Problem Set

Problem Set #2

Math 5322, Fall 2001

March 4, 2002

ANSWERS
i
 All of the problems are from Chapter 3 of the text.

Problem 1. [Problem 24, p. 100] If f ∈ L1loc and f is continuous at x, then

x is in the Lebesgue set of f .
Answer :
To show that x is in the Lebesgue set of f , we must show
Z
1
(∗) lim |f (y) − f (x)| dy = 0.
r→0 m(B(r, x)) B(r,x)

Let ε > 0 be given. Since f is continuous at x, there is a δ > 0 such that

|f (y) − f (x)| < ε for all y ∈ B(δ, x).
Hence, if 0 < r < δ, we have |f (y) − f (x)| < ε for all y ∈ B(r, x), so
Z Z
1 1
|f (y) − f (x)| dy ≤ ε dy = ε.
m(B(r, x)) B(r,x) m(B(r, x)) B(r,x)

Since ε was arbitrary, this shows that (∗) holds.

Problem 2. [Problem 28, p. 107] If F ∈ NBV, let G(x) = |µF |((−∞, x]).
Prove that |µF | = µTF by showing G = TF via the following
steps.
a. From the definition of TF , TF ≤ G.
b. |µF (E)| ≤ µTF (E) when E is an interval, and hence when E is a Borel set.
c. |µF | ≤ µTF and hence G ≤ TF . (Use Exercise 21.)
Answer :
The definition of TF is at the top of page 102 in the text. If we have points
−∞ < x0 < x1 < · · · < xn = x, then
n
X ∞
X
|F (xj ) − F (xj−1 )| = |µF ((xj−1 , xj ])|
j=1 j=1
Xn
≤ |µF |((xj−1 , xj ])
j=1

= |µF |((x0 , x])

≤ |µF |((−∞, x])
= G(x).

Taking the sup over all choices of the partition { xj }, we get TF (x) ≤ G(x).

1
 For an h-interval (a, b], we have

|µF ((a, b])| = |F (b) − F (a)|

≤ TF (b) − TF (a) by equation (3.24), p. 102
= µTF ((a, b])

Hence, we have the inequality

(∗) |µF (E)| ≤ µTF (E)

when E is an h-interval. We want to show this holds when E is a Borel set. It

seems to me that some work is required—the uniqueness theorem Theorem 1.14
(p. 31) doesn’t seem to help, since we are not proving the equality of two mea-
sures. S
A simple solution is to proceed as follows. If E = n In is the countable
union of a family of disjoint h-intervals then we have
∞ 
X 
|µF (E)| =  µF (In )
n=1
∞
X
≤ |µF (In )|
n=1
X∞
≤ µTF (In )
n=1
= µTF (E).

Hence (∗) holds for E ∈ A, where A is the algebra of all finite disjoint unions
of h-intervals. The σ-algebra generated by A is the Borel sets, but it doesn’t
seem too easy to show that the collection of Borel sets that satisfy (∗) is a
σ-algebra (the problem being to show it’s closed under taking complements).
We can get around this by using the Monotone Class Lemma (page 68). Let
C be the collection of Borel sets such that (∗) holds. We know that A ⊆ C.
We claim that C is a monotone class. To see this, suppose that S we have an
increasing family E1 ⊆ E2 ⊆ E3 ⊆ · · · of elements of C and E = n En . We
want to show that E ∈ C. We have

(∗∗) |µF (En )| ≤ µTF (En ),

for each n. Since continuity from below works for complex measures, we have

lim µF (En ) = µF (E)

n→∞
lim µTF (En ) = µTF (E),
n→∞

so letting n go to infinity in (∗∗) yields |µF (E)| ≤ µTF (E). Thus E ∈ C. A

similar argument shows that C is closed under countable decreasing intersections.
Thus, C is a monotone class.

