Chapter 3

Absolutely Continuous
Functions
Undergradese, III: “Hmmm, what do you mean by that?” Trans-
lation: “What’s the answer so we can all go home?”
— Jorge Cham, www.phdcomics.com

The Cantor function shows that for monotone functions the fundamental
theorem of calculus fails for Lebesgue integration. Indeed,
 1
u (t) dt = 0 < u (1) − u (0) = 1.
0

To recover it in the context of Lebesgue integration, we need to restrict

ourselves to a subclass of functions of bounded pointwise variation. This
leads us to the notion of absolute continuity.

3.1. AC (I) Versus BP V (I)

Definition 3.1. Let I ⊂ R be an interval. A function u : I → R is said to
be absolutely continuous on I if for every ε > 0 there exists δ > 0 such that


(3.1) |u (bk ) − u (ak )| ≤ ε
k=1

for every finite number of nonoverlapping intervals (ak , bk ), k = 1, . . . , ,

with [ak , bk ] ⊂ I and


(bk − ak ) ≤ δ.
k=1
The space of all absolutely continuous functions u : I → R is denoted by
AC (I).

73
74 3. Absolutely Continuous Functions

Remark 3.2. Note that since  is arbitrary, we can also take  = ∞, namely,
replace finite sums by series.

A function u : I → R is locally absolutely continuous if it is absolutely

continuous in [a, b] for every interval [a, b] ⊂ I. The space of all locally
absolutely continuous functions u : I → R is denoted by ACloc (I). Note
that
ACloc ([a, b]) = AC ([a, b]) .

If u : I → Rd , then we can define the notion of absolute continuity

exactly as in Definition 3.1, with the only difference that the absolute value
is now replaced by the norm in Rd . The space of all absolutely continuous

u :d I → R (respectively,
locally
functions d
 absolutely continuous) is denoted
by AC I; R (respectively, ACloc I; Rd ).
If Ω ⊂ R is an open set, then we define the notion of absolute continuity
for a function u : Ω → R as in Definition 3.1, with the only change that we
now require the intervals [ak , bk ] to be contained in Ω in place of I. The
space of all absolutely continuous functions u : Ω → R is denoted by AC (Ω).

Exercise 3.3. Let I ⊂ R be an interval and let u : I → R.

(i) Prove that u belongs to AC (I) if and only if for every ε > 0 there
exists δ > 0 such that
 
 
 
 (u (bk ) − u (ak )) ≤ ε
 
k=1

for every finite number of nonoverlapping intervals (ak , bk ), k =

1, . . . , , with [ak , bk ] ⊂ I and


(bk − ak ) ≤ δ.
k=1

(ii) Assume that for every ε > 0 there exists δ > 0 such that
 
 
 
 (u (bk ) − u (ak )) ≤ ε
 
k=1

for every finite number of intervals (ak , bk ), k = 1, . . . , , with

[ak , bk ] ⊂ I and


(bk − ak ) ≤ δ.
k=1
Prove that u is locally Lipschitz.
3.1. AC (I) Versus BP V (I) 75

Part (ii) of the previous exercise shows that in Definition 3.1 we cannot
remove the condition that the intervals (ak , bk ) are pairwise disjoint.
By taking  = 1 in Definition 3.1, it follows that an absolutely continuous
function u : I → R is uniformly continuous. The next exercise shows that
the converse is false.
Exercise 3.4. Let u : (0, 1] → R be defined by
1
u (x) := xa sin b ,
x
where a, b ∈ R. Study to see for which a, b the function u is absolutely
continuous. Prove that there exist a, b for which u is uniformly continuous
but not absolutely continuous.
Exercise 3.5. Let I ⊂ R be an interval and let u : I → R be differentiable
with bounded derivative. Prove that u belongs to AC (I).
Exercise 3.6. Let u, v ∈ AC ([a, b]). Prove the following.
(i) u ± v ∈ AC ([a, b]).
(ii) uv ∈ AC ([a, b]).
(iii) If v (x) > 0 for all x ∈ [a, b], then u
v ∈ AC ([a, b]).
(iv) What happens if the interval [a, b] is replaced by an arbitrary in-
terval I ⊂ R (possibly unbounded)?

We now turn to the relation between absolutely continuous functions and

functions of bounded pointwise variation. In Corollary 2.23 we have proved
that if u : I → R has bounded pointwise variation, then u is bounded and
u is Lebesgue integrable. However, the function u (x) := x, x ∈ R, is abso-
lutely continuous, but it is unbounded and u (x) = 1, which is not Lebesgue
integrable. Also the function u (x) := sin x, x ∈ R, is absolutely continuous,
bounded, but u is not Lebesgue integrable. These simple examples show
that an absolutely continuous function may not have bounded pointwise
variation. Proposition 3.8 below will show that this can happen only on
unbounded intervals.
Exercise 3.7. Let I ⊂ R be an interval and let u : I → R be uniformly
continuous.
(i) Prove that u may be extended uniquely to I in such a way that the
extended function is still uniformly continuous.
(ii) Prove

that if u belongs to AC (I), then its extension belongs to
AC I .
(iii) Prove that there exist A, B > 0 such that for all x ∈ I,
|u (x)| ≤ A + B |x| .
76 3. Absolutely Continuous Functions

The previous exercise shows that although an absolutely continuous

function may be unbounded, it cannot grow faster than linear when |x| → ∞.
Next we show that ACloc (I) ⊂ BP Vloc (I).
Proposition 3.8. Let I ⊂ R be an interval and let u ∈ ACloc (I) (respec-
tively, AC (I)). Then u belongs to BP Vloc (I) (respectively, BP V (J) for
every bounded subinterval J of I). In particular, u is differentiable L1 -a.e.
in I and u is locally Lebesgue integrable (respectively, Lebesgue integrable
on bounded subintervals of I).

Proof. Step 1: Assume that u ∈ ACloc (I) and let [a, b] ⊂ I. Take ε = 1,
and let δ > 0 be as in Definition 3.1. Let n be the integer part of 2(b−a)
δ and
b−a
partition [a, b] into n intervals [xi−1 , xi ] of equal length n ,
a = x0 < x1 < · · · < xn = b.

n ≤ δ, in view of (3.1), on each interval [xi−1 , xi ] we have that

Since b−a
Var[xi−1 ,xi ] u ≤ 1, and so by Remark 2.7,

n
2 (b − a)
Var[a,b] u = Var[xi−1 ,xi ] u ≤ n ≤ < ∞,
δ
i=1

where we have used the fact that b−a

n ≥ 2δ .
Step 2: Assume that u ∈ AC (I) and let J ⊂ I be a bounded interval. By

 extend u uniquely to a function u : J → R such that
Exercise 3.7 we may
u belongs to AC I . The previous step (applied to J in place of I) implies
that VarJ u < ∞, and so, in particular, VarJ u < ∞.
This completes the proof.

Corollary 3.9. Let I ⊂ R be a bounded interval. Then AC (I) ⊂ BP V (I).
In particular, if u ∈ AC (I), then u is differentiable L1 -a.e. in I and u is
Lebesgue integrable.
The converse of the previous corollary is false, since absolutely contin-
uous functions are always continuous while monotone functions may not
be. Even more, there exist continuous monotone functions that are not ab-
solutely continuous. The Cantor function and the function constructed in
Theorem 1.47 are such examples. What is missing for a continuous function
in BP V (I) to belong to AC (I) is the so-called (N ) property.
We recall the following definition.
Definition 3.10. If E ⊂ R is a Lebesgue measurable set and v : E → R is
a Lebesgue measurable function, then v is equi-integrable if for every ε > 0
there exists δ > 0 such that

|v (x)| dx ≤ ε
F
3.1. AC (I) Versus BP V (I) 77

for every Lebesgue measurable set F ⊂ E, with L1 (F ) ≤ δ.

Exercise 3.11. Let E ⊂ R be a Lebesgue measurable set, let 1 ≤ p ≤ ∞,

and let v ∈ Lp (E). Prove that v is equi-integrable. Prove that if we only
assume that v ∈ L1loc (E), then the result may no longer be true.

Theorem 3.12 (Lusin (N ) property). Let I ⊂ R be an interval. A function

u : I → R belongs to ACloc (I) if and only if
(i) u is continuous on I,
(ii) u is differentiable L1 -a.e. in I, and u ∈ L1loc (I),
(iii) u maps sets of Lebesgue measure zero into sets of Lebesgue measure
zero.

Property (iii) is called the Lusin (N ) property. We begin with some

preliminary results, which are of interest in themselves.

Lemma 3.13. Let I ⊂ R be an interval and let u : I → R. Assume that

there exist a set E ⊂ I (not necessarily measurable) and M ≥ 0 such that u
is differentiable for all x ∈ E, with
  
u (x) ≤ M for all x ∈ E.

Then
L1o (u (E)) ≤ M L1o (E) .

Proof. Without loss of generality we may assume that E ⊂ I ◦ . Fix ε > 0

and for each n ∈ N let En be the set of points x ∈ E such that

(3.2) L1o (u (J)) ≤ (M + ε) L1o (J)

for all intervals J ⊂ I such that x ∈ J and 0 < length J < 1

n. Note that
En ⊂ En+1 . We claim that
∞

(3.3) E= En .
n=1

Since each En is contained in E, to prove the claim, it suffices to prove that

each point x ∈ E belongs to some En . Fix x ∈ E. Since |u (x)| ≤ M and
u (y) − u (x)
lim = u (x) ,
y→x y−x
there exists δ > 0 such that

|u (y) − u (x)| ≤ (M + ε) |y − x|
78 3. Absolutely Continuous Functions

for all y ∈ I, with |y − x| < δ. Hence, if y, y  ∈ I, with |y − y  | < δ and

y < x < y,

 
 
u (y) − u y   ≤ |u (y) − u (x)| + u y  − u (x)


≤ (M + ε) (x − y) + (M + ε) y  − x


= (M + ε) y  − y .
This implies that x ∈ En for every integer n > 1δ , and so the claim is proved.
We now fix n ∈ N and we prove that


(3.4) L1o (u (En )) ≤ (M + ε) L1o (En ) + ε .
By the definition of L1o (En ) we may find an open set Un such that Un ⊃ En
and
L1 (Un ) ≤ L1o (En ) + ε.
By replacing Un with Un ∩ I ◦ , if necessary,
 we may suppose that Un ⊂ I ◦ .

(n)
Decompose Un as a countable family Jk of pairwise disjoint intervals
(n)
with 0 < length Jk < n1 . Define
 
(n)
I := k : Jk ∩ En = ∅ .

