Real Analysis I

Ilkka Holopainen

October 14, 2017

2 Real Analysis I

1 Lp-spaces
1.1 Measure space
Definition 1.2. Let X be an arbitrary set and P(X) = {A : A ⊂ X} its power set. A family
Γ ⊂ P(X) is called a σ-algebra (”sigma algebra”) on X if
(1) ∅ ∈ Γ;

(2) A ∈ Γ ⇒ X \ A ∈ Γ; (denote Ac = X \ A)
S
(3) Ai ∈ Γ, i ∈ N ⇒ ∞i=1 Ai ∈ Γ.

Definition 1.3. Let Γ be a σ-algebra on X. A function µ : Γ → [0, +∞] is a (positive) measure

on X (or on the σ-algebra Γ) if
(a) µ(∅) = 0;
S∞
P
(b) Ai ∈ Γ, i ∈ N, disjont ⇒ µ i=1 Ai = i∈N µ(Ai ). ”countably additivity”
The triple (X, Γ, µ) is a measure space (and Γ is the family of µ-measurable sets).
Example 1.4. 1. X = Rn , Γ = Leb Rn = the set of all Lebesgue measurable sets and µ = mn =
Lebesgue measure.

2. X = Rn , Γ = Bor Rn = the Borel σ-algebra and µ = mn | Bor Rn = the restriction of the

Lebesgue measure to the Borel σ-algebra. (Recall: Bor Rn = the smallest σ-algebra on Rn
that contains all closed subsets of Rn .)

3. Let X 6= ∅ be an arbitrary set. Fix x ∈ X and define, for all A ⊂ X,

(
1, if x ∈ A;
µ(A) =
0, if x 6∈ A.

Then µ : P(X) → [0, +∞] is a measure (so-called Dirac measure at x ∈ X). Usually it is
denoted by µ = δx .

1.5 Complete measures

Let (X, Γ, µ) be a measure space and F ∈ Γ. Suppose that P = P (x) is some property whose
validity depends on the point x ∈ X.
We say: The property P holds µ-a.e. in F (a.e. = almost everywhere) if there exists E ∈ Γ
such that µ(E) = 0, E ⊂ F and P holds in F \ E.
Sometimes we would like to be more specific: ”P holds in F with the execption of a set of
measure 0”. The problem is the case:

{x ∈ F : P (x) does not hold} 6∈ Γ

| {z }
=A

although A ⊂ E and µ(E) = 0. Note: In the case of a general measure space it is possible that

A ⊂ E, µ(E) = 0, but A 6∈ Γ

(i.e. A is not µ-measurable).

2017 3

Example 1.6. Consider a measure space (Rn , Bor Rn , µ), µ = mn | Bor Rn . Then there exists
B ∈ Bor Rn , µ(B) = 0, and A ⊂ B s.t. A 6∈ Bor Rn .
Let’s prove the case n ≥ 2: (n = 1 later). Let A ⊂ R be a non-Lebesgue measurable set and
f : R → Rn ,
f (x) = (x, 0, . . . , 0) .
Then f is continuous and


m∗n (f A) ≤ m∗n {(x1 , . . . , xn ) ∈ Rn : xi = 0 ∀i = 2, . . . , n} = 0,

hence f A ∈ Leb Rn .
Claim: f A 6∈ Bor Rn .
Assume on the contrary: f A ∈ Bor Rn . Then the pre-image f −1 (f A) = A is a Borel set since
f is continuous [see (1.8)]. This leads to a contradiction since A is not Lebesgue measurable.

Let G ∈ Bor Rn . We say that a mapping g : G → Rm is a Borel mapping (or just Borel) if

U ⊂ Rm open ⇒ g−1 U ∈ Bor Rn .

In particular, every continuous mapping g : G → Rm , G ∈ Bor Rn , is Borel because then g−1 U is

open in G for every open U ⊂ Rm . In other words, g−1 U = G ∩ V, where V ⊂ Rn is open, and so
g−1 U ∈ Bor Rn .

Lemma 1.7. Let G ⊂ Rn be a Borel set and g : G → Rm a Borel mapping. Then

(1.8) A ∈ Bor Rm ⇒ g−1 A ∈ Bor Rn .

Proof. Let Γ = {V ⊂ Rm : g−1 V ∈ Bor Rn }. Then Γ is a σ-algebra because:

(1) g −1 ∅ = ∅ ∈ Bor Rn ⇒ ∅ ∈ Γ;

G \ g−1 V ∈ Bor Rn ⇒ V c ∈ Γ;
(2) V ∈ Γ ⇒ g −1 V c = |{z}
| {z }
∈Bor Rn ∈Bor Rn
S
S S
(3) Vi ∈ Γ, i ∈ N ⇒ g −1 i∈N Vi = g −1 V ∈ Bor Rn . ⇒ Vi ∈ Γ.
i∈N | {z }i i∈N
∈Bor Rn

Moreover, Γ contains all open subsets of Rm since:

U ⊂ Rm open ⇒ g −1 U ∈ Bor Rn ⇒ U ∈ Γ .

