Bounded Derivatives Which Are Not Riemann Integrable

by

Elliot M. Granath

A thesis submitted in partial fulfillment of the requirements

for graduation with Honors in Mathematics.

Whitman College
2017
 Certificate of Approval

This is to certify that the accompanying thesis by Elliot M. Granath has

been accepted in partial fulfillment of the requirements for
graduation with Honors in Mathematics.

Russell A. Gordon

Whitman College
May 10, 2017
Contents

Acknowledgements iv

Abstract v

1 Finding an antiderivative of a continuous function without

integration 1
1.1 Lebesgue’s method of antidifferentiation . . . . . . . . . . . . 1
1.2 A version of the Fundamental Theorem of Calculus . . . . . . 9
1.3 Use of the Mean Value Theorem . . . . . . . . . . . . . . . . . 11

2 Bounded derivatives which are not Riemann integrable 13

2.1 Volterra’s function (an overview) . . . . . . . . . . . . . . . . 15
2.2 Pompeiu’s function . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Beginning our construction (defining f ) . . . . . . . . . 19
2.2.2 Differentiating our function (determining f 0 ) . . . . . . 20
2.2.3 Finding Pompeiu’s function (exploring h = f −1 ) . . . . 26
2.2.4 Further notes on Pompeiu’s derivative (h0 ) . . . . . . . 28

3 Absolute continuity and Baire class one 33

3.1 Absolute continuity and bounded derivatives . . . . . . . . . . 36
3.2 Some properties of Baire class one functions . . . . . . . . . . 38

4 The Henstock Integral 46

4.1 A worked example using the Henstock integral . . . . . . . . . 51
4.2 Further properties of the Henstock integral . . . . . . . . . . . 53

iii
Acknowledgements

Thank you to our teacher, Professor Schueller, to my advisor, Professor Gor-

don, and to my emotional support cats, James and Pancake.

iv
Abstract

In an elementary calculus course, we talk mostly, or exclusively, about in-

tegrating continuous, real-valued functions. Since continuous functions on
closed intervals are integrable, the Fundamental Theorem of Calculus gives us
a method to calculate these integrals (given that we can find an antiderivative).
Furthermore, the Fundamental Theorem of Calculus states that the integral
can be used to define an antiderivative of a continuous function. In this pa-
per, we will present a method for establishing the existence of antiderivatives
of continuous functions without using any integration theory. In addition,
we will explore the potentially counter-intuitive topic of derivatives which are
not Riemann integrable. It is easy to find a function whose derivative is un-
bounded, and thus not Riemann integrable; what is more surprising is that
even bounded derivatives are not necessarily Riemann integrable. We will
present two classes of functions, one conceived by Volterra and one by Pom-
peiu, which are differentiable on closed intervals, and whose derivatives are
not Riemann integrable. Finally, we will develop the Henstock integral as a
tool which integrates all derivatives.

v
1 Finding an antiderivative of a continuous

function without integration

When first learning about differentiation and integration, it is tempting to

think of “antiderivative” and “integral” as synonymous. In some sense, differ-
entiation “undoes” integration, and vice versa. Of course, the Fundamental
Theorem of Calculus canonizes this relationship: if f : [a, b] → R is integrable
and has an antiderivative F , then we can calculate its integral as F (b) − F (a).
Rx
Moreover, the Fundamental Theorems states that, if f is continuous, then a f
2
is an antiderivative of f . This shows that functions such as e−x indeed have
antiderivatives, even if they are impossible to write using elementary functions.

Given the intimate relationship between integrals and derivatives, it is

perhaps interesting that we can establish the existence of antiderivatives of
continuous functions without relying on any integration theory. Lebesgue gave
a proof of this fact [2]; it is recorded below in a modern format.

1.1 Lebesgue’s method of antidifferentiation

We prove that any continuous function f : [a, b] → R has an antiderivative

by estimating f using continuous piecewise linear functions. With some care,
we can find an antiderivative of each of these functions without invoking the
Fundamental Theorem of Calculus. Finally, we can show that these piece-
wise estimations, together with their antiderivatives, can be used to construct
an antiderivative of f . First, it requires proof that continuous piecewise lin-
ear functions have antiderivatives. Figure 1 gives us an intuition for how to
construct such an antiderivative.
First, we begin with some useful definitions.

1
 F (x)

f (x)

Figure 1: A piecewise linear, continuous function

together with an antiderivative we can find with-
out using the Fundamental Theorem of Calculus.
Note that F is still differentiable at the points
where f has a cusp.

Definition 1.1. A partition of [a, b] is a set P = {x0 , x1 , . . . , xn } such that

xi < xi+1 , a = x0 , and b = xn . The norm of a partition is denoted ||P ||, where
||P || = max{xi − xi−1 : 1 ≤ i ≤ n}. 

Definition 1.2. A function f : [a, b] → R is piecewise linear if there is some

partition {x0 , . . . , xn } of [a, b] such that, for any 1 ≤ i ≤ n, f is of the form
ai x + bi for all x ∈ [xi−1 , xi ], where ai , bi are real numbers. 

Theorem 1.3. If f : [a, b] → R is continuous and piecewise linear on [a, b],

then there exists a function F : [a, b] → R such that F 0 (x) = f (x) for all
x ∈ [a, b] (one-sided limits are used at the endpoints).

Proof. We proceed by induction on the number of linear segments of f . If

f has one linear segment, i.e. if f (x) = c1 x + c2 for x ∈ [a, b], then the
function F (x) = c1 x2 /2 + c2 x is an antiderivative for f on [a, b]. Our induction
hypothesis is that a continuous, piecewise linear function defined on a closed
interval with k linear segments has an antiderivative on that closed interval.

2
Consider the function


g(x),
 if a ≤ x < c;
f (x) =

d1 x + d2 , if c ≤ x ≤ b,


where d1 , d2 are constants, g(c) = d1 c + d2 , and g is a continuous, piecewise

linear function with k linear segments on [a, c]. By the induction hypothesis,
g has an antiderivative G on [a, c]. We claim that the function


G(x),
 if a ≤ x < c,
F (x) =
 1 d1 x2 + d2 x − 1
d c2



2 2 1
+ d2 c + G(c), if c ≤ x < b,

is an antiderivative for f on [a, b]. This is clearly true on [a, b] \ {c}, so it

remains to show that F 0 (c) = f (c). Note that

1
F (c + h) − F (c) d1 (c + h)2 + d2 (c + h) − 21 d1 c2 − d2 c
lim+ = lim+ 2
h→0 h h→0 h
1 2
d1 ch + 2 d1 h + d2 h
= lim+
h→0 h
= d1 c + d2 ,

and on the other hand

F (c + h) − F (c) G(c + h) − G(c)

lim− = lim−
h→0 h h→0 h
= G0 (c)

= d1 c + d2 .

Thus F is indeed an antiderivative for f on [a, b]. By induction, any con-

tinuous piecewise linear function on a closed interval has an antiderivative.

3
 Having established that continuous piecewise linear functions have an-
tiderivatives, we claim that any continuous function on a closed interval can
be estimated with a sequence of such functions. Figure 2 shows a simple con-
tinuous function, together with a linear approximation in six parts. As our
intuition suggests, we can choose increasingly fine linear approximations that
will converge to the desired function. Using the Mean Value Theorem, we
can show that the corresponding sequence of antiderivatives converges to an
antiderivative of f .

f (x)

Figure 2: We may approximate any continuous

function using a piecewise linear continuous func-
tion.

Theorem 1.4. If f : [a, b] → R is continuous on [a, b], then there exists a se-
quence {fn } of continuous piecewise linear functions that converges uniformly
to f on [a, b].

Proof. Since f is continuous on a closed and bounded interval, f is also uni-

formly continuous on [a, b]. Thus, for each positive integer n, there exists
δn > 0 such that if |x − y| < δn , then |f (x) − f (y)| < 1/n for all x, y ∈ [a, b].
Without loss of generality, {δn } is decreasing. Let Pn = {x0 , . . . , xkn } be
a partition of [a, b] with norm less than δn and, for each n, define fn as a
piecewise linear function joining the points (xi , f (xi )) where 0 ≤ i ≤ kn .

4
We claim that {fn } converges uniformly to f on [a, b]. Let  > 0 and pick
N such that 1/N < /2. Let x ∈ [a, b], and identify an interval [xi , xi+1 ]
containing x. Then |xi − x| < δn , and without loss of generality, for all
n ≥ N , δn < δN . Also note that, for all 1 ≤ i ≤ n, if x, y ∈ [xi−1 , xi ], then
|f (y)−f (x)| ≤ max{x : x ∈ [xi−1 , xi ]}−min{x ∈ [xi−1 , xi ]} = |f (xi )−f (xi−1 )|.
Using this fact together with the uniform continuity of f ,

|fn (x) − f (x)| ≤ |fn (x) − fn (xi )| + |fn (xi ) − f (x)|

= |fn (x) − fn (xi )| + |f (xi ) − f (x)|

≤ |fn (xi+1 ) − fn (xi )| + |f (xi ) − f (x)|

= |f (xi+1 ) − f (xi )| + |f (xi ) − f (x)|

< 1/N + 1/N

< .

It follows that {fn } converges uniformly to f on [a, b].

In general, it is not necessarily true that lim lim f (x, y) = lim lim f (x, y).
x→x0 y→y0 y→y0 x→x0

For example,
x + 2y x + 2y
1 = lim lim 6= lim lim = 2.
x→0 y→0 x + y y→0 x→0 x + y

Before continuing, we will establish Lemma 1.5, which shows that, in the
right circumstances, we may simply interchange limits. In the same way that
uniform continuity preserves continuity, it also preserves limits.

Lemma 1.5. Let {fn } be a sequence of functions that converges uniformly to

f on [a, b] \ {c} where c ∈ [a, b]. Then, provided each function fn and f has a
limit at c,
lim lim fn (x) = lim lim fn (x).
x→c n→∞ n→∞ x→c

5
Proof. Denote yn = lim fn (x). Since {fn } converges uniformly, given  > 0,
x→c

there exists N such that if n, m > N , then |fn (x) − fm (x)| <  for all points
x ∈ [a, b] \ {c} (this is the Cauchy Criterion for Uniform Convergence). Then,
since |fn (x) − fm (x)| < , we have |yn − ym | = | lim(fn (x) − fm (x))| ≤ . Thus,
x→c

{yn } is a Cauchy sequence converging to some y ∈ R.

Let  > 0. By the uniform convergence of f , pick N1 such that n ≥ N1

implies |f (x) − fn (x)| < . Pick N2 ≥ N1 such that if n ≥ N2 , we have
|yn − y| < . Since yN2 = lim fN2 (x), pick δ > 0 such that |x − c| < δ implies
x→c

|fN2 (x) − yN2 | < . If |x − c| < δ, we see that

|f (x) − y| ≤ |f (x) − fN2 (x)| + |fN2 (x) − yN2 | + |yN2 − y|

< 3.

