Lecture 5

[5.0.21] Affine independence

1. Fix S = {p0 , p1 , . . . , pk } ⊆ Rn . If {p1 − p0 , . . . , pk − p0 } is linearly independent, then we say that S is

affine independent, otherwise it is called affine dependent.
2. If S = {p0 , p1 , . . . , pk } is affine independent, then we call conv(S) a simplex of dimension k.

[5.0.22] Example

1. The sets S1 = {0, e1 , e2 } ⊆ R2 and S2 = {e1 , e2 , e3 } ⊆ R3 are affine independent.

2. Whereas related to our favorite set, the set {0, e1 , e2 , e1 + e2 } is affine dependent.
3. A point is regarded as a zero dimensional simplex.
4. A closed line segment is a 1-dimensional simplex.
5. A closed triangular plate is a 2-dimensional simplex.
6. A closed solid tetrahedron is a 3-dimensional simplex.

[5.0.23] Exercise(+) (Similar to linear dependence) The set S = {p0 , p1 , . . . , pk } is affine dependent
iff some pi is an affine combination of the remaining.

[5.0.24] Exercise(+) (Similar to linear independence) Let S = {p0 , p1 , . . . , pk } ⊆ Rn . Then S is

affine independent iff each x ∈ aff(S) can be written as an affine combination of pi ’s in a unique way.

The following is an immediate corollary. It will be used in the sequel.

[5.0.25] Corollary Let S = {p0 , p1 , . . . , pk } be affine independent. Then each x ∈ conv(S) can be written
as a convex combination of pi ’s in a unique way.!!

20
5.1 Some exercises

[5.1.1] Exercise(E) Related to our favorite set in R2 , the set {0, e1 , e2 , e1 + e2 } is affine dependent. So one
element should be an affine combination of the rest. Verify this.

[5.1.2] Exercise(E) (Not every element will be an affine combination of the rest) Take S =
{0, e1 , e2 , 2e1 }. It is affine dependent. But we cannot express e2 as an affine combination of the rest, as
e2 does not lie on the affine span of the other three. Does this contradict our previous results?

[5.1.3] Exercise(E) Write our favorite set as a union of two simplices of dimension 2, in two different ways.

[5.1.4] Exercise(E) Recall that our favorite set is aff(S), where S = {0, e1 , e2 , e1 + e2 } is affine dependent.
So, there should be a point x ∈ aff(S) which can be written as an affine combination of points of S in two
different ways. Find such a point.

[5.1.5] Exercise(E) (Pictures of unit balls) Draw B1 (0) in R2 with respect to the norms k · k1 , k · k2
and k · k∞ . Can you imagine the pictures of the unit ball in R3 with respect to the above norms?

6 Vertices, hyperplanes

[6.0.1] Vertex

1. A point x of a convex set S is called a vertex if it is not on any open line segment in S. That is,
∃
6 a, b ∈ S, a 6= b and λ ∈ (0, 1) such that x = λa + (1 − λ)b.
2. Equivalently, it means x ∈
/ conv(S \ {x}).
3. It also means, you cannot write x as a convex combination of two points different from x.

[6.0.2] Example If we take a 6= b in Rn , then by definition, (.1)a + (.9)b is not a vertex of [a, b].

[6.0.3] Fact Let ∅ 6= S ⊆ Rn . Then the vertices of conv(S) must be elements of S.

Proof. Let p ∈ conv(S) and p ∈

/ S. So S ⊆ conv(S) \ {p}. So p ∈ conv(S) ⊆ conv(conv(S) \ {p}). So p is not
a vertex of conv(S).

[6.0.4] Example So, our favorite set can have at most 4 vertices, namely, 0, e1 , e2 , e1 + e2 .

[6.0.5] Theorem Let S = {p0 , p1 , . . . , pk } ⊆ Rn be affine independent. Then pi are precisely the vertices of
conv(S).

21
 Proof. We know that vertices
P of conv(S)
Pmust be elements of S. Now, suppose p0 is not a vertex. So there
exist distinct points x = λi pi and y = µi pi in conv(S) such that for some λ ∈ (0, 1), we have

p0 = λx + (1 − λ)y = (λλ0 + (1 − λ)µ0 )p0 + (λλ1 + (1 − λ)µ1 )p1 + · · · + (λλk + (1 − λ)µk )pk .

As p0 = (1)p0 + (0)p1 + · · · + (0)pk , and S is affine independent, we have 1 = λλ0 + (1 − λ)µ0 and 0 =
λλ1 + (1 − λ)µ1 = · · · = λλk + (1 − λ)µk . Observe that, if a convex combination of two nonnegative numbers
is 0, then they both must be 0. Hence λ1 = µ1 = · · · = λk = µk = 0. So λ0 = µ0 = 1 and x = y = p0 , a
contradiction to the fact that x and y are distinct.

[6.0.6] Example It follows that a closed triangular plate has 3 vertices. As our favorite set as S =
{0, e1 , e2 , e1 + e2 } is affine dependent, the previous results only tell that the vertices of conv(S) are in S.

[6.0.7] Careful A nonempty convex cone may not have a vertex, for example take Rn itself. But if it has
one, then it is the point 0. This is because if x 6= 0, then it is the midpoint of 12 x and 32 x.

