LECTURE 7: SEPARATION, FACES OF CONVEX

SETS, EXTREMAL AND EXPOSED POINTS, THE

KREIN-MILMAN THEOREM

We start with two theorems regarding separation.

Theorem 1. Let K, L ⊂ Rn be convex sets with int(K) ̸= ∅ and
int(K) ∩ L = ∅. Then K and L can be separated by a hyperplane.
Proof. We have seen that if K is convex, then int(K) is convex (Exercise
3 on the rst worksheet). But then by last theorem on the last lecture,
the sets int(K) and L can be separated by a hyperplane. Since we
learned that if int(K) ̸= ∅, then K ⊂ cl(int(K)), and a hyperplane
separating int(K) and L separates also cl(int(K)) and L, the assertion
follows. 2
Theorem 2. If K, L ⊂ Rn are disjoint, convex sets, K is compact and
L is closed, then K and L can be strictly separated by a hyperplane.

Proof. We apply the idea of Theorem 4 in the rst lecture. Let x ∈ K

and y ∈ L be arbitrarily chosen points, and let r = ||y − x||. Let L0
be the set of the points of L whose distance from a point of K is at
most r; in other words, let L0 = L ∩ (K + rBn ), where Bn is the closed
unit ball centered at o. Then the distance between any points of L \ L0
and K is greater than r, yielding that dist(K, L) = dist(K, L0 ), where
dist(A, B) = inf{||a − b|| : a ∈ A, b ∈ B}. But both K and L0 are
compact sets, and hence, there are points x ∈ K and y ∈ L for which
dist(x, y) is minimal. Let H be the hyperplane bisecting the segment
[x, y]. Then H strictly separates K and L, as otherwise there are points
x′ ∈ K and y ′ ∈ L for which ||x′ − y ′ || < ||x − y||. 2

We have already seen (Corollary 4 in the rst lecture) that for every
boundary point of a convex set there is a hyperplane through the point
such that the set is contained in one of the two closed half spaces
bounded by the hyperplane. This is the motivation behind the following
denitions.
Denition 1. Let K ⊆ Rn
be a convex set. If H is a closed half space
satisfying K ⊆ H and whose boundary intersects the boundary of K ,
we say that H is a supporting half space of K , and the boundary of H
is a supporting hyperplane of K .
1
2

Denition 2. Let K ⊆ Rn be a closed, convex set and let H be a

supporting hyperplane of K . Then the set H ∩ K is called a proper
face of K . The empty set is called a not proper face of K . The 0-
dimensional faces (consisting of only one point) are called the exposed
points of K , and their set is denoted by ex(K).
Our rst observation implies the next remark in a natural way.
Remark 1. If K ⊆ Rn is closed and convex, and p ∈ Rn is a boundary
point of K , then K has a proper face F such that p ∈ F .

Problem 1. Construct closed, convex sets which have no exposed points.

Proposition 1. If F is a proper face of the closed, convex set K ⊆ Rn ,
then F is closed and convex.

Proof. Since every proper face F of K can be written as F = K ∩ H ,

where H is a supporting hyperplane of K , and a hyperplane is closed
and convex, the assertion follows from the fact that the intersection of
closed, convex sets is closed and convex. 2
Denition 3. K ⊆ Rn be closed and convex. If p ∈ bd K , and for
Let
every q, r ∈ K , p ∈ [q, r] we have p = q or p = r , then we say that p is
an extremal point of K . In other words, the extremal points of K are
the points of K that are not relative interior points of a segment in K .
The set of the extremal points of K is denoted by ext(K).

Proposition 2. If K ⊆ Rn is closed and convex, then ex(K) ⊆ ext(K).

Proof. Let p be an exposed point of K . Then there is a linear functional
f : R → R and a quantity α ∈ R such that K ⊂ f −1 ([α, ∞)), and
n

K ∩ f −1 (α) = {p}. Assume that q, r ∈ K and p ∈ [q, r]. Then there is

value t ∈ [0, 1] with p = tq+(1−t)r. By our conditions and the linearity
of f , we have α = f (p) = tf (q) + (1 − t)f (r) ≥ tα + (1 − t)α = α.
But here, inequality occurs if and only if t = 1 and f (q) = α, or t = 0
and f (r) = α, or f (q) = f (r) = α. But these yield p = q , p = r, and
p = q = r, respectively, implying the statement. 2

Example. Let K ⊆ R2 be the union of the unit square [0, 1]2 and the
circular region dened by the inequality (x − 1/2)2 + y 2 ≤ 1/4.then o
and the point (1, 0) are extremal points of K , but not exposed points of
L. Thus, there are closed, convex sets K for which ex(K) and ext(K)
do not coincide.
Our next theorem explores the connection between extremal points
and linear functionals.
 3

Theorem 3. Let K ⊆ Rn be a closed, convex set, and let f : Rn → R

be a linear functional whose minimal or maximal value on K is α. Let
F = K ∩ f −1 (α). Then p ∈ F is an extremal point of F if and only if it
is an extremal point of K . In other words, ext(F ) = ext(K) ∩ f −1 (α).

