Division of the Humanities

and Social Sciences

Ec 181 KC Border
Convex Analysis and Economic Theory AY 2019–2020

Topic 1: Convex sets and functions

1.1 Convex sets

Recall the affine combination function κ : X × X × R → X is defined by

κ(x, y, α) = (1 − α)x + αy.

1.1.1 Definition A subset C of a real vector space X is a convex set if it

includes the line segment joining any two of its points. That is, C is convex if for
every real α with 0 ⩽ α ⩽ 1 and every x, y ∈ C,

(1 − α)x + αy ∈ C.

If α = 0 or then (1 − α)x + αy = x and if α = 1, then (1 − α)x + αy = y, so

C ⊂ κ C × C × [0, 1] . Thus
   
C is convex ⇐⇒ κ C × C × [0, 1] = C ⇐⇒ κ C × C × (0, 1) ⊂ C.

Note that the empty set is convex.

1.1.2 Definition A convex combination is a linear combination αx+βy where

α, β ⩾ 0 and α + β = 1. More generally, a convex combination is a (finite) linear
P
combination α1 x1 + · · · + αk xk where each αi ⩾ 0 and ki=1 αi = 1.

1.1.3 Lemma If C is convex, then it is closed under general convex combinations.

P
That is, for any k ⩾ 2, if xi ∈ C and αi ⩾ 0, i = 1, . . . , k, and ki=1 αi = 1, then

α1 x1 + · · · + αk xk ∈ C.

1.1.4 Exercise Prove Lemma 1.1.3. □

Sample answer: We prove this by induction on k. Let P[k] denote the following
proposition:

If C is convex, and each xi ∈ C and αi ⩾ 0, for i = 1, . . . , k, and

Pk
i=1 αi = 1, then α1 x1 + · · · + αk xk ∈ C.

By definition, P[2] is true. We now show that P[k] =⇒ P[k + 1]: Assume xi ∈ C
P
and αi ⩾ 0, i = 1, . . . , k + 1, and k+1
i=1 αi = 1. We wish to prove that C contains

x = α1 x1 + · · · + αk+1 xk+1 .

KC Border: for Ec 181, 2019–2020 src: ConvexSets v. 2019.12.23::02.49

Ec 181 AY 2019–2020
KC Border Convex sets and functions 1–2

If αk+1 = 1, then x = xk+1 ∈ C and we are done. Otherwise we may write

!
α1 αk
α1 x1 + · · · + αk+1 xk+1 = (1 − αk+1 ) x1 + · · · + xk +αk+1 xk+1 .
1 − αk+1 1 − αk+1
| {z }
= x0
The term x0 in parentheses belongs to C by P[k]. So x = (1 − αk+1 )x0 + αk+1 xk+1 ,
where both x0 and xk+1 belong to C. Thus x ∈ C. This proves P[k + 1] and the
lemma is proved.

1.1.5 Definition (The unit m-simplex) In Rm+1 , the unit m-simplex

 X
m+1 
∆m = (α1 , . . . , αm+1 ) ∈ R m+1
: αi = 1, αi ⩾ 0, i = 1, . . . , m + 1
i=1

It is the set of all convex combinations of the unit coordinate vectors.

Sometimes it is convenient to number the indices of a vector in Rm+1 using
0, . . . , m instead of 1, . . . , m + 1, in which case we may write
 X
m 
∆m = (α0 , . . . , αm ) ∈ R m+1
: αi = 1, αi ⩾ 0, i = 0, . . . , m
i=0

1.1.6 Exercise (Examples of convex sets) Prove the following.

• A subset A of R is an interval if x, y ∈ A and x < z < y imply z ∈ A. A

subset of R is convex if and only if its an interval.

• The unit simplex ∆m is a compact convex set.

• In a normed vector space, the unit ball {x : kxk ⩽ 1} is convex.

• Let A be an m × m positive semidefinite matrix, and define the bilinear

form Q(x, y) = x · Ay where x, y ∈ Rm . Then for any α ⩾ 0, the set
Bα = {x ∈ Rm : Q(x, x) ⩽ α} is convex.
Hint (Fenchel [1, p. 34]): Note that

Q(λx + µy, λx + µy) = λ2 Q(x, x) + 2λµQ(x, y) + µ2 Q(y, y) ⩾ 0,

where the inequality follows from positive semidefiniteness.