Note that by (3.2) if k ∈ I, then

    
(n) (n)
L1o u Jk ≤ (M + ε) L1o Jk ,

and so
 
      
(n) (n)
L1o (u (En )) ≤ L1o u Jk ≤ L1o u Jk
k∈I k∈I
 
   
(n) (n)
≤ (M + ε) L 1
Jk ≤ (M + ε) L1o Jk
k∈I k


= (M + ε) L (Un ) ≤ (M + ε) L1o (En ) + ε ,
1

(n)
where we have used the fact that the intervals Jk are pairwise disjoint.
Hence, (3.4) holds.
Since L1o is a regular outer measure and {En } (and in turn {u (En )})
is an increasing sequence, by Proposition B.105 in Appendix B we may let
n → ∞ in the previous inequality to obtain


L1o (u (E)) ≤ (M + ε) L1o (E) + ε .
It now suffices to let ε → 0+ .

As a consequence of the previous lemma we have the following result.

3.1. AC (I) Versus BP V (I) 79

Corollary 3.14. Let I ⊂ R be an interval and let u : I → R. Assume that

there exists a set E ⊂ I such that u is differentiable for all x ∈ E. If the set
E has Lebesgue measure zero or if u = 0 in E, then L1 (u (E)) = 0.

Proof. Assume that E has Lebesgue measure zero. For every n ∈ N write
En := {x ∈ E : |u (x)| ≤ n}. Since u is differentiable in E, it follows that
∞

E= En ,
n=1

while by the previous lemma

L1o (u (En )) ≤ nL1o (En ) = 0.
By the countable subadditivity of L1o we obtain that L1 (u (E)) = 0.
If u = 0 in E, then we may take M := 0 in the previous lemma.

Remark 3.15. The Cantor function shows that the previous corollary does
not hold if we replace everywhere differentiability in E with L1 -a.e. differ-
entiability in E.

Another important consequence of Lemma 3.13 is the following.

Lemma 3.16. Let I ⊂ R be an interval, let u : I → R be a Lebesgue
measurable function, and let E ⊂ I be a Lebesgue measurable set on which
u is differentiable. Then u (E) is Lebesgue measurable and

  
(3.5) L1 (u (E)) ≤ u (x) dx.
E

Proof. Step 1: By the properties of the Lebesgue measure we may write E

as the union of a set of Lebesgue measure zero and countably many compact
sets, precisely,

E = E0 ∪ Kn ,
n
where Kn is compact and L1 (E0 )= 0. By the previous corollary L1 (u (E0 ))
= 0, while u (Kn ) is Lebesgue measurable since u : Kn → R is continuous
(since differentiable). Hence, u (E) is Lebesgue measurable.
Step 2: Assume that L1 (E) < ∞. Fix n ∈ N. For every k ∈ N write


k − 1    k
En := x ∈ E :
k
≤ u (x) < n .
2n 2
Then
∞

E= Enk ,
k=1
80 3. Absolutely Continuous Functions

and so by the previous lemma

∞ 
   ∞
   
L (u (E)) = L
1 1 k
u En ≤ L1 u Enk
k=1 k=1

k 1  k  k − 1 1  k 1  1  k
∞ ∞ ∞
≤ L En = L E n + L En
2n 2n 2n
k=1 k=1 k=1
∞  
    
u (x) dx + L (E) ≤ u (x) dx + L (E) ,
1 1
≤
Enk 2n E 2n
k=1

and it suffices to let n → ∞.

Step 3: If L1 (E) = ∞, for every k ∈ Z write

Ek := E ∩ [k, k + 1] .

Then by the previous step applied to Ek we have

  ∞  ∞
 
L (u (E)) = L u
1 1
Ek ≤ L1 (u (Ek ))
k=−∞ k=−∞
∞ 
 
     
≤ u (x) dx = u (x) dx,
k=−∞ Ek E

and the proof is complete.

Remark 3.17. Let I ⊂ R be an interval and let u : I → R. If u is

differentiable on a interval [a, b] ⊂ I, then, in particular, it is continuous
on [a, b], and so u ([a, b]) contains the interval of endpoints u (a) and u (b).
Hence by (3.5),
 b
  
|u (b) − u (a)| ≤ L (u ((a, b))) ≤
1 u (x) dx.
a

We now turn to the proof of Theorem 3.12.

Proof of Theorem 3.12. Step 1: Assume that u satisfies (i)–(iii) and fix

[a, b] ⊂ I. We claim that u belongs to AC ([a, b]). Let {(ak , bk )}k be a finite
number of nonoverlapping intervals of [a, b] and let
 
Ek := x ∈ (ak , bk ) : u (x) exists .

By (ii), L1 ((ak , bk ) \ Ek ) = 0, and so by (iii), L1 (u ((ak , bk ) \ Ek )) = 0.

Since u is continuous, by the intermediate value theorem it assumes all values
between u (ak ) and u (bk ), and so the open interval of endpoints u (ak ) and
3.1. AC (I) Versus BP V (I) 81

u (bk ) is contained in u ((ak , bk )). Therefore by Lemma 3.16,

  
|u (bk ) − u (ak )| ≤ L1 (u ((ak , bk ))) = L1 (u (Ek ))
k k k
      bk   
≤ u (x) dx = u (x) dx.
k Ek k ak

Since, by (ii), u is Lebesgue integrable in [a, b], in view of Exercise 3.11, u

is equi-integrable in [a, b]. Hence, given ε > 0, we may find δ > 0 such that
  * bk 
if (bk − ak ) ≤ δ, then ak |u (x)| dx ≤ ε. The absolute continuity of u
k k
in [a, b] follows from the previous inequality.
Step 2: Conversely, assume that u ∈ ACloc (I). Fix [a, b] ⊂ I. By Propo-
sition 3.8 and Corollary 2.23, u is differentiable L1 -a.e. in [a, b] and u is
Lebesgue integrable in [a, b]. It remains to show that u satisfies property
(iii). Thus, fix a Lebesgue measurable set E ⊂ [a, b] with L1 (E) = 0. Fix
ε > 0 and let δ > 0 be as in Definition 3.1. Since L1 (E) = 0, we may find an
open set A ⊃ E such that L1 (A) ≤ δ. Decompose A ∩ [a, b] into a countable
family {Jk } of pairwise disjoint intervals. Since u is continuous, for every k
we may find ak , bk ∈ Jk such that
u (ak ) = min u (x) , u (bk ) = max u (x) ,
x∈Jk x∈Jk

so that u (Jk ) = [u (ak ) , u (bk )]. Using the fact that L1 (A) ≤ δ, we have
that 
|bk − ak | ≤ δ,
k
and so, by the absolute continuity of u and Remark 3.2,

|u (bk ) − u (ak )| ≤ ε.
k
Hence,
L1 (u (E)) ≤ L1 (u (A ∩ [a, b]))
   
≤ L1 (u (Jk )) ≤ L1 u (Jk )
k k
 
= L ([u (ak ) , u (bk )]) =
1
|u (bk ) − u (ak )| ≤ ε.
k k

Given the arbitrariness of ε > 0, we conclude that L1 (u (E)) = 0.

Remark 3.18. Note that since the sets Ek defined in the first part of the
proof are Borel sets (why?), the previous theorem continues to hold if in
place of the (N ) property we only require that u map Borel sets of Lebesgue
measure zero into sets of Lebesgue measure zero.
82 3. Absolutely Continuous Functions

Remark 3.19. Note that the Cantor function does not satisfy the (N )
property since it sends a set of Lebesgue measure zero, the Cantor set D,
into the full interval [0, 1].
Exercise 3.20. Let u : [a, b] → R be continuous and strictly increasing.
Prove that u is absolutely continuous if and only if it maps the set
 
E := x ∈ [a, b] : u (x) = ∞
into a set of Lebesgue measure zero.
Exercise 3.21. Let u : [a, b] → R be continuous and strictly increasing.
Prove that its inverse u−1 : [u (a) , u (b)] → R is absolutely continuous if and
only if the set
 
E := x ∈ [a, b] : u (x) = 0
has Lebesgue measure zero.

In view of Remark 3.17 and Step 1 of the proof of Theorem 3.12, we

have the following.
Corollary 3.22. Let I ⊂ R be an interval. If u : I → R is everywhere
differentiable in I and u ∈ L1loc (I), then u belongs to ACloc (I).
Exercise 3.23. Prove that if u : [a, b] → R is continuous, differentiable on
[a, b] except for at most a countable number of points, and if u is Lebesgue
integrable, then u belongs to AC ([a, b]).
Corollary 3.24. Let I ⊂ R be an interval and let u : I → R be such that
(i) u is continuous on I,
(ii) u is differentiable L1 -a.e. in I, and u belongs to L1loc (I) and is
equi-integrable,
(iii) u maps sets of Lebesgue measure zero into sets of Lebesgue measure
zero.
Then u belongs to AC (I).

Proof. It is enough to repeat Step 1 of the proof of Theorem 3.12 word

for word, with the only differences that [a, b] should be replaced by I and
that Exercise 3.11 is no longer needed, since, by (ii), u is assumed to be
equi-integrable.

Remark 3.25. Note that if I is a bounded interval, then u : I → R belongs
to AC (I) if and only if (i)–(iii) of the previous corollary hold. Indeed, if
u ∈ AC (I), then by Corollary 3.9, u is integrable, and so equi-integrable by
Exercise 3.11. Thus, property (ii) of the previous corollary holds. Properties
(i) and (iii) follow from Theorem 3.12.
3.1. AC (I) Versus BP V (I) 83

We will prove later on (see Corollary 3.41) that conditions (i)–(iii) in

the previous corollary are actually necessary and sufficient for u : I → R to
be absolutely continuous.
The next corollary will be useful in the study of Sobolev spaces in Chap-
ter 7.
Corollary 3.26. Let I ⊂ R be an interval and let u : I → R be such that
(i) u is continuous on I,
(ii) u is differentiable L1 -a.e. in I, and u ∈ Lp (I) for some 1 ≤ p ≤
∞,
(iii) u maps sets of Lebesgue measure zero into sets of Lebesgue measure
zero.
Then u belongs to AC (I).

Proof. In view of the previous corollary, it remains to show that u is lo-

cally Lebesgue integrable and equi-integrable. The fact that u is locally
Lebesgue integrable follows by Hölder’s inequality, while equi-integrability
follows from Exercise 3.11.

As a consequence of Theorem 3.12 we can characterize those functions

with locally bounded pointwise variation that are locally absolutely contin-
uous functions. Precisely, we have the following result.
Corollary 3.27. Let I ⊂ R be an interval. A function u : I → R belongs
to ACloc (I) if and only if
(i) u is continuous on I,
(ii) u ∈ BP Vloc (I),
(iii) u maps sets of Lebesgue measure zero into sets of Lebesgue measure
zero.

Proof. In view of Corollary 2.23 we are in a position to apply Theorem

3.12.