2
 It M is the smallest monotone class containing A, we must have A ⊆ M ⊆
C ⊆ BR . But, by the Monotone Class Lemma, the monotone class M generated
by A is the same as the σ-algebra generated by A, which is BR . Thus, C = BR ,
so (∗) holds for all Borel sets.
We’re now ready for the final step. By Exercise 21 (p. 94), if E is a Borel
set, then
( ∞ ∞
)
X [
|µF |(E) = sup |µF (En )| | E = En , En ’s disjoint .
n=1 n=1

So, suppose that E is the union of a disjoint family { En }. We then have

∞
X ∞
X
|µF (En )| ≤ µTF (En )
n=1 n=1
= µTF (E).
Taking the sup over all Borel partitions of E, we get
|µF |(E) ≤ µTF (E).
Since this holds for all Borel sets, we can set E = (−∞, x], and then
G(x) = |µF |((−∞, x]) ≤ µTF ((−∞, x]) = TF (x) − TF (−∞) = TF (x).
Thus, we have G ≤ TF . Combined with our earlier result, this shows that
G = TF , and hence that |µF | = µG = µTF .

Problem 3. [Problem 30, p. 107] Construct an increasing function on R

whose set of discontinuities is Q.
Answer :
∞
Lets start with a general construction. Let A = { an }n=1 be a countable set
∞
of distinct points in R and suppose that { αn }n=1 is a sequence of nonnegative
numbers such that
X∞
αn < ∞.
n=1
Define a function gn by
(
0, x < an
gn (x) =
αn , an ≤ x.

Then we can define a function g by

∞
X
g(x) = gn (x),
n=1

where the series converges for each x.

3
Claim. The function g is nondecreasing and is continuous at all points x ∈
/ A.
Since each gn is nondecreasing, it should be clear that g is nondecreasing.
Suppose that p ∈/ A. We want to show that g is continuous at p. Let ε > 0 be
given. There is some N ∈ N such that
∞
X
αn < ε.
n=N +1

Consider the finite list of points a1 , a2 , . . . , aN . Since p is not in this list, we can
find a δ > 0 so that none of the points in the list is in the interval (p − δ, p + δ).
Suppose that p < x < p + δ. Then g(x) − g(p) is the sum of the αn ’s for which
the corresponding point an is in the interval (p, x]. None of the points an for
n ≤ N is in this interval, so
X ∞
X
g(x) − g(p) = αn ≤ αn < ε.
an ∈(p,x] n=N +1

Similarly, if p − δ < x < p, then g(p) − g(x) is the sum of the αn ’s such that
an is in (x, p]. By the same reasoning as above g(p) − g(x) < ε. This completes
the proof of the claim.
∞
We can now use this idea to complete the problem. Let { rn }n=1 be an
enumeration of the rationals. Define functions fn by
(
0, x < rn
fn (x) = 1
2n , rn ≤ x.

and define f by
∞
X
f (x) = fn (x).
n=1

By our previous work, f is nondecreasing. In fact f is strictly increasing since

if x < y, there is some rational rn so that x < rn < y. But then f (y) ≥
f (x) + 1/2n > f (x).
By our work above, f is continuous at the irrational points. It remains to
prove that f is discontinuous at the rationals. To see this, let rk be a rational
point. Define a function h by
X
h(x) = fn (x).
1≤n<∞
n6=k

By our claim above, h is continuous at rk and we have f (x) = fk (x) + h(x).

Then, f (rk +) = fk (rk +)+h(rk +) = 1/2k +h(rk ). Similarly, f (rk −) = 0+h(rk ).
Thus, f (rk +) − f (rk −) = (1/2k + h(rk )) − h(rk ) = 1/2k , so f has a jump
discontinuity at rk .
We’ve now shown that the set of discontinuities of f is exactly Q, and the
solution is complete.

4
Problem 4. [Problem 34, p. 108] Suppose that F, G ∈ NBV and −∞ <
a < b < ∞.
a. By adapting the proof of Theorem 3.36, show that
Z Z
F (x) + F (x−) G(x) + G(x−)
(∗) dG(x) + dF (x)
[a,b] 2 [a,b] 2
= F (b)G(b) − F (a−)G(a−).

b. If there are no points in [a, b] where both F and G are discontinuous, then
Z Z
(∗∗) F dG + G dF = F (b)G(b) − F (a−)G(a−).
[a,b] [a,b]

Answer :
We will use Exercise 1.28, page 39 (which was assigned last semester). In
particular, µH ([r, s]) = H(s) − H(r−).
Define Ω ⊆ [a, b] × [a, b] by

Ω = { (x, y) | a ≤ x ≤ y ≤ b } .