Thus lim f (x) = y = lim yn , as desired.

x→c n→∞

Perhaps contrary to our intuition, it is possible to give a sequence of func-

tions {Fn } which converges uniformly to F , and whose derivatives fn converge
pointwise to f , but where F is not an antiderivative of f .

Example 1.6. This example is adapted from Gordon [4]. For each n ∈ Z+ ,
define Fn : [0, 1] → R where

if x ≥ n1 ,

x,

Fn (x) =
 3 nx2 − 1 n3 x4 , if x < n1 .


2 2

It is clear that each Fn is differentiable away from 1/n. At x = 1/n, the

right-sided limit of Fn is 1/n and the left-sided limit is 32 n1 − 12 n1 = n1 .

6
Furthermore, the right-sided derivative is 1 and the left-sided derivative is

3
3n n1 − 2n3 n1


= 1, so Fn is differentiable on [0, 1]. For all n > 1, if
x ∈ [0, 1/n],

   
 3 2  1 3 4
|Fn (x) − x| ≤  nx  +  n x  + |x|
2 2
3 1 2
≤ + +
2n 2n 2n
3
= .
n

Since 3/n converges to 0, it follows that {Fn } converges uniformly to x on

[0, 1]. Furthermore, for each n, Fn0 (0) = 0. If x ∈ (0, 1], then there is some
N such that, if n > N , then Fn0 (x) = x. Thus {Fn0 } converges pointwise to
the function f : [0, 1] → R where f (0) = 0 and f (x) = 1 for all x ∈ (0, 1].
However, F 0 (x) = 1 for all x ∈ [0, 1], and thus F 0 6= f . (Additionally, since f
does not have the intermediate value property, it has no antiderivative). 

It thus requires proof that our sequence of functions which estimate an

antiderivative for a function indeed converges to the desired antiderivative.

Theorem 1.7. Let {Fn } be a sequence of functions that are differentiable

on [a, b], and suppose that {Fn (c)} converges for some c ∈ [a, b]. If {Fn0 }
converges uniformly to f on [a, b], then {Fn } converges uniformly to a function
F on [a, b]. Furthermore, F is differentiable on [a, b] and F 0 (x) = f (x) for all
x ∈ [a, b].

Proof. (Rudin [5]) Let  > 0. Choose N such that if n, m ≥ N , then we have
|Fn (c) − Fm (c)| <  and |Fn0 (x) − Fm0 (x)| <  for all x ∈ [a, b]. Let x, y ∈ [a, b]

7
where x < y. By the Mean Value Theorem, for some d ∈ (x, y),

 
 (Fn − Fm )(x) − (Fn − Fm )(y) 
|(Fn − Fm )(x) − (Fn − Fm )(y)| = |x − y|  
x−y 

(1)

= |x − y||(Fn − Fm )0 (d)|

< |x − y|

≤ (b − a).

Thus, if m, n ≥ N , for any x ∈ [a, b],

|Fn (x) − Fm (x)| ≤ |(Fn − Fm )(x) − (Fn − Fm )(c)| + |Fn (c) − Fm (c)|

< (b − a) + 

= (b − a + 1).

Thus {Fn } converges uniformly to some F on [a, b]. Fix c ∈ [a, b] and define
the functions

Fn (x) − Fn (c) F (x) − F (c)

φn (x) = ; φ(x) = ,
x−c x−c

where x 6= c. Similar to (1), for any n, m ≥ N ,

 
 (Fn − Fm )(x) − (Fn − Fm )(c) 
|φn (x) − φm (x)| = 
 <
x−c 

whenever x 6= c. Since {φn } converges pointwise to φ on [a, b]\{c}, it converges

8
uniformly to φ on [a, b] \ {c}. By Lemma 1.5,

F 0 (c) = lim lim φn (x)

x→c n→∞

= lim lim φn (x)

n→∞ x→c

= lim fn (c)
n→∞

= f (c).

Since c ∈ [a, b] was arbitrary, we have shown that F 0 = f on [a, b].

We now proceed to establish the existence of antiderivatives of continu-

ous functions without the use of the Fundamental Theorem of Calculus or
integration theory.

Theorem 1.8. If f : [a, b] → R is continuous on [a, b], then f has an an-

tiderivative on [a, b].

Proof. Let f be a continuous function on [a, b]. By Theorem 1.4, there exists
a sequence {fn } of continuous piecewise linear functions that converges uni-
formly to f on [a, b]. By Theorem 1.3, there exists a sequence {Fn } of functions
on [a, b] such that Fn0 (x) = fn (x) for all x ∈ [a, b]. For convenience, choose
each Fn such that Fn (a) = 0. Then by Theorem 1.7, {Fn } converges uniformly
to a function F on [a, b], F is differentiable on [a, b], and F 0 (x) = f (x) for all
x ∈ [a, b].

1.2 A version of the Fundamental Theorem of Calculus

Having established that we can find an antiderivative of a continuous function

f : [a, b] → R on [a, b], we can now prove a version of the Fundamental The-
orem of Calculus. A tagged partition of a closed interval [a, b] is a finite set
t
P = {(ti , [xi−1 , xi ]) : 1 ≤ i ≤ n}, where a = x0 , xn = b, and xi−1 < xi for

9
1 ≤ i ≤ n; additionally, ti ∈ [xi−1 , xi ] for 1 ≤ i ≤ n. Denote the norm of a
tagged partition tP by ||tP || = max{xi − xi−1 : 1 ≤ i ≤ n}. Finally, define the
Riemann sum of f with tagged partition tP by

n
X
S(f,tP ) = f (ti )(xi − xi−1 ).
i=1

We will use the following definition of what it means for a function to be

Riemann integrable. A similar definition applies for functions f : I → R where
I is an interval, but we are primarily concerned with closed intervals here.

Definition 1.9. A function f : [a, b] → R is Riemann integrable on [a, b] if

there is a number L satisfying the following: for each  > 0, there is δ > 0
for which, given any tagged partition tP of [a, b] with ||tP || < δ, we have
Rb
|S(f,t P ) − L| < . We denote the Riemann integral of f as a f . 

Theorem 1.10 (Fundamental Theorem of Calculus). If f : [a, b] → R is con-

tinuous on [a, b], then f has an antiderivative F , f is Riemann integrable, and
Rb
a
f = F (b) − F (a).

Proof. By Theorem 1.8, f has an antiderivative F on [a, b]. Let  > 0. Since
f is uniformly continuous, there is δ > 0 such that, if |y − x| < δ, then
|f (y) − f (x)| < . Let the set tP = {([xi−1 , xi ], ti ) : 1 ≤ i ≤ n} be a partition
of [a, b] where ||tP || < δ. By the Mean Value Theorem, for each i ∈ {1, . . . , n},
there is si ∈ (xi−1 , xi ) satisfying F (xi ) − F (xi−1 ) = f (si )(xi − xi−1 ). Then, for
each i ∈ {1, . . . , n}, we have

|f (ti )(xi − xi−1 ) − (F (xi ) − F (xi−1 ))| = |f (ti )(xi − xi−1 ) − f (si )(xi − xi−1 )|

= |(f (ti ) − f (si ))(xi − xi−1 )|

< (xi − xi−1 ).

10
It follows that
 
Xn n
X 
|S(f,tP ) − (F (b) − F (a))| =  f (ti )(xi − xi−1 ) − (F (xi ) − F (xi−1 ))
 
 
i=1 i=1
n
X
≤ |f (ti )(xi − xi−1 ) − (F (xi ) − F (xi−1 ))|
i=1
n
X
< (xi − xi−1 )
i=1

= (b − a).

Rb
Thus, f is Riemann integrable on [a, b] and a
f = F (b) − F (a).

Remark. Although Theorem 1.10 allows us to integrate any continuous func-

tion on a closed interval, it is weaker than the version which states “if f
is Riemann integrable on [a, b] and has an antiderivative F on [a, b], then
Rb
a
f = F (b) − F (a).” The difference may seem negligible, but Theorem 1.10
does not allow us to integrate the classes of functions which are not continuous
on [a, b] but which have an antiderivative on [a, b]. As an example, consider
the function f (x) = x2 sin πx if x ∈ (0, 1] and f (0) = 0. Then, with the details
omitted, f is differentiable on [0, 1], f 0 is not continuous on [0, 1], and f 0 is
R1
Riemann integrable on [0, 1]. In this case we have 0 f 0 = 0.

1.3 Use of the Mean Value Theorem

In the proof of Theorem 1.7, we used the Mean Value Theorem to show that, if
a sequence of functions {fn } uniformly converges to f , then an appropriately
chosen corresponding sequence of antiderivatives {Fn } uniformly converges to
F where F 0 = f . However, we can provide a similar version of Rudin’s proof
without using the Mean Value Theorem. We use instead the following result.

Theorem 1.11. If f is a piecewise linear, continuous function on [a, b] and

11
F is an antiderivative for f on [a, b], then

|F (b) − F (a)| ≤ (b − a) max{|f (x)| : x ∈ [a, b]}.

Proof. We proceed again by induction on the number of linear terms of f .

First, suppose f : [a, b] → R is linear and f (x) = mx + n. Then the maximum
value of |f (x)| on [a, b] is either |ma+n| or |mb+n|. Without loss of generality,
assume |f (x)| ≤ |mb + n| on [a, b]. If F is an antiderivative of f on [a, b], then
F (b) − F (a) = 21 m(b2 − a2 ) + n(b − a). Then

 
1 
|F (b) − F (a)| = (b − a)  m(a + b) + n

2
1
≤ (b − a) · · (|ma + n| + |mb + n|)
2
= (b − a)|mb + n|

as desired. Our induction hypothesis is that, if g is a piecewise linear, con-

tinuous function on a closed interval, say, [a, c], and G has k linear pieces,
then
|G(c) − G(a)| ≤ (c − a) max{|g(x)| : x ∈ [a, c]}

where G is an antiderivative of g.

Now consider a function f : [a, b] → R, where a < c < b, with k + 1 linear

parts defined by 

g(x),
 if a ≤ x ≤ c;
f (x) =

mx + n,
 if c < x ≤ b.

12
Then

|F (b) − F (a)| ≤ |F (b) − F (c)| + |F (c) − F (a)|

≤ (b − c) max{|f (x)| : x ∈ [c, b]} + (c − a) max{|f (x)| : x ∈ [a, c]}

≤ (b − c) max{|f (x)| : x ∈ [a, b]} + (c − a) max{|f (x)| : x ∈ [a, b]}

= (b − a) max{|f (x)| : x ∈ [a, b]}

which concludes the proof.

Although the Mean Value Theorem is often considered an elementary the-

orem to be taught in any calculus course, avoiding the use of it has some
advantages. Although the statement of the Mean Value Theorem is intuitive,
its proof is somewhat roundabout. The use of Rolle’s Theorem as a lemma is
potentially confusing and tedious for some students. Of course, we cannot say
whether the Mean Value Theorem or Theorem 1.11 is more understandable to
students.