The next exercise gives an example of a function that satisfies the (N )

property, but it is not of bounded pointwise variation in any interval of
(0, 1).
Exercise 3.28. Let {(an , bn )} be a base for the topology of (0, 1) and let
{rn } be a sequence of positive numbers such that ∞ n=1 n < ∞. We con-
r
struct inductively a sequence of functions un : [0, 1] → R and a sequence of
intervals (cn , dn ) as follows. Let u0 (x) := x. Assume that un−1 : [0, 1] → R
has been defined and that un is a continuous piecewise affine function. Let
(cn , dn ) ⊂ (an , bn ) be an interval such that un−1 restricted to (cn , dn ) is linear
84 3. Absolutely Continuous Functions

and L1 (un−1 ((cn , dn ))) < rn . We define the continuous function un to be

un−1 outside (cn , dn ), while in (cn , dn ) we define it as a continuous piecewise
affine function such that L1 (un ((cn , dn ))) < rn and Var(cn ,dn ) un > n.
(i) Prove that {un } converges uniformly to a continuous function u :
[0, 1] → R.
(ii) Prove that for every interval [a, b] ⊂ (0, 1) with a < b, Var[a,b] u =
∞.
(iii) Prove that u has the (N ) property. Hint: Let
∞
 ∞

E0 := [0, 1] \ (ci , di ) , En := (cn , dn ) \ (ci , di ) ,
i=1 i=n+1
and
∞ 
+ ∞
E∞ := (ci , di ) .
n=1 i=n
For every set E ⊂ [0, 1], with L1 (E) = 0, write
∞

E = (E ∩ E∞ ) ∪ (E ∩ En ) .
n=0

Exercise 3.29 (The Cantor set and the (N ) property). Let D be the Cantor
set.
(i) Prove that every number x ∈ [0, 2] can be written as
∞
cn
x=2 ,
3n
n=1
where cn ∈ {0, 1, 2}, and deduce that every element in [0, 2] can be
written as the sum of two elements of D.
(ii) Prove that if v : D → R is continuous and L1 (v (D)) = 0, then v
can be extended to a continuous function on [0, 1] that has the (N )
property. Hint: Make v differentiable outside D.
(iii) For each x ∈ D write
∞
 cn (x)
x=2 ,
3n
n=1
where cn ∈ {0, 1} and prove that there exist two continuous func-
tions u1 : [0, 1] → R and u2 : [0, 1] → R with the (N ) property and
such that for all x ∈ D,
∞
 ∞
c2n (x) c2n+1 (x)
u1 (x) = n
, u2 (x) = .
3 3n
n=1 n=1
(iv) Prove that u1 + u2 does not have the (N ) property.
3.1. AC (I) Versus BP V (I) 85

(v) Why is this example important for absolute continuity?

We now show that absolutely continuous functions can be characterized

as the family of all functions for which the fundamental theorem of calculus
holds (for the Lebesgue integral).

Theorem 3.30 (Fundamental theorem of calculus). Let I ⊂ R be an inter-

val. A function u : I → R belongs to ACloc (I) if and only if
(i) u is continuous in I,
(ii) u is differentiable L1 -a.e. in I, and u belongs to L1loc (I),
(iii) the fundamental theorem of calculus is valid; that is, for all x, x0 ∈
I,
 x
u (x) = u (x0 ) + u (t) dt.
x0

The proof hinges upon on a preliminary result, which is of independent

interest.

Lemma 3.31. Let I ⊂ R be an interval and let v : I → R be a Lebesgue

integrable function. Fix x0 ∈ I and let
 x
u (x) := v (t) dt, x ∈ I.
x0

Then the function u is absolutely continuous in I and u (x) = v (x) for

L1 -a.e. x ∈ I.

Proof. The facts that u is absolutely continuous and differentiable for L1 -

a.e. x ∈ I follow from Exercise 3.11 and Proposition 3.8, respectively. In
the remainder of the proof we show that u (x) = v (x) for L1 -a.e. x ∈ I.
Step 1: Assume first that v = χE for some Lebesgue measurable set E ⊂ R.
Fix a bounded open interval J ⊂ I containing x0 in its closure. We claim
that
u (x) = 1 for L1 -a.e. x ∈ E ∩ J.
By the definition of Lebesgue outer measure we may find a decreasing se-
quence {Un } of open sets such that Un ⊃ E ∩ J and
+
(3.6) L1 (U∞ \ (E ∩ J)) = 0, where U∞ := Un .
n=1

By replacing Un with Un ∩ J, we may assume that Un ⊂ J. Define

 x
un (x) := χUn (t) dt, x ∈ J.
x0
86 3. Absolutely Continuous Functions

Since U1 ⊂ J, which is bounded, we are in a position to apply Lebesgue’s

dominated convergence theorem and (3.6) to conclude that for all x ∈ J,
 x  x
lim un (x) = lim χUn (t) dt = χU∞ (t) dt
n→∞ n→∞ x x0
 x 0
 x
= χE∩J (t) dt = χE (t) dt = u (x) ,
x0 x0

where in the fourth equality we have used the fact that the open interval of
endpoints x and x0 is contained in J.
Hence, in the interval J we may write u in terms of the telescopic series
∞

u = u1 + (un+1 − un ) .
n=1

Note that since Un ⊃ Un+1 ,

 x
(un − un+1 ) (x) = χUn \Un+1 (t) dt, x ∈ J,
x0

and since χUn \Un+1 ≥ 0, the function un − un+1 is monotone in the intervals
(−∞, x0 )∩J and [x0 , ∞, )∩J. By Fubini’s theorem (applied in each interval)
we get that for L1 -a.e. x ∈ J,
∞


u (x) = u1 (x) + un+1 (x) − un (x) = lim un (x) .
n→∞
n=1

On the other hand, if x ∈ Un , then un (x) = 1 (why?), and so, if x ∈ U∞ ,

then un (x) = 1 for all n ∈ N. Hence, we have proved that u (x) = 1 for
L1 -a.e. x ∈ U∞ , and so, in particular, for L1 -a.e. x ∈ E ∩ J.
Next we show that u (x) = 0 for L1 -a.e. x ∈ J \ E. Since for all x ∈ J,
 x  x


u (x) = χE (t) dt = 1 − χJ\E (t) dt,
x0 x0

by applying what we just proved to χJ\E , we conclude that u (x) = 1−1 = 0

for L1 -a.e. x ∈ J \ E. Thus, u (x) = χE (x) for L1 -a.e. x ∈ J. By letting
J  I, we obtain the same result in I.
Step 2: By the linearity of the derivatives and Step 1 we conclude that
if v is a simple function, then u (x) = v (x) for L1 -a.e. x ∈ I. If v is
a nonnegative Lebesgue measurable function, then we may construct an
increasing sequence {sn } of nonnegative simple functions such that sn (x) 
v (x) for L1 -a.e. x ∈ I. Then by Lebesgue’s dominated convergence theorem
for all x ∈ I,
 x  x
lim sn (t) dt = v (t) dt = u (x) ,
n→∞ x x0
0
3.1. AC (I) Versus BP V (I) 87

and so we may proceed as in the first step (using telescopic series) to show
that u (x) = v (x) for L1 -a.e. x ∈ I.
In the general case, it suffices to write v = v + − v − .

We are now ready to prove Theorem 3.30.

Proof of Theorem 3.30. Assume that u ∈ ACloc (I). In view of Theorem

3.12, it remains to prove (iii). Let [a, b] ⊂ I be so large that x0 ∈ [a, b] and
define   x 
w (x) := u (x) − u (a) + u (t) dt , x ∈ [a, b] .
x0
By Lemma 3.31 and Theorem 3.12, there exists a Lebesgue measurable set
E ⊂ [a, b], with L1 (E) = 0, such that for all x ∈ [a, b] \ E the function
w is differentiable at x and w (x) = 0. By Corollary 3.14 we have that
L1 (w ([a, b] \ E)) = 0. On the other hand, since w is absolutely continu-
ous in [a, b] (see Exercise 3.6 and Lemma 3.31), by Theorem 3.12 it sends
sets of Lebesgue measure zero into sets of Lebesgue measure zero, and so
L1 (w (E)) = 0. Thus, we have shown that L1 (w ([a, b])) = 0. But since w is
a continuous function, by the intermediate value theorem w ([a, b]) is either
a point or a proper interval. Thus, it has to be a point. In conclusion, we
have proved that w (x) ≡ const. Since w (x0 ) = 0, it follows that w = 0,
and so (iii) holds for all x ∈ [a, b]. Given the arbitrariness of [a, b], we have
that (iii) holds for all x ∈ I.
Conversely, assume that (i)–(iii) are satisfied. Then again by Lemma
3.31, u belongs to ACloc (I) and the proof is complete.

The next corollary follows from the previous theorem and Corollary 3.22.
Corollary 3.32. Let I ⊂ R be an interval and let u : I → R be everywhere
differentiable. If u ∈ L1loc (I), then for all x, x0 ∈ I,
 x
u (x) = u (x0 ) + u (t) dt.
x0

Using the previous corollary, we are in a position to complete the proof

of the Katznelson–Stromberg theorem. We begin with some well-known
results on Riemann integration.
Exercise 3.33 (Riemann integration, I). Let u : [a, b] → R be a bounded
function and for x ∈ [a, b] define
ω (x) := lim sup {|u (x1 ) − u (x2 )| : x1 , x2 ∈ [a, b] ,
δ→0+
|x1 − x| ≤ δ, |x2 − x| ≤ δ}.
(i) Prove that u is continuous at x ∈ [a, b] if and only if ω (x) = 0.
88 3. Absolutely Continuous Functions

(ii) Prove that if the set

E := {x ∈ [a, b] : u is discontinuous at x}
has positive Lebesgue outer measure, then there exists a constant
α > 0 such that the set
E1 := {x ∈ E : ω (x) ≥ α}
has positive Lebesgue outer measure.
(iii) Let t := L1o (E1 ) > 0 and prove that for every partition P the
intervals containing points of E1 in their interior have total length
greater than or equal to t.
(iv) Deduce that if u is Riemann integrable, then E has Lebesgue mea-
sure zero.
Exercise 3.34 (Riemann integration, II). Let u : [a, b] → R be a bounded
function.
(i) Prove that if g : [a, b] → R is Lipschitz and g  (x) = 0 for L1 -a.e.
x ∈ [a, b], then g is constant.
(ii) Let
 x  x
g (x) := u (t) dt − u (t) dt, a < x ≤ b, g (a) := 0,
a a
*x *x
where a and a are the upper and lower Riemann integrals. Prove
that g is Lipschitz.
(iii) Prove that if u is continuous at x, then g is differentiable at x and
g  (x) = 0.
(iv) Deduce that if u is continuous L1 -a.e. in [a, b], then u is Riemann
integrable.

