gradients at that point. If the rank equals the dimension, it is an extreme point.

Uploaded by

Jerimiah
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
65 views

1 Extreme Points

This document summarizes key points from a lecture on convex optimization: 1) It discusses extreme points of convex sets and polyhedra. An extreme point cannot be written as a convex combination of other points in the set. 2) It presents two theorems showing that a point is an extreme point of a polyhedron if and only if it is a minimal face, where minimal means the face does not contain any other faces. 3) A test is given to check if a point is an extreme point by checking the rank of a submatrix formed from the active constraint gradients at that point. If the rank equals the dimension, it is an extreme point.

Uploaded by

Jerimiah
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

The University of Texas at Austin

Department of Electrical and Computer Engineering

EE381K: Convex Optimization — Fall 2020

Lecture 2

Aryan Mokhtari Monday, August 31, 2020

Goal: In this lecture we focus on extreme points and their properties.

1 Extreme Points
Definition 1. Let C ∈ Rn be a non-empty, closed convex set. Then, x̂ is an extreme point of C if
there are not two points y, z ∈ C and scalar λ ∈ (0, 1) such that x̂ = λy + (1 − λ)z.

This definition means that a point x̂ is an extreme point of a set if it is not between any two other
points of that set.
Can we say every polyhedron has at least one extreme point? The answer is no! For instance a
half-space is a polyhedron, but it does not have an extreme point.

Now there are two questions that we aim to answer in this lecture:

• How do we check if a point x̂ is an extreme point of a polyhedron P?

• How do we ensure if a polyhedron P has extreme points?

To answer the first question we need to first define Face and Minimal Face of a polyhedron.

Definition 2. Consider a polyhedron P = {x | Ax ≤ b, Dx = e}, where A ∈ Rm×n . For a given


set of indices J ⊂ {1, . . . , m}, define FJ as FJ = {x ∈ P | a>
i x = bi , for i ∈ J}. If FJ is nonempty
it is called a face of polyhedron P.

Note that for the points in set FJ (if it is non-empty) we have

a>
i x ≤ bi , for i ∈
/ J, a>
i x = bi , for i ∈ J, Dx = e. (1)

A few remarks follows:

• Remark 1. Faces of FJ are also faces of P.

• Remark 2. All faces of P have the same lineality space as P.

• Remark 3. The number of faces is finite and at least one. (P is a face of P for J = ∅.)

Definition 3. A face of polyhedron P is a minimal face if it does not contain another face of P.

1
Theorem 1. Consider a pointed polyhedron P = {x | Ax ≤ b, Dx = e} and a point x̂ which
belongs to P. Define J(x̂) = {i1 , . . . , ik } as the indices of the active constraints at x̂, i.e.,

a>
i x < bi , for i ∈
/ J(x̂), a>
i x = bi , for i ∈ J(x̂). (2)
 
AJ(x̂)
Then, x̂ is a minimal face of the pointed polyhedron P if and only if rank = n, where
D
 >
ai1
 .. 
AJ(x̂) is a sub-matrix of A defined as AJ(x̂) =  . .
a>ik

Proof. Note that according to the definition of a face, we can define the face FJ(x̂) as the set of
points that satisfy the following conditions:

a>
i x ≤ bi , for i ∈
/ J(x̂), a>
i x = bi , for i ∈ J(x̂), Dx = e. (3)

By definition and according to the expressions in (2), the point x̂ belongs to the face FJ(x̂) and
satisfies the conditions in (3).
(a) Now we prove (< −−) [If the rank condition holds, then x̂ is a minimal face].
If the rank condition holds, then x̂ is the only point that satisfies the conditions in (3), cause if
there was another pointz 6= x̂ that satisfies these conditions then their difference x − x̂ belongs to
AJ(x̂)
the nullspace of which would be a contradiction. Hence, x̂ is the only point in the face
D
FJ(x̂) and therefore the dimension of FJ(x̂) is zero and as a result it is a minimal face.
(b) Now we prove (−− >) [If x̂ is a minimal face, then the rank condition holds]. We prove the
equivalent statement [If the rank condition does not hold, then x̂ is not a minimal face]. If the
rank condition does not hold then there exists a vector v 6= 0 such that AJ(x̂) v = 0 and Dv = 0.
Note that for x̂ we have a> i x̂ < bi when i ∈/ J(x̂). Hence, for a sufficiently small t we can show
that a>i (x̂ + tv) ≤ bi and a > (x̂ − tv) ≤ b when i ∈
i i / J(x̂). Indeed, it is easy to verify that for both
x̂ + tv and x̂ − tv the second and third conditions in (3) are also satisfied. Hence, points x̂ + tv
and x̂ − tv also belong to the face FJ(x̂) and as a result the face FJ(x̂) is not minimal and the proof
is complete.

Theorem 2. A point is an extreme point if and only if it is a minimal face of a pointed polyhedron.