Before proving Theorem 3, we observe that if p ∈ ex(K), then there

is a linear functional f : Rn → R which attains its minimum on K
only at p. Thus, a consequence of this theorem is the containment
ex(K) ⊆ ext(K) for every closed, convex set K .
Proof. Assume that p ∈ ext(K) and p ∈ F . Then, by the denition of
extremal point, for any q, r ∈ K , p ∈ [q, r] we have q = p or r = p. In
particular, this holds also for any q, r ∈ F , implying that p ∈ ext(F ).
Now, let p ∈ ext(F ), and consider points q, r ∈ K with p ∈ [q, r].
If q ̸= p and r ̸= p, then for a suitable t ∈ (0, 1), p = tq + (1 − t)r.
But from this α = f (p) = f (tq + (1 − t)r) = tf (q) + (1 − t)f (r). As
f (q), f (r) ≥ α, there is equality if and only if f (q) = f (r) = α, azaz
ha q, r ∈ F . But as p ∈ ext(F ), thiy yields q = p or r = p, which is a
contradiction. 2
Our next theorem shows an important property of extremal points.
Theorem 4 (Krein, Milman). Any compact,convex set K ⊂ Rn is the
convex hull of its extremal points.

Proof. We prove the statement by induction on the dimension. Assume

that K ⊂ R is a compact, convex set. Then K is a closed segment,
whose extremal points are its endpoints, and the segment is the convex
hull of its endpoints. Thus, the assertion holds for n = 1.
Assume that the statement is true for any at most (n−1)-dimensional
compact, convex set, and let K be K an n-dimensional compact, convex
set. Let p ∈ K be arbitrary, and let L bi an arbitrary line through p.
According to our conditions, L ∩ K is a closed nite segment. Let
the endpoints of this segment be q and r, where these points may
not be distinct from each other or p. Then, by Remark 1, there are
faces Fq and Fr of K such that q ∈ Fq and r ∈ Fr . But as Fq and
Fr are convex subsets of the boundary of K , they have no interior
points, and thus, by Lemma 2 of the fourth lecture, they are at most
(n−1)-dimensional compact, convex sets. By the induction hypothesis,
we have q ∈ conv ext(Fq ) és r ∈ conv ext(Fr ). But by the denition
of face, there are linear functionals fq : Rn → R and fr : Rn → R
attaining their minima exactly at Fq and Fr , respectively, and thus, by
Theorem 3, the extremal points of Fq and Fr are extremal points of K .
But then p ∈ [q, r] ⊆ conv(ext(Fq ) ∪ ext(Fr )) ⊆ conv(ext(K)). 2
4

We have seen that the extremal points of a set are not necessarily ex-
posed points. On the other hand, it is true that they are accumulation
points of sequences of exposed points.
Theorem 5 (Straszevicz). For any compact, convex set K ⊂ Rn we
have K = cl(conv(ex(K))); or in other words, K is equal to the closure
of convex hull of its exposed points.

Proof. Let x ∈ ext(K) and ε > 0 be arbitrary. Let us consider the

compact, convex set Kε = conv(K \ int Bε (x)) ⊆ K , where Bε (x)
denotes the closed ball of radius ε and center x. If x ∈ Kε , then by the
Carathéodory theorem it is the convex combination of at most n + 1
points of (K \ int Bε (x)); that is, it is a relative interior point of a
segment in K . But this contradicts the assumption that x ∈ ext(K),
and thus, x ∈/ Kε .
Note that Kε is a compact, convex set, and thus, it can be strictly
separated from p. In other words, there is a hyperplane H such that
one of the closed half spaces bounded by it intersects K in a subset of
Bε (x), and this half space contains x in its interior. Let H + denote this
closed half space. Let L be the half line starting at x, perpendicular
to H and intersecting H . For any y ∈ L let z(y) be a farthest point
of K from y . Then z(y) ∈ ex(K) for any y ∈ L (see Problem sheet
5, Exercise 4). On the other hand, if y is suciently far from x, then
z(y) ∈ Bε (x). Thus x ∈ cl(ex(K)), from which ext(K) ⊆ cl(ex(K)).
By the containment relation conv(cl(X)) ⊆ cl(conv(X)), satised
for any set X ⊆ Rn , and by the Krein-Milman Theorem, we have
K ⊆ conv(ext(K)) ⊆ conv(cl(ex(K))) ⊆ cl(conv(ex(K))) ⊆ K,
that is, K = cl(conv(ex(K))). 2