130 Chapter 6 Differentiation and Integration

6.6 CONVEX FUNCTIONS

Throughout this section (a, b) is an open interval that may be bounded or unbounded.

Definition A real-valued function rp on (a, b) is said to be convex provided for each pair of
points Xl. X2 in (a, b) and each A with 0 ~ A ~ 1,

(38)

If we look at the graph of rp, the convexity inequality can be fonnulated geometrically by
saying that each point on the chord between (Xl, rp( Xl» and (X2, rp( X2) ) is above the graph
of rp.
Observe that for two points Xl < X2 in (a, b), each point X in (Xl, X2) may be expres-
sed as
X2 -x
X = Axl + (1- A)X2 where A = - - .
X2 - Xl
Thus the convexity inequality may be written as

rp(x) ~ X2 --
[- X ] rp(xt} + [ -
x-- Xl ] rp(X2) for Xl < X < X2 in (a, b).
X2 - Xl X2 -Xl

Regathering tenns)ms inequality may also be rewritten as

(39)

Therefore convexity may also be fonnulated geometrically by saying that for Xl < X < X2, the
slope of the chord from (Xl, rp(Xl» to (X, rp(x» is no greater than the slope of the chord
from (x, rp(x» to (X2, rp(X2».

Proposition 15 If rp is differentiable on (a, b) and its derivative rp' is increasing, then rp is

convex. In particular, rp is convex if it has a nonnegative second derivative rpl! on (a, b).

Proof Let Xl, X2 be in (a, b) with Xl < X2, and let X belong to (Xl. X2). We must show that

However, apply the Mean Value Theorem to the restriction of rp to each of the intervals
[Xl, X] and [x, X2] to choose points q E (Xl. x) and C2 E (x, X2) for which

'( ) rp(x) - rp(Xl) d '( ) rp(X2) - rp(x)

rpq= anrp c2= .
X -Xl X2-X

Thus, since rp' is increasing,

D
 Section 6.6 Convex Functions 131

Example Each of the following three functions is convex since each has a nonnegative
second derivative:
q>(X) = xP on (0, 00) for p 2: 1; q>(x) = IF on (-00, 00); q>(x) = In(1/x) on (0, 00).
The following final geometric reformulation of convexity will be useful in the estab-
lishment of differentiability properties of convex functions.
The Chordal Slope Lemma Let II' be convex on (a, b). If Xl < X < X2 belong to (a, b), then
for PI = (Xl> q>(xt}), P = (X, q>(x», P2 = (X2, q>(X2»,

Slope of PIP ~ slope of PIP2 ~ slope of PP2·

Proof Regather terms in the inequality (39) to rewrite it in the following two equivalent
forms:
q>(Xt)-q>(X) ~ q>(x2)-q>(xI) forxl <x<x2in(a, b);
Xl -X X2 -Xl
< X < X2 In
-.:. . :q>(,- ,X2:. :. )_---'-q>(-'---X=-:.l) _. (a, b) .
< 11'( X2) - 11'( X) fo r Xl D
X2 -Xl X2 - X
For a function g on an open interval (a, b), and point Xo E (a, b), if

Iim g( Xo + h) - g( xo) eXlS

.t d' fi .
s an IS rute,
h .... O. h <0 h

we denote this limit by g'(xo) and call it the left-hand derivative of gat Xo. Similarly, we
define g' (x6) and call it the right-hand derivative of g at Xo. Of course, g is differentiable at
Xo if and only if it has left-hand and right-hand derivatives at Xo that are equal. The continuity
and differentiability properties of convex functions follow from the following lemma, whose
proof follows directly from the Chordal Slope Lemma.

Lemma 16 Let II' be a convex function on (a, b). Then II' has left-hand and right-hand
derivatives at each point X E (a, b). Moreover, for points u, v in (a, b) with u < v, these
one-sided derivatives satisfy the following inequality:

q>'(u-) ~ q>'(u+) ~ q>(v) - q>(u) ~ q>'(v-) ~ q>'(v+). (40)

v-u
132 Chapter 6 Differentiation and Integration

CoroUary 17 Let cp be a convex function on (a, b). Then cp is Lipschitz, and therefore
absolutely continuous, on each closed, bounded subinterval [c, d] of (a, b).

Proof According to the preceding lemma, for c ::; u < v::; d,

(41)
and therefore
Icp(u) - cp( v)1 ::; Mlu - vi for all u, v E [c, d],
where M = max{lcp'(c+)I, Icp'(d-)I}. Thus the restriction of cp to [u, v] is Lipschitz. A
Lipschitz function on a closed, bounded interval is absolutely continuous. 0

We infer from the above corollary and Corollary 6 that any convex function defined
on an open interval is differentiable almost everywhere on its domain. In fact, much more
can be said.