2 Bounded derivatives which are not Riemann

integrable

It is tempting to think that the Fundamental Theorem of Calculus proves that

every derivative is Riemann integrable. That is, given a differentiable function
f : [a, b] → R, a calculus student might assume that the integral of f 0 over
[a, b] must be f (b) − f (a). However, this is only true if f 0 is integrable in the
first place. We can give some simple counterexamples. Take f (x) = x1/3 on
the interval (0, 1). Then, since f 0 (x) = 13 x−2/3 is unbounded on (0, 1), it is
R1
not Riemann integrable. In a sense, it is false to write 0 f 0 = f (x)|10 = 1,
although we commonly abuse this notation to refer to the so-called “improper

13
integral.” In this case, we might write

Z 1 Z 1
0
f = lim+ f0 = 1
0 a→0 a

This example uses an open interval, and takes advantage of the fact that
f 0 is unbounded at an endpoint. Suppose we restrict ourselves to functions
which are continuous and differentiable on a closed interval [a, b] (with one-
sided derivatives at a and b). It turns out that the derivative may still be
unbounded. Consider the function g : [−1, 1] → R where

xk sin 1 ,


x
if x 6= 0,
g(x) =

0,
 if x = 0,

for 1 < k < 2. Then, away from the origin, we can differentiate g using
elementary methods. At the origin, we have

hk sin h1
   
k−1 1
lim = lim h sin
h→0 h h→0 h
= 0.

Thus 
kxk−1 sin 1 − xk−2 cos 1 ,


x x
if x 6= 0,
g 0 (x) =

0,
 if x = 0.

On the domain [−1, 1], g 0 is defined. However, while the term kxk−1 sin x1 is
bounded, xk−2 cos x1 is not. Thus the derivative of g is not Riemann integrable,
despite the fact that g is differentiable everywhere.

In both of the examples we just explored, we find a way to ensure that the

14
derivative of a function is unbounded, which shows that the derivative is not
Riemann integrable without further thought. What if we add an additional
requirement that f 0 be bounded? Is it always integrable? The answer is
still no. The most widely known counter-example was constructed in 1881
by Vito Volterra [6]. We will provide an overview of his function, before
focusing rigorously on another example by Dimitrie Pompeiu with certain
other properties.

2.1 Volterra’s function (an overview)

In this section we will present a non-rigorous overview of Volterra’s bounded

derivative which is not Riemann integrable. Refer to Gordon for a simplified
version of Volterra’s function in full detail [3].
To make sense of Volterra’s construction of a bounded derivative which is
not Riemann integrable, it helps to start with a notion of exactly how we want
the derivative to “misbehave.” At the start of this section, we noted that it
is fairly easy to write down a function whose derivative is unbounded, and
thus not integrable. Another common requirement (which is also sufficient)
for Riemann integrability is that a function be bounded and continuous almost
everywhere. That is, if we can construct a derivative whose set of discontinu-
ities does not have measure zero, then it cannot be integrable. But how can
we easily introduce discontinuities in a derivative? All derivatives have the
Intermediate Value Property, so we cannot have a derivative that looks like,
for example, the indicator function f : [0, 1] → R mapping rationals to 1 and
irrationals to 0. Inspired by our earlier example, consider the function

x2 sin 1 ,


x
if x 6= 0,
f (x) =

0,
 if x = 0.

15
Then we have 
2x sin 1 − cos 1 ,


x x
if x 6= 0,
f 0 (x) =

0,
 if x = 0.

Note that f 0 is bounded on a finite interval containing the origin, but f 0 is not
continuous at 0. Now recall the Cantor set, which is constructed by taking the
interval [0, 1] and removing the middle third, then removing the middle third
of the remaining intervals, and so on. The resulting set has measure zero, so
we will use a modified version of the Cantor set that begins with removing the
middle fourth of the interval [0, 1], etc. (the details are omitted here). Note
that this so-called thick Cantor set does not have measure zero (the sum total
of the lengths of intervals removed is 1/2).

Now, construct a function h as follows. At each step of the creation of

our thick Cantor set, we will place two copies of f in each deleted interval, so
that h0 has discontinuities at each of their endpoints. For example, the first
removed interval has length 1/4, so find the largest value of x in [0, 1/8] such
that f 0 (x) = 0 and call it x0 . Define the functions





 0, if x = 0;


a(x) = x2 sin x1 , if 0 ≤ x ≤ x0 ;




x20 sin 1 , if x0 < x ≤ 81 .


x0

16
and define the reflection of a across x = 1/8:


x20 sin x10 , if 1/8 ≤ x ≤ 1
+ x0 ;


8




1

2
1 1
b(x) = 4
− x sin 1 , if 8
+ x0 < x < 14 ;


 4
−x


0, if x = 41 .


Finally, define the first term in our sequence f1 . Remember that a and b are
functions, not constants.

if 0 ≤ x ≤ 83 ;




 0,



a x − 3 , 3
< x ≤ 12 ;



8
if 8
f1 (x) =
3 1
< x ≤ 58 ;





 b x − 8
, if 2



5

0,
 if 8
≤ x ≤ 1.

Although cumbersome to write, Figure 3 provides a clearer picture of what

Volterra’s function begins to look like. Note that f1 is differentiable on (0, 1),
f10 is bounded, and f10 is discontinuous at 3/8 and 5/8. We continue this
process for each step of the construction of the thick Cantor set, obtaining
a limit function f with infinitely many discontinuities. As it turns out, f 0 is
discontinuous at every point of the thick Cantor set, which, as we noted, does
not have measure zero. Thus, the function f 0 is a bounded derivative which is
not Riemann integrable.

2.2 Pompeiu’s function

In addition to being cumbersome to write and relying on the notion of a perfect

nowhere dense set which does not have measure zero, which may be inacces-
sible to an undergraduate student, Volterra’s function is also integrable on

17
 f1 (x)

Figure 3: The first term in the construction of

Volterra’s function. At this scale, there appear
to be two non-differentiable cusps, but function is
differentiable everywhere.

many intervals. In particular, it is integrable on any interval removed during

the construction of the thick Cantor set (more trivially it is integrable on the
intervals where it is constant). In this sense, as a counter-example, Volterra’s
function “fails” on many subintervals of its domain. We turn our attention to
another example of a bounded, non-integrable derivative.

In Volterra’s function, the guiding idea was to find a function whose set
of discontinuities does not have have measure zero. Consider the fact that,
if a derivative f 0 is Riemann integrable on [a, b], then its integral must be
f (b) − f (a). On the other hand, given a sequence of partitions of [a, b] whose
norm converges to 0, we must also have that respective sequence of upper
and lower Riemann sums of f 0 converge to f (b) − f (a). Pompeiu’s example
hinges on the construction of a positive, strictly increasing function h whose
derivative vanishes on a dense subset. In this manner, if h0 were integrable,
then its integral must be the positive number h(b) − h(a). However, given any
partition of its domain, the lower Riemann sum of h0 is zero, since h0 (x) = 0
on a dense subset, and thus the integral must be 0. We conclude that h0 is not

18
Riemann integrable.

The general idea in the construction of our function is to create a positive,

strictly increasing function which has a “kink” at each point of a dense subset
where the derivative is 0. To achieve this, we will begin with a function whose
derivative is ∞ at every rational on [0, 1], and then take its inverse (we define
infinite derivative in Section 2.4). Such a function has a derivative which is
both bounded and not Riemann integrable on any subinterval of its domain.

This function was originally developed by Pompeiu, and the proof we use
here was outlined by Bruckner et al. [1]. Throughout Section 2.2, the meaning
of the functions f and h, will not change, nor will the sequences {ak }, {qk },
and {fk }.

2.2.1 Beginning our construction (defining f )

Consider the function referred to at the beginning of Section 2, p(x) = x1/3 .

At the origin, we have

p(h) h1/3
lim = lim =∞
h→0 h h→0 h

and, away from the origin, p0 (x) is simply the positive number 13 x−2/3 . Now
let {qk } be a listing of Q ∩ [0, 1], and for each k ∈ Z+ , define

fk (x) = (x − qk )1/3 .

∞
P 3/5
Let {ak } be a sequence of positive numbers such that ak converges. The
k=1
reason for the exponent 3/5 becomes evident in the proof of Theorem 2.2. For
∞
P
now, we simply care that the sum ak converges. This will be sufficient to
k=1

19
 ∞
P
prove that the sum ak fk (x) converges uniformly on [0, 1].
k=1

∞
P ∞
P
Lemma 2.1. If ak converges, then ak fk (x) converges uniformly on
k=1 k=1
[0, 1].

Proof. Since each fk is increasing, it is bounded above by its value at 1. Also,

since (x − qk )1/3 ≤ x1/3 , we have that fk (x) ≤ 1 for all x ∈ [0, 1]. In fact,
∞
P
|fk (x)| ≤ 1 on [0, 1]. Since |ak fk (x)| ≤ |ak | and ak converges, it follows
k=1
∞
P
that ak fk (x) converges uniformly on [0, 1] by the Weierstrass M -test.
k=1
∞
P
We have shown that ak fk (x) converges uniformly to some limit function
k=1
f . Note that since each fk is continuous, we know that f is continuous on [0, 1].
This will be relevant later in Section 2.2.3. Figure 4 may provide some intuition
for what f will look like.

1 1 1 1 1
6 5 4 3 2

Figure 4: An example of what the fifth partial sum

of f may look like. We have picked a convenient
geometric series for {ak } and used some arbitrary
rational numbers for the first few terms of {qk }.

2.2.2 Differentiating our function (determining f 0 )

The intention with the construction of f thus far was to ensure that the deriva-
tive of f is ∞ on a dense subset. Knowing that f 0 “blows up” on a dense subset,

20
the following result is surprising.
∞
P 3/5
Theorem 2.2. Let {ak } be a sequence of positive numbers such that ak
k=1
∞
ak fk0 (x)
P
converges. Then converges almost everywhere.
k=1
∞
ak fk0 (x) diverges. Let
P
Proof. Let D be the set of points x ∈ [a, b] where
k=1
 > 0 and define the interval

 
3/5 3/5
Ak = q k − ak , qk + ak .

∞
S
Let A = Ak and note that Q ∩ [0, 1] ⊆ A. Denoting the length of Ak by
k=1
3/5
`(Ak ), we have `(Ak ) = 2ak and

∞ ∞
3/5
X X
`(Ak ) = 2 ak = 2L,
k=1 k=1

3/5
where L is a real number. Let c ∈ [0, 1] \ A. Note that |c − qk | ≥ ak , so

3/5
ak ak a
2/3
≤ 3/5
= k2/3 .
3|c − qk | 3(ak )2/3 3

∞
P 1 3/5
Since a
32/3 k
converges and ak is positive for k ≥ 1, by the comparison
k=1
test,
∞ ∞
X ak X
= ak fk0 (c).
k=1
3(c − qk )2/3 k=1

∞
ak fk0 (c) converges;
P
converges. We have shown that, for any c ∈ [a, b] \ A,
k=1
thus, D ⊆ A. Since  > 0 was arbitrary and the sum of the lengths of the
∞
ak fk0 (x) converges almost
P
intervals in A is 2L, D has measure zero and
k=1
everywhere on [0, 1].