For the special case λ = 1 and µ = −1 we get

Q(x − y, x − y) = Q(x, x) − 2Q(x, y) + Q(y, y) ⩾ 0,

so
Q(x, x) + Q(y, y) ⩾ 2Q(x, y).

v. 2019.12.23::02.49 src: ConvexSets KC Border: for Ec 181, 2019–2020

Ec 181 AY 2019–2020
KC Border Convex sets and functions 1–3

Thus


Q (1 − λ)x + λy, (1 − λ)x + λy
= (1 − λ)2 Q(x, x) + 2(1 − λ)λQ(x, y) + λ2 Q(y, y)
 
⩽ (1 − λ)2 Q(x, x) + (1 − λ)λ Q(x, x) + Q(y, y) + λ2 Q(y, y)
= (1 − λ)Q(x, x) + λQ(y, y).
(This shows that x 7→ Q(x, x) is a convex function, which we shall define shortly.)
This shows that Bα is a convex set.

• Let A be an m × n matrix and let y ∈ Rm . Then a set of the form

{x ∈ Rn : Ax ≦ y}
is called a solution set. Every solution set is convex.
• An m × m matrix is a stochastic matrix if all its entries are nonnegative
and each row sums to one. The set of stochastic matrices is a convex set.
□

1.1.7 Exercise (Elementary properties of convex sets) Prove the follow-

ing.
1. The intersection of a family of convex sets is convex.
2. The sum of convex sets is convex.
3. Scalar multiples of convex sets are convex.
4. More generally, if T : X → Y is a linear transformation between vector
spaces and A is convex subset of X, then the image T (A) is a convex subset
of Y .
5. As a special case of the previous result, if C is a convex subset of X × Y ,
the projections
n o n o
x ∈ X : ( ∃y ∈ Y ) [ (x, y) ∈ C ] and y ∈ Y : ( ∃x ∈ X ) [ (x, y) ∈ C ]
are convex subsets of X and Y respectively.
6. A set C is convex if and only if
( ∀α, β > 0 ) [ αC + βC = (α + β)C ].

□
While the sum of two convex sets is necessarily convex, the sum of two non-
convex sets may also be convex. For example, let A be the set of rationals in R
and let B be the union of 0 and the irrationals. Neither set is convex, but their
sum is the set of all real numbers, which is of course convex.

KC Border: for Ec 181, 2019–2020 src: ConvexSets v. 2019.12.23::02.49

Ec 181 AY 2019–2020
KC Border Convex sets and functions 1–4

1.2 Convex functions

Geometrically, a function on a subset of a vector space is convex if the line segment
joining any two points on its graph lies above the graph.
 Given
 points x and  y in
the domain, a typical point on the segment joining x, f (x) and y, f (y) is of
 
the form (1 − α)x + αy, (1 − α)f (x) + αf (y) . Note that this requires that f be
defined at (1 − α)x + αy, so the geometry can be expressed as follows.

Figure 1.2.1. A (strictly) convex function.

1.2.1 Definition A function f : C → R on a convex subset C of a vector space

is convex if  
f (1 − α)x + αy ⩽ (1 − α)f (x) + αf (y) (C)
for all x, y in C with x 6= y and all 0 < α < 1.
Note that if x = y or α = 0 or α = 1, then the inequality above is automatically
true. This definition is one that you would typically find in a real analysis text,
such as Royden [8].