Proof of Theorem 2.26, continued. We claim that u cannot be Rie-

mann integrable on any closed interval [a, b]. Indeed, by the previous ex-
ercises, this would imply that u is continuous except for a set of Lebesgue
outer measure zero. But, since u is nowhere monotone, if u is continuous
at x, then necessarily u (x) = 0. Hence u = 0 L1 -a.e. on [a, b]. By the
previous corollary, we would get that u is a constant in [a, b], which is a
contradiction.

Remark 3.35. Note that the function constructed in Theorem 2.26 satis-
fies the hypotheses of Corollary 3.32. Hence, it provides an example of a
differentiable function for which the fundamental theorem of calculus holds
for Lebesgue integration but not for Riemann integration.
3.1. AC (I) Versus BP V (I) 89

Exercise 3.36. Let I ⊂ R be an interval and let f : I → R be a convex

function.
(i) Prove that the left and right derivatives f− (x) and f+ (x) exist in
R for all x ∈ I ◦ .
(ii) Prove that the functions f− and f+ are increasing.
(iii) Prove that
 y  y
(3.7) f (y) − f (x) = f− (t) dt = f+ (t) dt
x x

for all x, y ∈ I with x < y.

As a corollary of Theorem 3.30 we recover the formula for integration

by parts.

Corollary 3.37 (Integration by parts). Let I ⊂ R be an interval and let

u, v ∈ ACloc (I). Then for all x, x0 ∈ I,
 x  x

uv dt = u (x) v (x) − u (x0 ) v (x0 ) − u v dt.
x0 x0

Proof. Since u, v ∈ ACloc (I), by Exercise 3.6, uv ∈ ACloc (I), and thus, by
part (iii) of Theorem 3.30,
 x
u (x) v (x) − u (x0 ) v (x0 ) = (uv) dt.
x0

Since the functions u, v, and uv are differentiable L1 -a.e. in I, the standard

calculus rule for a product now gives the desired result.

Another consequence of Theorem 3.30 is the following result, which says

that for monotone functions, to prove absolute continuity in a fixed interval
[a, b], it is sufficient to test the fundamental theorem of calculus only at the
endpoints a and b.

Corollary 3.38 (Tonelli). Let I ⊂ R be an interval, let u : I → R be a

monotone function, and let [a, b] ⊂ I. Then u belongs to AC ([a, b]) if and
only if
 b
  
u (x) dx = |u (b) − u (a)| .
a
Moreover, if u is bounded, then u belongs to AC (I) if and only if

  
u (x) dx = sup u − inf u.
I I I
90 3. Absolutely Continuous Functions

Proof. Without loss of generality we may assume that u is increasing.

Step 1: If u ∈ AC ([a, b]), then by Theorem 3.30 we have
 b
(3.8) u (x) dx = u (b) − u (a) .
a
Conversely, assume that (3.8) holds and let
 x
w (x) := u (t) dt, x ∈ [a, b] .
a
If a ≤ x < y ≤ b, then by Corollary 1.37,
 y
(3.9) w (y) − w (x) = u (t) dt ≤ u (y) − u (x) ,
x
and so the function u − w is increasing in [a, b]. But by (3.8),
 b
u (b) − u (a) = u (x) dx = w (b) − w (a) ,
a
and so (u − w) (b) = (u − w) (a). This implies that the increasing function
u − w must be constant in [a, b]. Since w ∈ AC ([a, b]) by Lemma 3.31, it
follows that u ∈ AC ([a, b]).
Step 2: If u ∈ AC (I), then by the previous step for every inf I < a < b <
sup I, we have
 b
u (x) dx = u (b) − u (a) .
a
Letting a → (inf I) and b → (sup I)− and using the Lebesgue monotone
+

convergence theorem and the fact that u is continuous and increasing gives

(3.10) u (x) dx = sup u − inf u.
I I I

Conversely, assume that (3.10) holds. Then u is Lebesgue integrable. More-

over, by Exercise 1.39, u is continuous in I. Fix c ∈ R and define
 x
w (x) := c + u (t) dt, x ∈ I.
inf I
As in the previous step we have that u − w is increasing (and bounded)
and that (3.9) holds for all inf I < x < y < sup I. Letting x → (inf I)+
and y → (sup I)− in (3.9) and using the Lebesgue monotone convergence
theorem, the monotonicity and the continuity of u and w, we obtain that
 sup I
lim u (x) − lim u (x) = u (x) dx
x→(sup I)− x→(inf I)+ inf I
= lim w (x) − lim w (x) ,
x→(sup I)− x→(inf I)+
3.1. AC (I) Versus BP V (I) 91

and so, since u − w is increasing and continuous,

sup (u (x) − w (x)) = lim (u (x) − w (x))
x∈I x→(sup I)−
= lim (u (x) − w (x))
x→(inf I)+
= inf (u (x) − w (x)) ,
x∈I

which implies as before that u − w must be constant. By choosing c appro-

priately, we have that u = w, and so u belongs to AC (I), since w does (see
Lemma 3.31).

Since a function with bounded pointwise variation is the difference of two

bounded monotone functions, the previous corollary implies the following
result.
Theorem 3.39 (Tonelli). Let I ⊂ R be an interval, let u ∈ BP Vloc (I), and
let [a, b] ⊂ I. Then u belongs to AC ([a, b]) if and only if
 b
 
(3.11) u  dx = Var[a,b] u.
a

In addition, if u belongs to BP V (I), then u belongs to AC (I) if and only if


 
(3.12) u  dx = Var u.
I

Proof. Step 1: Let V be the increasing function defined in (2.2). By

(2.12),
 b  b
   
(3.13) u  dx ≤ V  dx ≤ V (b) − V (a) = Var[a,b] u.
a a

Hence, if
 b 
u  dx = Var[a,b] u,
a
then all the previous inequalities are equalities, and so
 b
 
V  dx = V (b) − V (a) .
a

Since V is increasing, it follows by Corollary 3.38 that the function V be-

longs to AC ([a, b]). In view of (2.3),
 
|u (bk ) − u (ak )| ≤ |V (bk ) − V (ak )| ,
k k

and so u ∈ AC ([a, b]).

92 3. Absolutely Continuous Functions

Conversely, if u ∈ AC ([a, b]), then by Theorem 3.30 for every partition

{x0 , . . . , xn } of [a, b],
 
 n  xi
   n  xi  b
n
     
|u (xi ) − u (xi−1 )| =  
u dx ≤  
u dx = u  dx.
 xi−1  xi−1 a
i=1 i=1 i=1

Taking the supremum over all partitions, we get

 b
 
Var[a,b] u ≤ u  dx,
a

which, together with (3.13) yields the desired equality.

Step 2: Assume that u ∈ BP V (I). Then by (2.10),
 
    
u (x) dx ≤ V  (x) dx ≤ sup V − inf V = Var u,
I I I I

and so if (3.12) holds, then


  
V (x) dx = sup V − inf V .
I I I

In turn, by the previous corollary, V is absolutely continuous. As in the

previous step we conclude that u is absolutely continuous.
Conversely, if u ∈ AC (I), then by the previous step (3.11) holds for
every [a, b] ⊂ I. If inf I ∈ I, define an :≡ inf I, and otherwise construct a
sequence an  inf I. Similarly, if sup I ∈ I, define bn :≡ sup I, and otherwise
construct a sequence bn  sup I. It suffices to apply the previous step in
[an , bn ] and then to let n → ∞ using Proposition 2.6 and the Lebesgue
monotone convergence theorem.

Remark 3.40. From the previous proof it follows that if I ⊂ R is an
interval, [a, b] ⊂ I, and u ∈ BP Vloc (I) (respectively, BP V (I)), then u ∈
AC ([a, b]) (respectively, AC (I)) if and only if V ∈ AC ([a, b]) (respectively,
V ∈ AC (I)).

As a consequence of Tonelli’s theorem we can prove the converse of

Corollary 3.24.
Corollary 3.41. Let I ⊂ R be an interval and let u : I → R. Then u
belongs to AC (I) if and only if
(i) u is continuous on I,
(ii) u is differentiable L1 -a.e. in I, and u belongs to L1loc (I) and is
equi-integrable,
(iii) u maps sets of Lebesgue measure zero into sets of Lebesgue measure
zero.
3.1. AC (I) Versus BP V (I) 93

Proof. In view of Theorem 3.12 and of Corollary 3.24, it remains to show

that if u ∈ AC (I), then u is equi-integrable. By Exercise 3.7 we may
assume, without loss of regularity, that I is closed. Fix ε > 0 and let δ > 0
be as in Definition 3.1. Consider a Lebesgue measurable set E ⊂ I, with
L1 (E) ≤ 2δ . By the outer regularity of the Lebesgue measure we may find an
open set A ⊃ E such that L1 (A) < δ. Decompose A into a countable family
{Jk } of pairwise disjoint intervals. By replacing each Jk with Jk ∩ I, we may
assume that Jk ⊂ I. Let Jk = [ak , bk ]. Using the fact that L1 (A) < δ, we
have that 
|bk − ak | < δ.
k
 
(k) (k)
Consider a partition Pk = x0 , . . . , xmk of [ak , bk ]. Since
mk 
  
 (k) (k) 
xi − xi−1  = |bk − ak | < δ,
k i=1 k

it follows from the fact that u ∈ AC (I) that

mk  
   
 (k) (k) 
 u xi − u xi−1  ≤ ε.
k i=1

Taking the supremum over every partition Pk of [ak , bk ] for each k, we have
that 
Var[ak ,bk ] u ≤ ε.
k
Hence, by Tonelli’s theorem applied to each interval [ak , bk ],
   
     
u  dx ≤ u  dx = u  dx = Var[ak ,bk ] u ≤ ε.
E A∩I k [ak ,bk ] k

This concludes the proof.

In the discussion before Exercise 3.7 we have shown that an absolutely

continuous function defined in an unbounded interval may not have bounded
pointwise variation. As a corollary of the previous theorem we can now char-
acterize the absolutely continuous functions that have bounded pointwise
variation.
Corollary 3.42. Let I ⊂ R be an interval and let u ∈ ACloc (I). Then u
belongs to BP V (I) if and only if u belongs to L1 (I). In this case u belongs
to AC (I).

Proof. In view of Corollary 2.23, it remains to show that if u is Lebesgue

integrable, then u belongs to BP V (I). By Proposition 3.8 we have that
94 3. Absolutely Continuous Functions

u ∈ BP Vloc (I). Hence by the Theorem 3.39, for every [a, b] ⊂ I,

 b
 
u  dx = Var[a,b] u.
a
Taking an and bn as in Step 2 of the proof of Theorem 3.39 and using
Proposition 2.6 (note that in that proposition the function u is completely
arbitrary), Exercise 2.8, and the Lebesgue monotone convergence theorem,
we have that 
 
u  dx = VarI u.
I
Since u is Lebesgue integrable, it follows that VarI u < ∞.
The last statement follows from Corollary 3.26.

