HUL212: Curvature of ICs

TA: Bristi Das

January 29, 2025

For the case of R2 ; that is, y = f (x1 , x2 ). The first-order total differential for y = f (x1 , x2 ) is:

∂f (x1 , x2 ) ∂f (x1 , x2 )
dy = dx1 + dx2
∂x1 ∂x2
or

dy = f1 dx1 + f2 dx2
The second-order total differential is the total differential of dy; that is, d[dy] or d2 y where

∂[dy] ∂[dy]
d[dy] = dx1 + dx2
∂x1 ∂x2
∂[f1 dx1 + f2 dx2 ] ∂[f1 dx1 + f2 dx2 ]
= dx1 + dx2
∂x1 ∂x2
= [f11 dx1 + f12 dx2 ]dx1 + [f12 dx1 + f22 dx2 ]dx2
= f11 dx21 + 2f12 dx1 dx2 + f22 dx22

since f21 = f12 by Young’s theorem.

Theorem 1.1

If the function y = f (x1 , x2 ) defined on R2 is twice continuously differentiable and

d2 y = f11 dx21 + 2f12 dx1 dx2 + f22 dx22 > 0

whenever at least one of dx1 or dx2 is nonzero, then y = f (x1 , x2 ) is a strictly convex function.

Theorem 1.2

If the function y = f (x1 , x2 ) defined on R2 is twice continuously differentiable and

d2 y = f11 dx21 + 2f12 dx1 dx2 + f22 dx22 < 0

whenever at least one of dx1 or dx2 is nonzero, then y = f (x1 , x2 ) is a strictly concave function.

Note: Weaker conditions of convexity or concavity ≥ 0 or ≤ 0 allow for linear functions or functions
with linear segments.

Example 1
Use the sign of the second-order total differential to show that the function y = 5 − (x1 + x2 )2 is concave.

1
 Figure 1: Linear segment on the surface of the graph

Solution
First- and second-order partial derivatives of this function are:

f1 = −2(x1 + x2 ), f2 = −2(x1 + x2 )

f11 = −2, f22 = −2, f12 = −2

It follows that
d2 y = f11 dx21 + 2f12 dx1 dx2 + f22 dx22
= −2dx21 − 4dx1 dx2 − 2dx22
= −2(dx1 + dx2 )2 ≤ 0
and so the function is concave. To see that this function is not strictly concave, notice that along the set
of (x1 , x2 ) values satisfying x2 = a − x1 , where a is any constant, we have

y = 5 − (x1 + a − x1 )2 = 5 − a2 .

These (x1 , x2 ) values generate horizontal linear segments of the graph of f as illustrated in figure.

2
 This analysis concerning concavity and convexity carries over to functions of more than two variables.
Recall that the first-order total differential for y = f (x1 , x2 , . . . , xn ) is
n
X
dy = f1 dx1 + f2 dx2 + · · · + fn dxn = fi dxi
i=1

Since each of the partial derivatives fi is a function of (x1 , x2 , . . . , xn ), dy is also a function of (x1 , x2 , . . . , xn ),
and so we can find its total differential d[dy] or d2 y as follows:

∂[dy] ∂[dy] ∂[dy]

d2 y = dx1 + dx2 + · · · + dxn
∂x1 ∂x2 ∂xn
Pn Pn
∂ [ i=1 fi dxi ] ∂ [ i=1 fi dxi ]
= dx1 + dx2
∂x1 ∂x2
Pn
∂ [ i=1 fi dxi ]
+··· + dxn .
∂xn
(1)
Each term in equation (1) represents n partial derivatives. For example,
Pn
∂ [ i=1 fi dxi ] ∂ [f1 dx1 + f2 dx2 + · · · + fn dxn ]
=
∂x1 ∂x1
∂f1 ∂f2 ∂fn
= dx1 + dx2 + · · · + dxn
∂x1 ∂x1 ∂x1

= f11 dx1 + f21 dx2 + · · · + fn1 dxn

It follows that since each of the n terms in equation (1) represents n terms, there are n2 terms altogether.
It is tedious to write these out in full. However, by using summation notation, we get
n  Pn  n X
n
X ∂[ fi dxi ] X
d2 y = i=1
dxj = fij dxi dxj
j=1
∂xj j=1 i=1

which in turn indicates how we can write d2 y in matrix notation

d2 y = dxT ∇2 F dx

(2)

where dxT = [dx1 dx2 . . . dxn ] is the vector of changes dxi and
 
f11 f12 ... f1n
 f21 f22 ... f2n 
∇2 F =  .
 