We calculate (µF × µG )(Ω) two ways, by the Fubini-Tonelli Theorem. First we

have
Z
(µF × µG )(Ω) = µF ([a, y]) dµG (y)
[a,b]
Z
(A) = [F (y) − F (a−)] dG(y)
[a,b]
Z
= F (y) dG(y) − F (a−)[G(b) − G(a−)].
[a,b]

Similarily, we have
Z
(µF × µG )(Ω) = µG ([x, b]) dµF (x)
[a,b]
Z
(B) = [G(b) − G(x−)] dF (x)
[a,b]
Z
= G(b)[F (b) − F (a−)] − G(x−) dF (x).
[a,b]

Change that name of the variable of integration in (A) to x and equate the
right-hand sides of (A) and (B). After a little algebra, this yields,
Z Z
(C) F (x) dG(x) + G(x−) dF (x) = F (b)G(b) − F (a−)G(a−).
[a,b] [a,b]

5
If we repeat this argument exchanging the roles of F and G (i.e., compute
(µG × µF )(Ω)), we will get
Z Z
(D) F (x−) dG(x) + G(x) dF (x) = F (b)G(b) − F (a−)G(a−).
[a,b] [a,b]

If we add the equations (C) and (D) and divide by 2, we get (∗).
Now, for the second part of the problem, assume that there are no points
where F and G are both discontinuous. If we let DF be the set of points of
disconinuity for F , and DG the set of points of discontinuoity of G, we have
DF ∩ DG = ∅.
In general, we have µF ({ x }) = F (x) − F (x−). Thus, µF ({ x }) = 0 if F is
continuous at x. Since F is continuous at each point of DG , each point in DG
has µF -measure zero. Since DG is at most countable (because G ∈ NBV), we
have µF (DG ) = 0. We have

G(x) + G(x−) G(x) + G(x)

= = G(x)
2 2
if x ∈
/ DG . Thus, we have

G(x) + G(x−)
= G(x), for µF -almost all x,
2
and so
Z Z
G(x) + G(x−)
(E) dF (x) = G(x) dF (x).
[a,b] 2 [a,b]

The same argument with the roles of F and G reveresed gives

Z Z
F (x) + F (x−)
(F) dG(x) = F (x) dG(x).
[a,b] 2 [a,b]

Subsituting (E) and (F) in (∗) gives us (∗∗).

Problem 5. [Problem 35, p. 108] If F and G are absolutely continuous on

[a, b], then so is F G and
Z b
(F 0 G + F G0 )(x) dx = F (b)G(b) − F (a)G(b).
a

Answer :
An absolutely continuous function is continuous and [a, b] is compact, so F and
G are bounded on [a, b]. Thus, we can find some constant M so that |F | ≤ M
and |G| ≤ M on [a, b]. Let H = F G.

6
 Let ε > 0 be given. Since F is absolutely continuous, there is some δ1 > 0
n
such that if { (aj , bj ) }j=1 is a collection of disjoint intervals in [a, b],
n
X n
X
(bj − aj ) < δ1 =⇒ |F (bj ) − F (aj )| < ε.
j=1 j=1

Similarly, there is a δ2 > 0 such that

n
X n
X
(bj − aj ) < δ2 =⇒ |G(bj ) − G(aj )| < ε.
j=1 j=1
P
Let δ = min(δ1 , δ2 ) and suppose that j (bj − aj ) < δ. Then we have
n
X n
X
|H(bj ) − H(aj )| = |F (bj )G(bj ) − F (aj )G(aj )|
j=1 j=1
Xn
= |F (bj )G(bj ) − F (bj )G(aj ) + F (bj )G(aj ) − F (aj )G(aj )|
j=1
Xn n
X
≤ |F (bj )G(bj ) − F (bj )G(aj )| + |F (bj )G(aj ) − F (aj )G(aj )|
j=1 j=1
Xn n
X
= |F (bj )||G(bj ) − G(aj )| + |G(aj )||F (bj ) − F (aj )|
j=1 j=1
n
X n
X
≤M |G(bj ) − G(aj )| + M |F (bj ) − F (aj )|
j=1 j=1

< M ε + M ε = 2M ε.