Remark. It follows from Theorem 3.5 that we can use many different se-
quences for {ak }. We can use any geometric sequence {rk }, where 0 < r < 1,

21
since (rk )3/5 = (r3/5 )k and 0 < r3/5 < 1. When outlining the basics of Pom-
peiu’s function, then, we can simply pick ak = 1/2k . Additionally, we can use
the sequence {1/np }∞
n=1 , for any p > 5/3.

At this juncture, we would hope to simply differentiate f term by term

∞
to show that f 0 indeed is the series ak fk0 (x). However, we noted that
P
k=1
f 0 (x) = ∞ on a dense subset, so we cannot use this method. As it turns
out, the function f 0 is indeed the function we would expect it to be if we could
differentiate term-by-term. In other words, the naive method indeed gives the
correct result, but proving this requires care. In this section, we will at times
use the following definition.

Definition 2.3. Let f : [a, b] → R be a function and let c ∈ (a, b). We say
that f 0 (c) = ∞ if, for all M > 0, there exists δ > 0 such that

f (x) − f (c)
≥M
x−c

for all x satisfying 0 < |x − c| < δ. 

Theorem 2.4. The function f is differentiable on [0, 1] (its derivative is ∞

at rationals), and
∞
X
0
f (x) = ak fk0 (x).
k=1

Proof. Let c ∈ [0, 1]. Our proof consists of three claims:

∞
(1) If c ∈ Q, then both f 0 (c) and ak fk0 (c) are ∞.
P
k=1
∞
ak fk0 (c) converges to ∞, then f 0 (c) = ∞. (The existence
P
(2) If c ∈
/ Q and
k=1
∞
ak fk0 (x) converges to ∞ is not obvious and
P
of points not in Q where
k=1
will be commented on later.)
∞
ak fk0 (c) converges to a real number r, then f 0 (c) = r.
P
(3) If c ∈
/ Q and
k=1

22
For claim (1), note that, for some j ∈ Z+ , c = qj . Since fj0 (c) = ∞, for any
M > 0, there is δ > 0 such that 0 < |x − c| < δ implies

fj (x) − fj (c)
≥ M.
x−c

Then, since each fk is increasing,

∞
f (x) − f (c) X fk (x) − fk (c)
=
x−c k=1
x−c
fj (c + h) − fj (c)
≥
h
≥ M.

This shows that f 0 (c) = ∞. Since each fk is increasing, fk0 (x) is positive for
all k ∈ Z+ and x ∈ [0, 1]. Additionally, recall that each ak is also positive. It
∞
ak fk0 (c) = ∞.
P
follows that
k=1

∞
ak fk0 (c) converges to ∞. Let
P
For claim (2), suppose that c ∈
/ Q and
k=1
N
ak fk0 (c) > M . Then
P
M > 0. There exists a positive integer N such that
k=1
there exists δ > 0 such that if 0 < |x − c| < δ, then

N
X fk (x) − fk (c)
ak > M.
k=1
x−c

Again, since each fk is increasing,

∞
f (x) − f (c) X fk (x) − fk (c)
= ak
x−c k=1
x−c
N
X fk (x) − fk (c)
≥ > M.
k=1
x−c

Thus shows that f 0 (c) = ∞.

23
 ∞
ak fk0 (c) converges to a real number rk . Note
P
For claim (3), suppose that
k=1
that c must not be a rational number. Since c is neither 0 nor 1, we may talk
about the two-sided derivative of f at c. Thus, there is some η > 0 such that
[c − η, c + η] ⊆ [0, 1]. Define the function φ : [−η, η] → R by

∞
f (c + x) − f (c) X (c + x − qk )1/3 − (c − qk )1/3
φ(x) = = ak . (2)
x k=1
x

Note that, if f 0 (c) exists, then lim φ(x) = f 0 (c). We will show that the sum
x→0

in (2) converges uniformly on its domain and apply Lemma 1.5. Note that

x = (c + x − qk ) − (c − qk )

and
a1/3 − b1/3 1
= 2/3
a−b a + a1/3 b1/3 + b2/3

which follows from factoring a difference of cubes. Applying both of these

facts, we have

∞
X (c + x − qk )1/3 − (c − qk )1/3
φ(x) = · ak
k=1
(c + x − qk ) − (c − qk )
∞
X ak
= . (3)
k=1
(c + x − qk )2/3 + (c + x − qk )1/3 (c − qk )1/3 + (c − qk )2/3

For convenience, let

(c + x − qk )1/3
rk = ,
(c − qk )1/3

and then expression (3) becomes

∞
X ak 1
2/3
· 2 .
k=1
(c − qk ) rk + rk + 1

24
Now, since rk is a real number, by considering rk2 + rk + 1 as a polynomial in
rk , we see that rk2 + rk + 1 has a global minimum of 3/4. Thus, for all k ∈ Z+ ,

   
 a k 1  4 a k 
 (c − qk )2/3 · r2 + rk + 1  ≤ 3  (c − qk )2/3  .
   
k

Recall that our beginning assumption was that

∞ ∞
X 1X ak
ak fk0 (c) =
k=1
3 k=1 (c − qk )2/3

converges. Since the terms are positive, the sum converges absolutely, and
 
P∞ 4 ak 
  < ∞. By the Weierstrass M -test,
2/3
3 (c − qk ) 

k=1

∞
X (c + x − qk )1/3 − (c − qk )1/3
ak
k=1
x

converges uniformly to φ on [c − η, c + η].

Since each term in (2) is continuous, φ is continuous. Thus, by Lemma 1.5,

f 0 (c) = lim φ(x)

x→0
N
X fk (c + x) − fk (c)
= lim lim ak
x→0 N →∞
k=1
x
N
X fk (c + x) − fk (c)
= lim lim ak
N →∞ x→0
k=1
x
N
X
= lim ak fk0 (c)
N →∞
k=1
∞
X
= ak fk0 (c)
k=1

∞
as desired. We have shown that f 0 (x) = ak fk0 (x) for all x ∈ [0, 1].
P
k=1

25
 Recall that our goal is to create a strictly increasing function whose deriva-
tive vanishes at some point in every interval. So far we have created a function
whose derivative is infinite at some point on every interval. We turn now to
the inverse of f .

2.2.3 Finding Pompeiu’s function (exploring h = f −1 )

Following Lemma 2.1, we noted that f is continuous. By the Extreme and

Intermediate Value theorems, the range of f is a closed interval. Call it [a, b].
Since f is a strictly increasing function, f has an inverse function. Call it h.
Graphically, the inverse of a function R → R looks like a reflection across the
line y = x. Thus, it is intuitive that, at points where f 0 (c) = ∞, we have
h0 (f (c)) = 0.

Theorem 2.5. Let c ∈ Q ∩ [0, 1] and let d = f (c). Then h0 (d) = 0.

Proof. Let {wn } be any sequence in [a, b] \ {d} converging to d and let {vn } be
the corresponding sequence in [0, 1] converging to c where f (vn ) = wn . Then

h(wn ) − h(d) vn − c
lim = lim . (4)
n→∞ wn − d n→∞ f (vn ) − f (c)

Since the limit of the reciprocal in (4) converges to infinity (since f 0 (c) = ∞),
we have that h0 (d) = 0.

We already know our derivative h0 vanishes at countably many points at

least. However, it bears proof that this set is indeed dense in [0, 1]. Here,
we call a set A dense in [a, b] if, for every open interval O ⊂ [a, b], A ∩ O is
nonempty.

Theorem 2.6. Let S = f (Q ∩ [0, 1]). Then S is dense in [a, b].

26
Proof. Let c ∈ (a, b) and let V be an open interval in (a, b) containing c. Since
f is continuous, the preimage of V is open. Call it U . Since the rationals are
dense in [0, 1], U contains a rational number q. Then f (q) is in S and V , so S
is dense in [a, b].

We next examine some properties of h0 .

Theorem 2.7. The function h is differentiable on [a, b] and h0 is bounded on

[a, b].

Proof. It follows from properties of inverse functions that if f is differentiable

at c then h is differentiable at f (c), and that h0 (f (c)) = 1/f 0 (c). But f is
strictly increasing, so f 0 (x) > 0 for all x ∈ [0, 1], which establishes that h is
differentiable at f (c) whenever f 0 (c) exists. If f 0 (c) = ∞, then by Theorem
2.5, h0 (f (c)) = 0. Thus h is indeed differentiable on its domain [a, b]. Again,
since f is strictly increasing, its inverse h is strictly increasing. Because h0 is
nonnegative, to show that h0 is bounded, it suffices to find an upper bound.
Since h0 (f (c)) = 1/f 0 (c), it suffices to find a lower bound for f 0 .

By Theorem 2.4, the k-th term in f 0 is

ak
ak fk0 (x) = .
3(x − qk )2/3

But (x − qk )2/3 is a positive number bounded by 1, so

ak
ak fk0 (x) ≥ .
3

Since each ak is positive, we may use, for example, a1 /3 as a lower bound for
f 0 . This shows that h0 is bounded.

We are now ready to turn to the main result of Section 2.2.

27
Theorem 2.8. The function h0 is not Riemann integrable on any subinterval
of its domain.

Proof. Suppose that h0 is Riemann integrable on [c, d] ⊆ [a, b]. By the Funda-
Rd
mental Theorem of Calculus, c h0 = h(d) − h(c) > 0 (recall that h is strictly
increasing). On the other hand, h0 vanishes on a dense set of [c, d]. Thus, for
any partition of [c, d], the lower Riemann sum of h0 is 0. Thus, we must have
Rd 0
c
h = 0, a contradiction. Therefore h0 is not integrable on [c, d].

This concludes our main example of a bounded derivative which is not

Riemann integrable. Although we cannot graph h, it seems that h is somewhat
more intuitive to visualize than Volterra’s function using copies of x2 sin x1 . If
we start with any partial sum of f , then its inverse will be integrable, but we
may picture h as an increasing function with a horizontal “kink” on a dense
set (refer to Figure 4 for an idea of what f looks like, and recall that the plot
of h is simply the reflection about y = x). Then h0 , for its part, will be a
nonnegative function which, as we note in Section 2.2.4, is positive on a dense
set and zero on another dense set. This gives some intuition for why, in some
sense, it is difficult to ascribe a value to the “area” under h0 .

2.2.4 Further notes on Pompeiu’s derivative (h0 )

As noted in Theorems 2.5 and 2.6, h0 (x) = 0 on a dense subset of [a, b].
However, the fact that h0 is not integrable on any subinterval leads to the
following observation.

Theorem 2.9. The function h0 attains the value 0 on a dense subset, attains
a positive value on another dense subset, and, where h0 is positive, it is also
discontinuous.

