Convexity-Print Version PDF
Convexity-Print Version PDF
• Thus a set is convex if the straight line joining any two points in D is completely
contained in D i.e. if for all x and y in D and λ ∈ (0, 1) it is the case that
λx + (1 − λ)y is a subset of D.
y y
epi f
sub f
x x
Figure 2: The subgraph and epigraph of f .
• Note concave and convex functions are required to have convex domains.
• The following theorem provides an alternative definition of concave and convex
functions.
2
f (λx + (1 − λ)y)
x λx + (1 − λ)y y
Figure 3: A function f is concave iff the secant line connecting any two points on the graph of f lies below the graph.
• We say f is strictly convex if for all x, y ∈ D with x 6= y, and all λ ∈ (0, 1), we
have
f (λx + (1 − λ)y) < λf (x) + (1 − λ)f (y). N
• The previous result allows us to easily apply all results about concave functions
to convex functions
• Another valuable property of concave functions is that they behave well under
addition and scalar multiplication by positive numbers.
a1 f1 + · · · + ak fk
is a concave function.
Proof. Simply apply the definition of a concave function.
3
• Second, every concave function is differentiable “almost everywhere”.
• Note that, if we let z = y−x, we can rewrite (1) to say f is concave iff Df (x)z+
f (x) ≥ f (x + z) for all x, z ∈ D.
• Thus a function is concave iff the tangent line lies above the graph of the func-
tion.
f 0 (x)z + f (x)
f (x + z)
f (x)
x x+z
Figure 4: A function is concave iff the tangent line lies above the graph of the function.
4
• In the next theorem, the concavity or convexity of a C 2 function is characterized
using the second derivative.
• The theorem also gives a sufficient condition which can be used to identify
strictly concave and strictly convex functions.
• It is important to note that parts (3) and (4) of the theorem are only sufficient
conditions. For example, part (3) does not say that if f is strictly concave on D,
then D2 f (x) is a negative definite matrix for all x ∈ D.
• Our next example illustrates the importance of the theorem for simplifying the
identification of concavity in practice.
• Similarly f is convex, if for all (x, y) and (x̂, ŷ) in R2++ and any λ ∈ (0, 1), we
have
5
• The latter only requires us to identify the definiteness of the following matrix:
• We now present some results which indicate the importance of convexity for
optimization theory.
• But first some terminology.
Definition.
• We refer to a maximization problem as a convex maximization problem if the
constraint set is convex and the objective function is concave.
• Similarly, we refer to a minimization problem as a convex minimization problem
if the constraint set is convex and the objective function is convex.
• More generally, we refer to an optimization problem as a convex optimization
problem if it is either of the above. N
• The first result establishes that in convex optimization problems, all local optima
must also be global optima.
• Thus, to find a global optimum in such problems, it is sufficient to identify a
local optimum.
6
• For example, in the utility maximization problem with two perfect substitutes, ei-
ther the solution is a unique corner solution or there are infinitely many solutions
along the budget constraint.
• The second result shows that if a strictly convex optimization problem has a
solution, then the solution is unique.
• We can combine this result with the Weierstrass theorem to establish the exis-
tence of a unique global optimum in a convex optimization problem in which the
objective function is continuous and the constraint set is compact.
• We have seen that convexity has powerful implications for optimization prob-
lems. However, convexity is a very restrictive assumption, which is important
when we come to applications.
• For example, we saw that the Cobb-Douglas function production f (x, y) = xa y b
(a, b > 0) is not concave unless a + b ≤ 1.
• So, we will now look at optimization under a weakening of the condition of
convexity, called quasiconvexity.
• Thus a function is quasiconcave if its upper contour sets are convex sets.
• Similarly, a function is quasiconvex if its lower contour sets are convex sets.
• As is the case with concave and convex functions, it is also true for quasicon-
cave and quasiconvex functions that a relationship exists between the value of a
function at two points and the value of the function at a convex combination.
7
Uf (a)
x
λx + (1 − λ)y
y {x̂ ∈ D | f (x̂) = a}
{x̂ ∈ D | f (x̂) = b}
Figure 5: The level sets of a strictly quasiconcave function (a > b). The upper contour set of f at a is Uf (a) = {x̂ ∈
D | f (x̂) ≥ a}.
• A similar result holds for quasiconvex functions, with the inequalities reversed
and “min” replaced with “max”.
Definition. Let D be a convex subset of Rn and let f : D → R be a function.
• We say f is strictly quasiconcave if for all x, y ∈ D with x 6= y, and all λ ∈
(0, 1), we have
f (λx + (1 − λ)y) > min{f (x), f (y)}.