Since ε was arbitrary, we conclude that H is absolutely continuous.

Thus, H is differentiable almost everywhere. Except for an exceptional set
of measure zero, the functions H, F and G are all differentiable. At a point x
where they are all differentiable, we have H 0 (x) = F 0 (x)G(x) + F (x)G0 (x) by
the ordinary product rule. Thus, H 0 = F 0 G + F G0 a.e. From the Fundamental
Theorem of Calculus for Lebesgue Integrals, we have
Z b Z b
H(b) − H(a) = H 0 (x) dx = (F 0 G + F G0 )(x) dx,
a a
which completes the solution of the problem.

Problem 6. [Problem 37, p. 108] Suppose F : R → C. There is a constant

M such that |F (x) − F (y)| ≤ M |x − y| for all x, y ∈ R (that is,
F is Lipschitz continuous) iff F is absolutely continuous and |F 0 | ≤ M a.e.

7
Answer :
Suppose first that F is Lipschitz, with |F (x) − F (y)| ≤ M |x − y|. To prove
that F is absolutely continuous, let ε > 0 be given. Choose δ > 0 so small that
n
M δ < ε. Suppose that { (aj , bj ) }j=1 is a finite collection of disjoint intervals so
that
X n
(bj − aj ) < δ.
j=1

Then we have
n
X ∞
X
|F (bj ) − F (aj )| ≤ M |bj − aj |
j=1 j=1
n
X
=M (bj − aj )
j=1

< M δ < ε.

Thus, F is absolutely continuous. Let p be a point where F is differentiable.

Then
F (x) − F (p)
F 0 (p) = lim ,
x→p x−p
but  
 F (x) − F (p)  M |x − p|
 ≤ = M,
 x−p  |x − p|
so |F 0 (p)| ≤ M .
For the second part of the proof, suppose that F is absolutely continuous and
|F 0 | ≤ M a.e. Suppose that x < y. By the Fundamental Theorem of Calculus
for Lebesgue integrals,
Z y
F (y) − F (x) = F 0 dm,
x

and so
Z y 
F 0 dm
 
|F (x) − F (y)| = 
Z xy
≤ |F 0 | dm
Zxy
≤ M dm
x
= M (y − x)
= M |x − y|,

so F is Lipschitz.

8
Problem 7. [Problem 42, p. 109] A function F : (a, b) → R (−∞ ≤ a <
b ≤ ∞) is called convex if

(∗) F (λs + (1 − λ)t) ≤ λF (s) + (1 − λ)F (t)

for all s, t ∈ (a, b) and λ ∈ (0, 1) Actually, you can say all λ ∈ [0, 1]]. . . .
a. F is convex iff for all s, t, s0 , t0 ∈ (a, b) such that s ≤ s0 < t0 and s < t ≤ t0 ,
F (t) − F (s) F (t0 ) − F (s0 )
(∗∗) ≤ .
t−s t0 − s0

b. F is convex iff F is absolutely continuous on every compact subinterval of

(a, b) and F 0 is increasing (on the set where it is defined).
c. If F is convex and t0 ∈ (a, b), there exists β ∈ R such that F (t) − F (t0 ) ≥
β(t − t0 ) for all t ∈ (a, b).
d. (Jensen’s Inequality) If (X, M, µ) is a measure space with µ(X) = 1,
g : X → (a, b) is in L1 (µ) and F is convex on (a, b), then
Z  Z
F g dµ ≤ F ◦ g dµ.

