Concave and Quasiconcave Functions

Reference: Simon & Blum, chap. 21, Mas-Collel’s mathematical appendix

MC, MD.

1 Convex Sets
• A set U ⊆ Rn is convex if whenever two points x,y ∈ U then the line
connecting x, y completely lies in U .

• Some examples:

• Formally U ⊆ Rn is convex if whenever x, y ∈ U then for every 0 ≤ α ≤ 1

αx + (1 − α)y ∈ U

2 Concave Functions
• Do not confuse concavity/convexity of functions with that of sets!

• Why we care about concave functions

– concave functions represent diminishing marginal returns which makes

sense for utility or some production functions.

∗ u0 (x) ≈ u(x + 1) − u(x) represents marginal utility, if we expect

the marginal utility to decrease then we should have u0 be a
decreasing function, meaning u00 < 0, implying the concavity of
u.

1
 – concave utility functions represent risk aversity which a reasonable
assumption for human behavior

∗ A risk averse person prefers a definite outcome to a risky out-

come even if the risky outcome has the same expectation. For
example, consider you can choose one of these cases: case one,
1 1
with probability 2 you get $100 and with probability 2 you get
nothing, or case 2 , you can get $50 for sure. Usually people
choose the not risky case. Meaning if they have utility function
u then
1 1
u($50) ≥ u($100) + u($0)
2 2
Concave functions satisfy these kind of inequalities.

• Definition: A real valued function f defined on a convex subset U ⊆ Rn

is concave if for all x, y ∈ U and for all 0 ≤ α ≤ 1

f (αx + (1 − α)y) ≥ αf (x) + (1 − α)f (y)

similarly f is called convex if

f (αx + (1 − α)y) ≤ αf (x) + (1 − α)f (y)

• What the definitions means graphically

check the meaning of above definitions graphically for

– f (x) = x2 convex

– f (x) = ln x concave

2
 Some Properties of concave functions

• If f, g : U ⊆ Rn → R are concave(convex) then f + g is concave(convex).

Proof: based on definition, we need to show for every x, y ∈ U and 0 ≤ α ≤ 1

we have

(f + g)(αx + (1 − α)y) ≥ α(f + g)(x) + (1 − α)(f + g)(y)

that follows directly by summing up corresponding inequalities for f, g :

f (αx + (1 − α)y) ≥ αf (x) + (1 − α)f (y)

g(αx + (1 − α)y) ≥ αg(x) + (1 − α)g(y)

• If f is concave then −f is convex.

Proof: for all x, y ∈ U and 0 ≤ α ≤ 1 we have

f (αx + (1 − α)y) ≥ αf (x) + (1 − α)f (y)

by taking each term to the opposite side we get

α(−f )(x) + (1 − α)(−f )(y) ≥ (−f )(αx + (1 − α)y)

which means −f is convex.

• For specific functions , not general cases, when we have the formula of the
function in hand, it is easier to use other criteria for concavity/convexity
than the basic definition. We have the following theorem

Theorem 1 (Criteria for concavity ) Suppose f : U ⊆ R → R is C 2 ,i.e. f 00

is continuos, then f is concave if and only if f 00 (x) ≤ 0 for all x ∈ U . Similarly
f is convex if and only if f 00 (x) ≥ 0 for all x ∈ U .

• There are other characterizations of concavity, for example the following

theorem. usually we do not use the following theorem as a way to show
concavity. However we state and prove it here because it implies some
important properties of concave functions in optimization.

3
Theorem 2 Let f : U ⊆ R → R be a C 1 function, i.e. continuously differen-
tiable. Then f is concave if and only if

f (y) − f (x) ≤ f 0 (x)(y − x) for all x, y ∈ U (1)

the function f is convex if and only if

f (y) − f (x) ≥ f 0 (x)(y − x) for all x, y ∈ U

Proof. We prove the claim about concave functions only. The convex part
is similar. First suppose f is concave and we want to show (??) holds. Let
x, y ∈ U and 0 < α ≤ 1 then we have

αf (y) + (1 − α)f (x) ≤ f (αy + (1 − α)x)

or
αf (y) − αf (x) ≤ f (αy + (1 − α)x) − f (x)

then

f (αy + (1 − α)x) − f (x)

f (y) − f (x) ≤
α
f (αy + (1 − α)x) − f (x)
= (x − y)
α(x − y)

now letting α → 0 we get

f (αy + (1 − α)x) − f (x)

→ f 0 (x)
α(x − y)

and therefore condition (??) follows.