Exercise 3.43. Let p ≥ 1 and let ACp ([a, b]) be the class of all functions
u : [a, b] → R such that for every ε > 0 there exists δ > 0 such that
 1/p

|u (bk ) − u (ak )|p ≤ε
k=1
for every finite number of nonoverlapping intervals (ak , bk ), k = 1, . . . , ,
with [ak , bk ] ⊂ I and
 1/p

(bk − ak )p ≤ δ.
k=1
(i) Prove that if u ∈ ACp ([a, b]), then Varp u < ∞ (see Exercise 2.29).
(ii) Prove that the function
∞
1
u (x) = n/p
cos 2n πx, x ∈ [0, 1] ,
n=0
2
is such that Varp u < ∞, but it does not belong to ACp ([0, 1]).

3.2. Chain Rule and Change of Variables

Next we discuss the validity of the chain rule and of the change of variables
for absolutely continuous functions. The next result establishes the validity
of the chain rule under very weak hypotheses.
Theorem 3.44 (Chain rule). Let I, J ⊂ R be two intervals and let f : J →
R and u : I → J be such that f, u, and f ◦ u are differentiable L1 -a.e. in
their respective domains. If f maps sets of Lebesgue measure zero into sets
of Lebesgue measure zero, then for L1 -a.e. x ∈ I,
(3.14) (f ◦ u) (x) = f  (u (x)) u (x) ,
where f  (u (x)) u (x) is interpreted to be zero whenever u (x) = 0 (even if
f is not differentiable at u (x)).
3.2. Chain Rule and Change of Variables 95

In the proof we will show that (f ◦ u) (x) = 0 and u (x) = 0 for L1 -a.e.
x ∈ I such that f is not differentiable at u (x).
To prove Theorem 3.44, we need an auxiliary result, which is a converse
of Corollary 3.14.
Lemma 3.45. Let I ⊂ R be an interval and let u : I → R. Assume
that u has derivative (finite or infinite) on a set E ⊂ I (not necessarily
measurable), with L1 (u (E)) = 0. Then u (x) = 0 for L1 -a.e. x ∈ E.

Proof. Let E ∗ := {x ∈ E : |u (x)| > 0}. We claim that L1 (E ∗ ) = 0. For

every integer k ∈ N let


∗ ∗ |x − y|

Ek := x ∈ E : |u (x) − u (y)| ≥ for all y ∈ x − k , x + k ∩ I .
1 1
k
Noting that
∞
∗
E = Ek∗ ,
k=1
we fix k and we let F := J ∩ Ek∗ , where J is an interval of length less than
∗
k . To prove that L (E ) = 0, it suffices to show that L (F ) = 0. Since
1 1 1

L (u (E)) = 0 and F ⊂ E, for every ε > 0 we may find a sequence of

1

intervals {Jn } such that

∞
 ∞
u (F ) ⊂ Jn , L1 (Jn ) < ε.
n=1 n=1
Let En := u−1 (J n) ∩ F . Since {En } covers F , we have
∞ ∞

L1o (F ) ≤ L1o (En ) ≤ sup |x − y|
n=1 n=1 x,y∈En
∞
≤ k sup |u (x) − u (y)| =: I,
n=1 x,y∈En

where we have used the fact that En ⊂ J ∩ Ek∗ . Since u (En ) ⊂ Jn , we have
sup |u (x) − u (y)| ≤ L1 (Jn ) ,
x,y∈En

and so
∞

I≤k L1 (Jn ) < kε.
n=1
It now suffices to let ε → 0+ .

Corollary 3.46. Let I ⊂ R be an interval and let u : I → R, v : I → R.
Assume that there exists a set E ⊂ I such that u and v are differentiable for
all x ∈ E and u (x) = v (x) for all x ∈ E. Then u (x) = v  (x) for L1 -a.e.
x ∈ E.
96 3. Absolutely Continuous Functions

Proof. Let w := u − v. Then w (E) = {0}, and so L1 (w (E)) = 0. By the

previous lemma, u (x) − v  (x) = w (x) = 0 for L1 -a.e. x ∈ E.

We turn to the proof of Theorem 3.44.

Proof of Theorem 3.44. Let

G := {z ∈ J : f is not differentiable at z} ,
F := {x ∈ I : u or f ◦ u is not differentiable at x} .
By hypothesis L1 (G) = L1 (F ) = 0. Let
E := {x ∈ I ◦ \ F : u (x) ∈ G} .
Since u (E) ⊂ G, we have L1 (u (E)) = 0, and since f maps sets of Lebesgue
measure zero into sets of Lebesgue measure zero, we obtain that
L1 ((f ◦ u) (E)) = 0.
By Lemma 3.45 applied to u and to f ◦ u, we conclude that
u (x) = (f ◦ u) (x) = 0
for L1 -a.e. x ∈ E.
On the other hand, if x ∈ I ◦ \ F and u (x) ∈
/ G, then we may apply
the standard chain rule to conclude that f ◦ u is differentiable at x with
(f ◦ u) (x) = f  (u (x)) u (x).

The next example shows the importance of the (N ) property.

Example 3.47. Let u : [0, 1] → R be a strictly increasing function such that
u (x) = 0 for L1 -a.e. x ∈ [0, 1] (see Theorem 1.47) and let f := u−1 . Note
that f is strictly increasing, and so by Lebesgue’s theorem it is differentiable
for L1 -a.e. x ∈ [0, 1], despite the fact that u (x) = 0 for L1 -a.e. x ∈ [0, 1].
Moreover, (f ◦ u) (x) = x for all x ∈ [0, 1], and so (f ◦ u) (x) = 1 for all
x ∈ [0, 1], while f  (u (x)) u (x) = 0 for L1 -a.e. x ∈ [0, 1], since u (x) = 0 for
L1 -a.e. x ∈ [0, 1].
Corollary 3.48. Let I, J ⊂ R be two intervals and let f : J → R and
u : I → J be such that f and u are differentiable L1 -a.e. in their respective
domains. Suppose that u is zero at most on a set of Lebesgue measure zero.
Then f ◦ u is differentiable L1 -a.e. in I and the chain rule (3.14) holds.

Proof. Let G and F be as in the proof of Theorem 3.44 and let

E := {x ∈ I ◦ : u (x) ∈ G} .
Since u (E) ⊂ G, we have L1 (u (E)) = 0, and so by Lemma 3.45 we conclude
that u (x) = 0 for L1 -a.e. x ∈ E. But then, according to our assumption
that u is zero at most on a set of Lebesgue measure zero, E must have
3.2. Chain Rule and Change of Variables 97

Lebesgue measure zero. If x ∈ I ◦ \ E, then f is differentiable at u (x), and

so the chain rule (3.14) holds L1 -a.e. in I ◦ \ E.

Corollary 3.49. Let I, J ⊂ R be two intervals and let f : J → R and
u : I → J be such that u and f ◦ u are differentiable L1 -a.e. in their
respective domains. If f ∈ ACloc (J), then the chain rule (3.14) holds.

Proof. In view of Theorem 3.12, all the hypotheses of Theorem 3.44 are
satisfied.

A less trivial consequence of Theorem 3.44 is the following result.

Corollary 3.50. Let I, J ⊂ R be two intervals, let f ∈ ACloc (J), and let
u : I → J be monotone. Then f ◦ u is differentiable L1 -a.e. in I and the
chain rule (3.14) holds.

Proof. Note that the composite f ◦ u belongs to BP Vloc (I) by Exercise

2.21, and so by Corollary 2.23, f ◦ u is differentiable L1 -a.e. in I. We are
now in a position to apply the previous corollary.

Exercise 3.51. Let I, J ⊂ R be two intervals, let f ∈ ACloc (J), and let
u : I → J be monotone and ACloc (I). Prove that f ◦ u belongs to ACloc (I).
Corollary 3.52. Let I, J ⊂ R be two intervals, let f : J → R be locally
Lipschitz, and let u ∈ BP Vloc (I). Then f ◦ u is differentiable L1 -a.e. in I
and the chain rule (3.14) holds.

Proof. By Theorem 2.31, we have that f ◦ u ∈ BP Vloc (I). Hence we can

apply Corollary 2.23 to conclude that f ◦ u is differentiable L1 -a.e. in I.
Since f is locally Lipschitz, it is locally absolutely continuous, and so the
result follows from Corollary 3.49.

Before moving to the next topic, we observe that

while all the results
proved before Theorem 3.44 continue to hold in AC I; Rd (respectively,


ACloc I; Rd ) (see Chapter 4), this is not the case for Theorem 3.44 and its
corollaries. Indeed, if f : Rd → R is a Lipschitz continuous function with
d > 1 and if u : I → Rd is absolutely continuous, then f ◦ u ∈ AC (I) (see
Step 1 of the proof of Theorem 3.68 below), but the analog of (3.14), which
is

d
∂f
(3.15) (f ◦ u) (x) = (u (x)) ui (x) ,
∂ui
i=1

where ∂u∂f
i
(u (x)) ui (x) is interpreted to be zero whenever ui (x) = 0, may
fail. This is illustrated by the next example.
98 3. Absolutely Continuous Functions

Example 3.53. Let d = 2, and consider the functions

f (z) = f (z1 , z2 ) := max {z1 , z2 } , z ∈ R2 ,
and u (x) := (x, x) for x ∈ R. Then v (x) := (f ◦ u) (x) = x so that v  (x) =
1, while the right-hand side of (3.15) is nowhere defined, since u (x) = (1, 1).

We will discuss this problem in more detail at the end of Section 4.3.
As a corollary of Theorem 3.44 we have the following change of variables
formula.
Theorem 3.54 (Change of variables). Let g : [c, d] → R be an integrable
function and let u : [a, b] → [c, d] be differentiable L1 -a.e. in [a, b]. Then
(g ◦ u) u is integrable and the change of variables
 u(β)  β
(3.16) g (t) dt = g (u (x)) u (x) dx
u(α) α

holds for all α, β ∈ [a, b] if and only if the function f ◦u belongs to AC ([a, b]),
where  z
f (z) := g (t) dt, z ∈ [c, d] .
c

Proof. If f ◦ u ∈ AC ([a, b]), then since f is absolutely continuous (see

Lemma 3.31), we can apply Corollary 3.49 to obtain the chain rule formula
(3.17) (f ◦ u) (x) = g (u (x)) u (x)
for L1 -a.e. x ∈ [a, b]. Since f ◦ u ∈ AC ([a, b]), it follows from Corollary 3.9
and (3.17) that (g ◦ u) u is integrable, and by the fundamental theorem of
calculus (see Theorem 3.30), for all α, β ∈ [a, b],
 u(β)
g (t) dt = (f ◦ u) (β) − (f ◦ u) (α)
u(α)
 β  β

= (f ◦ u) (x) dx = g (u (x)) u (x) dx.
α α
Conversely, if (g ◦ u) uis integrable and the identity
 β
(f ◦ u) (β) − (f ◦ u) (α) = g (u (x)) u (x) dx
α
holds for all α, β ∈ I, then, since the right-hand side is absolutely continuous
by Lemma 3.31, it follows that f ◦ u is absolutely continuous.

