Convex Geometry Carsten Sch¨ utt November 25, 2006
Convex Geometry
Carsten Schutt
November 25, 2006
2
Contents
0.1 Convex sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40.2 Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90.3 Extreme points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150.4 Blaschke selection principle . . . . . . . . . . . . . . . . . . . . . . . 180.5 Polytopes and polyhedra . . . . . . . . . . . . . . . . . . . . . . . . . 230.6 Volume and Surface Area of a Convex Body . . . . . . . . . . . . . . 310.7 Steiner and Schwarz Symmetrizations . . . . . . . . . . . . . . . . . . 380.8 A Characterization of the Ellipsoid . . . . . . . . . . . . . . . . . . . 470.9 Theorem of F. John . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500.10 Blaschke-Santalo inequality . . . . . . . . . . . . . . . . . . . . . . . 570.11 Packing of the Euclidean Sphere . . . . . . . . . . . . . . . . . . . . . 580.12 Approximation of the Euclidean Ball . . . . . . . . . . . . . . . . . . 62
3
4 CONTENTS
0.1 Convex sets
As a norm on Rn we consider the Euclidean norm
‖x‖ =
√√√√ n∑i=1
|xi|2
The ball with center x and radius r is
B(x, r) = y|‖x − y‖ ≤ r
The boundary of a set A is ∂A = A\A. A set K in a real vector space is said to be
convex if we have for all x, y ∈ K that
tx + (1 − t)y| t ∈ [0, 1] ⊆ K
Examples of convex sets are the n-dimensional unit balls Bnp of lnp .
Bnp =
x ∈ Rn
∣∣∣∣∣‖x‖p = (n∑
i=1
|xi|p)1p ≤ 1
Let x1, . . . , xn be elements of a real vector space and t1, . . . tn real numbers from theinterval [0, 1] with
∑ni=1 ti = 1. Then
n∑i=1
tixi
a convex combination of the vectors x1, . . . , xn.
Lemma 1 Let C be a convex set in Rn, k ∈ N, and x1, . . . , xk ∈ C. Then for allt ∈ Rn with 0 ≤ ti ≤ 1 and
∑ki=1 ti = 1 we have
k∑i=1
tixi ∈ C
Proof. We use induction over the number of vectors x1, . . . , xk. The case k = 2follows from the definition of the convexity of C.
k+1∑i=1
tixi = (k∑
j=1
tj)
(k∑
i=1
ti∑kj=1 tj
xi
)+ tk+1xk+1
By induction hypotheses we have
k∑i=1
ti∑kj=1 tj
xi ∈ C
0.1. CONVEX SETS 5
Lemma 2 An arbitrary intersection of convex sets is again convex.
A hyperplane in Rn is a set x| < x, ξ >= t. There are two halfspaces associatedwith a hyperplane
H+ = x| < x, ξ >≤ t H− = x| < x, ξ >≥ t
Halfspaces are assumed to be closed sets.A supporting hyperplane H of a convex set C is a hyperplane such that one of
its halfspaces contains C, C ⊆ H+, and such that C is not contained in a halfspacethat is properly contained in H+.
Lemma 3 Let A and B be convex sets in Rn. Then
A + B = x + y|x ∈ A, y ∈ B
is a convex set. If A and B are compact, then A + B is also compact.
Proof. Let x, y ∈ A + B. Then there are x1, y1 ∈ A and x2, y2 ∈ B with
x = x1 + x2 y = y1 + y2
Then
tx + (1 − t)y = t(x1 + x2) + (1 − t)(y1 + y2) = (tx1 + (1 − t)y1) + (tx2 + (1 − t)y2)
Lemma 4 The closure and the interior of a convex set K in Rn are convex.
Proof. Let x, y ∈ K. Then there are sequences xi ∈ K, i ∈ N, and yi ∈ K, i ∈ N,with limi→∞ xi = x and limi→∞ yi = y. By the convexity of K for all i ∈ N and allt with 0 < t < 1
txi + (1 − t)yi ∈ K
Thereforetx + (1 − t)y = lim
i→∞(txi + (1 − t)yi) ∈ K
Now we show that the interior of a convex set is convex. Let x and y be interiorpoints of K. Then there is ε > 0 such that
B(x, ε) ⊆ K B(y, ε) ⊆ K
We have to show that [x, y] ⊆K, i.e. the points tx+(1− t)y|0 ≤ t ≤ 1 are interior
points. Since K is convex we have for all z, w ∈ B(0, ε)
t(z + x) + (1 − t)(w + y) = tz + (1 − t)w + (tx + (1 − t)y) ∈ K
6 CONTENTS
Thus
(tx+(1−t)y)+tB(0, ε)+(1−t)B(0, ε) = tz+(1−t)w+(tx+(1−t)y)|z, w ∈ B(0, ε) ⊆ K
It followsB(0, ε) + (tx + (1 − t)y) ⊆ K
Lemma 5 For all x, y ∈ Rn we have(i) | < x, y > | ≤ ‖x‖‖y‖(ii) ‖x + y‖ ≤ ‖x‖ + ‖y‖(iii) |‖x‖ − ‖y‖| ≤ ‖x − y‖(iv) If there is ε > 0 such that for all t with 0 < t < ε we have ‖x‖ ≤ ‖x + ty‖, then0 ≤< x, y >.