Hence Γ ⊃ Bor Rm (= the smallest σ-algebra on Rn that contains open sets).

The situation as in Example 1.6 (i.e. A ⊂ E, µ(E) = 0, A 6∈ Γ) does not occur if µ is so called
complete measure.

Definition 1.9. Let (X, Γ, µ) be a measure space. The measure µ is complete if

E ∈ Γ, µ(E) = 0, F ⊂ E ⇒ F ∈ Γ .

Remark 1.10. µ monotonic ⇒ µ(F ) = 0.

Completeness = ”subsets of a set of measure 0 are measurable and of measure 0”.

Example 1.11. 1. (Rn , Leb Rn , mn ), the Lebesgue measure mn is complete.

4 Real Analysis I

2. (Rn , Bor Rn , µ), µ = mn | Bor Rn is not complete.

This is a minor problem and usually does not cause problems since a (non compete) measure µ
can be completed:

Theorem 1.12. Let (X, Γ, µ) be a measure space. Define Γ̄ ⊂ P(X) by setting

Γ̄ = {A ∪ F : A ∈ Γ and F ⊂ E for some E ∈ Γ, µ(E) = 0} .

and define µ̄ : Γ̄ → [0, +∞],

µ̄(A ∪ F ) = µ(A) ,

where A and F are as above. Then

(1) Γ̄ is a σ-algebra on X;

(2) µ̄ is a complete measure;

(3) µ = µ̄|Γ.

µ̄ is called the completion of µ (and (X, Γ̄, µ̄) is the completion of (X, Γ, µ)).

A∈Γ
E ∈ Γ, µ(E) = 0
F ⊂E

Proof. Let’s proof some claims (the rest are left as an exercise).
(1) (i): ∅ ∈ Γ̄.
(ii): Let B ∈ Γ̄, B = A ∪ F, where A ∈ Γ, F ⊂ E ∈ Γ and µ(E) = 0. Claim: X \ B ∈ Γ̄. Proof:
Since



X \ B = X \ (A ∪ F ) = X \ (A ∪ E) ∪ E \ (A ∪ F )
| {z } | {z }
∈Γ ⊂E

is of desired form, we have X \ B ∈ Γ̄.

X \ (A ∪ F )
E

A F

S
(iii): If Bi ∈ Γ̄, i ∈ N, then clearly i∈N Bi ∈ Γ̄ (Exerc.).
2017 5

(2) (i): µ̄ is well-defined: Let B = A1 ∪ F1 = A2 ∪ F2 , where Ai ∈ Γ, Fi ⊂ Ei , µ(Ei ) = 0, i =

1, 2. Then

A1 ⊂ A1 ∪ F1 = A2 ∪ F2 ⊂ A2 ∪ E2

⇒ µ(A1 ) ≤ µ(A2 ) + µ(E2 ) = µ(A2 ) .

| {z }
=0

Similarly,
µ(A2 ) ≤ µ(A1 ) ,
and therefore µ(A1 ) = µ(A2 ) = µ̄(B).
(ii): µ̄ is a measure. (Exerc.)
(iii): µ̄ is complete. (Exerc.)
(3) (Exerc.)

Example 1.13. (Rn , Γ, µ), Γ = Bor Rn , µ = mn | Bor Rn .

Claim: Γ̄ = Leb Rn , µ̄ = mn . (Exerc.)

Example 1.14. Let fj : Rn → R, j ∈ N, be Borel functions, i.e.

U ⊂ R open ⇒ fj−1 U ∈ Bor Rn ,

µ = m| Bor Rn and suppose that fj → f µ-a.e., i.e.

{x ∈ Rn : fj (x) 6→ f (x)} ⊂ E ∈ Bor Rn , µ(E) = 0 .

Then we can not conclude that f is a Borel function. Instead, we can conclude that f is Lebesgue
measurable (see [Ho, L. 2.23 and 2.27]). (Reason: m complete.)

Remark 1.15. If a measure µ is complete, it makes sense to talk about measurability of functions
that are defined µ-a.e.

1.16 Cantor set in R

Let I = [0, 1] and let p = (p1 , p2 , . . .) be a sequence of real numbers 0 < pi < 1. We erase (exactly)
from the middle of I the open interval I1,1 whose length is p1 .