[6.0.8] Hyperplanes

1. Take 0 6= c ∈ Rn and α ∈ R. Then the set H = {x ∈ Rn | ct x = α} is called a hyperplane.

2. We often write H : ct x = α, to mean the hyperplane H = {x ∈ Rn | ct x = α}. In such cases we
understand that c 6= 0.
3. When α = 0, we call H a linear hyperplane.

[6.0.9] Example In R3 , a hyperplane is a plane. In R2 , a hyperplane is a line. In R1 , a point is a hyperplane.

[6.0.10] Fact (Any linear hyperplane in Rn can be viewed as Rn−1.) Let n > 1. Let H : ct x = 0
be a linear hyperplane. Then H = c⊥ . So it is an n − 1 dimensional subspace. So it is isomorphic to
S = {y | y(n) = 0} as S also has dimension n − 1. There is a natural isomorphism between S and Rn−1 .
Hence any linear hyperplane in Rn can be viewed as Rn−1 .

[6.0.11] Fact Any hyperplane in Rn is a translated linear hyperplane. So it is an affine subspace. But in
general, affine subspaces can have dimensions other than n − 1.

|α|
[6.0.12] Fact (The distance from 0 to the hyperplane H : ct x = α is kck .) For any x ∈ H, we have
αc αc 2 αc 2 αc 2
kxk2 = kx − 2
+ 2
k = kx − 2
k +k k ,
kck kck kck kck2
αc αc 2 αc |α|
as hx − kck2
, ci = 0. So kxk2 ≥ k kck2 k . Equality is attained at x = kck2
. So the distance is kxk = kck .

22
[6.0.13] Linearly independent hyperplanes Hyperplanes cti x = αi , i = 1, . . . , k are called linearly
independent if {c1 , . . . , ck } is linearly independent.

[6.0.14] Example The hyperplanes x − y = 1 and 2x + y = 8 are linearly independent in R2 , whereas the
hyperplanes x − y = 1 and x − y = 2 are not linearly independent.

[6.0.15] Halfspaces Let H : ct x = α be a hyperplane. Then the set H+ = {x | ct x ≥ α} is called the

positive closed halfspace of H. The set H+ ◦ = {x | ct x > α} is called the positive open halfspace of H.

Negative closed and open halfspaces are defined similarly.

[6.0.16] Example

1. Take c = [1 −1]t and α = 0. Then the linear hyperplane H : ct x = 0 in R2 is the line x1 − x2 = 0. The
positive and negative closed halfspaces are shown below.

H H+ H−

2. But if we take c = [−1 1]t , then H+ and H− change.

[6.0.17] Fact Our favorite set is the intersection of 4 closed halfspaces namely, x1 ≤ 1, x1 ≥ 0, x2 ≤ 1 and
x2 ≥ 0!!

[6.0.18] Fact Hyperplanes, halfspaces are convex. Linear hyperplanes, linear closed halfspaces are convex
cones!!

6.1 Some exercises

[6.1.1] Exercise(M+) (Largest ball inside) Let S = {±ei | i = 1, · · · , 9} ⊆ R9 . Find the largest
value of δ such that Bδ (0) ⊆ conv(S).

[6.1.2] Exercise(M) (Minimal generating set) Let S ⊆ Rn be nonempty and finite. Throw out, one by
one, points from S which are convex combinations of the others and let T = {x1 , . . . , xk } be the set which
remains at the end. a) Is conv(T ) = conv(S)? b) Notice that x1 = 1x1 + 0x2 + · · · + 0xk . Argue that there
cannot be another way to express x1 as a convex combination of points in T . c) Hence prove that x1 is a
vertex of conv(S).

23
 Lecture 6

7 Sums of sets

[7.0.1] Definition Let C and D be two subsets of Rn and α ∈ R. Then we define αC:= {αx | x ∈ C} and
we define C + D:= {x + y | x ∈ C, y ∈ D}.

[7.0.2] Example (Warning! C +C = 2C may not be correct) Take the set C = {e1 , e2 }. Then C +C =
{2e1 , 2e2 , e1 + e2 } 6= 2C.

[7.0.3] Fact It is easy to show that, if C is convex, then C + C = 2C.

[7.0.4] Example We have

+ =

{(0, y) | y ∈ [0, 1]} {(x, 0) | x ∈ [0, 1]} {(x, y) | x, y ∈ [0, 1]}

[7.0.5] Fact Let C and D be convex, and λ ∈ R. Then λC, C (closure), C + D, and C × D are convex.!!

7.1 Some exercises

[7.1.1] Exercise(E) In R2 , let S = conv(e1 + e2 , e1 − e2 , −e1 + e2 , −e1 − e2 ) and T be the closed disc of
radius 21 centered at 0. Can you draw S + T ?

[7.1.2] Exercise(E+) (From our past knowledge) Let x, y ∈ Rn be linearly independent. Show that there
is a nonzero w ∈ span(x, y) such that the sum of the entries of w is zero. Do this in three ways: a) by a direct
formula for w using x, y, b) by using dimensions, c) by using intermediate value theorem.

8 Caratheodory theorem and its applications

Why do we need Caratheodory theorem? Caratheodory theorem proves that a polytope is compact. That
will be used to prove that a polytope is a bounded polyhedron. That result is used to prove Farka’s lemma,
which is very useful. For example, it is used to prove the existence of the generalized Lagrange multipliers in
nonlinear optimization.

24