Proof.‖x‖2 ≤ ‖x + ty‖2 = ‖x‖2 + 2t < x, y > +t2‖y‖2
0 ≤< x, y > +1
2t‖y‖2
Taking the limit of t towards 00 ≤< x, y >
The convex hull CH(A) of a set A is the smallest convex set containing A, i.e.the intersection of all convex sets containing A.
Lemma 6 Let A be a subset of Rn. Then
CH(A) =
k∑
i=1
tixi
∣∣∣∣∣ k ∈ N, xi ∈ A, ti ≥ 0, i = 1, . . . , k,k∑
i=1
ti = 1
Proof. We put
A =
k∑
i=1
tixi
∣∣∣∣∣ k ∈ N, xi ∈ A, ti ≥ 0, i = 1, . . . , k,k∑
i=1
ti = 1
Since CH(A) is a convex set that contains A we get by Lemma 1
CH(A) ⊇ A
We show now that A is convex. Let x, y ∈ A. Then there are xi ∈ A, yj ∈ A suchthat
x =k∑
i=1
tixi y =m∑
j=1
sjyj
0.1. CONVEX SETS 7
Moreover, let r ∈ R, 0 ≤ r ≤ 1. Then
k∑i=1
rti +m∑
j=1
(1 − r)sj = r + (1 − r) = 1
Consequently,
rx + (1 − r)y ∈
k∑i=1
tixi
∣∣∣∣∣ k ∈ N, xi ∈ A, ti ≥ 0, i = 1, . . . , k,k∑
i=1
ti = 1
Lemma 7 (Caratheodory) Let A be a subset of Rn. Then
CH(A) =
n+1∑i=1
tixi
∣∣∣∣∣ xi ∈ A, ti ≥ 0, i = 1, . . . , n + 1,n+1∑i=1
ti = 1
The important point in the lemma of Caratheodory is that we need at most n+1to write a point of the convex hull as a convex combination.
Proof. By Lemma 6 for every x ∈ CH(A) there are k ∈ N, x1, . . . , xk ∈ A andt1, . . . , tk ≥ 0 with
x =k∑
i=1
tixi andk∑
i=1
ti = 1
We consider now that this representation has the smallest possible k. (We assumethat all ti are strictly bigger than 0.) We show that k ≤ n + 1. Suppose thatk ≥ n + 2. Then, since x1, . . . , xn+2 are linearly dependent there are real numberss1, . . . , sn+2 such that
(1)k∑
i=1
sixi = 0k∑
i=1
si = 0
and not all of the coefficients si are 0. We prove this. Since x1, . . . , xk−1 are linearlyindependent there are real numbers r1, . . . , rk−1 such that not all of the coefficientsare 0 and
k−1∑i=1
rixi = 0
If∑k−1
i=1 ri = 0 we are done. Therefore we may suppose that∑k−1
i=1 ri = 0. Moreover,we may assume that r1 = 0. (If not, we rearrange the vectors.) Again, sincex2, . . . , xk are linearly dependent there are coefficients t2, . . . , tks such that not all ofthe ti are 0 and
k∑i=2
tixi = 0
8 CONTENTS
If∑k
i=2 ti = 0 we are done. If not, we may assume that
n+1∑i=1
ri =n+2∑i=2
ti
Now we choose as our coefficients
s1 = r1, s2 = r2 + t2, . . . , sk−1 = rk−1 + tk−1, sk = tk
Clearly, these coefficients sum up to 0 and not all of them equal 0 since r1 = 0. Thuswe have shown (1).
Now we consider
x =k∑
i=1
(ti + λsi)xi
We can choose λ > 0 in such a way that all the coefficients ti − λsi are nonnegativeand at least one of them equals 0. Therefore k was not chosen to be minimal.
As the volume of a convex, compact C set we take the Lebesgue measure ofC which is in this case the same as the Riemann integral over the characteristicfunction of C.
Example 1voln(Bn
p ) =
Example 2 [StR] Let Cnp be the Schatten class of dimension n. Then the volume
of the unit ball equals
2−n
(n∏
k=1
kπ
k2
Γ(k2
+ 1)
)2 ∫Ωp
∏1≤j<k≤n
(s2k − s2
j)ds1 · · · dsn
where
Ωp =
s
∣∣∣∣∣n∑
j=1
spj < 1
The center of gravity or barycenter of a convex body is
cg(K) =1
voln(K)
∫K
xdx
0.2. SEPARATION 9
0.2 Separation
Let A be a closed nonempty set in Rn and x ∈ Rn. We call a point a0 ∈ A nearestpoint to x or best approximation of x if
‖x − a0‖ = inf‖x − a‖|a ∈ A
If x ∈ A then the nearest point is x itself.
Lemma 8 Let A be a closed, convex, nonempty set in Rn and x ∈ Rn.(i) There is a nearest point a0 of x in A.(ii) For all a ∈ A we have
< x − a0, a − a0 >≤ 0
(Geometrically this means that A is contained in the halfspace y| < x− a0, y >≤<x − a0, a0 >.)(iii) The nearest point a0 is unique.