Remark 3.55. The previous proof shows, in particular, that the function
x ∈ I → g (u (x)) u (x) is measurable (see also Exercise 1.41), since it is the
derivative of the function f ◦ u, but this does not imply that the function
x ∈ I → g (u (x)) is measurable (see the next exercise).
3.2. Chain Rule and Change of Variables 99

Exercise 3.56. Let I, J ⊂ R be intervals. Prove that there exist a con-

tinuous increasing function u : I → J and a Lebesgue measurable function
g : J → R such that g ◦ u = I → R is not Lebesgue measurable. Hint: See
Exercise 1.45.
Corollary 3.57. Assume that g : [c, d] → R is an integrable function and
that u : [a, b] → [c, d] is monotone and absolutely continuous. Then (g ◦ u) u
is integrable and the change of variables formula (3.16) holds.
Exercise 3.58. Prove that under the hypotheses of the previous corollary
the function f ◦ u is absolutely continuous and then prove the corollary.
Corollary 3.59. Assume that g : [c, d] → R is a measurable, bounded func-
tion and that u : [a, b] → [c, d] is absolutely continuous. Then (g ◦ u) u is
integrable and the change of variables formula (3.16) holds.
Exercise 3.60. Prove that under the hypotheses of the previous corollary
the function f ◦ u is absolutely continuous and then prove the corollary.
Corollary 3.61. Assume that g : [c, d] → R is an integrable function, that
u : [a, b] → [c, d] is absolutely continuous, and that (g ◦ u) u is integrable.
Then the change of variables formula (3.16) holds.

Proof. Let ⎧
⎨ n if g (z) > n,
gn (z) := g (z) if − n ≤ g (z) ≤ n,
⎩
−n if g (z) < −n.
Applying the Lebesgue dominated convergence theorem and the previous
corollary, we obtain
 u(β)  u(β)
g (z) dz = lim gn (z) dz
u(α) n→∞ u(α)
 β  β
= lim gn (u (x)) u (x) dx = g (u (x)) u (x) dx.
n→∞ α α

To extend the previous results to arbitrary intervals, we consider two

functions g : J → R and u : I → J and we assume that there exist in R the
limits
lim u (x) = , lim u (x) = L.
x→(inf I)+ x→(sup I)−
In this case, the analog of (3.16) becomes
 L  sup I
(3.18) g (z) dz = g (u (x)) u (x) dx.
inf I
Indeed, we have the following result.
100 3. Absolutely Continuous Functions

Theorem 3.62. Let I, J ⊂ R be two intervals, let g : J → R be an integrable

function, and let u : I → J be differentiable L1 -a.e. in I and such that there
exist in R the limits
(3.19) lim u (x) = , lim u (x) = L.
x→(inf I)+ x→(sup I)−

Then (g ◦ u) u is integrable on I and the changes of variables (3.16) and

(3.18) hold for all [α, β] ⊂ I if and only if the function f ◦ u belongs to
AC (I) ∩ BP V (I), where
 z
f (z) := g (t) dt, z ∈ J.
inf J

Proof. Assume that f ◦ u ∈ AC (I) ∩ BP V (I). Then, we can proceed as in

the proof of Theorem 3.54 to show that (3.16) holds for all [α, β] ⊂ I. Since
f ◦ u ∈ BP V (I), by Corollary 2.23 and (3.17) its derivative (g ◦ u) u is
integrable on I. Hence, by taking limits as α and β approach the endpoints
of I and using (3.19) and the Lebesgue dominated convergence theorem, we
conclude that (3.18) holds.
Conversely, if (g ◦ u) u is integrable and (3.16) and (3.18) hold for all
[α, β] ⊂ I, then
 β
(f ◦ u) (β) − (f ◦ u) (α) = g (u (x)) u (x) dx
α

for all α, β ∈ I. As in Theorem 3.54 we deduce that f ◦u ∈ ACloc (I). In turn,

by Corollaries 3.41 and 3.42, it follows that f ◦ u ∈ AC (I) ∩ BP V (I).

Closely related to the change of variables formula is the area formula,

which will be discussed next.

Definition 3.63. Let X be a nonempty set and let ψ : X → [0, ∞] be a

function. We define the infinite sum of ψ over X as
 
 
ψ (t) := sup ψ (t) : Y ⊂ X finite .
t∈X t∈Y

Exercise 3.64. Let Xbe a nonempty set and let ψ : X → [0, ∞] be a

function. Prove that if t∈X ψ (t) < ∞, then the set
{t ∈ X : ψ (t) > 0}
is countable.

Theorem 3.65 (Area formula). Let I ⊂ R be an interval, let ψ : I → [0, ∞]

be a Borel function, and let u : I → R be differentiable L1 -a.e. in I and
3.2. Chain Rule and Change of Variables 101

such that u maps sets of Lebesgue measure zero into sets of Lebesgue measure
zero. Then
  
 
(3.20) ψ (t) dy = ψ (x) u (x) dx.
R I
t∈u−1 ({y})

Proof. Step 1: Assume first that I = (a, b) and that u ∈ Cc1 (I). Con-
sider the open set A := {x ∈ I : u (x) = 0} and let {(ak , bk )} be the count-
able family of connected components of A. If u > 0 in (ak , bk ) and ψ ∈
L∞ ((ak , bk )), then by Corollary 3.59 (applied in (ak , bk ) to the functions u
 −1
and g := ψ ◦ v, where v := u|(ak ,bk ) ) we get
 bk  u(bk )  −1 

(3.21) ψ (x) u (x) dx = ψ u|(ak ,bk ) (y) dy.
ak u(ak )

On the other hand, since u|(ak ,bk ) is strictly decreasing and continuous, for
every y ∈ (u (ak ) , u (bk )) there is one and only one t ∈ (ak , bk ) such that
u (t) = y, so that (ak , bk ) ∩ u−1 ({y}) = {t}, while, if y ∈ R \ (u (ak ) , u (bk )),
then (ak , bk ) ∩ u−1 ({y}) = ∅. This shows that
   u(bk ) 
ψ (t) dy = ψ (t) dy
R u(ak )
t∈(ak ,bk )∩u−1 ({y}) t∈(ak ,bk )∩u−1 ({y})
 u(bk )  −1 
= ψ u|(ak ,bk ) (y) dy.
u(ak )

Combining this equality with (3.21) gives

 bk  

(3.22) ψ (x) u (x) dx = ψ (t) dy.
ak R
t∈(ak ,bk )∩u−1 ({y})

To remove the additional assumption that ψ ∈ L∞ ((ak , bk )), it suffices

to apply (3.22) to ψn := min {ψ, n} and to use the Lebesgue monotone
convergence theorem.
A similar argument shows that if u < 0 in (ak , bk ), then
 bk  
  
 
ψ (x) u (x) dx = ψ (t) dy.
ak R
t∈(ak ,bk )∩u−1 ({y})

Adding over k, we obtain

  
 
ψ (x) u (x) dx = ψ (t) dy.
A R
t∈A∩u−1 ({y})
102 3. Absolutely Continuous Functions

Since u (x) = 0 in I \ A, by Corollary 3.14, L1 (u (I \ A)) = 0. Hence, the

previous equality can be rewritten as
  
  
 
ψ (x) u (x) dx = ψ (t) dy.
I R\u(I\A)
t∈A∩u−1 ({y})

On the other hand, if y ∈ R \ u (I \ A), then



u−1 ({y}) = A ∩ u−1 ({y}) ∪ (I \ A) ∩ u−1 ({y})
= A ∩ u−1 ({y}) ,
and so
   
ψ (t) dy = ψ (t) dy
R\u(I\A) R\u(I\A)
t∈A∩u−1 ({y}) t∈u−1 ({y})
 
= ψ (t) dy,
R
t∈u−1 ({y})

where in the last equality we have used the fact that L1 (u (I \ A)) = 0 once
more.
This shows that (3.20) holds.
Step 2: Assume next that I = (a, b), that there exists a compact set K ⊂ I
such that ψ = 0 on I \ K, that u is differentiable for all x in K, and that
u (y) − u (x)
lim = u (x) uniformly for x ∈ K.
y∈K, y→x y−x
Then by Exercise 3.66 below there exists a function v ∈ Cc1 (I) such that
v = u and v  = u on K. Applying the previous step to v and with ψ
replaced by χK ψ, we obtain
 
    
ψ (x) u (x) dx = χK (x) ψ (x) v  (x) dx
 
K
I 
= χK (t) ψ (t) dy
R
t∈v −1 ({y})
   
= ψ (t) dy = ψ (t) dy.
R R
t∈K∩v −1 ({y}) t∈K∩u−1 ({y})

Since ψ = 0 on I \ K, we have that (3.20) holds.

Step 3: Assume that I = (a, b). Since u is differentiable L1 -a.e. in I, the
sequence of functions
 
 u (y) − u (x) 
un (x) := sup  − u
(x) , x ∈ I,
 y − x 
y∈((x− n ,x+ n )∩I )\{x}
1 1
3.2. Chain Rule and Change of Variables 103

converges to 0 for L1 -a.e. x ∈ I. It follows by Egoroff’s theorem that there

exists an increasing sequence of compact sets {Kj } ⊂ I such that
⎛ ⎞
∞

L1 ⎝I \ Kj ⎠ = 0
j=1

and such that {un } converges to zero uniformly in Kj for every j ∈ N. In

particular,
u (y) − u (x)
lim = u (x) uniformly as x ∈ Kj
y∈Kj , y→x y−x
for all j ∈ N. By the previous step with ψ replaced by ψχKj ,
  
  
 
ψ (x) u (x) dx = ψ (t) dy.
Kj R
t∈Kj ∩u−1 ({y})

Letting j → ∞, it follows by the Lebesgue monotone convergence theorem

that  
   
ψ (x) u (x) dx = ψ (t) dy.
S ∞
Kj R S
j=1 t∈ ∞j=1 Kj ∩u
−1 ({y})

  
Since L1 I \ ∞ K
j=1 j = 0, by hypothesis we have that
⎛ ⎛ ⎞⎞
∞

L1 ⎝u ⎝I \ Kj ⎠⎠ = 0,
j=1

and so we obtain (3.20).