.. .. .. 
 .. . . . 
fn1 fn2 ... fnn
is the Hessian matrix of f .

3
 Theorem 1.3

For any function y = f (x), x ∈ Rn which is twice continuously differentiable with Hessian H, it
follows that:

1. The function f is strictly convex on Rn if H is positive definite for all x ∈ Rn (i.e., d2 y =

dxT H dx > 0).

2. The function f is strictly concave on Rn if H is negative definite for all x ∈ Rn (i.e., d2 y =

dxT H dx < 0).
3. The function f is convex on Rn if and only if H is positive semi-definite for all x ∈ Rn (i.e.,
d2 y = dxT H dx ≥ 0).

4. The function f is concave on Rn if and only if H is negative semi-definite for all x ∈ Rn (i.e.,
d2 y = dxT H dx ≤ 0).

The conditions for positive (negative) definiteness depend on the signs of the leading principle minors of
the matrix. (Proofs can be found in any linear algebra book)
Leading principal minors are:

f11 f12
|H1 | = f11 , |H2 | = ...
f21 f22

f11 f12 ... f1n

f21 f22 ... f2n
|Hn | = |H| = . .. .. ..
.. . . .
fn1 fn2 ... fnn
Principal minors are:

f11 f12 f11 f13 f22 f23

|H1∗ | = f11 , f22 , f33 |H2∗ | = , , ...
f21 f22 f31 f33 f32 f33

4
 Theorem 1.4

Let H be the Hessian matrix associated with a twice continuously differentiable function y = f (x),
x ∈ Rn . It follows that:

1. H is positive definite on Rn if and only if its leading principal minors are positive:

|H1 | > 0, |H2 | > 0, |H3 | > 0, . . . , |Hn | = |H| > 0

for x ∈ Rn . In this case, d2 y > 0 and so f is strictly convex.

2. H is negative definite on Rn if and only if its leading principal minors alternate in sign
beginning with a negative value for k = 1:
(
> 0, if n is even
|H1 | < 0, |H2 | > 0, . . . , |Hn | = |H|
< 0, if n is odd

for x ∈ Rn . In this case, d2 y < 0 and so f is strictly concave.

3. H is positive semidefinite on Rn if and only if all of its principal minors are positive or zero:

|H1∗ | ≥ 0, |H2∗ | ≥ 0, |H3∗ | ≥ 0, . . . , |Hn∗ | = |H| ≥ 0

for x ∈ Rn . In this case, d2 y ≥ 0 and so f is convex. Moreover, if f is convex, this set of

conditions must hold.
4. H is negative semidefinite on Rn if and only if all of its principal minors alternate in sign
beginning with a negative or zero value for k = 1:
(
∗ ∗ ∗ ≥ 0, if n is even
|H1 | ≤ 0, |H2 | ≥ 0, . . . , |Hn | = |H|
≤ 0, if n is odd

for x ∈ Rn . In this case, d2 y ≤ 0 and so f is concave. Moreover, if f is concave, this set of

conditions must hold.

Example 2
Determine the convexity/concavity property of the function

y = f (x1 , x2 ) = (x1 + x2 )1/2

defined on x ∈ R2++ .

Solution
The second-order partial derivatives are
1
f11 = (x1 + x2 )−3/2
4
1
f12 = f21 = − (x1 + x2 )−3/2
4
1
f22 = (x1 + x2 )−3/2
4
Since all of these are negative, we check first for strict concavity of f :

5
 |H1 | = f11 < 0.

f11 f12 2
|H2 | = = f11 f22 − f12 = 0.
f21 f22
Note that |H2 | = 0, and so we cannot apply result 2 of theorem 1.4. Therefore we check for (weak)
concavity.

|H1∗ | = f11 , f22 , both of which are negative.

|H2∗ | = |H2 | ≥ 0 (because |H2 | = 0).

Therefore f is concave.
Note: Can also be proved in terms of signs of eigenvalues.

Convex Set
A set is convex iff, it contains every convex combination of its points.
Here a convex combination of the finite number of points x1 , x2 , . . . , xk is defined as a point x = θ1 x1 +
θ2 x2 + · · · + θk xk , where θ1 + θ2 · · · + θk = 1 and θi ≥ 0 for all i = 1, . . . , k

Figure 2: Convex Set Figure 3: Non- convex set

Convex/ concave function

The function f is concave if
f (x̄) ≥ λf (x′ ) + (1 − λ)f (x′′ )
where x̄ = λx′ + (1 − λ)x′′ and λ ∈ [0, 1]. It is strictly concave if the strict inequality holds when λ ∈ (0, 1).