Proof. (i) If x ∈ A then we choose a0 = x. Suppose now that x /∈ A. Then thereis a sequence an ∈ A, n ∈ N, such that
limn→∞
‖x − an‖ = inf‖x − a‖|a ∈ A
Almost all an are contained in the ball B(x, 2ρ) where
ρ = inf‖x − a‖|a ∈ A
SinceA ∩ B(x, 2ρ)
is a compact set there is a convergent subsequence ani, i ∈ N. We have
limi→∞
ani= a0
limi→∞
‖x − ani‖ = inf‖x − a‖|a ∈ A
This implies‖x − a0‖ = inf‖x − a‖|a ∈ A
(ii) Since A is convex we have for all t with 0 < t < 1 and a ∈ A that (1−t)a0+ta ∈ A.Since a0 is nearest point to x
‖x − a0‖2 ≤ ‖x − ((1 − t)a0 + ta)‖2 = ‖(x − a0) + t(a0 − a)‖2
By Lemma 50 ≤< x − a0, a0 − a >
10 CONTENTS
(iii) Suppose there is another point a1 such that
‖x − a1‖ = inf‖x − a‖|a ∈ A
By (ii) we get
< x − a0, a1 − a0 >≤ 0
< x − a1, a0 − a1 >≤ 0
We add the two inequalities
0 ≥< x − a0, a1 − a0 > + < x − a1, a0 − a1 >= ‖a0 − a1‖2
Let A be a nonempty, closed, convex set. The metric projection pA : Rn → Amaps a point x onto its nearest point in A.
Lemma 9 Let A be a nonempty, closed, convex set in Rn. The metric projectionpA is a 1-Lipschitz map, i.e. for all x, y ∈ Rn
‖pA(x) − pA(y)‖ ≤ ‖x − y‖
Proof. pA is well defined since by Lemma 8 for every x ∈ Rn there is a unique pointof minimal distance.
Let x, y ∈ Rn. By Lemma 8 for all a ∈ A
< x − pA(x), a − pA(x) >≤ 0
< y − pA(y), a − pA(y) >≤ 0
In particular
< x − pA(x), pA(y) − pA(x) >≤ 0
< y − pA(y), pA(x) − pA(y) >≤ 0
Thus
< (x − pA(x)) − (y − pA(y)), pA(x) − pA(y) >≥ 0
With this
‖x − y‖2 = ‖(x − pA(x)) − (y − pA(y)) + (pA(x) − pA(y))‖2
= ‖(x − pA(x)) − (y − pA(y))‖2
+2 < (x − pA(x)) − (y − pA(y)), pA(x) − pA(y) > +‖pA(x) − pA(y)‖2
≥ ‖pA(x) − pA(y)‖2
0.2. SEPARATION 11
Proposition 1 Let A and B be nonempty, disjoint, convex sets in Rn. Moreover, letA be closed and B compact. Then A and B can be strictly separated by a hyperplane,i.e. there exists ξ ∈ Rn such that
supa∈A
< ξ, a >< infb∈B
< ξ, b >
Proof. There are two points a0 ∈ A and b0 ∈ B such that
‖a0 − b0‖ = inf‖a − b‖|a ∈ A, b ∈ B
We show this. Clearly, there are two sequences an ∈ A and bn ∈ B, n ∈ N, suchthat
limn→∞
‖an − bn‖ = inf‖a − b‖|a ∈ A, b ∈ B
By compactness of B there is a subsequence bni, i ∈ N, that converges
limi→∞
bni= b0
We have
‖ani− bni
‖ − ‖bni− b0‖ ≤ ‖ani
− b0‖ ≤ ‖ani− bni
‖ + ‖bni− b0‖
It follows thatlimi→∞
‖ani− b0‖ = inf‖a − b‖|a ∈ A, b ∈ B
Almost all of ani, i ∈ N, are elements of the ball B(b0, 2ρ) with
ρ = inf‖a − b‖|a ∈ A, b ∈ B
thus the sequence is contained in a compact set and has a convergent subsequence.The limit is denoted by a0.
Now we choose ξ = a0 − b0. Since a0 is nearest point for b0 and vice versa wehave by Lemma 8 for all a ∈ A and b ∈ B
< a0 − b0, a − b0 >≤ 0
< b0 − a0, b − a0 >≤ 0
It follows
< a0 − b0, b > ≥< a0 − b0, a0 >= ‖a0‖2− < b0, a0 >
= 12‖a0‖2 − ‖b0‖2 + ‖a0 − b0‖2
> 12‖a0‖2 − ‖b0‖2
> 12‖a0‖2 − ‖b0‖2 − ‖a0 − b0‖2
=< a0, b0 > −‖b0‖2 =< a0 − b0, b0 >≥< a0 − b0, a >
Therefore
infb∈B
< a0 − b0, b >> 12‖a0‖2 − ‖b0‖2 > sup
a∈A< a0 − b0, a >
12 CONTENTS
Corollary 1 Let A be a nonempty, closed, convex set in Rn and x ∈ Rn with x /∈ A.Then there is a hyperplane that separates A and x strictly.
Corollary 2 Let A be a closed convex set in Rn. Then A is the intersection of allclosed halfspaces H+ containing A, i.e.
A =⋂
A⊆H+
H+
Proof. Let x ∈ Rn with x /∈ A. Then there is a hyperplane separating x and Astrictly.
Lemma 10 Let K be a nonempty, closed, convex set in Rn and let x ∈ ∂K. Thenthere is y /∈ K such that p(y) = x.
