Advanced Math

Concavity and Quasiconcavity

Christoph Vanberg

University of Heidelberg
Department of Economics

0 of 21
Concavity and Quasiconcavity
▶ Concave (and convex) functions in R and Rn
▶ Sufficiency result for concave f and convex hj
▶ Quasiconcave (and quasiconvex) functions in Rn
▶ Sufficiency result for quasiconcave f and quasiconvex hj

In many economic applications, it is natural to assume that objective functions are con-
cave (or something similar) and constraint functions are convex (or something similar).
Under these conditions (and with certain technical caveats), the first order conditions
from the Lagrange formulation become sufficient. Our goal in this lecture is to develop
a basic understanding of the properties of concave (convex) functions that lead to these
results, and to state them formally.

The material presented in this lecture is based on Chapter 21 of the textbook:

Simon, Carl and Lawrence Blume (1994) Mathematics for Economists. Norton.

1 of 21
Definition: A set U ⊂ Rn is convex if

x0 , x1 ∈ U ⇒ tx0 + (1 − t)x1 ∈ U ∀t ∈ [0, 1].

2 of 21
Concave functions in R

Definition: A function f defined on an interval I ⊂ R1 is concave if ∀x0 , x1 ∈

I and ∀t ∈ [0, 1]

f(tx0 + (1 − t)x1 ) ≥ tf(x0 ) + (1 − t)f(x1 )

3 of 21
Concave functions in R
Note: The graph of a concave function in R has two interesting properties:
▶ By definition: Secant lines lie below
▶ If differentiable: Tangent lines lie above

Theorem: f : I ⊂ R → R and C1 is concave on I if and only if ∀x0 , x1 ∈ I,

f′ (x0 )(x1 − x0 ) ≥ f(x1 ) − f(x0 )

Proof: See textbook.

Exercise: Draw a similar figure in which x1 < x0 to verify that the statement
goes through.
4 of 21
Concave functions in R

Observe:: If f is concave, then the set {x : f(x) ≥ f(x0 )} is convex.

(Note: This is not an ‘if and only if’ statement. Many functions have this
property and are not concave.)

5 of 21
Concave functions in R
Summary
▶ By definition: Secant lines lie below the graph.
▶ Implication 1: Tangent lines lie above the graph.
▶ Implication 2: The set of x’s such that f(x) ≥ f(x0 ) is convex.

6 of 21
Convex functions in R

Definition: A function f defined on an interval I ⊂ R1 is convex if ∀x0 , x1 ∈

I and ∀t ∈ [0, 1]
f(tx0 + (1 − t)x1 ) ≤ tf(x0 ) + (1 − t)f(x1 )

Then everything is simply flipped:

▶ By definition: Secant lines lie above the graph.
▶ Implication 1: Tangent lines lie below the graph.
▶ Implication 2: The set of x’s such that f(x) ≤ f(x0 ) is convex.

7 of 21
Concave functions in Rn

Definition: Let f : U → R where U is a convex subset of Rn . We say that f is

concave if ∀x0 , x1 ∈ U and ∀t ∈ [0, 1]

f(tx0 + (1 − t)x1 ) ≥ tf(x0 ) + (1 − t)f(x1 )

Fact:: f is concave if and only if its restriction to any line segment in U is a

concave function of one variable.
(Proof: see textbook)

8 of 21
Concave functions
Recall: For any two points x0 and x1 , Df(x)(x1 − x0 ) is the directional derivative
along the line segment connecting the two points. Using this and the fact that
a concave function is concave on any line segment, we can generalize our results
for f in R to f in Rn .
Theorem: Let f be a C1 function on an open and convex set U ⊂ Rn . Then f
is concave on U if and only if ∀x0 , x1 ∈ U,

Df(x0 )(x1 − x0 ) ≥ f(x1 ) − f(x0 )

that is
Xn
∂f 0 1
(x )(xi − x0i ) ≥ f(x1 ) − f(x0 )
i=1
∂xi

9 of 21
Concave functions
If for any x1 and x0 we have
Df(x0 )(x1 − x0 ) ≥ f(x1 ) − f(x0 )
then the following useful result follows immediately:
Corollary:: If f is a concave C1 function on a convex set U and if x0 ∈ U, then
Df(x0 )(x1 − x0 ) ≤ 0 ⇒ f(x0 ) ≥ f(x1 )

That is: If the directional derivative at x0 in the direction to x1 is non-positive,

then f(x0 ) ≥ f(x1 ).

10 of 21
Concave functions
The following Theorems follow immediately from the Corollary:
Theorem: Let f be a concave function on an open, convex subset U of Rn . If
x0 ∈ U and Df(x0 ) = 0, then x0 is a global maximizer of f on U.