28
Proof. We have seen that h0 (x) = 0 on a dense subset, and we know from the
proof of Theorem 2.7 that h0 is nonnegative. Suppose that h0 (x) is identically
0 on an interval. Then h0 is integrable over that interval, which contradicts
Theorem 2.8. Thus h0 attains a positive value on every interval, so h0 (x) is
positive on a dense subset of [a, b].

Next we claim that, if h0 (c) > 0, then h0 is discontinuous at c. For all

n ∈ Z+ , there is some xn ∈ (c − 1/n, c + 1/n) such that h0 (xn ) = 0. Then {xn }
converges to c and {h0 (xn )} converges to 0. Since h0 (c) > 0, h0 is discontinuous
at c.

Recall that we constructed the function f with the intention of having an

infinite derivative at every rational in [0, 1]. Furthermore, Theorem 2.2 to-
gether with Theorem 2.4 shows that f 0 exists (in a sense) almost everywhere
on [0, 1]. It is perhaps surprising, then, that there are irrational points in [0, 1]
where f 0 is ∞. This fact is not obvious, and relies on some concepts that may
not be taught in undergraduate analysis. A Baire class one function is any
function which is the pointwise limit of a sequence of continuous functions.
(Recall that the uniform limit of a sequence of continuous functions is contin-
uous, but pointwise convergence need not preserve this property.) As it turns
out, all derivatives are Baire class one functions.

Theorem 2.10. If f is differentiable on [a, b], then f 0 is a Baire class one

function.

Proof. First, extend f to a function f˜: [a, b + 1] → R where



f (x),
 if a ≤ x ≤ b;
f˜(x) =
f 0 (b)(x − b) + f (b),

 if b < x ≤ b + 1.

29
Now define
f˜(x + 1/n) − f˜(x)
φn (x) = .
1/n

Then each φn is continuous on [a, b] since both f and f˜ are continuous on [a, b].
Since lim φn (x) = f 0 (x) for all x ∈ [a, b], f 0 is a Baire class one function.
n→∞

A well-known property of continuous functions is that the preimage of open

sets under a continuous function is open. (In fact, this property is equivalent
to continuity.) As it turns out, there is a similar result for Baire class one
functions. First, we need some new terms. While any union of open sets is
open, the same is not necessarily true for an infinite intersection of open sets.
We will call a set S an Fσ set if S is a countable union of closed sets; we will
call S a Gδ set if S is a countable intersection of open sets. The following basic
properties of Fσ and Gδ sets will be useful, although we will not prove them
here.

(1) The complement of an Fσ set is a Gδ set.

(2) The complement of a Gδ set is an Fσ set.

(3) A finite intersection or countable union of Fσ sets is an Fσ set.

(4) A finite union or countable intersection of Gδ sets is a Gδ set.

Theorem 2.11. If f : [a, b] → R is a Baire class one function, then the preim-
age of any open set under f is an Fσ set.

Proof. Let {fp } be a sequence of continuous functions which converge point-

wise to f on [a, b]. We will show that, for all r ∈ R, the preimage of (−∞, r),
and similarly (r, ∞), are Fσ sets. Namely, we will show that

∞ [
[ ∞ \
∞
−1
f ((−∞, r)) = fp−1 ((−∞, r − 1/k]). (5)
k=1 n=1 p=n

30
Since each fp is continuous, we know fp−1 ((−∞, r − 1/k]) is closed for every
p, k ∈ Z+ . Their intersection is closed, and the countable union of these closed
sets is an Fσ set. That is, if equation (5) is true, then f −1 ((−∞, r)) is an Fσ set.

Suppose that x ∈ f −1 ((−∞, r)). Then f (x) ∈ (−∞, r), and, for some
K ∈ Z+ , we have f (x) ∈ (−∞, r − 1/K). Since {fp (x)} converges to f (x),
there is some N such that, if p ≥ N , we have fp (x) ∈ (−∞, r − 1/K). That
is,
∞
\ ∞
\
x∈ fp−1 ((−∞, r − 1/K)) ⊆ fp−1 ((−∞, r − 1/K])
p=N p=N

and this shows

∞ [
[ ∞ \
∞
x∈ fp−1 ((−∞, r − 1/k]). (6)
k=1 n=1 p=n

Conversely, begin by supposing (6) is true. Then there exist positive integers
K and N such that

∞
\
x∈ fp−1 ((−∞, r − 1/K)).
p=N

This means that fp (x) ∈ (−∞, r − 1/K) for all p ≥ N . Since {fp (x)}∞
p=N con-

verges to f (x), we have that f (x) ≤ r − 1/K < r, and thus x ∈ f −1 ((−∞, r)).
This proves equation (5). As we noted, this establishes that f −1 (−∞, r)), and
similarly that f −1 ((r, ∞)), are Fσ sets.

Now, given any open set U , we can write U as the countable union of open
intervals (ai , bi ) where i ≥ 1, and thus

∞
[
U= ((−∞, bi ) ∩ (ai , ∞)) . (7)
i=1

31
Using some basic properties of inverse set functions, note the following:

∞ ∞
!
[ [
f −1 f −1 ((−∞, bi )) ∩ f −1 ((ai , ∞)) .


(−∞, bi ) ∩ (ai , ∞) = (8)
i=1 i=1

Now, since the intersection of two Fσ sets is an Fσ set, and the countable union
of Fσ sets is an Fσ set, it follows from (7) and (8) that f −1 (U ) is an Fσ set.

Before we can prove the existence of irrational points where f 0 converges

to ∞, we need Theorem 2.12, which we will not prove here. A limit point
of a set A is a number x such that, for all r > 0, the set A ∩ (x − r, x + r)
is nonempty. The set A is nowhere dense if A ∪ {x : x is a limit point of A}
contains no open intervals.

Theorem 2.12 (Baire Category Theorem [4]). If {En } is a sequence of nowhere

dense sets, then any interval [a, b] contains a point x where x is not contained
in any En .

Recall that S is the set of points x ∈ [a, b] where f −1 (x) is rational. In

Theorem 2.5, we showed that h0 (x) = ∞ for all x ∈ S. We are now ready to
prove the following.

/ S such that h0 (x) = 0.

Theorem 2.13. There exists a point x ∈

Proof. By Theorem 2.10, we know that h0 is a Baire class one function. Us-
ing Theorem 2.11, we see that h0−1 ((−∞, 0) ∪ (0, ∞)) is an Fσ set, and thus
h0−1 ({0}) is a Gδ set. As we noted previously (see Theorems 2.5 and 2.6),
h0 (S) = {0}, and so S ⊆ h0−1 ({0}). If we can show that S is not a Gδ set,
then S is a proper subset of h0−1 ({0}), which guarantees some x ∈
/ S which
maps to 0 under h0 . By way of contradiction, suppose S is a Gδ set and thus

32
that [a, b] \ S is an Fσ set. There is a sequence {En } of closed sets such that

∞
[
[a, b] \ S = En .
n=1

Now, since S is dense in [a, b], none of the sets En may contain an open interval.
Thus, {En } is a sequence of nowhere dense sets. But then

∞ ∞
! !
[ [
[a, b] = ([a, b] \ S) ∪ S = En ∪ {f (qn )} .
n=1 n=1

However, the Baire Category Theorem guarantees a point x ∈ [a, b] not con-
tained in any of the nowhere dense sets En or {f (qk )}. Thus, S is not a Gδ
/ S such that h0 (x) = 0. This corresponds to a
set. There must be a point x ∈
/ [0, 1] ∩ Q such that f 0 (t) = ∞.
point t ∈

3 Absolute continuity and Baire class one

We saw that h0 vanishes on dense set which is at least countably infinite.

Similarly, it is positive on a dense set. It turns out that h0 is positive on a set
which does not have measure zero. In order to prove that claim, we will need
the following result.

Theorem 3.1. If f 0 is bounded, then f maps any set of measure zero to a set
of measure zero.

Proof. Let f be a differentiable function and let S be a set of measure zero in

the domain of f . Suppose M > 0 is a bound for f 0 , and fix  > 0. Then there
exists some sequence of intervals {(ak , bk )} such that

∞ ∞
[ X 
S⊆ (ak , bk ) and (bk − ak ) < .
k=1 k=1
M +1

33
Since each x ∈ S is contained in some interval (aj , bj ), which is in turn con-
tained in f ((aj , bj )),
∞
[
f (S) ⊆ f ((ak , bk )).
k=1

Using the Mean Value Theorem, if we take any two points x, y ∈ (ak , bk ), we
have
|f (y) − f (x)| ≤ |y − x|M ≤ (bk − ak )M

and thus, for each k ∈ Z+ , we can cover f ((ak , bk )) with an open interval of
length (M + 1)(bk − ak ). Since

∞
X
(M + 1)(bk − ak ) < ,
k=1

we have shown that f (S) has measure zero.

Now we can prove our desired result.

Theorem 3.2. The set E = {x : h0 (x) 6= 0} does not have measure zero.

Proof. Suppose that E has measure zero. Since h0 is bounded, by Theorem

3.1, the image of E under h also has measure zero. Then, for any x ∈ [0, 1],
x is either in h(E), or not in h(E). If x 6∈ h(E), then, since h is a bijection,
there is some y ∈ [a, b] \ E such that h(y) = x. But this means h0 (y) = x = 0,
which, as we noted previously, implies f 0 (x) = ∞. But we already have shown
that the set of points where f 0 (x) = ∞ has measure zero. We can then write
[0, 1] as a union of two sets of measure zero, a contradiction. Thus, the set E
does not have measure zero.

The proof of Theorem 3.2 relies on the fact that functions with bounded
derivatives map sets of measure to sets of measure zero. We can prove a more
general result using the notion of absolute continuity.

34
Definition 3.3. A function f on an interval I is absolutely continuous if, for
each  > 0, there exists δ > 0 satisfying the following condition: if the set
{[ci , di ] : 1 ≤ i ≤ n} is a finite collection of disjoint intervals in I that satisfies
n
P Pn
the inequality (di − ci ) < δ, then |f (di ) − f (ci )| < . 
i=1 i=1

It is easy to see from Definition 3.3 that absolute continuity implies uniform
continuity (simply take n = 1 as it is used in the definition). What is less
obvious is that there are uniformly continuous functions that are not absolutely
continuous. The Cantor function, which we won’t explore here, is a canonical
example, but we will now give an example of a more elementary function which
is uniformly continuous, but not absolutely continuous.

Example 3.4. Consider the function f : [0, 1] → R defined by


x sin π ,


x
if x 6= 0,
f (x) =

0,
 if x = 0.