Theorem 18 Let cp be a convex function on (a, b). Then cp is differentiable except at a

---
countable number of points and its derivative cp' is an increasing function.
.
Proof We infer from the ineqUalities (40) that the functions
X 1-+ J'(x-) and X 1-+ J'(x+)
are increasing real-valued functions on (a, b). But, according to Theorem 1, an increasing
real-valued function is continuous except at a countable number of points. Thus, except
on a countable subset C of (a, b), both the left-hand and right-hand derivatives of cp are
continuous. Let xo belong to (a, b) '" C. Choose a sequence {xn } of points greater than
Xo that converges to Xo. Apply Lemma 16, with Xo = u and Xn = v, and take limits
to conclude that
cp'(x o)::; CP'(x6) ::; cp'(xo)·
Then cp' (xo) = cp' (x6) so that cp is differentiable at xo. To show that cp' is an increasing
function on (a, b) "'C, let u, v belong to (a, b) '" C with u < v. Then by Lemma 16,

cp'(u)::; cp(u) - cp(v) ::; cp'(v). 0

u-v
Let cp be a convex function on (a, b) and Xo belong to (a, b). For a real number m,
the line y = m(x - xo) + cp(xo), which passes through the point (xo, cp(xo)), is called a
supporting line at Xo for the graph of cp provided this line always lies below the graph of cp,
that is, if
cp(x) ~ m(x - xo) + cp(xo) for all x E (a, b).
It follows from Lemma 16 that such a line is supporting if and only if its slope m lies between
the left- and right-hand derivatives of cp at Xo. Thus, in particular, there is always at least one
supporting line at each point. This notion enables us to give a short proof of the following
inequality:
 Section 6.6 Convex Functions 133

Jensen's Inequality Let Ip be a convex function on ( -00, 00 ), f an integrable function over

[0, 11 and Ip 0 f also integrable over [0, 1]. Then

(42)

Proof Define a = f~ f(x)dx. Choose m to lie between the left-hand and right-hand
derivative of Ip at the point a. Then y = m( t - a) + Ip( a) is the equation of a supporting line
at (a, Ip( a» for the graph of Ip. Hence

Ip(t} :::: m( t - a) + Ip( a) for all t E R.

Since f is integrable over [0, 1], it is finite a.e.on [0, 1] and therefore, substituting f( x) for t
in this inequality, we have

Ip(f(x}) :::: m(f(x) - a) + Ip(a} for almost all x E [0, 1].

Integrate across this inequality, using the monotonicity of the Lebesgue integral and the
assumption that both/! and Ip 0 f are integrable over [a, b], to obtain

f Ip(f(x)}dx:::: f [m(f(x) - a) +Ip(a}] dx

= [f
m f(x)dx - a] + Ip(a} = Ip(a}. D

A few words regarding the assumption, for Jensen's Inequality, of the integrability of
Ip 0 f over [0, 1] are in order. We have shown that a convex function is continuous and
therefore Proposition 7 of Chapter 3 tells us that the composition Ip 0 f is measurable if Ip is
convex and f is integrable. If Ip 0 f is nonnegative, then it is unnecessary to assume the Ip 0 f
is integrable since equality (42) trivially holds if the right-hand integral equals +00. In the
case Ip 0 f fails to be nonnegative, if there are constants ct and C2 for which

(43)

then we infer from the integral comparison test that Ip 0 f is integrable over [0, 1] if f is. In
the absence of the growth assumption (43), the function Ip 0 f may not be integrable over
[0, 1] (see Problem 71).

PROBLEMS
61. Show that a real-valued function Ip on (a, b) is convex if and only if for points Xl, ... , xn in
(a, b) and nonnegative numbers At. ... , An such that ~LI Ak = 1,

Use this to directly prove Jensen's Inequality for f a simple function.

134 Chapter 6 Differentiation and Integration

62. Show that a continuous function on (a, b) is convex if and only if

63. A function on a general interval I is said to be convex provided it is continuous on I and (38)
holds for all Xl, X2 E I. Is a convex function on a closed, bounded interval [a, b] necessarily
Lipschitz on [a, b]?
64. Let Ip have a second derivative at each point in (a, b). Show that Ip is convex if and only if Ip"
is nonn~ive.
°
65. Suppose a :::: and b > 0. Show that the function Ip( t) = (a + bt)P is convex on [0, 00) for
1 S p < 00.
66. For what functions Ip is Jensen's Inequality always an equality?
67. State and prove a version of Jensen's Inequality on a general closed, bounded interval [a, b].
68. Let f be integrable over [0, 1]. Show that

exp[[ f(X)dX] S [eXP(f(X»dX

69. Let {an} be a sequence of nonnegative numbers whose sum is 1 and Un} a sequence of positive
numbers. Show that
00 00

IT ~. S ~ an Cn·
n=l n=l

70. Let g be a positive measurable function on [0, 1]. Show that log (J~ g( X ) dx) :::: J~ log (g( x ) ) dx
whenever each side is defined.
71. (Nemytskii) Let Ip be a continuous function on R. Show that if there are constants for which
(45) holds, then Ip 0 f is integrable over [0, 1] whenever f is. Then show that if Ip 0 f is
integrable over [0, 1] whenever f is, then there are constants C1 and C2 for which (45) holds.