Proof: If Df(x0 ) = 0 then Df(x0 )(x1 − x0 ) = 0 for any x1 ∈ U, and so the

Corollary gives f(x0 ) ≥ f(x1 ) for all x1 ∈ U.

11 of 21
Concave functions
If the set U has a boundary (e.g. it is defined by inequality constraints), then if
x0 is at the boundary, we obtain a weaker condition:
Theorem: Let f be a C1 function defined on a convex (not necessarily open)
subset U ⊂ Rn . If x0 ∈ U satisfies Df(x0 )(x1 − x0 ) ≤ 0 for all x1 ∈ U, then x0 is
a global maximizer of f on U.

Note: The second theorem includes the first as a special case (when x0 is in the
interior of U).

12 of 21
Concave functions
Discussion
▶ We have learned: If a function is concave, then the first order
(directional) derivative can be used to state sufficient conditions for a
point to maximize f on a (convex) set.
▶ These arguments can be used to derive conditions under which the first
order conditions for constrained maximization problems become sufficient.
▶ Such conditions essentially involve the concavity of f and the convexity
of the constraint functions h1 , ..., hk .
▶ The latter condition guarantees that the constraint set is a convex set,
such that the results we just discussed are applicable.

13 of 21
Consider the constrained maximization problem
maxn f(x) s.t. h1 (x) ≤ c1 , ..., hk (x) ≤ ck
x∈R

Theorem: (Karush-Kuhn-Tucker) Suppose that

▶ f is C1 and concave and all hj are C1 and convex, and
▶ One of the following CQ’s is satisfied:
(i) The gradients of the binding constraints are linearly independent at
the solution x∗
(ii) ∃x ∈ Rn such that hj (x) < 0∀j (Slater’s Condition)
P
Form the Lagrangian L(x1 , ..., xn , µ1 , ..., µk ) = f(x) − kj=1 µj (hj (x) − cj ). Then
x∗ ∈ Rn solves the constrained maximization problem if and only if there exist
µ∗1 , ..., µ∗k , all µ∗j ≥ 0, such that

∂L ∗ ∗
(x , µ ) = 0 ∀j = 1, ..., n
∂xj
µ∗j (hj (x∗ ) − cj ) = 0 j = 1, ...k
hj (x∗ ) ≤ cj j = 1, ...k

∑ ∗ ∗
Comments: Recall that the first of the three equations says that ∇f(x) = j µj ∇hj (x ). Also note that
the constraint qualification is only required to make the Lagrange conditions necessary. Even without this, these
conditions are sufficient. (However without checking the CQ we cannot know whether there are additional solutions.)

14 of 21
Concave functions: Second derivative criterion

Theorem: Let f be a C2 function on an open convex U ⊂ Rn . Then f is concave

on U if and only if the Hessian D2 f(x) is negative semi-definite for all x ∈ U.
Definition: An n×n matrix H is negative semidefinite if vT Hv ≤ 0∀v ̸= 0 ∈ Rn .

Test: An n×n matrix H is negative semidefinite if its 2n −1 principal minors (not

just ‘leading’!) alternate in sign such that odd order p.m.’s are ≤ 0 and even
order p.m.’s are ≥ 0. (The kth order principal minors are the determinants of all
submatrices formed by deleting (n − k) rows and the same (n − k) columns.)
Example  
a11 a12 a23
A =  a21 a22 a23 
a31 a32 a33
3rd order p.m.: |A|   be ≤ 0)
 (This should   
 a22 a23   a11 a23   a11 a12 

2nd order p.m.:  , , (All should be ≥ 0)
a32 a33   a31 a33   a21 a22 
1st order p.m.: a11 , a22 , a33 (All should be ≤ 0)

15 of 21
Cardinal vs. ordinal properties

▶ Economists often distinguish between cardinal and ordinal properties of

functions.
▶ Cardinal properties are sensitive to ‘monotonic transformations’ of a
function: If g is a strictly increasing function of one variable, then g(u(x))
is a monotonic transformation of the function u.
▶ Monotonic transformations have no effect on the shapes of the level sets
of a function, only on the numbers assigned to them.
▶ Ordinal properties are those which are unaffected by monotonic
transformations, i.e. all properties that have to do only with the shapes of
a function’s level sets.
▶ When economists use utility functions, we do not attach meaning to the
numbers attached to level sets. If one utility function represents an
individual’s preferences, any monotonic transformation represents the
same preferences.
▶ Unfortunately, concavity is a cardinal property of functions. Therefore if
the preceding result is only applicable to concave functions, it might not
be useful in the context of utility theory.