Proof. There is a sequence xi ∈ Rn \ K, i ∈ N, with
limi→∞
xi = x
We put
yi = p(xi) +xi − p(xi)
‖xi − p(xi)‖We have
‖yi − p(xi)‖ =
∥∥∥∥ xi − p(xi)
‖xi − p(xi)‖
∥∥∥∥ = 1
andp(yi) = p(xi)
We verify that p(yi) = p(xi). Since p(xi) is the nearest point to xi and p(yi) to yi
we have by Lemma ?? that for all ξ, η ∈ K
< xi − p(xi), ξ − p(xi) >≤ 0
< yi − p(yi), η − p(yi) >≤ 0
Choosing ξ = p(yi) and η = p(xi) we get
< xi − p(xi), p(yi) − p(xi) >≤ 0
and
0 ≥< yi − p(yi), p(xi) − p(yi) >
=< p(xi) − p(yi) +xi − p(xi)
‖xi − p(xi)‖, p(xi) − p(yi) >
= ‖p(xi) − p(yi)‖2 +
⟨xi − p(xi)
‖xi − p(xi)‖, p(xi) − p(yi)
⟩
0.2. SEPARATION 13
Thus we get
0 ≥⟨
xi − p(xi)
‖xi − p(xi)‖, p(yi) − p(xi)
⟩≥ ‖p(xi) − p(yi)‖2
which implies p(xi) = p(yi). Therefore
p(x) = p(
limi→∞
xi
)= lim
i→∞p(xi) = lim
i→∞p(yi)
The sequence yi, i ∈ N, is bounded since
‖yi‖ =
∥∥∥∥p(xi) +xi − p(xi)
‖xi − p(xi)‖
∥∥∥∥ ≤ 1 + ‖p(xi)‖
and xi, i ∈ N, is a convergent sequence. Therefore the sequence yi, i ∈ N, has aconvergent subsequence yij , j ∈ N,
y = limj→∞
yij
We have
p(y) = limj→∞
p(yij) = limj→∞
p(xij) = p( limj→∞
yij) = p(x) = x
Moreover we have x = y. Indeed,
‖x − y‖ = limj→∞
‖xij − yij‖ = limj→∞
∥∥∥∥xij − p(xij) +xij − p(xij)
‖xij − p(xij)‖
∥∥∥∥Since
limj→∞
xij − p(xij) = 0
we get ‖x − y‖ = 1.
Lemma 11 Let K be a nonempty, convex set in Rn that does not contain the origin0. Then there is a hyperplane H with
0 ∈ H− and K ⊆ H+
Proof. Suppose 0 /∈ K. By Lemma ?? K is convex. Now we can apply Corollary1.
Now suppose that 0 ∈ K. By Lemma ?? there is x ∈ Rn \ K with p(x) = 0. Asa hyperplane we choose now H = ξ| < x, ξ >= 0. Clearly, 0 ∈ H and therefore0 ∈ H− and 0 ∈ H+. Since 0 is the nearest point to x we have by Lemma 8 for allξ ∈ K
< ξ, x >≤ 0
14 CONTENTS
Proposition 2 Let K and C be nonempty, convex, disjoint sets in Rn. Then thereis a hyperplane separating K and C.
If one of the sets is lower dimensional then it might happen that the set iscontained in the separating hyperplane.
Proof. The setC − K = x − y|x ∈ C, y ∈ K
is convex and 0 /∈ C−K. We verify this. Let x1−y1, x2−y2 ∈ C−K and 0 < t < 1.Then
t(x1 − y1) + (1 − t)(x2 − y2) = (tx1 + (1 − t)x2) − (ty1 + (1 − t)y2) ∈ C − K
We have 0 /∈ C −K because C and K are disjoint. Therefore, there is a hyperplaneseparating 0 from C − K, i.e. there is a ξ ∈ Rn, ξ = 0, such that for all x ∈ C − K
0 ≤< ξ, x >
Thus there is ξ such that for all y ∈ K and z ∈ C
0 ≤< ξ, y − z >
and consequently< ξ, z >≤< ξ, y >
A hyperplane H is called a support hyperplane to the convex set K if H∩K = ∅and K ⊆ H+ or K ⊆ H−.
Proposition 3 Let K be a nonempty, convex set in Rn and x ∈ ∂K. Then thereis a supporting hyperplane of K that contains x.
Proof. Suppose that K has no interior point. Then every point of K is a boundarypoint and K is contained in a hyperplane. We verify this.
Suppose that there is no hyperplane that contains K. We may assume that0 ∈ K. We choose a point x1 ∈ K with 0 = x1. Now we choose a point x2 ∈ Kthat is not contained in the subspace spanned by x1. Such a point exists becauseotherwise K would be contained in this subset. By induction we get that thereare n linearly independent points x1, . . . , xn that are contained in K. Therefore[0, x1, . . . , xn] ⊆ K and [0, x1, . . . , xn] contains an interior point.
Suppose now that K has an interior point and let x be a boundary point of K.
Then we consider the disjoint, convex setsK and x. By Proposition 2 there is a
separating hyperplane H, i.e.
H+ ⊇K and x ∈ H−
It follows H+ ⊇ K and x ∈ H.
