Chapter #5: Convexity
Math Econ (MEF)
Raúl Bajo (
[email protected])
Affine subspaces
1 p n
Pp
• An affine
Ppcombination of the vectors x , · · · , x ∈ R is a sum k=1 λp ·xp
where k=1 λk = 1, λk ∈ R for all k.
• An affine subspace of Rn is a subset X that is stable by affine combinations:
1 Pp Pp
x , · · · , xp ∈ X, k=1 λk = 1 ⇒ k=1 λk xk ∈ X
• The vectors x1 , · · · , xp are called affinely indepedent if and only if for all
λk ∈ R for all k:
Pp k
Pp
k=1 λk x = 0 and k=1 λk = 0 ⇒ {λ1 = 0, · · · , λp = 0}
• Some properties
. {x1 , · · · , xp affinely independent} ⇒ {x2 −x1 , x3 −x1 , · · · , xp −x1 linearly independent}
. In R we can have at most n + 1 affinely independent vectors.
. X is an affine subspace ⇔ ∀a ∈ X we have X − {a} is a linear
subspace.
• The affine dimension of X is defined as the (linear) dimension of X. This,
in an affine subspace of dimension l one can find at most l + 1 affinely
indepedent vectors.
Examples:
. A point X = {a} is an affine subspace of dimension 0.
. A line borne by b and going through a, X = {x | ∃λ ∈ R x = a + λb}
is an affine subspace of dimension 2.
. In general, the affine hyperplane X = {x | x · a = λ} where a ∈ Rn
and λ ∈ R is of affine dimension n − 1.
Convex sets
1 p n
Pp
• A convex
Ppcombination of the vectors x , · · · , x ∈ R is a sum k=1 λk xk
where k=1 λk = 1 and λk ≥ 0 for k = 1, · · · , p.
• A subset C ⊆ Rn is convex if it is stable be convex combinations:
1 Pp Pp
x , · · · , xp ∈ C, λ1 , · · · , λp ∈ R+ k=1 λk = 1 ⇒ k
k=1 λk x ∈ C
. Equivalent definition: {x, y ∈ C, 0 ≤ λ ≤ 1} ⇒ {λx + (1 − λ)y ∈ C}
1
. E.g. the convex sets of R are the intervals of R; a closed ball or an
open ball for any norm in Rn .
• Stability properties of convex sets:
. A (finite or infinite) intersection of convex sets is convex.
. Corollary: a set defined by an arbitrary number of linear inequalities
is convex. That is
x ∈ C ⇔ A · x ≤ b where A is m × n and b ∈ Rm
. Convexity is preserved by any affine transformation (e.g. scalar mul-
tiplication, translation, etc.)
. Convexity is preserved by addition: C, D convex in Rn ⇒ C + D
convex.
. Convexity is preserved by cartesian product: {Ci convex in Rni , i =
1, 2} ⇒ C × D convex in Rn1 +n2
• The convex hull of a subset X ⊆ Rn is denoted CO(X) and defined in
two equivalent ways:
CO(X) is the intersection of all the convex sets containing X, or
Pn k 1 p n
CO(X) = k=1 λk x | x , · · · , x ∈ R is the set of all convex combi-
nations of points in X.
• The dimension of a convex set is the affine dimension of the affine subset
generated by X, namely the set of affine combinations of vectors in X.
1
. If x , · · · , xp are p affinely independent vectors in Rn , their convex
hull CO[x1 , · · · , xp ] is called a simplex of dimension p − 1.
• Topological properties of convex sets:
. If C ⊆ Rn is a convex set, x ∈ int C (interior point of C) and y ∈ C
(a limit point of C), then λx+(1−λ)y ∈ int C for all λ s.t. 0 < λ ≤ 1.
. If C is convex, its interior int C and closure Cl C are both convex.
. The convex hull of a compact set is compact.
. The convex hull of a closed set may not be closed.
• The closed convex hull of a subset X is the set, denoted CO(X), inter-
section of all closed convex sets containing X. It is also the closure of
CO(X): CO(X) = CO(X) ⊇ CO(X).
Convex, concave functions
• If C ⊆ Rn is convex, and f is a function from C into R, we say that
f is convex (or concave up) if it satisfies the two following equivalent
properties:
Pp Pp
∀x1 , · · · , xp ∈ C, ∀λ1 , · · · , λp ∈ R+ s.t. k=1 λk = 1 ⇒ f k=1 λk x
k
≤
P p k
k=1 λ k f (x ), or
∀x, y ∈ C ∀λ ∈ [0, 1] we have that f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y)
2
• The function f from C into Rn is strictly convex iff
∀x, y ∈ C ∀λ ∈ [0, 1] we have that f (λx + (1 − λ)y) < λf (x) + (1 − λ)f (y)
• The function f from C into Rn is concave (or concave down) if −f is
convex.
. E.g. f (x) = xα is strictly convex if α > 1 and strictly concave for
0 < α < 1.
• A function f is both concave and convex on C ⊆ Rn iff it is affine on C:
f (x) = a · x + b for some a ∈ Rn , b ∈ R.
• Operations on convex functions:
. If f : C → R is convex, and h : R → R is convex and non-decreasing,
then g ◦ f is convex on C.
. Convexity is respected by positive linear combinations:
Pp
f1 , · · · , fp convex and λ1 , · · · , λp ∈ R+ ⇒ k=1 λk fk is convex.
. The supremum of any set of convex functions is convex: fi convex
⇒ g(x) = supi∈I fi (x) is convex (similarly, the infimum of concave
functions is concave).
• Continuity of convex functions:
. A convex function f0 from C ⊆ Rn into R is continuous on int C.
. A convex function f from the bounded convex set C ⊆ Rn into R is
bounded from below: inf x∈C f (x) > −∞
• Quasi-convexity: the function f from C into R is quasi-convex iff it satisfies
one of the two following equivalent properties:
∀λ ∈ R {x ∈ C | f (x) ≤ λ} is convex, or
∀x, y ∈ C ∀λ ∈ [0, 1] we have f (λx + (1 − λ)y) ≤ sup{f (x), f (y)}
• The function f is called quasi-concave iff −f is quasi-convex.
. E.g. if C ⊆ R, then f is quasi-concave iff f is monotone on C, or
f is “single-peaked” that is, there exists c ∈ C such that f is non
decreasing to the left of c and non-increasing to its right.
• Operations on quasi-convex/quasi-concave functions:
. If f : C → R is quasi-convex and h : R → R is non-decreasing, then
g = h ◦ f is quasi-convex.
. The supremum of any set of quasi-convex functios is quasi-convex.
. The infimum of any set of quasi-concave functios is quasi-concave.
3
Separation theorems
• Strict separation theorem:
If C1 , C2 are both convex, if C1 is closed and C2 is compact (closed and
bounded), and if C1 ∩ C2 = ∅, there exists a ∈ Rn such that:
supx∈C1 a · x < inf y∈C2 a · y
This means that there exists an affine hyperplane H such that C1 and C2
are in the two opposite open halfspaces defined by H.
• Proper separation theorem: if C1 and C2 are both convex in Rn , and if
C1 ∩ C2 = ∅ the can be properly separated. That is, the linear form
x → a · x separates properly the two sets X, Y iff:
sup a · x ≤ inf a · y and inf a · x < sup a · y
X Y X Y
• There are many useful applications of the separation theorems. E.g. we
can define a closed convex set as the intersection of all halfspaces contain-
ing it. The well-known Farkas lemma (not required) is also a consequence
of the separation theorems.