Answer :
We begin with the first part of the problem by supposing that F is convex. We
are given s, t, s0 , t0 with s ≤ s0 < t0 and s < t ≤ t0 . In other words, s0 , t ∈ [s, t0 ]
and no ordering between s0 and t is specified, except that they can’t be equal
to the endpoint with the other letter. To derive (∗∗), we proceed as follows.
Since t ∈ [s, t0 ] we can find a λ1 so that

t = (1 − λ1 )s + λ2 t0 ,

since s 6= t, we have λ1 ∈ (0, 1]. Solving for λ1 gives

t−s
λ1 = .
t0 − s
Now, by (∗), we have

F (t) = F ((1 − λ1 )s + λ1 t0 )
≤ (1 − λ1 )F (s) + λ1 F (t0 )
= F (s) − λ1 F (s) + λ1 F (t0 ),

which yields
F (t) − F (s) ≤ λ1 [F (t0 ) − F (s)].
Plugging in the value of λ1 yields
t−s
F (t) − F (s) = [F (t0 ) − F (s)].
t0 − s

9
Dividing both sides by the postive number t − s, we get
F (t) − F (s) F (t0 ) − F (s)
(A) ≤
t−s t0 − s
Similarly, we can write
s0 = λ2 s + (1 − λ2 )t0
for some λ2 ∈ (0, 1]. A little algebra yields
s0 − t0
λ2 = .
s − t0
Thus, we have

F (s0 ) = F (λ2 s + (1 − λ2 )t0 )

≤ λ2 F (s) + (1 − λ2 )F (t0 )
= λ2 F (s) + F (t0 ) − λ2 F (t0 ),

and so
F (s0 ) − F (t0 ) ≤ λ2 [F (s) − F (t0 )].
Putting in the value of λ2 gives
s0 − t0
F (s0 ) − F (t0 ) ≤ [F (s) − F (t0 )].
s−t
Dividing both sides of this inequality by the negative number s0 − t0 gives
F (s0 ) − F (t0 ) F (s) − F (t0 )
≥ .
s0 − t0 s − t0
Using the symmetry of the difference quotients, we can rewrite this as
F (t0 ) − F (s) F (t0 ) − F (s0 )
(B) ≤ .
t0 − s t0 − s0
Combining (A) and (B) yields (∗∗).
To see the geometric significance of the inequalities set
F (t) − F (s)
m1 =
t−s
F (t0 ) − F (s)
m2 =
t0 − s
F (t ) − F (s0 )
0
m3 = ,
t0 − s0
so we proven that m1 ≤ m2 ≤ m3 , and consider Figure 1
Conversely, suppose that (∗∗) holds. Let s < t0 be two points in (a, b) and
let λ ∈ (0, 1) be given. Apply (∗∗) with

t = s0 = λs + (1 − λ)t0 .

10

 

 
 
y

 

 
 

Figure 1: Slope inequalities for a convex function

11
A little algebra yields

t − s = (1 − λ)(t0 − s)
t0 − s0 = λ(t0 − s),

and so (∗∗) becomes

F (t) − F (s) F (t0 ) − F (t)
≤ .
(1 − λ)(t0 − s) λ(t0 − s)
Since λ, 1 − λ and t0 − s are positive, cross multiplication gives

λ[F (t) − F (s)] ≤ (1 − λ)[F (t0 ) − F (t)]

which simplifies easily to

F (t) ≤ λF (s) + (1 − λ)F (t0 ).

Thus, F is convex on (a, b).

Next, we do the second part of the problem. First, suppose that F is convex
on (a, b); we want to show that F is absolutely continuous on every compact
subinterval of (a, b) and F 0 is increasing.
Let [q, r] be a compact subinterval of (a, b). We can choose points p and s
with a < p < q < r < s < b. Suppose that x, y ∈ [q, r] and x < y. With two
applications of (∗∗) (First: s = p, t = q, s0 = x, t0 = y; Second: s = x, t = y,
s0 = r, t0 = s) we get
F (q) − F (p) F (y) − F (x) F (s) − F (r)
≤ ≤ .
q−p y−x s−r
Since the two outer quotients don’t involve x and y, we conclude that there are
constants m1 and m2 so that
F (y) − F (x)
m1 ≤ ≤ m2
y−x
for all x, y ∈ [q, r]. If we set M = max(|m1 |, |m2 |) then we have
 
 F (y) − F (x) 
 ≤M
 y−x 

for all x, y ∈ [q, r]. Thus, we have

|F (y) − F (x)| ≤ M |x − y| ∀x, y ∈ [q, r].