0.3. EXTREME POINTS 15
0.3 Extreme points
A point x of a convex set C in Rn is not an extreme point of C if there are y, z ∈ C,y = z, such that
x =y + z
2
In fact this is equivalent to the following. A point x of a convex set C in Rn is notan extreme point of C if there are y, z ∈ C, y = z, and t with 0 < t < 1 such that
x = ty + (1 − t)z
We show this. We may suppose that 12≤ t ≤ 1. We choose the points y and 2x− y.
Then we havex = 1
2y + 1
2(2x − y)
and 2x − y ∈ C because
2x − y = 2ty + 2(1 − t)z − y = (2t − 1)y + (2 − 2t)z
where 2t − 1 ≥ 0 and (2t − 1) + (2 − 2t) = 1.These considerations can be extended further. Suppose x ∈ C, x1, . . . , xk ∈ C
and t1, . . . , tk > 0 with∑k
i=1 ti = 1 and
x =k∑
i=1
tixi
Then x is not an extreme point. We choose y = x1 and
z =k∑
i=2
ti∑ki=2 ti
xi
Then
x = t1y +
(k∑
i=2
ti
)z
Therefore, x is not an extreme point. The following theorem is a special case of theKrein-Milman theorem.
Theorem 0.3.1 A convex, compact subset of Rn equals the closed, convex hull ofits extreme points.
This result can be improved. In fact, it can be shown that in Rn a convex,compact set equals the convex hull of its extreme points. For infinite dimensionalvector spaces one has to take the closed convex hull.
Lemma 12 Let C be a convex subset of Rn and H a supporting hyperplane of C.The the extreme points of C∩H are exactly the extreme points of C that are containedin H.
16 CONTENTS
Proof. Clearly, all extreme points of C that are contained in H are also extremepoints of C ∩ H.
Let x be an extreme point of C ∩H and not an extreme point of C. Then thereare y and z with y, z ∈ C and x = y+z
2. We show that y, z ∈ H.
Let H be given by < ·, ξ >= t and suppose that C is a subset of v| < v, ξ >≤ t.Then
t =< x, ξ >=1
2(< y, ξ > + < z, ξ >) ≤ t
This means that we have equality in the above equation which means y, z ∈ H.
Lemma 13 Let C be a nonempty, compact, convex set in Rn. Then C has anextreme point.
Proof. If C consists of one point only, this point is also extreme point of C. Thuswe may assume that C consists of more than just one point.
Since C is compact, there is z ∈ C such that the Euclidean norm is maximal,i.e.
‖z‖ = maxx∈C
‖x‖
We claim that z is an extreme point. We show this. Suppose that z is not anextreme point of C. Then there are x, y ∈ C and t, 0 < t < 1, with
z = tx + (1 − t)y
We have ‖x‖ = ‖y‖ = ‖z‖ because if ‖x‖ < ‖z‖ or ‖y‖ < ‖z‖ we get
‖z‖ = ‖tx + (1 − t)y‖ ≤ t‖x‖ + (1 − t)‖y‖ < ‖z‖
Therefore, we get
‖z‖2 = ‖tx + (1 − t)y‖2
= t2‖x‖2 + 2t(1 − t) < x, y > +(1 − t)2‖y‖2
≤ t2‖x‖2 + 2t(1 − t)‖x‖‖y‖ + (1 − t)2‖y‖2
= (t‖x‖ + (1 − t)‖y‖)2 = ‖z‖2
Thus we have2t(1 − t) < x, y >= 2t(1 − t)‖x‖‖y‖
Since 0 < t < 1 we get< x, y >= ‖x‖‖y‖
This implies that there is s with x = sy. Since ‖x‖ = ‖y‖ we have x = y or x = −y.Since < x, y >≥ 0 we conclude x = y and thus z is an extreme point.
Lemma 14 Let C be a compact, convex set in Rn. Then the set of extreme points ofC is not empty and every support hyperplane H of C contains at least one extremepoint of C.
0.3. EXTREME POINTS 17
Proof. Let H be a support hyperplane of C. C ∩ H is a compact, convex set. ByLemma 13 C ∩ H has an extreme point. By Lemma 12 this is an extreme point ofC.
Proof of Theorem 0.3.1 We denote the set of extreme points of C by
Ext(C)
Since C is compact and closed we have
CH(Ext(C)) ⊆ C
Now we show the opposite inclusion, namely
CH(Ext(C)) ⊇ C
Suppose that this inclusion does not hold, then there is z ∈ C with z /∈ CH(Ext(C)).Since CH(Ext(C)) is a closed convex set by Corollary 1 there is a hyperplane H =x| < ξ, x >= t that separates CH(Ext(C)) and z strictly, i.e.