1.2.2 Exercise Prove by induction on n that a function f is convex if and only

if X 
n Xn
f λi xi ⩽ λi f (xi )
i=1 i=1
Pn
for every convex combination i=1 λ i xi . □

1.2.3 Exercise A function f : C → R on a subset C of a vector space is convex

if and only if its epigraph is a convex set. □

Sample answer: ( =⇒ ) Let f satisfy (C), let (x,

 β) and (y, γ) belong to epi f,
and let 0 < α < 1. We must show that the point (1 − α)x + αy, (1 − α)β + αγ
belongs to epi f .
Now by the definition of epigraph we have
β ⩾ f (x) and γ ⩾ f (y),
so multiplying by the positive numbers 1 − α and α we have
(1 − α)β ⩾ (1 − α)f (x) and αγ ⩾ αf (y).

v. 2019.12.23::02.49 src: ConvexSets KC Border: for Ec 181, 2019–2020

Ec 181 AY 2019–2020
KC Border Convex sets and functions 1–5

Thus  
(1 − α)β + αγ ⩾ (1 − α)f (x) + αf (y) ⩾ f (1 − α)x + αy ,
the first inequality is the sum of the previous two, and the second inequality
is just (C) (unless x = y, in which case it is trivial). But this asserts that

(1 − α)x + αy, (1 − α)β + αγ belongs to epi f .
   
( ⇐= ) Since x, f (x) and y, f (y) belong to epi f , the convexity of epi f
 
implies that (1 − λ)x + λy, (1 − λ)f (x) + λf (y) ∈ epi f for 0 ⩽ λ ⩽ 1. By
 
definition of epi f , we have (1 − λ)f (x) + λf (y) ⩾ f (1 − λ)x + λy . That is, f
satisfies (C).
This result suggests an alternative definition of a convex function, which is
the one preferred by convex analysts. (For instance, Rockafellar [7].) It applies
equally well to extended-real valued functions.

1.2.4 Definition A function f : X → R♯ on a vector space is convex if its

epigraph is a convex set.
One of the virtues of the convex analysts’ definition is that we can extend a
conventional real-valued convex function f defined on a subset C of a vector space
X to an extended real-valued function f˜ defined on all of X by setting

f (x) x∈C
f˜(x) = 
∞ x∈
/ C.

Note that epi f˜ = epi f , so f˜ is convex in the convex analyst’s sense if and only if
f is convex in the conventional sense.

1.2.5 Definition The effective domain of an extended real-valued convex func-

tion f is defined to be
dom f = {x ∈ X : f (x) < ∞},
which is the projection of the epigraph onto X.
When f˜ is defined by extending a convex function f as above, then dom f˜ = C.
There is a caveat in this regard. What happens when an extended real-valued
function takes on both infinite values, say f (x) = ∞ and also f (y) = −∞. Then
(1 − α)f (x) + αf (y) is undefined, so condition (C) makes no sense. The epigraph
is still well defined, so Definition 1.2.1 can still be used to define convexity of
f . Some authors, e.g., Rockafellar [7], are happy to allow such functions, but
call them improper. Other authors, such as Hiriart-Urruty and Lemaréchal [5]
simply rule out by definition functions that assume the value −∞ as being convex.
There are arguments to be made on both sides, but I will allow consideration of
improper convex functions.
A further qualification is this: The function that is identically ∞ is also con-
sidered to be improper.

KC Border: for Ec 181, 2019–2020 src: ConvexSets v. 2019.12.23::02.49

 Ec 181 AY 2019–2020
KC Border Convex sets and functions 1–6

1.2.6 Definition A convex function is proper if it never assumes the value −∞

and its effective domain is nonempty.
For a proper convex function, the usual rules for operating with extended
real numbers (Section A.1) allow (C) to be used as an equivalent definition of
convexity.

1.2.7 Proposition Let A be a convex subset of X ×R and let C be the projection

of A on X. Then C is convex and the function f defined on C by

f (x) = inf{α : (x, α) ∈ A}

is a (possibly extended real-valued) convex function on C.

Draw a picture
.

1.2.8 Exercise Prove Proposition 1.2.7 □

1.3 Related concepts

1.3.1 Definition A function f : C → R on a convex subset C of a vector space
is:
 
convex if f (1 − λ)x + λy ⩽ (1 − λ)f (x) + λf (y)
 
strictly convex if f (1 − λ)x + λy < (1 − λ)f (x) + λf (y)
 
concave if f (1 − λ)x + λy ⩾ (1 − λ)f (x) + λf (y)
 
strictly concave if f (1 − λ)x + λy > (1 − λ)f (x) + λf (y)

for all x, y in C with x 6= y and all 0 < λ < 1.