Thus, F is Lipschitz on [q, r] and so absolutely continuous. Next, we want to

show that the derivatives is increasing. To see this, let s < t0 be two points at
which the derivative exits. From (∗∗) we have
F (t) − F (s) F (t0 ) − F (s0 )
≤
t−s t0 − s0

12
where s ≤ s0 < t0 and s < t ≤ t0 . If we hold s0 fixed and take the limit as t ↓ s,
we get
F (t0 ) − F (s0 )
F 0 (s) ≤
t0 − s0
and taking the limit of this as s0 ↑ t0 gives F 0 (s) ≤ F 0 (t0 ). Thus, F 0 is increasing.
Next, we want to do the converse. Thus, we assume that F is absolutely
continuous on every compact subinterval of (a, b) and that F 0 is increasing (on
the set where it is defined).
We need the following Lemma.
Lemma. Let G : [c, d] → C be absolutely continuous. Let ϕ : R → R : x 7→
αx + β, where α > 0. Suppose that ϕ maps [a, b] onto [c, d]. Then G ◦ ϕ is
absolutely continuous on [a, b] and (G ◦ ϕ)0 (x) = αG0 (αx + β) a.e.
Proof of Lemma. We want to prove that G ◦ ϕ is absolutely continuous. Let
n
ε > 0 be given. Then there is a δ > 0 so that if { (cj , dj ) }j=1 is a collection of
intervals in [c, d], then
n
X n
X
(C) (dj − cj ) < δ =⇒ |G(dj ) − G(cj )| < ε.
j=1 j=1

n
Choose η > 0 so small that αη < δ. Let { (aj , bj ) }j=1 be a collection of
disjoint intervals in [a, b]. The map ϕ maps the interval [aj , bj ] onto the interval
[ϕ(aj ), ϕ(bj )] which has length α(bj − aj ). Suppose that
n
X
(bj − aj ) < η
j=1

Then { (ϕ(aj ), ϕ(bj )) } is a collection of disjoint intervals in [c, d] and

n
X n
X
(ϕ(bj ) − ϕ(aj )) = α (bj − aj ) < αη < δ.
j=1 j=1

Thus, from (C), we have

n
X
|G(ϕ(bj )) − G(ϕ(aj ))| < ε.
j=1

This shows that G ◦ ϕ is absolutely continuous. The formula for the derivatives
follows from the usual chain rule at points x such that G is differentiable at
ϕ(x).
To apply this Lemma to our situation, suppose that x < y, fix λ ∈ (0, 1) and
let p = (1 − λ)x + λy. Consider the function g(t) = F (x + tλ(y − x)). By our

13
Lemma, this is absolutely continuous, and g 0 (t) = F 0 (x + tλ(y − x))λ(y − x).
By the Fundamental Theorem of Calculus for Lebesgue integrals, we have
Z 1
g(1) − g(0) = g 0 (t) dt,
0

which we can write as

Z 1
F (p) − F (x) = F 0 (x + tλ(y − x))λ(y − x) dt.
0

On the other hand, consider h(t) = F (x+t(y−x)) and apply the same reasoning.
This gives Z 1
F (y) − F (x) = F 0 (x + t(y − x))(y − x) dt.
0

Now, for t ∈ [0, 1], we have x + tλ(y − x) ≤ x + t(y − x), since 0 < λ < 1. Since
F 0 is increasing, we have F 0 (x + tλ(y − x)) ≤ F 0 (x + t(y − x)) for almost all t.
Thus,
Z 1
F (p) − F (x) = λ F 0 (x + tλ(y − x))(y − x) dt
0
Z 1
≤λ F 0 (x + t(y − x))(y − x) dt
0
= λ[F (y) − F (x)].