maxx∈CH(Ext(C))
< ξ, x ><< ξ, z >
Consider now the hyperplane
L = x| < ξ, x >= maxy∈C
< ξ, y >
This hyperplane is a support hyperplane of C and it contains by Lemma ?? anextreme point v of C. In particular,
v ∈ L ∩ CH(Ext(C))
It follows< ξ, v ><< ξ, z >
On the other hand, since z ∈ C
< ξ, z >≤ maxx∈C
< ξ, x >=< ξ, v ><< ξ, z >
18 CONTENTS
0.4 Blaschke selection principle
A convex body in Rn is a convex, compact set with an interior point. The supportfunction hC : Rn → R of a convex body C is defined by
hC(ξ) = supx∈C
< ξ, x >
The Hausdorff distance between two convex, compact sets K and C is
dH(C, K) = max infρ|C ⊆ K + Bn2 (0, ρ), infρ|K ⊆ C + Bn
2 (0, ρ)
Lemma 15 Let C and K be two convex, compact sets in Rn. Then we have(i)
dH(C, K) = max
supx∈C
infy∈K
‖x − y‖, supy∈K
infx∈C
‖x − y‖
(ii)dH(C, K) = sup
‖ξ‖=1
|hC(ξ) − hK(ξ)|
Proof. (i) Suppose we have C ⊆ K + Bn2 (0, ρ). Then, for every x ∈ C there are
y ∈ K and z ∈ Bn2 (0, ρ) with x = y + z. This implies ‖x− y‖ = ‖z‖ ≤ ρ. Therefore,
for every x ∈ C we have infy∈K ‖x − y‖ ≤ ρ and subsequently
supx∈C
infy∈K
‖x − y‖ ≤ ρ
and thereforesupx∈C
infy∈K
‖x − y‖ ≤ infρ|C ⊆ K + Bn2 (0, ρ)
It follows that
dH(C, K) ≥ max
supx∈C
infy∈K
‖x − y‖, supy∈K
infx∈C
‖x − y‖
(ii)C ⊆ K + Bn
2 (0, ρ) and K ⊆ C + Bn2 (0, ρ)
is equivalent to: For all ξ, ‖ξ‖ = 1, we have
hC(ξ) ≤ hK(ξ) + ρ and hK(ξ) ≤ hC(ξ) + ρ
We prove the equivalence.
hC(ξ) = supx∈C
< ξ, x >≤ supx∈K+Bn
2 (0,ρ)
< ξ, x >= supy ∈ K
z ∈ Bn2 (0, ρ)
< ξ, y + z >
= supy∈K
< ξ, y > + supz∈Bn
2 (0,ρ)
< ξ, z >= hK(ξ) + ρ
0.4. BLASCHKE SELECTION PRINCIPLE 19
On the other hand, if C K + Bn2 (0, ρ) then there is x ∈ C but x /∈ K + Bn
2 (0, ρ).By the Theorem of Hahn-Banach there is a hyperplane H = y| < ξ, y >= t thatseparates x and K + Bn
2 (0, ρ), i.e.
supy∈K+Bn
2 (0,ρ)
< ξ, y ><< ξ, x > .
It follows that hK(ξ) + ρ < hC(ξ).
The symmetric difference distance is
dS(C, K) = voln((C \ K) ∪ (K \ C))
Lemma 16 (i) dH is a metric on the space K of all compact, convex subsets of Rn.(ii) dS is a metric on the space K of all compact, convex subsets of Rn.
Proof. For compact sets C and K we have
dH(C, K) = max‖ξ‖2=1
|hC(ξ) − hK(ξ)|
From this it follows immediately that dH(C, C) = 0, dH(C, K) ≥ 0, and dH(C, K) ≤dH(C, M) + dH(M, K). We have that dH(C, K) > 0 if C = K since two compact,convex sets are equal provided that their support functions are equal. We verifythis. Suppose that C = K. Then, w.l.o.g there is x ∈ C with x /∈ K. Then byCorollary 1 there is ξ ∈ Rn with
supy∈K
< ξ, y ><< ξ, x >
This implies
hK(ξ) = supy∈K
< ξ, y ><< ξ, x >≤ hC(ξ)
Lemma 17 Let Ki, i ∈ N, be a sequence of compact, nonempty, convex sets in Rn
such that we have for all i ∈ N
Ki+1 ⊆ Ki
Then∞⋂i=1
Ki
is a convex, compact, nonempty set and the sequence Ki, i ∈ N, converges to it withrespect to the Hausdorff metric.
20 CONTENTS
Proof. We have to show that for all ε > 0 there is l ∈ N such that we have for allj ≥ l
dH
( ∞⋂i=1
Ki, Kj
)< ε
This means
max
sup
x∈⋂∞
i=1 Ki
infy∈Kj
‖x − y‖, supy∈Kj
infx∈
⋂∞i=1 Ki
‖x − y‖
= supy∈Kj
infx∈
⋂∞i=1 Ki
‖x − y‖ < ε
Suppose this is not true. Then there is a subsequence Kjm , m ∈ N and elementsxjm ∈ Kjm such that
infx∈∩∞
i=1Ki
‖xjm − x‖ ≥ ε
Then there is a subsequence of xjm , m ∈ N, that converges to an element z. Wehave
z ∈∞⋂i=1
Ki and infx∈∩∞
i=1Ki
‖z − x‖ ≥ ε
This cannot be.
Lemma 18 The space K of all nonempty, compact, convex subsets of Rn with themetric dH is complete.
Proof. Let Ki, i ∈ N, be a Cauchy sequence. We put
Cj = CH
( ∞⋃i=j
Ki
)
The sequence Cj, j ∈ N, is a decreasing sequence of compact, convex sets. The setsCj are obviously convex and closed. We show that they are bounded. Since thesequence Ki, i ∈ N, is a Cauchy sequence we have that there is i0 such that for alli ≥ i0
dH(Ki0 , Ki) ≤ 1
This means that for all i ≥ i0
Ki ⊆ Ki0 + Bn2 (0, 1)
Since Ki0 + Bn2 (0, 1) is a bounded set, the set
⋃i≥i0
Ki
is also bounded.