From some elementary properties of continuous functions, it is clear that f is

continuous on (0, 1]. By the Squeeze Theorem,

π
lim+ x sin = 0,
x→0 x

and so f is continuous on [0, 1]. This also shows that f is uniformly continuous
on its domain. We note that when

π π
= + 2πn,
x 2

for n ∈ Z, then
2
f (x) = x = .
4n + 1

35
 π π 2
Similarly, if = − + 2πn, then f (y) = −y = − . For the same values
y 2 4n − 1
x and y, we have

2(4n − 1) + 2(4n + 1) 16n 1

|f (y) − f (x)| = x + y = = > .
16n2 − 1 16n2 − 1 n

∞
1
P
Now let  = 1 and let δ > 0. Since n
is the divergent harmonic series, each
n=1
∞ ∞
1 2 1
P P
tail n
is also divergent. Take N such that 4n−1
< δ. Since n
diverges,
n=N n=N
M
1
P
there is some positive integer M such that n
> 1 = . Consider the set of
n=N
disjoint intervals {[xn , yn ] : N ≤ n ≤ M } where

2 2
xn = , and yn = .
4n + 1 4n − 1

M M M
1
P P P
Then (yi − xi ) < δ and |f (yi ) − f (xi )| > n
> 1 = , so f is not
n=N n=N n=N
absolutely continuous. 

3.1 Absolute continuity and bounded derivatives

Theorem 3.1 stated that a function f which has a bounded derivative maps
any set of measure zero to a set of measure zero. Here, we will prove a similar
result for absolutely continuous functions. Finally, we will give an alternate
proof of Theorem 3.1 by showing that if f has a bounded derivative on [a, b],
then f is absolutely continuous on [a, b].

Theorem 3.5. If f : [a, b] → R is absolutely continuous and A has measure

zero, then B = f (A) has measure zero.

Proof. Let  > 0. Since f is absolutely continuous, there is some δ > 0 such
∞
P
that, for any sequence {[rk , sk ]} of disjoint intervals satisfying (sk − rk ) < δ,
k=1
∞
P
we have |f (sk ) − f (rk )| < . Since A has measure zero, there exists a
k=1

36
 ∞
S ∞
P
sequence of open sets {(ak , bk )} such that A ⊆ (ak , bk ) and (bk − ak ) < δ.
k=1 k=1
Since any open set is the union of disjoint open intervals, we may assume that
the (ak , bk )’s are disjoint. By the absolute continuity of f ,

∞
X
|f (bk ) − f (ak )| ≤ .
k=1

Since f is continuous, by the Extreme and Intermediate Value Theorems, we

know that f maps closed intervals to closed intervals. Define the intervals
[ck , dk ] = f ([ak , bk ]). Again, using the continuity of f , the preimage of each
[ck , dk ] is a closed interval whose endpoints must fall in [ak , bk ]. This shows
P∞
that |dk − ck | ≤ , and since
k=1

∞
[ ∞
[
B⊆ f ([ak , bk ]) = [ck , dk ],
k=1 k=1

we may also write

∞
[
B⊆ (ck − /2k , dk + /2k ).
k=1

∞
(dk − ck + /2k−1 ) ≤  + 2, it follows that
P
Because  > 0 was arbitrary and
k=1
B has measure zero.

We have shown that if a function f has a bounded derivative, or if f is

absolutely continuous, then f maps sets of measure zero to sets of measure
zero. Next we will show that, if f 0 is bounded, then f is absolutely continuous.
Since it follows from Theorem 3.5 that f maps sets of measure zero to sets of
measure zero, Theorems 3.5 and 3.6 provide an alternate proof to Theorem
3.1.

Theorem 3.6. If f : [a, b] → R is differentiable on [a, b] and f 0 is bounded,

then f is absolutely continuous on [a, b].

37
Proof. Let  > 0 and let M > 0 be a bound for f 0 . Suppose that the set
{[ci , di ] : 1 ≤ i ≤ n} is a finite collection of disjoint intervals in [a, b] that
Pn
satisfies the inequality (di − ci ) < . By the Mean Value Theorem, for each
i=1
i ∈ {1, . . . , n}, there is some ti ∈ (ci , di ) such that |f (di ) − f (ci )| < M (di − ci ).
Then
n
X n
X
|f (di ) − f (ci )| < M (di − ci ) < M,
i=1 i=1

and so f is absolutely continuous.

3.2 Some properties of Baire class one functions

In Section 2.2, we constructed a strictly increasing, differentiable function

whose derivative vanished on a dense subset of its domain. Using some ideas
about Fσ sets and Baire class one functions, we showed that Pompeiu’s non-
integrable derivative vanishes at even more points than that dense subset.
To expand on this, we will explore some more general properties of Fσ sets
and Baire class one functions. The goal of this section will be to arrive at a
necessary and sufficient condition for a function to be Baire class 1. Many
readers will recognize the following fact, which is common in a real analysis
course:

• The function f : [a, b] → R is continuous if and only if the preimage of any

open set under f is an open set.

If we think of Baire class one functions as one level of abstraction removed

from continuous functions, and Fσ sets as one level removed from open sets,

38
then the following analogy, which we will prove, is fitting:

• The function f : [a, b] → R is a Baire class one function if and only if the

preimage of any open set under f is an Fσ set.

Before we turn our attention to this main idea, we first prove another fact
which alludes back to a common fact about uniform continuity. In an analysis
course, one might learn that uniform convergence “preserves” continuity. That
is, if {fn } converges uniformly to f , and each fn is continuous, then f is
continuous. Again, in a pleasing analogy, it turns out that, if each fn is a
Baire class one function, then f is a Baire class one function. To prove this,
we will first consider the sequence {fn } in a series representation.

Lemma 3.7. Suppose that f : [a, b] → R is a Baire class one function. If

|f (x)| ≤ M for all x ∈ [a, b], then there is a sequence of continuous functions
{hk } which converge pointwise to f on [a, b] and each hk is bounded by M .

Proof. Since f is a Baire class one function, there is a sequence of continuous

functions {fn } converging pointwise to f on [a, b]. For each k ∈ Z+ , define the
function hk as follows:





 M, if fk (x) > M,


hk (x) = fk (x), if |fk (x)| ≤ M,





−M,
 if fk (x) < −M.

Now let c ∈ [a, b]. If |fk (c)| < M , then on some open interval containing c,
hk (x) is equal to fk (x), and so hk is continuous at c. Similarly, if |fk (c)| > M ,
then hk is constant on some neighborhood of c. If fk (c) = M , then for any
 > 0, there is δ > 0 such that |x − c| < δ implies |fk (x) − M | < . Then,

39
if |x − c| < δ and fk (x) > M , we have |hk (x) − hk (c)| = |M − M | < .
If −M < fk (x) ≤ M , we have |hk (x) − hk (c)| = |fk (x) − M | < . The
case where fk (c) = −M is similar. Thus hk is continuous for each k. For
each x ∈ [a, b], {fk (x)} converges to f (x). If |f (x)| < M , there is some K
such that, if k > K, then |fk (x)| < M . Thus {hk (x)} converges to f (x). If
f (x) = M , then for any  > 0, there is some K such that, if k > K, then
|fk (x) − K| < . Then, if fk (x) > M , |hk (x) − M | = 0, and, if fk (x) < M ,
then |hk (x) − M | = |fk (x) − M | < . We have shown that {hk } is a sequence
of continuous functions converging pointwise to f on [a, b], where each hk is
bounded by M .

As we mentioned earlier, we first prove the following result for series. Af-
terwards, we will relate our findings back to sequences.
∞
P
Theorem 3.8. Suppose that the series gj converges to h on [a, b], and,
j=1
for each j ≥ 1, there is a positive constant Nj such that |gj (x)| < Nj for all
∞
P
x ∈ [a, b], where Nj converges. If each gj is a Baire class one function,
j=1
then h is a Baire class one function.
∞
P
Proof. Suppose that gj converges to h on [a, b], where each gj is a Baire class
j=1
∞
P
one function and bounded by Nj > 0, where Nj = N . By the Weierstrass
j=1
M -test, the convergence is uniform. Each gj is bounded by Mj = Nj + 1/2j ,
∞
P
and Mj converges to M = N + 1. By Lemma 3.7, for each j ≥ 1, there
j=1
is a sequence of continuous functions {fj,i }∞
i=1 which converge pointwise to gj

on [a, b], and where each fj,i satisfies |fj,i (x)| < Mj for all x ∈ [a, b] and for
∞
P
all i ≥ 1. Note: to avoid confusion, we emphasize here that the series gj
j=1
converges uniformly to h, and, for each j, the sequence {fj,i }∞
i=1 converges

40
pointwise to gj . We claim that the sequence

( p
)∞
X
{f1,1 , f1,2 + f2,2 , f1,3 + f2,3 + f3,3 , . . .} = fj,p (9)
j=1 p=1

converges pointwise to h on [a, b]. Since each term in this sequence is a finite
sum of continuous functions, this will show that h is a Baire class one function.

∞
P
Fix x ∈ [a, b] and let  > 0. Since the series Mj converges, there is a
j=1
positive integer Q such that, if n > Q, then

∞
X
Mj < . (10)
j=n

For each j ∈ {1, . . . , Q}, there is a positive integer Pj such that, if p > Pj ,
then
|fj,p (x) − gj (x)| < /n. (11)

Using (10) and (11), if n > max{Q, P1 , . . . , PQ },

 n   n ∞

X  X X 
fj,n (x) − h(x) =  fj,n (x) − gj (x)
   

   
j=1 j=1 j=1
 Q Q
  ∞   n 
X X   X   X 
≤ fj,n (x) − gj (x) +  gj (x) +  fj,n (x)
     
     
j=1 j=1 j=Q+1 j=Q+1
n
X ∞
X n
X
≤ |fj,n (x) − gj (x)| + |gj (x)| + |fj,n (x)|
j=1 j=n+1 j=Q+1

< 3.

This shows that the sequence in (9), when evaluated at a fixed value of
x, indeed converges to h(x). This suffices to show that h is a Baire class one
function.

41
 We now prove a similar result for sequences of functions. Corollary 3.9 will
be instrumental in our main result (Theorem 3.13).

Corollary 3.9. If the sequence {hk } converges uniformly to h on [a, b] and

each hk is a Baire class one function, then h is a Baire class one function.

Proof. Suppose {hk } is a sequence of Baire class one functions converging

uniformly to h on [a, b]. Choose a subsequence {hki } of {hk }, such that, for all
i ≥ 1 and x ∈ [a, b], we have |hki (x) − h(x)| < /2i . Define g1 : [a, b] → R by
g1 (x) = hk1 (x) and if i > 1, define gi : [a, b] → R by gi (x) = hki (x) − hki−1 (x).
Pn
Then we have gi (x) = hki (x), and for all i > 1 and x ∈ [a, b],
i=1

|gi (x)| ≤ |hki (x) − h(x)| + |h(x) − hki−1 (x)|

 
< i
+ i−1
2 2

< i−2 .
2

Since g1 is also bounded by some M > 0, it follows that h is a Baire class one
function.

We now can begin working towards the main result of this section, which
gives a necessary and sufficient condition for a function to be Baire class one.
Recall that Theorem 2.11 gave a necessary condition; it remains to show that,
if the preimage of any open set under f is an open set, then f is a Baire
class one function. To that end, we will partition the range of f , rather than
the domain, into open intervals. The preimage of these open intervals are Fσ
sets, and we will construct a sequence converging to f by using these sets.
Readers familiar with concepts like Lebesgue measure, Lebesgue integration,
and simple sets may find some of these concepts familiar. As our discussion
of Fσ sets concerns unions of closed sets, the following canonical result will be

42
useful, and we will not prove it here.