Step 4: If I is an arbitrary interval, let (an , bn ) ⊂ I be such that an →
(inf I)+ , bn → (sup I)− . By the previous step, (3.20) holds in each (an , bn ).
Formula (3.20) now follows in I ◦ from the Lebesgue monotone convergence
theorem. If one or both endpoints of I belong to I, we can proceed as in
last part of Step 1 to show that (3.20) holds in I.

Choosing ψ (x) := g (u (x)) in (3.20), where g : R → [0, ∞] is a Borel

function, yields
 
 
g (y) Nu (y; I) dy = g (u (x)) u (x) dx,
R I

where, we recall, Nu (·; I) is the Banach indicatrix of u. In particular, for

g = 1, we get the analog of Banach’s theorem (see (2.48)), that is,
 
 
Nu (y; I) dy = u (x) dx.
R I
104 3. Absolutely Continuous Functions

Exercise 3.66. Let K ⊂ (a, b) be a compact set and let u : K → R be such

that u is differentiable on K and
u (y) − u (x)
lim = u (x) uniformly for x ∈ K.
y∈K, y→x y−x
Prove that there exists a function v : (a, b) → R, with v ∈ Cc1 ((a, b)), such
that v = u and v  = u on K. Hint: On each connected component (ak , bk )
of (a, b) \ K define v to be a suitable third-order polynomial.

We conclude this section by discussing the analog of Theorem 2.31. The

following exercise (see Exercise 2.30) shows that the composition of abso-
lutely continuous functions is not absolutely continuous (see however Exer-
cise 3.51).
Exercise 3.67. Let f : R → R be defined by
⎧
⎨ 1 if z ≤ −1,
f (z) := |z| if − 1 < z < 1,
⎩
1 if z ≥ 1,
and let u : [−1, 1] → R be the function

2
x sin x12 if x = 0,
u (x) :=
0 if x = 0.
Prove that f and u are absolutely continuous but their composition f ◦ u is
not.

The next result gives necessary and sufficient conditions on f : R → R

for f ◦ u to be absolutely continuous for all absolutely continuous functions
u : [a, b] → R.
Theorem 3.68 (Superposition). Let I ⊂ R be an interval and let f : R →
R. Then f ◦ u ∈ ACloc (I) for all functions u ∈ ACloc (I) if and only if f is
locally Lipschitz. In particular, if f is locally Lipschitz and u ∈ ACloc (I),
then the chain rule (3.14) holds.

Proof. Step 1: Assume that f is locally Lipschitz and let u ∈ ACloc (I).
Fix an interval [a, b]. In particular, |u| is bounded in [a, b] by some constant
, and so there exists L > 0 such that
(3.23) |f (z1 ) − f (z2 )| ≤ L |z1 − z2 |
for all z1 , z2 ∈ [−, ]. We claim that f ◦ u is absolutely continuous in [a, b].
Indeed, since u ∈ AC ([a, b]), for every ε > 0 there exists δ > 0 such that
 ε
|u (bk ) − u (ak )| ≤
L
k
3.2. Chain Rule and Change of Variables 105

for every finite number of nonoverlapping intervals (ak , bk ) ⊂ [a, b], with

(bk − ak ) ≤ δ.
k

Hence, by (3.23),
 
|(f ◦ u) (bk ) − (f ◦ u) (ak )| ≤ L |u (bk ) − u (ak )| ≤ ε,
k k

which proves the claim.

The validity of the chain rule follows from Corollary 3.52.
Step 2: Assume that f ◦ u ∈ ACloc (I) for all functions u ∈ ACloc (I). We
claim that f is locally Lipschitz. The proof follows closely that of Theorem
2.31, with the only difference that instead of discontinuous functions u (see
(2.30), (2.35)) we will use piecewise affine functions. Fix [a, b] ⊂ I. We begin
by showing that f is locally bounded. Consider an interval [−r, r], where
r > 0. We claim that f is bounded in [−r, r]. Indeed, for every z0 ∈ [−r, r]
consider the function
⎧
⎨ z0 + x − 2 if x ∈ [a, b] ,
a+b

u (x) := z0 − 2
b−a
if x < a,
⎩ b−a
z0 + 2 if x > b.
Since u ∈ ACloc (I), by hypothesis (f ◦ u) ∈ AC ([a, b]). In particular, it is
bounded in [a, b]. Thus, there exists a constant Mz0 = Mz0 (a, b) > 0 such
that   
 
f z0 + x − a + b  ≤ Mt0
 2 
for all x ∈ [a, b], which implies that
|f (z)| ≤ My0


for all z ∈ z0 − b−a b−a
2 , z0 + 2 . A compactness argument shows that f is
bounded in [−r, r] by some constant Mr > 0.
Next we claim that f is Lipschitz in [−r, r]. Indeed, assume by con-
tradiction that this is not the case. Then we may find two sequences
{sn } , {tn } ⊂ [−r, r] such that sn = tn and
|f (sn ) − f (tn )|

(3.24) > 2 n2 + n
|sn − tn |
for all n ∈ N. Since {sn } is bounded, we may extract a subsequence (not
relabeled) such that sn → s∞ . Take a further subsequence (not relabeled)
such that (3.24) continues to hold and
1
(3.25) |sn − s∞ | < .
(n + 1)2
106 3. Absolutely Continuous Functions

Since f is bounded in [−r, r] by Mr , by (3.24) for all n ∈ N we have


(3.26) 2Mr ≥ |f (sn ) − f (tn )| > 2 n2 + n |sn − tn | .

Hence,
|sn − tn | (b − a) (b − a)
0 < δn := < → 0.
2Mr 2 (n2 + n)
$ %
For every n ∈ N, we divide the interval a + n+1
b−a
, a + b−a
n into subintervals
of length δn . Thus, let

diam In b−a 2Mr

(3.27) n := = = >2
δn 2
δn (n + n) |sn − tn | (n2 + n)

and set mn := max {j ∈ N0 : j < n }. Since n > 2, we have

n
(3.28) ≤ m n < n .
2
$ %
b−a b−a
Consider the partition Pn of a + n+1 , a + n given by



b−a 1 b−a
Pn := a + + jδn : j = 0, . . . , 2mn ∪ a +
n+1 2 n
 
(n) (n)
= : x0 , . . . , x2mn +1 .

We now define the piecewise affine function u : I → R in the following way.

Set s0 :=$ s1 . Define u (x)  := s∞ if x ≤ a, u (x) := s0 if x ≥ b, while in each
interval a + n+1 , a + n , n ∈ N, set
b−a b−a

(3.29) ⎧   (n) (n)

⎪
⎪ 2(sn −tn ) (n) if x2i−1 ≤ x ≤ x2i ,
⎪
⎪ x − x + t n
⎪
⎪
δn 2i−1
1 ≤ i ≤ mn − 1,
⎨   (n) (n)
u (x) := 2(tn −sn ) (n) if x2i ≤ x ≤ x2i+1 ,
⎪ x − x + sn
⎪
⎪
δn 2i
  0 ≤ i ≤ mn − 1,
⎪
⎪
⎪
⎩ (n) sn−1 −tn (n) (n) (n)
x − x2mn −1 + tn if x2mn −1 ≤ x ≤ x2mn +1 .
(n)
x2mn +1 −x2mn −1

(n) (n)
Note that δ2n ≤ x2mn +1 − x2mn −1 ≤ 2δn . We claim that u ∈ AC (I). Since
sn → s∞ and tn → s∞ , we have that u is continuous at x = a. Hence, the
function u is continuous and differentiable except for a countable number of
points, and so, in view of Exercise 3.23, to prove that it is locally absolutely
continuous in I, it remains to show that u is integrable. By (3.25), (3.26),
3.3. Singular Functions 107

and (3.28) we have

  b ∞

   
u  dx = u  dx ≤ (2mn |sn − tn | + |tn − sn−1 |)
I a n=1
∞

≤ (2n |sn − tn | + |tn − sn | + |sn − s∞ | + |s∞ − sn−1 |)
n=1
∞  2Mr + 2 ∞
Mr Mr 2
≤ 2
+ 2 + 2 ≤ < ∞.
n +n n +n n n2
n=1 n=1
∞
Since the series 1
n=1 n2 converges, we have that u is integrable in I.
On the other hand, by (3.26)–(3.28),

2m n +1       
 (n) (n) 
f u xi − f u xi−1  ≥ 2mn |f (sn ) − f (tn )|
i=1


> 2n n2 + n |sn − tn | = 4Mr ,
and so for every n ∈ N by Remark 2.7 we obtain that

n
Var[a,b] (f ◦ u) ≥ VarIk (f ◦ u) ≥ 4Mr n → ∞
k=1
as n → ∞. Hence, we have obtained a contradiction.

Remark 3.69. Note that in the necessity part of the theorem we have
actually proved a much stronger result, namely that if f : R → R is such
that f ◦ u ∈ BP Vloc (I) for all functions u ∈ AC (I) ⊂ ACloc (I), then f is
locally Lipschitz.

f : dR → R;
Remark 3.70. Note that the previous proof continues to hold if d

namely f ◦ u belongs to ACloc (I) for all functions u ∈ ACloc I; R if and

only if f is locally Lipschitz.

3.3. Singular Functions

In this section we prove that every function of bounded pointwise variation
may be decomposed into the sum of an absolutely continuous function and
a singular function.
Definition 3.71. Let I ⊂ R be an interval. A nonconstant function u : I →
R is said to be singular if it is differentiable at L1 -a.e. x ∈ I with u (x) = 0
for L1 -a.e. x ∈ I.

The jump function uJ of a function u ∈ BP Vloc (I) is an example of a

singular function. Another example is the Cantor function or the function
given in Theorem 1.47.
108 3. Absolutely Continuous Functions

The following theorem provides a characterization of singular functions.

Theorem 3.72 (Singular functions). Let I ⊂ R be an interval and let

u : I → R be a nonconstant function such that u (x) exists (possibly infinite)
for L1 -a.e. x ∈ I. Then u is a singular function if and only if there exists a
Lebesgue measurable set E ⊂ I such that L1 (I \ E) = 0 and L1 (u (E)) = 0.

Proof. Assume that u is singular and let E := {x ∈ I : u (x) = 0}. Then

L1 (I \ E) = 0. By Corollary 3.14 we have that L1 (u (E)) = 0. Con-
versely, assume that there exists a Lebesgue measurable set E ⊂ I such that
L1 (I \ E) = 0 and L1 (u (E)) = 0. Then by Lemma 3.45, u (x) = 0 for
L1 -a.e. x ∈ E. Since L1 (I \ E) = 0, we have that u (x) = 0 for L1 -a.e.
x ∈ I.