0.4. BLASCHKE SELECTION PRINCIPLE 21
By Lemma 17 the sequence Cj, j ∈ N, converges in the metric space. We claimthat the sequence Ki, i ∈ N, converges to the same limit. It suffices to show thatfor every ε > 0 there is j0 so that we have for all j ≥ j0
dH(Cj, Kj) < ε
Since Ki, i ∈ N, is a Cauchy sequence there is i0 so that we have for all l, m ≥ i0
dH(Kl, Km) < ε
orKl ⊆ Km + Bn
2 (0, ε)
Therefore we get for all j ≥ i0
Kj ⊆ Cj = CH
( ∞⋃i=j
Ki
)⊆ Kj + Bn
2 (0, ε) = Kj + Bn2 (0, ε)
ord(Kj, Cj) ≤ ε
Lemma 19 (Blaschke selection principle) Let Ki, i ∈ N, be a sequence of nonempty,compact, convex sets in a bounded subset of Rn. Then there is a subsequence Kij ,j ∈ N, that converges with respect to the Hausdorff distance to a nonempty, compact,convex set.
Moreover, if all sets Ki, i ∈ N, are contained in a compact set C, then the limitis contained in C.
Proof. We show first that for any ε > 0 and any sequence Ki, i ∈ N, in a boundedset there is a subsequence Kij , j ∈ N, such that we have for all j, j′ ∈ N
dH(Kij , Kij′ ) < ε
By assumption the sequence Ki, i ∈ N, is contained in a set [a, b]n. We may assumethat a = 0 and b = 1. We choose k ∈ N such that
√n2−k < ε and we put
Ql =n∏
i=1
[li − 1
2k,
li2k
]
where 1 ≤ li ≤ 2k, l = (l1, . . . , ln) and 1 ≤ i ≤ n. Then we have
[0, 1]n =⋃
l
Ql
We considerIi = l|Ql ∩ Ki = ∅
22 CONTENTS
There are 2kn different vectors . Therefore there are at most 22kndifferent sets Ii.
Thus there are only finitely many different sets Ii and therefore there is a subsequenceij, j ∈ N, such that we have for all j, j′ ∈ N
Iij = Iij′
Let x ∈ Kij . Then there is Ql with x ∈ Ql. Since Iij = Iij′ there is y ∈ Ql ∩ Kij′ .
Therefore ‖x − y‖2 ≤√
n2k < ε, or x ∈ Bn
2 (y, ε). Thus we get
Kij ⊆ Kij′ + Bn2 (0, ε)
We concludedH(Kij , Kij′ ) ≤ ε
Now we select from the sequence Ki, i ∈ N, a subsequence that is a Cauchy sequence.From the sequence Ki, i ∈ N, we choose a subsequence K1,j such that we havedH(K1,j, K1,j′) < 1
2for all j ∈ N. After having chosen the subsequence K(k,j), j ∈ N,
we choose the sequence Kk+1,j, j ∈ N, as a subsequence of Kk,j, j ∈ N, such thatwe have for all j ∈ N
dH(Kk+1,j, Kk+1,j′) <1
2k+1
This implies that the sequence Kj,j, j ∈ N, is a Cauchy sequence.
0.5. POLYTOPES AND POLYHEDRA 23
0.5 Polytopes and polyhedra
A polytope P is the convex hull of of finitely many points in Rn
P = [x1, . . . , xk]
We have thatExt(P ) ⊆ x1, . . . , xk
We verify this. P is the convex hull of the points x1, . . . , xk. By Lemma ?? theconvex hull consists of all convex combinations
∑ki=1 tixi of the points x1, . . . , xk.
All points that are convex combinations∑k
i=1 tixi where at least 2 coefficients ti arestrictly larger than 0 are not extreme points. That leaves all those points whereexactly one coefficient is strictly larger than 0 and thus 1.
The extreme points of a polytope are also called vertices.A face of a convex body is the intersection of a support hyperplane with the
boundary of K. The dimension of a face is the smallest dimension of all affinespaces containing that face.
A simplex in Rn is the convex hull of n+1 points. A regular simplex is a simplexfor which all the distances between two vertices are equal.
A polyhedral set is the intersection of finitely many closed halfspaces in Rn.
Lemma 20 Each face of a polytope P is again a polytope and its vertices are exactlythe vertices of P that are contained in the hyperplane defining the face.
Proof. Let P = [x1, . . . , xk] and let H be a support hyperplane of P . Let usnote that P ∩ H is a convex set. Then Ext(P ) ⊆ x1, . . . , xk. By choosing a newnumbering we my assume that x1, . . . , xm, m ≤ k, are the extreme points of P . Bythe Theorem of Krein-Milman we have P = [x1, . . . , xm]. By Lemma ?? the extremepoints of P ∩ H are the extreme points of P contained in H. By the Theorem ofKrein-Milman P ∩ H is the convex hull of the extreme points of P contained in H.
Corollary 3 Let P be a convex polytope in Rn. Then P has only finitely manyfaces.