If f is improper and assumes both values ∞ and −∞, then this definition of
strict convexity is unusable, so we would have to define a strictly convex function
as one with no line segment included in its graph. Just as convex analysts define
convex functions in terms of their epigraphs, they define concave functions in
terms of their hypographs.

1.3.2 Definition An extended real-valued function is concave if its hypograph

is a convex subset of X × R. Given a concave function f : X → R♯ , its effective
domain, dom f , is the projection of its hypograph on X, that is,

dom f = {x ∈ X : f (x) > −∞}.

The effective domain of a concave function is a (possibly empty) convex set.

v. 2019.12.23::02.49 src: ConvexSets KC Border: for Ec 181, 2019–2020

Ec 181 AY 2019–2020
KC Border Convex sets and functions 1–7

1.3.3 Exercise Prove the following.

1. The sum of convex functions is convex. The sum of concave functions is

concave.

2. A nonnegative multiple of a convex function is convex. A nonnegative mul-

tiple of a concave function is concave.

3. Consequently, the collection of convex functions is a convex cone (see Def-

inition 3.1.1 below) in the vector space of functions. Also, the collection of
concave functions is a convex cone in the vector space of functions.

4. The pointwise limit of a sequence of convex functions is convex. The point-

wise limit of a sequence of concave functions is concave.

5. The set of points where the pointwise supremum of a family of convex func-
tions is finite is a convex set. On this set the supremum is a convex function.
Similarly, the pointwise infimum of a family of concave functions is concave.

6. Every affine function is both convex and concave.

□

1.4 Complements
1.4.1 Exercise (Alternate criteria for convexity)

1. A function f on a convex set C is midpoint-convex if for all x, y ∈ C, it

satisfies  
f 12 x + 12 y ⩽ 12 f (x) + 12 f (y).
Prove that a continuous midpoint-convex function is convex. (This is true
in any topological vector space.)

2. A stronger result was pointed out to me by Tomasz Strzalecki. He attributes

it to Hardy, Littlewood, and Polya. But they [4, § 3.7, p. 73] in turn attribute
it to both Reisz [6] 1 and to Jessen. 2 It is a clever way of using continuity to
replace the “for every λ” clause in the definition of convexity with a “there
exists some λ” clause.
1
The point is made in the last paragraph of section 8, on p. 471.
2
Hardy, et al., give the Jessen reference simply as: Om Uligheder imellem Potensmid-
delværdier, Mat. Tidsskr. B, 1931 (no volume or page numbers). The Jessen archives at
http://www.math.ku.dk/arkivet/jessen/bjpap3.htm list a different date: Om Uligheder
imellem Potensmiddelværdier, Matematisk Tidsskrift, B, 1933, pp. 1–19. The archives also list
Bemærkninger om konvekse Funktioner og Uligheder imellem Middelværdier, I and II, Matem-
atisk Tidsskrift, B, 1931, pp. 17–28 and 84–95. When I get the time to learn Danish and find a
library that has the journal, I may try to track down the proper citation. Or perhaps I should
just ask John Chipman.

KC Border: for Ec 181, 2019–2020 src: ConvexSets v. 2019.12.23::02.49

 Ec 181 AY 2019–2020
KC Border Convex sets and functions 1–8

Suppose that f is defined on a convex set C of a tvs and has the following
existential property:
h   i
( ∀ x, y ∈ C ) ( ∃ 0 < λ < 1 ) f (1 − λ)x + λy ⩽ (1 − λ)f (x) + λf (y) .
(E)
Prove that if f is also continuous, then f is convex.