Theorem 3.10 (Tietze Extension Theorem). Let K be a closed set. If the

function f : K → R is continuous on K, then there is a continuous function
g : R → R such that f (x) = g(x) for all x ∈ K.

As we discussed earlier, the proof of Theorem 3.13 consists of “partitioning”

the range of a function f , taking preimages of those sets and then reconstruct-
ing f . This “reconstruction” will be done using indicator functions, which we
define now. Define the indicator function of a set S by XS : S → {0, 1} by


0,
 if x ∈
/ S,
XS (x) =

1,
 if x ∈ S.

Lemma 3.11 will streamline the proof of Theorem 3.13.

Lemma 3.11. Suppose that {A1 , . . . , An } is a set of pairwise disjoint Fσ sets

whose union contains [a, b]. Then the function f : [a, b] → R, defined by

n
X
f (x) = ak XAk (x),
k=1

is a Baire class one function, where each ak is a real constant.

Proof. Since each Ak is an Fσ set, we can write

∞
[
Ak = Bik
i=1

where each Bik is closed. Now define the function fj on a subset of [a, b] such
j
Bik . Note that fj is a piecewise function with j
S
that fj (x) = ak if x ∈
i=1
“pieces,” and it is defined on a finite union of closed sets. Then, since the
domain of fj is closed, by the Tietze Extension Theorem (Theorem 3.10), for

43
each j, there is a continuous function gj : [a, b] → R such that gj (x) = fj (x)
for any x in the domain of fj . We claim that the sequence {gj } converges
n
S
pointwise to f . Let x ∈ [a, b]. Since [a, b] ⊆ Ak , x must be contained in
k=1
J
Bik . It follows that, for
S
some Ak . Then there is an integer J such that x ∈
i=1
any j ≥ J, gj (x) = ak . Thus f is a Baire class one function.

We now turn our attention to the main proof of this section. We will first
give a necessary and sufficient condition for bounded functions to be Baire
class one functions, and then prove a similar result for all Baire class one
functions.

Theorem 3.12. Let f : [a, b] → R be a bounded function. Then f is a Baire

class one function if and only if the preimage of any open set under f is an
Fσ set.

Proof. For convenience, this proof uses superscript when listing both sets and
real numbers; they do not represent exponents.

Theorem 2.11 shows that, if f is a Baire class one function, then the preim-
age of any open set under f is an Fσ set. It remains to prove the converse.

Suppose f : [a, b] → R is a function bounded by M > 0 such that, if U is

open, then f −1 (U ) is an Fσ set. For positive integers n, k, where 0 ≤ k ≤ n,
define
2M
ykn = −M + k.
n

Then, for each k ∈ {1, . . . , n − 1}, define

Bkn = f −1 ((yk−1
n n
, yk+1 n
)) = {x ∈ [a, b] : yk−1 n
< f (x) < yk+1 }.

44
By our hypothesis, each Bkn is an Fσ set. Define An1 = B1n and, for each integer
k−1
k ∈ {2, . . . , n − 1}, define Ank = Bkn \
S n
Al−1 . Since the set difference of two
l=1
Fσ sets is an Fσ set, the set {An1 , . . . , Ann−1 } contains disjoint Fσ sets whose
union contains [a, b]. For each positive integer n, define fn : [a, b] → R where

n−1
X
fn (x) = ykn XAnk (x).
k=1

By Lemma 3.11, each fn is a Baire class one function. We will show that {fn }
converges uniformly to f on [a, b].

2M
Let  > 0 and pick a positive integer N such that N
< . Then let
x ∈ [a, b]. For some k ∈ {1, . . . , n − 1}, x ∈ Ak . Note that fn (x) = ykn and
n n
f (x) ∈ (yk−1 , yk+1 ). For any n ≥ N ,

2M 2M
|fn (x) − f (x)| = |ykn − f (x)| < ≤ < .
n N

Thus {fn } converges uniformly to f on [a, b]. By Corollary 3.9, f is a Baire

class one function.

Theorem 3.13. Let f : [a, b] → R be any real-valued function. Then f is a

Baire class one function if and only if the preimage of any open set under f
is an Fσ set.

Proof. If f : [a, b] → R is a Baire class one function, then, by Theorem 2.11,

the preimage of any open set under f is an Fσ set. Theorem 3.12 shows that,
if f is bounded, then it is a Baire class one function; it remains to show that,
if f is unbounded and the preimage of any open set is an Fσ set, then f is a
Baire class one function.

45
 Suppose that f : [a, b] → R is an unbounded real valued function for which
the preimage of any open set is an Fσ set. Define h : R → (0, 1) to be a con-
tinuous, strictly increasing function mapping the real line into (0, 1). Then
h ◦ f : [a, b] → (0, 1). If U is any open set, then, since h is continuous, h−1 (U )
is open, and thus, by our hypothesis, (h ◦ f )−1 (U ) = f −1 ◦ h−1 (U ) is an Fσ
set. Since h◦f is bounded, by Theorem 3.12, h◦f is a Baire class one function.

Since h maps an interval to an interval and is continuous and strictly in-

creasing, h−1 is continuous. Let {fn } be a sequence of continuous functions
converging pointwise to h ◦ f on [a, b]. Then, since the composition of two
continuous functions is continuous, h−1 ◦ fn is continuous for each n. Further-
more, we claim that {h−1 ◦ fn } converges pointwise to h−1 ◦ h ◦ f = f on [a, b].
Let c ∈ [a, b]. Since h−1 is continuous,

lim h−1 (fn (c)) = h−1 lim fn (c) = (h−1 ◦ h ◦ f )(c) = f (c)
n→∞ n→∞

Thus, h−1 ◦ fn is a sequence of continuous functions converging uniformly to

f on [a, b]. We have shown that f is a Baire class one function if and only if
the preimage of any open set under f is an Fσ set.

4 The Henstock Integral

As we have already seen, there are some functions which are not Riemann
integrable, but perhaps feel like they “should” be integrable. The simplest
example uses the so-called improper Riemann integral, such as the limit de-
scribing the area under the curve y = x−2/3 between x = 0 and 1. In Section
2.2, we discovered a function on a closed interval which is bounded, has an
antiderivative, and yet is not Riemann integrable. In this section, we present a

46
new integral, consistent with the Riemann integral on all Riemann integrable
functions, which allows us to integrate such derivatives.

We say that a function is Riemann integrable on [a, b] if there is some

L ∈ R satisfying the following: for each  > 0, we can find a real number δ > 0
such that, if tP is any tagged partition of [a, b] with norm less than δ, we have
|S(f,t P ) − L| < . We will define the Henstock integral a similar way, but
replacing the real number δ > 0 with a function δ : [a, b] → (0, ∞). Rather
than a tagged partition of [a, b], with a requirement about its norm, we will
use a δ-fine tagged partition, whose definition follows.

Definition 4.1. Given the closed interval [a, b] and a positive valued function
δ : [a, b] → (0, ∞), a δ-fine tagged partition of [a, b] is a set

t
Pδ = {([xi−1 , xi ], ti ) : 1 ≤ i ≤ n}

where x0 = a, xn = b, xi−1 < xi , and ti ∈ [xi−1 , xi ] for all 1 ≤ i ≤ n.

Furthermore, for all 1 ≤ i ≤ n, [xi−1 , xi ] ⊆ (ti − δ(ti ), ti + δ(ti )). 

Now we can formally define the Henstock integral.

Definition 4.2. A function f : [a, b] → R is Henstock integrable on [a, b] if

there is a real number L satisfying the following: for each  > 0, there exists
a function δ : [a, b] → (0, ∞), such that for any δ-fine tagged partition tPδ of
[a, b], we have |S(f,t Pδ ) − L| < . If such a number L exists, we call L the
Rb
Henstock integral of f over [a, b], denoted L = (H) a f . 

Rb
In this section, we will use the notation (R) a f to denote the Riemann
Rb
integral of f , and, if not specified, we assume that a f denotes the Riemann
integral of f .

47
 It follows easily from the definition that Riemann integrable functions are
Henstock integrable.

Theorem 4.3. If f : [a, b] → R is Riemann integrable on [a, b], then f is

Rb Rb
Henstock integrable on [a, b], and (R) a f = (H) a f .

Proof. Because f is Riemann integrable, there is some L such that, for each
 > 0, there is some δ0 > 0 such that, for any tagged partition tP of [a, b] with
norm smaller then δ0 , we have

|S(f,tP ) − L| < .

Then let δ : [a, b] → (0, ∞) be defined by δ(x) = δ0 /2. If tPδ is a δ-fine tagged
partition of [a, b], then for all 1 ≤ i ≤ n, [xi−1 , xi ] ⊆ (ti − δ0 /2, ti + δ0 /2), and
thus |tPn | < δ0 . It follows that |S(f,t Pn ) − L| < , and thus

Z b Z b
(R) f = L = (H) f,
a a

which completes the proof.

Before proceeding, we make the following observation: if, given a function

f , for each  > 0, we could construct a function δ : [a, b] → (0, ∞) for which
there be no δ-fine tagged partition of [a, b], then f would vacuously satisfy
the criterion for Henstock integrability. As it turns out, we can always find a
δ-fine tagged partition.

Theorem 4.4. For any interval [a, b] and function δ : [a, b] → (0, ∞), there
exists a δ-fine tagged partition of [a, b].

Proof. Let D be the set of all numbers x, where a < x ≤ b, for which there
is a δ-fine tagged partition of [a, x]. By taking x = a + δ(a)/2, we see that

48
{([a, a + δ(a)/2], a)} is a δ-fine tagged partition of [a, a + δ(a)/2]. Thus, D
is non-empty. Since D is bounded above by b, it has a least upper bound —
call it β. We claim that β ∈ D and β = b, and conclude that there is a δ-fine
tagged partition of [a, b].

First we must show that β ∈ D. Since β is the least upper bound of D,

there is some r ∈ (0, δ(β)/2) such that we have a δ-fine tagged partition of
[a, β − r]. If necessary, we can remove tagged intervals or part of tagged inter-
vals from this partition to obtain a δ-fine tagged partition of [a, β − δ(β)/2].
By adjoining the singleton {([β − δ(β)/2, β], β)}, we have a δ-fine tagged par-
tition of [a, β].

Now suppose that β < b, and let tPδ = {([xi−1 , xi ], ti ) : 1 ≤ i ≤ n} be a

δ-fine tagged partition of [a, β]. Then since δ(β) is positive, we can extend
t
Pδ to the tagged partition tPδ ∪ {([β, β + δ(β)/2], β)}. Since this is a δ-fine
tagged partition of [a, β + δ(β)/2)], we have a contradiction (since we said β
is an upper bound for D and have shown that β + δ(β)/2 ∈ D). Thus β = b,
as desired.