1 − p1
I = J1,1 ∪ I1,1 ∪ J1,2 , where J1,1 and J1,2 are closed intervals of length .
2
J1,1 I1,1 J1,2
E1

J2,1 J2,2 J2,3 J2,4

I2,1 I2,2
E2

E3
.. ..
. .
6 Real Analysis I

Next we erase from the middle of each J1,k the open interval I2,k whose length is = p2 ℓ(J1,k ) =
p2 (1−p1 )
2 .
What remains is the set

I \ (I1,1 ∪ I2,1 ∪ I2,2 ) = J2,1 ∪ J2,2 ∪ J2,3 ∪ J2,4

1 − p2 1 − p1 1
length of J2,k = · (< 2 ) .
2 2 2
′
total length of J2,k s = (1 − p1 )(1 − p2 ) .

We continue the process . . .:

Finally, what remains is the set:

∞ 2[ j−1 ∞ [
2 j
[ \
E=I\ Ij,k = Jj,k or equivalently
j=1 k=1 j=1 k=1
∞ 2 j
\ [
E= Ej , where Ej = Jj,k is compact .
j=1 k=1

In this course, we say that E = E(p) = is the Cantor set determined by the sequence p.

[Ho, L. 1.60]
E1 ⊃ E2 ⊃ · · · , m(E1 ) < ∞ =⇒
m(E) = lim m(Ej ) = lim (1 − p1 )(1 − p2 ) · · · (1 − pj )
j→∞ j→∞
Y∞
= (1 − pj ) .
j=1

If pj = 1/3 ∀j, the corresponding Cantor set E is called the Cantor 13 -set, in which case

2
j
m(E) = lim = 0.
j→∞ 3

Remark 1.17. The numbers pj can be chosen such that m(E) takes any given value on the
interval [0, 1[. (Exerc.) [Hint: By taking log from the (infinite) product that gives m(E) we obtain
an infinite series. Then choose the numbers pj so that this series is a geometric series (whose sum
we can, of couse, compute). Or more simpler: Let a = m(E) ∈ ]0, 1[ . Choose 0 < p1 < 1 s.t.
a < 1 − p1 < a + 1, p2 ∈ ]0, 1[ s.t. a < (1 − p1 )(1 − p2 ) < a + 1/2 etc.]

Theorem 1.18. The Cantor set E = E(p) satisfies:

(a) E is compact and it does not contain any open interval.

(b) If E(p) and E(q) are Cantor sets determined by sequences p and q, then there exists a home-
omorphism
f : E(p) → E(q).

(c) E is uncountable.

Remark 1.19. (i) E is closed and does not contain any (non-empty) open sets ⇒ E is nowhere
dense. [Def. A ⊂ Rn is nowhere dense if int Ā = ∅.]
2017 7

(ii) There exists a strictly increasing continuous bijection f : R → R s.t. f (Ep ) = Eq .

(iii) Recall: f : A → B is a homeomorphism if f is a continuous bijection whose inverse map f −1

is also continuous.

Proof. (a): 
Ej closed ⇒ E closed.
⇒ E compact
E ⊂ I, I compact
The construction ⇒ E does not contain open intervals.
(b): If x ∈ E, the clearly ∃ a unique sequence of closed intervals I ⊃ J1,k1 ⊃ J2,k2 ⊃ · · · s.t.
∞
\
Jj,kj = {x} .
j=1

Conversely: If I ⊃ J1,k1 ⊃ J2,k2 ⊃ · · · is a sequence of closed intervals, the intersection

∞
\
Jj,kj is a singleton, i.e. contains exactly one point,
j=1

j→∞
because m Jj,kj −−−→ 0.
Define f : E(p) → E(q) as follows:
∞
\ ∞
\
if {x} = Jj,kj (p) , then {f (x)} = Jj,kj (q) .
j=1 j=1

Note: the indexes j, kj corresponding to x and f (x) are the same.

The construction ⇒ f is bijective.
Let’s prove next that f is continuous:
Denote


δj (p) = min{m Ii,k (p) : i ≤ j} , and so

j→∞
δj (p) −−−→ 0 .

I3,1 I2,1 I1,1

E3

If x, y ∈ E(p) and |x − y| < δj (p), then necessarily x and y belong to the same interval Jj,k (p), and
therefore

f (x), f (y) ∈ Jj,k (q) (same indexes)

1
⇒ |f (x) − f (y)| ≤ m Jj,k (q) < j (the number of disjoint intervals Jj,k is 2j )
2
⇒ f continuous.
8 Real Analysis I

Concluding similarly we get that f −1 is continuos (or: f continuous bijection and E(p) compact
⇒ f homeomorphism).
(c): If m(E) > 0, then E must
be uncountable. Let then E = E(p) be such that m(E) = 0.
Choose a sequence q s.t. m E(q) > 0.