 3. Show that the results above need not hold for discontinuous f . □
For even more general notions along these lines, see the paper by Green and
Gustin [2].
Sample answer: Clearly part (2) implies part (1), so I shall prove part (2), using
a proof by contradiction. So assume by way of contradiction that f is continuous
and satisfies
 (E), but is not convex. Then there exist x, y, and 0 < λ̄ < 1 such
that f (1 − λ̄)x + λ̄y > (1 − λ̄)f (x) + λ̄f (y). By continuity, the set A = {λ :
 
f (1 − λ)x + λy > (1 − λ)f (x) + λf (y)} is open, and by hypothesis contains λ̄.
But 0, 1 ∈ / A, so λ̄ is contained in a maximal open interval included in A. That
is, there exist some
 0 ⩽ α < λ̄ < β ⩽ 1 such
 that for all λ in (α, β), we have
f (1−λ)x+λy > (1−λ)f (x)+λf (y), but f (1−α)x+αy = (1−α)f (x)+αf (y)
 
and f (1 − β)x + βy = (1 − β)f (x) + βf (y). (This argument works in any tvs).
Now consider x0 = (1 − α)x + αy and y 0 = (1 − β)x + βy. By construction, for
every
 0 < λ < 1, the point (1 − λ)x0 + λy 0 strictly between x0 and y 0 satisfies
f (1 − λ)x0 + λy 0 > (1 − λ)f (x0 ) + λf (y 0 ). But this violates (E) applied to the
points x0 and y 0 , a contradiction. Therefore f must be convex.
 (3) This has got to be a pretty weird function, and its description takes us pretty
far afield, but here goes.
Recall (or just take my word for it) that the real numbers R can be thought of
as a vector space X over the field Q of rational numbers. (After all, a rational linear
combination of real numbers is another real number.) As such, it has a Hamel basis,
that is, a set B of vectors (real numbers) such that every vector (real number) is a
unique (finite) rational linear combination of elements of B. (This makes X an infinite
dimensional vector space over Q.) Moreover we may find such a basis that contains
the vector 1 ∈ X. Uncountably many of these basis elements are irrational. Fix some
irrational α in B and consider the set of rational linear combinations of basis vectors of
the form n1 + mα where n, m are integers. Such sets are discussed in section 6 of my
notes on the Kolmogorov Extension Theorem, or see Halmos [3, Theorem 16.C, p. 69],
and it follows from Proposition 11 in my notes that in every interval A, for every M ,
there is a point n + mα ∈ A with m > M .
Now define f : R → Q so that f (x) is the coefficient on the vector α in the unique
representation of x as a linear combination of basis vectors. It follows from the remark
above that f is unbounded on every interval, and consequently not continuous. Moreover
f is linear with respect to linear combinations with coefficients in Q, and since 1/2 is
rational, f is mid-point convex, but it is not convex! It is also not Lebesgue measurable,
but that’s a different story.

v. 2019.12.23::02.49 src: ConvexSets KC Border: for Ec 181, 2019–2020

Ec 181 AY 2019–2020
KC Border Convex sets and functions 1–9

References
[1] W. Fenchel. 1953. Convex cones, sets, and functions. Lecture notes, Princeton
University, Department of Mathematics. From notes taken by D. W. Blackett,
Spring 1951.

[2] J. W. Green and W. Gustin. 1950. Quasiconvex sets. Canadian Journal of

Mathematics 2:489–507.
http://cms.math.ca/cjm/v2/cjm1950v02.0489-0507.pdf

[3] P. R. Halmos. 1974. Measure theory. Graduate Texts in Mathematics. New

York: Springer–Verlag. Reprint of the edition published by Van Nostrand,
1950.

[4] G. H. Hardy, J. E. Littlewood, and G. Pólya. 1952. Inequalities, 2d. ed.

Cambridge: Cambridge University Press.

[5] J.-B. Hiriart-Urruty and C. Lemaréchal. 2001. Fundamentals of convex anal-

ysis. Grundlehren Text Editions. Berlin: Springer–Verlag.

[6] M. Riesz. 1927. Sure les maxima des formes bilinéaires et sur les fonctionelles
linéaires. Acta Mathematica 49(3–4):465–497. DOI: 10.1007/BF02564121

[7] R. T. Rockafellar. 1970. Convex analysis. Number 28 in Princeton Mathe-

matical Series. Princeton: Princeton University Press.

[8] H. L. Royden. 1988. Real analysis, 3d. ed. New York: Macmillan.

KC Border: for Ec 181, 2019–2020 src: ConvexSets v. 2019.12.23::02.49