Lemma 21 Let P be a polyhedral set in Rn that has an interior point
P =N⋂
i=1
x| < ξi, x >≤ ti
Suppose that none of the halfspaces can be dropped, i.e. the intersection of thehalfspaces would be different from P . Then the sets
Fi = P ∩ x| < ξi, x >= ti
24 CONTENTS
have the dimension n − 1 and
∂P =N⋃
i=1
Fi
Proof. We show first that
P=
N⋂i=1
x| < ξi, x >< ti
Clearly, it is an open set contained in P and thus
P⊇
N⋂i=1
x| < ξi, x >< ti
On the other hand, let x be such that there is i with < x, ξi >= ti. Then in everyneighborhood of x there is a point not in P . The point x + εξi is not an element ofP :
< x + εξi, ξi >= ti + ε
Thus we have shown that
P=
N⋂i=1
x| < ξi, x >< ti
and, consequently,
∂P = P\P= P ∩ (
P )c = P ∩
N⋃i=1
x| < ξi, x >= ti =N⋃
i=1
Fi
We show now that the faces Fi have dimension equal to n − 1. Since none of thehalfspaces can be omitted there are xi /∈ P , 1 ≤ i ≤ N , such that
< xi, ξj >≤ tj 1 ≤ j = i ≤ N
< xi, ξi >> ti 1 ≤ i ≤ N
Let x0 be an interior point of P . Then there is ε with Bn2 (x0, ε) ⊆ P . By continuity
there is a point yi ∈ ∂P with < ξi, yi >= ti. Indeed, we consider all points sx0 +(1 − s)xi with 0 ≤ s ≤ 1. By convexity we have
< sx0 + (1 − s)xi, ξj >≤ tj
so that yi ∈ P . By convexity
CH(xi, Bn2 (x0, ε)) ⊆
⋂j =i
x| < ξj, x >< tj
0.5. POLYTOPES AND POLYHEDRA 25
It follows that there is δ > 0
Bn2 (yi, δ) ⊆
⋂j =i
x| < ξj, x >< tj
ThusBn
2 (yi, δ) ∩ x| < ξi, x >= ti ⊆ P
Therefore the boundary of P contains a n − 1-dimensional Euclidean ball and thecorresponding face is n − 1-dimensional.
Lemma 22 A set in Rn is a polytope if and only if it is a bounded polyhedral set.
Proof. We may assume that the set has an interior point. Otherwise, we restrictourselves to the smallest affine subspace containing the set.
Let P be a poytope. According to the lemma following the Theorem of Krein-Milman each face of P is the closed, convex hull of extreme points of P . There areonly finitely many extreme points and thus there are only finitely many faces. LetFi, i = 1, . . . , N , be all the faces of P and let x| < x, ξi >= ti, i = 1, . . . , N , bethe hyperplanes giving the faces
Fi = P ∩ x| < x, ξi >= ti
The normals ξi are oriented in such a way that
P ⊆ x| < x, ξi >≤ ti
Clearly, we have
P ⊆N⋂
i=1
x| < x, ξi >≤ ti
We show that the sets are equal. Suppose that they are not equal. Then there isx0 /∈ P with
x0 ∈N⋂
i=1
x| < x, ξi >≤ ti
Moreover, there is an interior point x1 of P . Therefore there is y0 ∈ [x0, x1] withy0 ∈ ∂P , i.e.
y0 = λx0 + (1 − λ)x1 0 ≤ λ ≤ 1
Since x0 /∈ P we have y0 = x0 and since x1 is an interior point of P we have y0 = x1.Thus
y0 = λx0 + (1 − λ)x1 0 < λ < 1
By the Theorem of Hahn-Banach there is a supporting hyperplane of P at y0. There-fore y0 is an element of a face of P . Therefore, there is i0 such that
ti0 =< y0, ξi0 >= λ < x0, ξi0 > +(1 − λ) < x1, ξi0 >
26 CONTENTS
Since x1 is an interior point of P we have < x1, ξi0 < ti0 . Moreover, 0 < λ < 1.Thus
ti0 =< y0, ξi0 >< ti0
We show now that a bounded polyhedral set is a polytope. For this we show thatan extreme point e of
⋂Ni=1x| < x, ξi >≤ ti satisfies
e =⋂
<e,ξi>=ti
x| < x, ξi >= ti
Since e is an extreme point it is an element of the boundary. There are sets I andJ such that
I = i| < e, ξi >= ti J = i| < e, ξi >< tiSuppose that
e =⋂
<e,ξi>=ti
x| < x, ξi >= ti
Then there are at least two different elements e and f in the set⋂
<e,ξi>=tix| <
x, ξi >= ti. Both are also elements of
N⋂i=1
x| < x, ξi >= ti
For all λ ∈ R we have
(1 − λ)e + λf ∈⋂
<e,ξi>=ti
x| < x, ξi >= ti
Indeed, for all i ∈ I
< (1 − λ)e + λf, ξi >= (1 − λ) < e, ξ > +λ < f, ξi >= ti
Moreover, there is ε > 0 such that for all i ∈ J
< (1 − ε)e + εf, ξi >= (1 − ε) < e, ξi > +ε < f, ξi >< ti
< (1 + ε)e − εf, ξi >= (1 + ε) < e, ξi > −ε < f, ξi >< ti
Thus we have
e =(1 − ε)e + εf
2+
(1 + ε)e − εf
2
and e is not an extreme point.Therefore, every extreme point is the intersection of some of the hyperplanes
defining the halfspaces. Since there are only N there are at most 2N extreme points.By the Theorem of Minkowski/Krein-Milman a compact convex set is the closedconvex hull of its extreme points. The convex hull of finitely many points is apolytope (which is a closed set).