16 of 21
Quasiconcave functions
Luckily, the following property of concave functions is ordinal (unaffected by
monotonic transformations):
▶ Upper contour sets Ca = {x ∈ U : f(x) ≥ a} are convex.
It turns out that this ordinal property is enough to achieve the kind of sufficiency
result we just established for concave functions. Thus we define the following
class of functions using just the (weaker) ordinal property:
Definition: Let f : U → R where U ⊂ Rn is convex. f is quasiconcave if ∀a ∈ R,

a ≡ {x ∈ U : f(x) ≥ a} is a convex set

C+

We say that this is a weaker condition because all concave functions are quasi-
concave, but not all quasiconcave functions are concave.
Similarly, quasiconvexity is an ordinal property satisfied by all convex functions
but also other functions:
Definition: Let f : U → R where U ⊂ Rn is convex. f is quasiconvex if ∀a ∈ R,

C−
a ≡ {x ∈ U : f(x) ≤ a} is a convex set

17 of 21
Quasiconcave functions

Theorem: Let f : U → R where U ⊂ Rn is convex. f is quasiconcave if and

only if ∀x0 , x1 ∈ U and t ∈ [0, 1]
f(tx0 + (1 − t)x1 ) ≥ min{f(x0 ), f(x1 )}

Proof: Let a = min{f(x0 ), f(x1 )}. Then both x0 and x1 are in C+ a ≡ {x ∈ U : f(x) ≥ a}.
Since this is a convex set, the point tx0 + (1 − t)x1 is also in C+a .

Theorem: Let f be a C1 function on U ⊂ Rn (open and convex). Then f is

quasiconcave if and only if
f(x1 ) ≥ f(x0 ) ⇒ Df(x0 )(x1 − x0 ) ≥ 0

Proof: If f(x1 ) > f(x0 ) then by the previous theorem f(tx1 + (1 − t)x0 ) ≥ f(x0 ). Thus
for any t > 0 we have
f(x0 + t(x1 − x0 )) − f(x0 )
≥0
t∥x1 − x0 ∥
Then if we take the limit for t → 0 we get
Df(x0 )(x1 − x0 ) ≥ 0

18 of 21
Quasiconcave functions
We have established
Theorem: Let f be a C1 function on U ⊂ Rn (open and convex). Then f is
quasiconcave if and only if
f(x1 ) ≥ f(x0 ) ⇒ Df(x0 )(x1 − x0 ) ≥ 0

Exercise: State the contrapositive of the last statement in the theorem. (Recall:
the contrapositive of A ⇒ B is ∼B ⇒ ∼A)
Answer:
Df(x0 )(x1 − x0 ) < 0 ⇒ f(x1 ) < f(x0 )

Note
▶ This is very close to the useful ‘Corollary’ we obtained in our analysis of
concave functions.
▶ Again, certain statements about first order directional derivatives at a
point x0 will be sufficient to guarantee that x0 maximizes f on a convex
set.
▶ As a consequence, we obtain a very similar sufficiency theorem for
quasiconcave objective functions and quasiconvex constraints.

19 of 21
Consider the constrained maximization problem

maxn f(x) s.t. h1 (x) ≤ c1 , ..., hk (x) ≤ ck

x∈R

Theorem: (Arrow-Enthoven) Suppose that

▶ f is C1 and quasiconcave and all hj are C1 and quasiconvex.
P
Form the Lagrangian L(x1 , ..., xn , µ1 , ..., µk ) = f(x)− kj=1 µj (hj (x)−cj ). Suppose
x∗ and µ∗1 , ..., µ∗k , all µ∗j ≥ 0 satisfy the familiar necessary conditions

∂L ∗ ∗
(x , µ ) = 0 ∀j = 1, ..., n
∂xj
µ∗j (hj (x∗ ) − cj ) = 0 j = 1, ...k
hj (x∗ ) ≤ cj j = 1, ...k

Then x∗ solves the constained maximization problem provided that at least one
of the following conditions holds:
▶ ∇f(x∗ ) ̸= 0
▶ f is concave

20 of 21
Summary

▶ Concavity is a cardinal property of functions which renders first order

conditions necessary and sufficient for maximization on a convex set.
▶ Convexity of constraint functions guarantees that the constraint set
defined by a set of inequality constraints of the form hj (x) ≤ cj is convex.
(Because each constraint defines a convex set and the intersection of
convex sets is convex.)
▶ These result together (along with additional technical details) yield the
Karush-Kuhn-Tucker Theorem.
▶ The ordinal analogues of convave and convex functions are quasiconcave
and quasiconvex functions. These concepts yield nearly the same
properties for the purpose of maximization. This gives us the
Arrow-Endhoven Theorem.
▶ Both theorems tell us that under conditions that are usually satisfied in
economic applications, the first order Lagrange conditions are both
necessary and sufficient.

21 of 21