Now we have
F (p) − F (x) ≤ λ[F (y) − F (x)],
which is easily rearranged to give

F (p) ≤ (1 − λ)F (x) + λF (y).

This shows that (∗) holds, so F is convex.

Next, consider the third part of the problem. Thus, we assume that F is
convex on (a, b). Let t0 ∈ (a, b). Let s and t0 be points such that s < t0 < t0
(s, t0 ∈ (a, b)). Apply (∗∗) with s0 = t = t0 . This gives

F (t0 ) − F (s) F (t0 ) − F (t0 )

(D) ≤ .
t0 − s t0 − t0
The set  
F (t0 ) − F (s)
| a < s < t0
t0 − s
is bounded above by the right-hand side of (D) for any t0 > t0 . Thus, if we
define  
F (t0 ) − F (s)
β1 = sup | a < s < t0
t0 − s

14
We have
F (t0 ) − F (t0 )
β1 ≤
t0 − t 0
for all t0 > t0 . Thus, if we set

F (t0 ) − F (t0 )
 
0
β2 = inf | t 0 < t < b ,
t0 − t0

we have β1 ≤ β2 . Choose β so that β1 ≤ β ≤ β2 .

If t > t0 , then the quotient

F (t) − F (t0 )
(E)
t − t0
is an element of the set in the definition of β2 , so we have

F (t) − F (t0 )
β≤ ,
t − t0
which implies
β(t − t0 ) ≤ F (t) − F (t0 ).
On the other hand, if t < t0 , then the quotient (E) is an element of the set
defining β1 , so we have
F (t) − F (t0 )
β≥ .
t − t0
which gives
β(t − t0 ) ≤ F (t) − F (t0 ),
since t − t0 < 0. This completes the solution of the third part of the problem.
Finally, we prove Jensen’s inequality. Thus we assume that F is convex on
(a, b), g : X → (a, b) is in L1 (µ) and µ(X) = 1. Since a < g(x) < b for all x, we
can integrate this to get
Z Z Z
a = aµ(X) = a dµ ≤ g dµ ≤ b dµ = bµ(X) = b.
R
Thus, if we set t0 = g dµ, we have t0 ∈ (a, b). From the previous part of the
problem, there is some β so that

F (t) − F (t0 ) ≥ β(t − t0 ).

Since g(x) ∈ (a, b), we can set t = g(x) in this inequality to get

F (g(x)) − F (t0 ) ≥ β(g(x) − t0 ).

Integrate this inequality, using µ(X) = 1. This gives

Z Z 
F (g(x)) dµ(x) − F (t0 ) ≥ β g(x) dµ(x) − t0 = 0,

15
 R
since t0 = g dµ. Thus, we have
Z  Z
F g dµ = F (t0 ) ≤ F (g(x)) dµ(x),

R we’re done. Well, almost done. Why is F ◦ g

which is Jensen’s inequality and
measurable? Does the integral F ◦ g dµ make sense?

Remark. Some people tried to do Exercise 37 and Exercise 42 using the Mean
Value Theorem. This would be exactly the right thing to do if our functions
were differentiable. Unfortunately, the Mean Value Theorem does not work for
absolutely continuous functions that are only differentiable almost everywhere.
For a counterexample, consider F : R → R : x 7→ |x|. By the triangle inequality
we have
|F (x) − F (y)| = ||x| − |y|| ≤ |x − y|
so F is Lipschitz, and hence absolutely continuous. It’s differentiable except at
0, and hence differentiable almost everywhere, but not everywhere. Consider
the cord of the graph that connects the two points (−1, 1) and (1, 1). The slope
of this cord is zero, but there is no point on the graph where the derivative is
zero (or even anywhere close to zero). Thus, it is not true that

F (1) − F (−1) = F 0 (c)(1 − (−1))

for some point c in (−1, 1). The workaround for the lack of the Mean Value
Theorem is to use the Fundamental Theorem of Calculus for Lebesgue Integrals
instead.

16