The function f is convex if

f (x̄) ≤ λf (x′ ) + (1 − λ)f (x′′ )
where x̄ = λx′ + (1 − λ)x′′ and λ ∈ [0, 1]. It is strictly convex if the strict inequality holds when λ ∈ (0, 1).

6
 Figure 4: Strictly concave function
Figure 5: Strictly convex function

Quasi- convexity (concavity)

The function f (x), x ∈ Rn defined on a convex set S is quasiconvex if every lower contour set of f is convex.
(That is, Pa = x ∈ S : f (x) ≤ a is convex for every value of a.)

The function f (x), x ∈ Rn defined on a convex set S is quasiconcave if every upper contour set of f is convex.
(That is, Pa = x ∈ S : f (x) ≥ a is convex for every value of a.)

Proposition I

A concave function is quasiconcave. A convex function is quasiconvex.

Proof: Denote the function by f , and the (convex) set on which it is defined by S. Let a be a real
number and let x and y be points in the upper contour set Pa : x ∈ Pa and y ∈ Pa . We need to show that
Pa is convex.
That is, we need to show that for every λ ∈ [0, 1] we have (1 − λ)x + λy ∈ Pa .
First note that the set S on which f is defined is convex, so we have (1 − λ)x + λy ∈ S, and thus f is
defined at the point (1 − λ)x + λy.
Now, the concavity of f implies that

f ((1 − λ)x + λy) ≥ (1 − λ)f (x) + λf (y). (1)

Further, the fact that x ∈ Pa means that f (x) ≥ a, and the fact that y ∈ Pa means that f (y) ≥ a, so
that
(1 − λ)f (x) + λf (y) ≥ (1 − λ)a + λa = a. (2)
Combining the last two inequalities, we have

f ((1 − λ)x + λy) ≥ a, (3)

so that (1 − λ)x + λy ∈ Pa . Thus every upper level set is convex and hence f is quasiconcave.

Note: The converse may not be true.

Result: Average is preferred to extreme

Suppose a function f : S → R defined on convex set S ⊂ Rn ,

f is quasiconcave ⇐⇒ ∀ x, x′ ∈ S, and ∀ λ ∈ [0, 1], we have f (λx + (1 − λ)x′ ≥ min{f (x), f (x′ )}

7
 Proof:
• To Show: f is quasiconcave =⇒ f (λx + (1 − λ)x′ ) ≥ min{f (x), f (x′ )}

If f is quasiconcave, Px = {y ∈ S : f (y) ≥ a} is a convex set.

Pick arbitrary x, x′ ∈ S and λ ∈ [0, 1]
We know, f (x) ≥ min{f (x), f (x′ )} and f (x′ ) ≥ min{f (x), f (x′ )}
Hence, we get x ∈ Pmin{f (x),f (x′ )} and x′ ∈ Pmin{f (x),f (x′ )}
We get λx + (1 − λ)x′ ∈ Pmin{f (x),f (x′ )}
Using definition of Px , we get f (λx + (1 − λ)x′ ) ≥ min{f (x), f (x′ )}
Therefore, f satisfies “Average is better than extremes” property.
• f (λx + (1 − λ)x′ ) ≥ min{f (x), f (x′ )} =⇒ f is quasiconcave

If f satisfies “Average is better than extremes” property i.e., f (λx + (1 − λ)x′ ) ≥ min{f (x), f (x′ )}
Pick any arbitrary x, x′ ∈ Pa , and arbitrary a ∈ R
We get f (x) ≥ a and f (x′ ) ≥ a
We get min(f (x), f (x′ ) ≥ a
Hence,f (λx + (1 − λ)x′ ) ≥ a
By definition of Pa we get λx + (1 − λ)x′ ∈ Pa i.e Pa is a convex set
Therefore, f is quasiconcave.

Note: Curvature (convex/ concave) of IC [U (x, y) = constant =⇒ y = f (x)] is different than the quasi-
concavity or quasiconvexity of the utility function, U (x, y).

Note: Quasiconvexity implies f (λx1 + (1 − λ)x2 ) ≤ max(f (x1 ), f (x2 )) Example: U = x2 + y 2 is quasiconvex
and extreme is preferred to average.

Note: Convex preference means desire for diversification or average is preferred to extreme.