On the other hand, suppose (??) holds for all x, y ∈ U , then

f (x) − f ((1 − α)x + αy) ≤ f 0 ((1 − α)x + αy)(x − (1 − α)x − αy)

= −αf 0 ((1 − α)x + αy)(y − x)

similarly

f (y) − f ((1 − α)x + αy) ≤ (1 − α)f 0 ((1 − α)x + αy)(y − x)

4
multiply the first inequality by (1 − α) and the second inequality by α , we get

(1 − α)f (x) − (1 − α)f ((1 − α)x + αy) ≤ −α(1 − α)f 0 ((1 − α)x + αy)(y − x)

αf (y) − αf ((1 − α)x + αy) ≤ α(1 − α)f 0 ((1 − α)x + αy)(y − x)

then add them up to get

(1 − α)f (x) + αf (y) ≤ f ((1 − α)x + αy)

The natural generalization of this result to f : Rn → R is

Theorem 3 Let f : U ⊆ Rn → R be continuously differentiable, then f is

concave if and only if

f (y) − f (x) ≤ Df (x)(y − x) for all x, y ∈ U

equivalently
h i
f (y) − f (x) ≤ ∂f ∂f ∂f (y − x)
∂x1 (x) ∂x2 (x) ... ∂xn (x)
∂f ∂f
= (x)(y1 − x1 ) + ... + (x)(yn − xn )
∂x1 ∂xn

One important implication of this theorem is the following

Corollary 4 Let f : U ⊆ Rn → R be concave and continuously differentiable.

If x0 ∈ U is such that Df (x0 ) = 0, i.e.

∂f
(x0 ) = 0 for i = 1, 2, ..., n
∂xi

then x0 is a global maximum of f in the domain.

Proof. Using the previous theorem for all y ∈ U we have

f (y) − f (x0 ) ≤ Df (x0 )(y − x0 ) = 0

therefore for all y ∈ U we have f (y) ≤ f (x0 ) which means x0 is the global
maximum.

5
• The importance of the result is that in an optimization of a concave func-
tion if we only let the first order conditions to be equal to 0 then we are
sure that we are in a maximum and we do not need to check the second
order conditions.

• Suppose f : U ⊆ Rn → R is concave and continuously differentiable.

Consider a point x0 ∈ U , if Df (x0 )(y − x0 ) ≤ 0 for some y ∈ U then
since f (y) − f (x0 ) ≤ Df (x0 )(y − x0 ) we get

Df (x0 )(y − x0 ) ≤ 0 =⇒ f (y) ≤ f (x0 )

Notice that
Df (x0 )(y − x0 ) = ∇f (x0 ) · (y − x0 )

so Df (x0 )(y − x0 ) ≤ 0 means that the two vectors ∇f (x0 ) and (y − x0 )

make an obtuse angle. This implies the upper contour set

{z : f (z) ≥ f (x0 )}

lies above the tangent hyperplane. This result is called the separating
hyperplane result. Graphically:

6
 • Similar to concavity for f : R → R meaning f 00 ≤ 0, there is some criteria
about the Hessian matrix of f : Rn → R to guarantee the concavity. It
requires D2 f be a negative semi definite matrix. You will see what this
means in detail in Catherine’s lecture.

3 Quasiconcavity
• Although concavity has desirable implications for optimization, there
are weaker conditions on a functions that also imply nice properties with
regard to optimization.

• having convex upper contour sets is a nice property of a function with

regard to optimization. For example consider the problem of

max u(x1 , x2 ) subject to p1 x1 + p2 x2 ≤ w

x1 ,x2

graphically:

• Definition: A function f : Rn → R is quasiconcave if

f (αx + (1 − α)y) ≥ min{f (x), f (y)} for all x, y ∈ Rn and 0 ≤ α ≤ 1

A function f is quasiconvex if −f is quasiconcave.

The important property of quasiconcave functions is

Theorem 5 f : Rn → R is quasiconcave if and only if the upper contour sets

of f are convex, i.e. the sets like

{x ∈ Rn : f (x) ≥ a} (2)

7
Proof. First suppose f is quasiconcave. To show A = {x ∈ Rn : f (x) ≥ a}
is convex, let’s x, y ∈ A, then

f (x) ≥ a and f (y) ≥ a

for 0 ≤ α ≤ 1 we have

f (αx + (1 − α)y) ≥ min{f (x), f (y)} ≥ a

which implies A is convex. On the other hand, suppose each set of form (??)
is a convex set. To show quasiconcavity, consider x, y ∈ Rn and 0 ≤ α ≤ 1.
Without loss of generality suppose a = f (x) = min{f (x), f (y)} and let A =
{x ∈ Rn : f (x) ≥ a}. Obviously x, y ∈ A and by concavity of A,

αx + (1 − α)y ∈ A

hence by definition of A we have

f (αx + (1 − α)y) ≥ a = min{f (x), f (y)}

this shows f is quasiconcave.

Other properties

• if f, g quasiconcave then f + g is not necessarily quasiconcave.

• If f is concave then f is quasiconcave.

• If f : R → R is monotone then f is both quasiconcave and quasiconvex,

• A C 2 function f : R2 → R is quasiconcave if and only if

2fx fy fxy − fx2 fyy − fy2 fxx ≥ 0

• In general, f : Rn → R is quasiconcave if and only if

whenever f (y) ≥ f (x) then Df (x)(y − x) ≥ 0

• some examples of quasiconcave functions(check it):

8
– f (x) = ln x (also concave)

– any monotone function f : R → R

– f (x, y) = ln x + ln y
1 1
– f (x, y) = x 4 y 4

– f (x, y) = xa y b for 0 < a, b < 1 and a + b ≤ 1

– f (x, y) = −xy