Proof. Considering the result of Theorem 1, we need to prove that for a polyhedron P = {x|Ax ≤
b, Dx = e} the following statements for a point x are equivalent.

 
AJ(x)
rank( )=n ⇐⇒ there are no y, z ∈ P and λ ∈ (0, 1) s.t. x = λy + (1 − λ)z
D

First we prove (→) Suppose the claim does not hold and there exist y and z in P such that
x = λy + (1 − λ)z where λ ∈ (0, 1). (we will show that it will lead to a contradiction.)
Now note that a>i x = bi for all i ∈ J(x). This implies that

λa> >
i y + (1 − λ)ai z = bi

At the same time we know that


a>
i y ≤ bi , a>
i z ≤ bi

2
Therefore, these conditions are satisfied if and only if

a>
i y = bi , a>
i z = bi

Therefore, for any ai where i ∈ Jx we can show that

a>
i (y − z) = 0

Hence, y − z belongs to the null space of AJ(x) . It is also easy to check that D(y − z) = 0 and
 
AJ(x)
y − z also belongs to the null space of D. Hence, y − z belongs to the null space of which
D
 
A
contradicts with the fact that rank( J(x) ) = n. Hence, the proof for (→) is complete.
D
Now we proceed to prove (←)  
A
Suppose the claim does not hold and rank( J(x) ) < n which means that there exists a vector v
D
such that AJ(x) v = 0 and Dv = 0. Remember that for x we know that

a>
i x = bi for i ∈ J(x), a>
i x < bi for i ∈
/ J(x), Dx = e

We also know that

a>
i v = 0 if i ∈ J(x), Dv = 0.

Then, it can be shown that for sufficiently small  we can show that y = x ± v satisfy the following
conditions

a>
i y = bi if i ∈ J(x), a>
i y ≤ bi if i ∈
/ J(x), Dx = e.

Therefore, both y1 = x + v and y2 = x − v are in P. But, note that it can be checked that x can
be written as a convex combination of y1 and y2 , i.e, x = 0.5y1 + 0.5y2 which is a contradiction.
The proof is complete.

Corollary 1. Consider x̂ which belongs  polyhedron P = {x | Ax ≤ b, Dx = e}. x̂ is an


 to the
A
extreme point of P if and only if rank( J(x̂) ) = n.
D

2 How should one check if a point is an extreme point? (Rank


Test)
 us a way to check if a given point x̂ is an extreme point of a polyhedron P is to
Corollary 1 gives
A
check if rank( J(x̂) ) = n holds or not. If it holds then x̂ is an extreme point of P, and if not
D
then x̂ is not an extreme point.

Example: Consider the polyhedron P = {(x1 , x2 , x3 ) | x1 ≥ 0, x2 ≥ 0, x3 ≥ 0, x1 + x2 + x3 = 1}.


   
AJ(x̂) 0 0 −1
Question 1 : Is (0.5, 0.5, 0) an extreme point? rank( ) = rank( ) = 2 < 3. Hence,
D 1 1 1
it’s not.

3
 
  0 −1 0
A
Question 2 : Is (1, 0, 0) an extreme point? rank( J(x̂) ) = rank(0 0 −1) = 3. Hence, it’s
D
1 1 1
an extreme point.

Proposition 1. Consider the polyhedron P = {x | x ≥ 0, Dx = e}. The following statements


hold:

• (a) If x̂ ∈ P and rank([di1 , . . . , dik ]) = k, where dj is column j of D and {i1 , . . . , ik } =


{i | x̂i > 0}, then x̂ is an extreme point.

• (b) An extreme point x̂ has at most rank(D) nonzero elements.

Proof. (a) WLOG we assume that the first k inequalities that are not active x̂, and the last n − k
are active. Hence, we can write
 
0 ... 0 −1 0 ... 0
   0 ... 0
 0 −1 ... 0
AJ(x̂)

 .. .
.. .
..
rank( ) = rank( . )=n

D  0 ... ... 0
 0 ... 0 0 0 ... −1 
d1 . . . dk dk+1 . . . ... dn

Note that the rank is n is due to the following observations: 1) The first k columns are linearly
independent, 2) The last n − k columns are linear independent, 3) any column of the last n − k
columns is linearly independent from the first k columns. Hence, all the columns are linearly
independent. As a result, x̂ is an extreme point.
(b) We prove this claim by contradiction. Let’s call rank(D) = k. This means that k rows of D are
linearly independent. If the number of nonzero elements of x̂ is more than k, then it means that
the number of active constraints at x̂ is strictly less than n − k. Hence, rank of AJ(x̂) is strictly
 
AJ(x̂)
less than n − k which means that the rank of is strictly less than n. This contradicts with
D
the fact that x̂ is an extreme point. Hence, the proof is complete.