As an application of Lemma 3.31 we obtain the standard decomposition

of a monotone function into an absolutely continuous monotone function
and a singular monotone function.

Theorem 3.73. Let I ⊂ R be an interval and let u : I → R be an increasing

function. Then u may be decomposed as the sum of three increasing functions
(3.30) u = uAC + uC + uJ ,
where uAC ∈ ACloc (I), uC is continuous and singular, and uJ is the jump
function of u.

Proof. Define v := u − uJ . By Exercises 1.5 and 1.50 we have that v is

increasing, continuous, and v  (x) = u (x) for L1 -a.e. x ∈ I. Fix x0 ∈ I and
for every x ∈ I define
 x  x
(3.31) uAC (x) := v  (t) dt = u (t) dt, uC (x) := v (x) − uAC (x) .
x0 x0

Then the decomposition (3.30) holds. Moreover by Lemma 3.31 we have

uC (x) = 0 for L1 -a.e. x ∈ I. It remains to show that uC is increasing. Let
x, y ∈ I, with x < y. By Corollary 1.37,
 y
uAC (y) − uAC (x) = v  (t) dt ≤ v (y) − v (x) ,
x

and so uC (y) ≥ uC (x) by (3.31)2 .

The function uC is called the Cantor part of u.

Since every function with bounded pointwise variation may be written
as a difference of two increasing functions, an analogous result holds for
functions of bounded pointwise variation.
3.3. Singular Functions 109

Corollary 3.74. Let I ⊂ R be an interval and let u ∈ BP Vloc (I). Then u

may be decomposed as the sum of three functions in BP Vloc (I), i.e.,
(3.32) u = uAC + uC + uJ ,
where uAC ∈ ACloc (I), uC is continuous and singular, and

(3.33) uJ (x) := (u+ (y) − u− (y)) + u (x) − u− (x) .
y∈I, y<x

Moreover for every interval [a, b] ⊂ I,

(3.34) Var[a,b] u = Var[a,b] uAC + Var[a,b] uC + Var[a,b] uJ ,
where
 b 
(3.35) Var[a,b] uAC = u (x) dx,
a

(3.36) Var[a,b] uJ = (|u+ (x) − u (x)| + |u (x) − u− (x)|)
x∈(a,b)
+ |u (a) − u+ (a)| + |u (b) − u− (b)| .
If, in addition, u ∈ BP V (I), then
Var u = Var uAC + Var uC + Var uJ
 
 
= u (x) dx + Var uC + (|u+ (x) − u (x)| + |u (x) − u− (x)|) ,
I x∈I

where u− (inf I) := u (inf I) if inf I ∈ I and u+ (sup I) := u (sup I) if sup I ∈

I.

Proof. The decomposition (3.32) follows either by modifying the proof of

the previous theorem or by writing u as a difference of two increasing func-
tions (see Theorem 2.18) and applying the previous theorem to each increas-
ing function. We leave the details as an exercise.
By (2.7) for every interval J ⊂ I we have
VarJ u ≤ VarJ uAC + VarJ uC + VarJ uJ .
Thus, to prove (3.34), it remains to show
(3.37) Var[a,b] u ≥ Var[a,b] uAC + Var[a,b] uC + Var[a,b] uJ .
We divide the proof of (3.37) into five steps.
Step 1: Assume first that u is continuous, so that uJ = 0. We claim that
for every interval [α, β] ⊂ I,
(3.38) Var[α,β] u ≥ Var[α,β] uAC + |uC (β) − uC (α)| .
To see this, let  
E := x ∈ (α, β) : uC (x) = 0 .
110 3. Absolutely Continuous Functions

Fix ε > 0 and let δ > 0 be as in Definition 3.1 for the absolutely continuous
function uAC . Using the definition of differentiability, for every x ∈ E
we may find an interval (ax , bx ) ⊂ [α, β] such that ax and bx are rational
numbers, and if ax < x1 < x < x2 < bx , then
ε
(3.39) |uC (x2 ) − uC (x1 )| ≤ |x2 − x1 | .
β−α
Since, L1 (E) = β −α, from the countable cover {(ax , bx )}x∈E we may choose
a finite subcollection such that
 n 

(3.40) L 1
(axi , bxi ) ≥ β − α − δ.
i=1
By relabeling the points, if necessary, we may assume that
x1 < x2 < · · · < xn ,
and, by shortening the intervals where necessary, that bxi−1 ≤ axi for all
i = 2, . . . , n. For simplicity of notation we write ai := axi , bi := bxi for all
i = 1, . . . , n , b0 := α, and an+1 := β. Consider now the partition
P = {b0 , a1 , b1 , . . . , an , bn , an+1 }
of [α, β]. By Remark 2.7,

n 
n+1
(3.41) Var[α,β] u = Var[ai ,bi ] u + Var[bi−1 ,ai ] u.
i=1 i=1

By Corollary 2.23, the fact that (uAC ) = u L1 -a.e. in I, and Theorem

3.39, in this order,
 n  bi

n
 
(3.42) Var[ai ,bi ] u ≥ u  dx
i=1 i=1 ai
n  bi 
  n
= (uAC )  dx = Var[ai ,bi ] uAC .
i=1 ai i=1
Using (2.7), we obtain that
Var[bi−1 ,ai ] u ≥ Var[bi−1 ,ai ] uC − Var[bi−1 ,ai ] uAC
≥ |uC (ai ) − uC (bi−1 )| − Var[bi−1 ,ai ] uAC ,
which, together with (3.41) and (3.42), yields

n 
n+1
(3.43) Var[α,β] u ≥ Var[ai ,bi ] uAC + |uC (ai ) − uC (bi−1 )|
i=1 i=1

n+1
− Var[bi−1 ,ai ] uAC .
i=1
3.3. Singular Functions 111

By (3.40) we have

n+1 
n
(ai − bi−1 ) = β − α − (bi − ai ) ≤ δ.
i=1 i=1

Hence (see the proof of Corollary 3.41),


n+1
(3.44) Var[bi−1 ,ai ] uAC ≤ ε.
i=1

On the other hand, by (3.39),


n+1 
n
|uC (ai ) − uC (bi−1 )| ≥ |uC (β) − uC (α)| − |uC (bi ) − uC (ai )|
i=1 i=1

ε 
n
(3.45) ≥ |uC (β) − uC (α)| − (bi − ai )
β−α
i=1
≥ |uC (β) − uC (α)| − ε.
Combining (3.43), (3.44), and (3.45) and using Remark 2.7 for uAC , we
obtain
Var[α,β] u ≥ Var[α,β] uAC + |uC (β) − uC (α)| − 3ε.
By letting ε → 0+ , we obtain (3.38).
Step 2: Fix an interval [a, b] ⊂ I and consider a partition P of [a, b], with
a = y0 < y1 < · · · < ym = b.
Applying (3.38) in each interval [yi−1 , yi ] and using Proposition 2.6 yields

m
Var[a,b] u = Var[yi−1 ,yi ] u
i=1

m 
m
≥ Var[yi−1 ,yi ] uAC + |uC (yi ) − uC (yi−1 )|
i=1 i=1

m
= Var[a,b] uAC + |uC (yi ) − uC (yi−1 )| .
i=1

Taking the supremum over all partitions of [a, b] gives

Var[a,b] u ≥ Var[a,b] uAC + Var[a,b] uC .
Thus, we have proved that (3.37) holds if u is continuous.
Step 3: Assume that u has a finite number of discontinuity points in (a, b),
say
a < t1 < · · · < t < b.
112 3. Absolutely Continuous Functions

Let t0 := a, t +1 := b and fix δ > 0 so small that

0 < δ < min {ti − ti−1 : i = 1, . . . ,  + 1} .

Using Remark 2.7, we have


+1 
+1
Var[a,b] u = Var[ti−1 +δ,ti −δ] u + Var[ti −δ,ti +δ]∩[a,b] u
i=1 i=0

+1
≥ Var[ti−1 +δ,ti −δ] u
i=1


+ (|u (ti + δ) − u (ti )| + |u (ti ) − u (ti − δ)|)
i=1
+ |u (a + δ) − u (a)| + |u (b) − u (b − δ)|

+1


≥ Var[ti−1 +δ,ti −δ] uAC + Var[ti−1 +δ,ti −δ] uC
i=1


+ (|u (ti + δ) − u (ti )| + |u (ti ) − u (ti − δ)|)
i=1
+ |u (a + δ) − u (a)| + |u (b) − u (b − δ)| ,

where we have used the previous step together with the fact that in each
interval [ti−1 + δ, ti − δ] the jump function uJ is constant, and so the varia-
tion of u in those intervals reduces to the one of uC + uAC . Letting δ → 0+
in the previous inequality and using Remark 2.7 and Exercise 2.8 yields


+1


Var[a,b] u ≥ Var[ti−1 ,ti ] uAC + Var[ti−1 ,ti ] uC
i=1


+ (|u+ (ti ) − u (ti )| + |u (ti ) − u− (ti )|)
i=1
+ |u+ (a) − u (a)| + |u (b) − u− (b)|
= Var[a,b] uAC + Var[a,b] uC + Var[a,b] uJ .

Step 4: Finally, if u has an infinite number of discontinuity points in (a, b),

say {ti }, consider the saltus function uJ,k corresponding to the points ti with
i < k. For each k ∈ N and x ∈ [a, b] define

uk (x) := uAC (x) + uC (x) + uJ,k .

3.3. Singular Functions 113

Since the discontinuity points of uk in (a, b) are {t1 , . . . , tk }, by the previous

step
Var[a,b] uk = Var[a,b] uAC + Var[a,b] uC

k
+ (|u+ (ti ) − u (ti )| + |u (ti ) − u− (ti )|)
i=1
+ |u+ (a) − u (a)| + |u (b) − u− (b)| .
Since
Var[a,b] uk = Var[a,b] (uk − u + u) ≤ Var[a,b] (uk − u) + Var[a,b] u
∞

≤ (|u+ (ti ) − u (ti )| + |u (ti ) − u− (ti )|) + Var[a,b] u,
i=k
by the previous equality we have

k
Var[a,b] uAC + Var[a,b] uC + (|u+ (ti ) − u (ti )| + |u (ti ) − u− (ti )|)
i=1
+ |u+ (a) − u (a)| + |u (b) − u− (b)|
∞

≤ (|u+ (ti ) − u (ti )| + |u (ti ) − u− (ti )|) + Var[a,b] u.
i=k
Letting k → ∞, we obtain (3.37).
Step 5: If, in addition, u ∈ BP V (I), taking an and bn as in Step 2 of the
proof of Theorem 3.39, we apply (3.34) in [an , bn ] and use Proposition 2.6
and the Lebesgue monotone convergence theorem.