• We say f is strictly quasiconvex if for all x, y ∈ D with x 6= y, and all λ ∈ (0, 1),
we have
f (λx + (1 − λ)y) < max{f (x), f (y)}. N
Theorem 10. Let D be a convex subset of Rn and let f : D → R be a function. Then
1. f is quasiconcave iff the function −f is quasiconvex.
2. f is strictly quasiconcave iff the function −f is strictly quasiconvex.
8
Quasiconvexity as a Generalization of Convexity.
• To show this, consider any x, y ∈ R and any λ ∈ (0, 1). Assume, without loss
of generality, that x > y. Then
• Since f (x) = max{f (x), f (y)}, the first inequality shows that f is quasiconvex.
• Similarly, since f (y) = min{f (x), f (y)}, the second inequality shows that f is
quasiconcave.
• Since it is always possible to choose a nondecreasing function f that is neither
concave nor convex on R (say f (x) = x3 ), we have shown that not every quasi-
concave function is concave and not every quasiconvex function is convex.
• The next theorem elaborates on the relationship between concave and quasicon-
cave functions.
• The converse of this theorem is not true. That is, we cannot say that every quasi-
concave function is an increasing transformation of some concave function. See
Sundaram pp207-209 for two concrete examples of quasiconcave functions that
are not increasing transformations of any concave function.
9
Quasiconvexity and the Properties of the Derivative.
• The condition (1) is illustrated in the figure. If we think of Df (x)T as the gradi-
ent vector ∇f (x), then the theorem says that the angle between the gradient and
the vector y − x is acute (or right).
x
y
Figure 6: The condition (1) says that the angle between the vector y − x and ∇f (x) is acute.
10
1. If f is quasiconcave on D, then, for all x ∈ D, (−1)r−1 |H r | ≥ 0 for r =
2, . . . , n + 1.
2. If f is quasiconvex on D, then, for all x ∈ D, |H r | ≤ 0 for r = 2, . . . , n + 1.
3. If (−1)r−1 |H r | > 0 for all r = 2, . . . , n + 1, then f is quasiconcave on D.
4. If |H r | < 0 for all r = 2, . . . , n + 1, then f is quasiconvex on D.
• Part (3) requires the signs of the leading principal minors to alternate, starting
with negative for the 2 × 2 matrix H 2 .
• Compare this theorem with the corresponding theorem on concavity.
• There are two important differences.
– In theorem (6), a weak inequality i.e. the negative semidefiniteness of D2 f
was both necessary and sufficient to establish concavity.
– However, in the result above, the weak inequality is only a necessary condi-
tion for quasiconcavity. The sufficient condition involves a strict inequality.
– Second, the theorem does not give a test for strict quasiconcavity.
Example 4. Let f : R2++ → R be given by f (x, y) = xa y b , a, b > 0.
• We saw that f is strictly concave on if a + b < 1, concave if a + b = 1, and
neither concave nor convex if a + b > 1.
• We will show that f is quasiconcave for all a, b > 0.
• To show this directly, using the definition of quasiconcavity, requires us to prove
that
[λx + (1 − λ)x̂]a [λy + (1 − λ)ŷ]b ≥ min{xa y b , x̂a , ŷ b }
holds for all (x, y) 6= (x̂, ŷ) in R2++ and for all λ ∈ (0, 1).
• Compare checking for quasiconcavity f using this inequality to checking using
the second derivative test.
• We have to show that |H 3 (x, y)| < 0 and |H 3 (x, y)| > 0 for all x, y ∈ R2++ ,
where
axa−1 y b
0
H 2 (x, y) = ,
axa−1 y b a(a − 1)xa−2 y b
axa−1 y b bxa y b−1
0
H 3 (x, y) = axa−1 y b a(a − 1)xa−2 y b abxa−1 y b−1 .
a b−1 a−1 b
bx y abx y b(b − 1)xa y b−2
• Calculating the determinants we find
|H 2 (x, y)| = −a2 x2(a−1) y 2b < 0
|H 3 (x, y)| = ab(a + b)x3a−2 y 3b−2 > 0,
for all (x, y) ∈ R2++ .
• Thus f is quasiconcave on (x, y) ∈ R2++ .
11
Quasiconvexity and Optimization.
• We saw that local maxima of quasiconcave functions need not be global maxima.
• However, when the function is strictly quasiconcave, there is a result identical to
that for strictly concave functions.
12
• A similar result holds for strictly quasiconvex functions in minimization prob-
lems.
• This is significant because it says that the weaker property of strict quasiconcav-
ity is enough to guarantee uniqueness of the solution (if there is one).
13