Theorem 3. Consider the set of doubly stochastic matrices

P = {X ∈ Rn×n | xij ≥ 0 for all i, j, X1 = 1, X> 1 = 1}.

Show that any extreme point of this polyhedron is a permutation matrix and any permutation matrix
is an extreme point of this polyhedron.

(a permutation matrix is a doubly stochastic matrix with elements 0 or 1.)

Proof. Homework.

3 When does a polyhedron have extreme points?


As we mentioned earlier, not all polyhedrons have extreme points. For, instance a half space does
not have any extreme point. In fact, if a polyhedron contains a line it doesn’t have any extreme
point. Only pointed polyhedrons have extreme points. We prove this in the following theorem.

4
Theorem 4. Consider a polyhedron P = {x | Ax ≤ b, Dx = e}. The following statements are
equivalent:

• (a) P has at least one extreme point.

• (b) P does not contain a line.


 
A
• (c) P is a pointed polyhedron. (rank = n)
D

Proof. We prove these statements are equivalent by showing (a) ⇒ (c) ⇒ (b) ⇒ (a)

(a) ⇒ (c) : Note that if (a) holds then P has at least one extreme point which  we denote by x̂.
AJ(x̂)
According to Theorem 1 and Theorem 2, if x̂ is an extreme point then rank = n. Since
D
     
A AJ(x̂) A
rank ≥ rank then we obtain that rank = n and hence P is a pointed
D D D
polyhedron.
 
A
(c) ⇒ (b) : If P is a pointed polyhedron, then n rows of are linearly independent. Let’s
D
 
A
denote {â1 , . . . , âk , d̂k+1 , . . . , d̂n } as the set of n linearly independent rows of . Suppose the
D
polyhedron P contains a line of the form x0 + λv0 , where v0 6= 0. Then, for any λ ∈ R we have

â>
i (x0 + λv0 ) ≤ b̂i for i = 1, . . . , k

This condition only holds if â> >


i x0 ≤ b̂i and âi v0 = 0 for i = 1, . . . , k.
Further, for any λ ∈ R we have

d̂>
i (x0 + λv0 ) = êi for i = k + 1, . . . , n

This condition only holds if d̂> >


i x0 = êi and d̂i v0 = 0 for i = 1, . . . , k.
Note that vectors {â1 , . . . , âk , d̂k+1 , . . . , d̂n } are linearly independent (create a basis for Rn ), but
their inner product with vector v0 is zero. This implies that vector v0 should be 0 which is a
contradiction.

(b) ⇒ (a) : Assume that P does not contain a line. Pick a point x0 in the polyhedron. Define
J(x0 ) as the indices of active constraints at x0 , i.e., J(x0 ) = {i | a> i x0 = bi }. If span({ai | i ∈
J(x )} ∪ {d , . . . , d }) = R n then the proof is complete, since the span of a subset of rows of matrix
 0 1 p
A n
is R , hence, it has n linearly independent rows and hence its rank is n and therefore x0 is an
D
extreme point. Now suppose span({ai | i ∈ J(x0 )} ∪ {d1 , . . . , dp }) 6= Rn , then there exists a vector
v such that its inner product with all of the vectors {ai | i ∈ J(x0 )} and {d1 , . . . , dp } is zero. Now
consider the following line: {y | y = x0 + λv, for any λ ∈ R}. Note that for the points on this line
we have
a>i y = bi , for i ∈ J(x0 ), d>
j y = ej , for j = 1, . . . , p.

Hence, all active constraints remain active for any. point in this line, but since polyhedron P does
not contain any line, there should be a λ∗ such that x0 + λ∗ v hits a new constraint a> k x ≤ bk
> ∗
meaning ak (x0 + λ v) = bk , where k ∈ / J(x0 ). It can be shown that ak does not belong to

5
span({ai | i ∈ J(x0 )} ∪ {d1 , . . . , dp }), since if it belongs to it then we have a> k v = 0 which
implies a>
k x 0 = bk and it contradicts with the fact that k ∈
/ J(x 0 ). Hence, a k does not belong to
span({ai | i ∈ J(x0 )} ∪ {d1 , . . . , dp }). Therefore, if we define a new point in the polyhedron x1 as
x1 = x0 + λ∗ v, we can conclude that

dim (span({ai | i ∈ J(x0 )} ∪ {d1 , . . . , dp })) < dim (span({ai | i ∈ J(x1 )} ∪ {d1 , . . . , dp }))

If we do this process at most n times then we reach a point x̂ such that dim (span({ai | i ∈ J(x0 )})) =
n and therefore x̂ is an extreme point.

Remark 1. According to the above theorem for any polyhedron P = {x | Ax ≤ b, Dx  =  e}, we


A
can easily check if it has extreme points or not by following the rank test of matrix . If the
D
rank of this matrix is n, then it has at least one extreme point. If the rank is smaller than n, then
it does not have any extreme point.

You might also like