Theorem 4.5. If f is differentiable on [a, b], then f 0 is Henstock integrable

Rb
on [a, b] and (H) a f 0 = f (b) − f (a).

Proof. Let f be differentiable on [a, b] and let  > 0. For each x ∈ [a, b], define
δ(x) as a positive number for which, if 0 < |y − x| < d(x), then

 
 f (y) − f (x) 0


 y−x − f (x)  < .
 (12)

(The existence of such a δ(x) follows from the fact that f is differentiable at

49
x.) Note that we may rewrite (12) as

|f 0 (x)(y − x) − (f (y) − f (x))| < |y − x|. (13)

Now suppose that tPδ = {([xi−1 , xi ], ti ) : 1 ≤ i ≤ n} is a δ-fine tagged

partition of [a, b] (and such a partition does exist, from Theorem 4.4). Using
the inequality in (13), for each of the tagged intervals ([xi−1 , xi ], ti ), we have
the following:

|(xi −xi−1 )f 0 (ti ) − (f (xi ) − f (xi−1 ))|

= |(xi − ti + ti − xi−1 )f 0 (ti ) − (f (xi ) − f (ti ) + f (ti ) − f (xi−1 ))|

≤ |f 0 (ti )(xi − ti ) − (f (xi ) − f (ti ))| (14)

+ |f 0 (ti )(ti − xi−1 ) − (f (ti ) − f (xi−1 ))|

< (xi − ti ) + (ti − xi−1 )

= (xi − xi−1 ) (15)

Then, using (15), if we take the Riemann sum of f 0 with the partition tPδ , we
find
 
Xn  
S(f 0 ,tPδ ) − (f (b) − f (a)) =  (xi − xi−1 )f 0 (ti ) − (f (xi ) − f (xi−1 )) 
  
 
i=1
n
X
< (xi − xi−1 )
i=1

= (b − a).

Rb
Thus f 0 is Henstock integrable on [a, b] and (H) a
f 0 = f (b) − f (a).

Corollary 4.6. Pompeiu’s derivative h0 , as defined in Section 2.2, is not Rie-

mann integrable on any subinterval of [a, b], yet it is Henstock integrable on

50
every subinterval of [a, b]. Furthermore, for any [c, d] ⊆ [a, b],

Z d
(H) h0 = h(d) − h(c),
c

which is a real (positive) number.

4.1 A worked example using the Henstock integral

We have successfully engineered a new integral which can integrate derivatives.

However, Pompeiu’s function and its derivative are difficult (if not impossible)
to visualize, and the proof of Theorem 4.5 may not provide much intuition
for the problem. We turn to a more elementary example of a function which
is Henstock integrable but not Riemann integrable, and hopefully gain some
intuition into the difference between the two integration techniques.

Example 4.7. Define the function f : [−a, a] → R where a > 0 and


x2 cos 1 ,

 if x 6= 0,
f (x) = x2

0,
 if x = 0.

We will show that f is differentiable on [−a, a], f 0 is not Riemann integrable

Ra
on [−a, a], and (H) −a f 0 = 0. It is clear that f is differentiable away from the
origin. At the origin, we have

h2 cos h12
lim =0
h→0 h

by the Squeeze Theorem. Thus,

    
1 2 1
2x cos 2 + sin 2 , if x 6= 0,


f 0 (x) = x x x

0,

if x = 0.

51
 Note that f 0 is unbounded near the origin, and thus f 0 is not Riemann
integrable on [−a, a]. However, since f is even, we know from Theorem 4.5
Ra
that (H) a f 0 = 0. But how do two reasonable definitions of the integral arrive
at different conclusions? Let’s consider the function we are trying to integrate,
shown in Figure 5.

f 0 (x)
1 1
−√ √
π π

Figure 5: A partial graph of f 0 . Note that f 0 is

unbounded near the origin and that f 0 is an odd
function.

Recall that, if f 0 is Riemann integrable and we take any sequence of parti-

tions of [−a, a] whose norm converges to 0, then any corresponding sequence
of Riemann sums of f 0 must converge to some number. Let’s consider the
sequence of partitions {Pn }. The distance between two consecutive points in
Pn is greater than some M > 0, and so, for some neighborhood of the origin,
each interval in Pn will contain a relative maximum and minimum. Intu-
itively, then, the upper and lower Riemann sums will always differ by some
amount. Furthermore, since f 0 is unbounded in each direction near the origin,
the amount by which the upper and lower sums differ does not converge to 0.
This gives us some insight into why f 0 is not Riemann integrable. On the other
hand, f 0 is an odd function, and, as long as we ignore a symmetric interval
around the origin, one expects the integral of f 0 on its domain to be 0. We

52
can ascribe meaning to this by writing

 Z −r Z a 
0 0
lim+ (R) f + (R) f = 0.
r→0 −a r

Ra
However, we still cannot write (R) −a f 0 = 0. The Henstock integral solves
this problem for us by allowing δ to vary throughout the domain. Rather than
using a fixed-value δ > 0, we use a positive-valued function δ to construct a
δ-fine tagged partition of the domain. Intuitively, one can imagine that, as we
approach the origin, the values of δ get smaller so as to closely estimate the
“waves” of the curve. Then, when we take a Riemann sum using this partition,
the “evenness” of the function come into play, leaving a value which converges
to 0. The question remains: how does this approach solve the issue at the
origin, with unbounded magnitude and oscillation? Let’s consider the interval
of our δ-fine tagged partition which contains the origin. Call it ([xi−1 , xi ], ti ).
Then ti cannot be nonzero, because we constructed δ in a such a way that
[xi−1 , xi ] closely estimated a part of the “wave” of the curve containing ti
(imagine that each interval is restricted to one single “hump” of the sinusoid).
Thus, we must have ti = 0. This gives us some clearer understanding for how
the Henstock integral can solve problems that the Riemann integral cannot.

4.2 Further properties of the Henstock integral

In this paper, we developed the Henstock integral with the problem of integrat-
ing derivatives in mind. In Theorem 4.5, we showed that, with the Henstock
integral, we can indeed integrate all derivatives by calculating the difference of
the endpoints in the antiderivative. In Theorem 4.3, we saw that, for Riemann
integrable functions, the Riemann and Henstock integrals give the same result.
The Henstock integral thus solves a problem that the Riemann integral has,

53
and aligns with our intuition vis a vis the Fundamental Theorem of Calculus.

In this section, we explore some further properties of the Henstock integral,

focusing on classes of functions which may not be Riemann integrable but
which are Henstock integrable. We already saw that a derivative is Henstock
but not necessarily Riemann integrable. Next, we prove the stronger result
that a continuous function may be not differentiable at infinitely many places,
and its “derivative” will still be Henstock integrable. In this section, we will
say that a property holds nearly everywhere on S if it holds everywhere but a
countable subset of S.

Theorem 4.8. If F : [a, b] → R is continuous and differentiable nearly every-

where, we define f : [a, b] → R as

F 0 (x),

 if F is differentiable at x,
f (x) =

0,
 if F is not differentiable at x.

Then
Z b
(H) f = F (b) − F (a).
a

Proof. Let F : [a, b] → R be a continuous function which is differentiable on

[a, b] \ {qk : k ∈ Z+ }, where each qk ∈ [a, b]. Define f as in the statement of
Theorem 4.8. Let  > 0, and define the function δ : [a, b] → (0, ∞) as follows:
if x 6= qk for all k ≥ 1, δ(x) is a positive number such that, if 0 < |y −x| < δ(x),
then
 
 F (y) − F (x) 
 − f (x)  < ;
 y−x 

if x = qk , then define δ(x) as the positive number such that, if |y − x| < δ,

then |F (y) − F (x)| < /2k . Now let tPδ be a δ-fine tagged partition of [a, b],

54
where tPδ = {([xi−1 , xi ], ti ) : 1 ≤ i ≤ n}. We now split tPδ into two sets, based
on whether or not F is differentiable at ti . That is, define tRδ as the set

t
Rδ = {([yi−1 , yi ], ri ) : 1 ≤ i ≤ p},

where ([yi−1 , yi ], ri ) = ([xj−1 , xj ], tj ) for some i, j, and F is differentiable at ti ;

define t Sδ as the set

t
Sδ = {([zi−1 , zi ], si ) : 1 ≤ i ≤ q},

where ([zi−1 , zi ], si ) = ([xj−1 , xj ], tj ) for some i, j, and F is not differentiable

at ti .

Recall the inequality in line (15) of the proof of Theorem 4.5, which says
that, for 1 ≤ i ≤ p,

|(yi − yi−1 )f (ti ) − (F (yi ) − F (yi−1 ))| < (yi − yi−1 ). (16)

Then note that, because of the continuity of F and the way we defined δ,
together with the fact that f (ti ) = 0, for 1 ≤ i ≤ q,

|(zi − zi−1 )f (ti ) − (F (zi ) − F (zi−1 ))| = |F (zi ) − F (ti ) + F (ti ) − F (zi−1 )|

≤ |F (zi ) − F (ti )| + |F (ti ) − F (zi−1 )|

< /2k−1 . (17)

55
Incorporating (16) and (17),
 
X n 
t
|S(f, Pδ ) − (F (b) − F (a))| =  (xi − xi−1 )f (ti ) − (F (xi ) − F (xi−1 ))
 
 
 i=1
p

X 
≤  (yi − yi−1 )f (ti ) − (F (yi ) − F (yi−1 ))
 
 
i=1
 q 
X 
+  (zi − zi−1 )f (ti ) − (F (zi ) − F (zi−1 ))
 
 
 p i=1   q 
X  X  
≤ (yi − yi−1 ) + 
   
k−1
2 

  
i=1 i=1

≤ (b − a + 2).

Rb
Therefore f is Henstock integrable on [a, b], and (H) a f = F (b) − F (a).

Remark. A similar result to Theorem 4.8 is the following: if F : [a, b] → R is

differentiable on [a, b] \ S and continuous on [a, b], where S is a finite set, and
we define its “derivative” f : [a, b] → R where f (x) = F 0 (x) for all x ∈ [a, b] \ S
and f (x) = 0 for all x ∈ S, then f is Riemann integrable, and the integral has
the expected value f (b) − f (a).

56
References

[1] A. Bruckner, J. Bruckner, T. Bruckner, B. Thomson, Elementary Real

Analysis, Prentice Hall, 2nd Ed., 2001.

[2] H. Lebesgue, Leçons sur L’intégration et la Recherche des Fonctions

Primitives, Chelsea Publishing Company, New York, 3rd Ed., 1973.

[3] R. Gordon, A Bounded Derivative That Is Not Riemann Integrable, Math-

ematics Magazine, Vol. 89, No. 5, 2016.

[4] R. Gordon, Real Analysis: A First Course, Addison-Wesley Higher Math-

ematics, 2nd Ed., 2002.

[5] W. Rudin, Principles of Mathematical Analysis, McGraw-Hill, New York,

3rd Ed., 1976.

[6] V. Volterra, Sui Principii del Calcolo Integrale, Giornale di Matematiche,

Vol. XIX, Naples, 1881.

57