(b) ⇒ ∃ homeomorphism f : E(p) → E(q)
⇒ E(p) uncountable.
E(q) uncountable

Example 1.20. Let us now prove the claim in Example 1.6 in the remaining case n = 1. In other
words, there exists a Lebesgue measurable subset of R that is not Borel.
Proof: Choose Cantor sets E and E ′ s.t. m(E) > 0 and m(E ′ ) = 0. Then there exists a non-
Lebesgue measurable set F ⊂ E.

Thm. 1.18 ⇒ ∃ homeomorphism f : E → E ′

f F ⊂ E ′ ⇒ m∗ (f F ) = 0 ⇒ f F ∈ Leb R.

Suppose that f F is a Borel set.


E ∈ Bor R 
L. 1.7
f : E → R continuous =⇒ f −1 (f F ) = F ∈ Bor R .

f F ∈ Bor R

Contradiction since F 6∈ Leb R. Hence f F is Lebesgue measurable but not Borel.

Example 1.21. Let E be the Cantor 1/3-set. Define f : I → I, I = [0, 1], first by setting
1
f (x) = , ∀x ∈ I1,1 ,
2
1
f (x) = , ∀x ∈ I2,1 ,
4
1
f (x) = 1 − , ∀x ∈ I2,2 ,
4
..
.
1 + 2(k − 1)
f (x) = , ∀x ∈ Ij,k ,
2j
..
.
2017 9

Now
∞ 2[j−1
[
f: Ij,k → I
j=1 k=1
| {z }
=A

is increasing and ∀y ∈]0, 1[:

lim f (x) = lim f (x) .
x → y+ x → y−
x∈A x∈A

Define f (y) as the above limit at points y ∈ E \{0, 1} and as the one-sided limit at points y ∈ {0, 1}.
We obtain f : I → I that satisfies:
(a) f is a continuous and increasing surjection;

(b) f ′ (x) = 0 a.e. x ∈ I (since f ′ (x) = 0 ∀x ∈ Ij,k and m(E) = 0);

(c) f E = I (because f is constant on each open interval Ij,k ⊂ I \ E and their end points belong
to E).
The function f is called the Cantor 1/3-function ( the Cantor ternary function, ”Devil’s Staircase”).
We will return to this function later.

1.22 The space L1

Let (X, Γ, µ) be a complete measure space, i.e. µ is complete. The theory of (Lebesgue) measurable
functions and the theory of Lebesgue integration that were developed in the course ”Mitta ja
integraali” work (almost verbatim) in this more general context. Furthermore, no problems occur
with the concept of ”almost everywhere” if the measure µ is complete. Also the convergence
theorems (Monotone convergence theorem, Dominated convergence theorem, Fatou’s lemma) and
their proofs are the same as in the case of Lebegue integration. Instead Fubini’s theorems require
different proofs in the case of a general product measure µ × ν (in X × Y ) than in the case of
Lebesgue measure (see e.g. [Ru, s. 136-142]).
Remark 1.23. The notions like ”an open set, a continuous function, etc.” require a topological
space X, similarly Bor X (= the smallest σ-algebra on X containing all closed subsets of X).
Let A ⊂ X be µ-measurable (abbr. measurable), i.e. A ∈ Γ. Note: It is, in fact, misleading
to talk about µ-measurability because it is a property that depends on the σ-algebra Γ, not on
the measure µ : Γ → [0, +∞]. A function f : A → Ṙ is said to be measurable (or µ-measurable,
Γ-measuable) if

f −1 (−∞) ∈ Γ, f −1 (+∞) ∈ Γ, and f −1 U ∈ Γ ∀U ⊂ R open.

Let us define
Z
1
L (A) = {f : A → Ṙ | f measurable and |f |dµ < ∞} .
A

We also denote L1 (A, µ) and L1 = L1 (µ) = L1 (X).

Remark 1.24. If A ⊂ X is measurable and ΓA = {B ∩ A : B ∈ Γ}, then (A, ΓA , µ|ΓA ) is also
a complete measure space. Therefore it suffices (usually) to study the ”whole” measure space
(X, Γ, µ).
10 Real Analysis I

Recall: a measurable function f : X → Ṙ is integrable (over X) if

Z Z
f + dµ < ∞ and f − dµ < ∞ .
X X
In that case
Z Z Z Z Z Z
+ − +
f dµ = f dµ − f dµ and |f |dµ = f dµ + f − dµ .
X X X X X X

Thus: f ∈ L1 ⇐⇒ f : X → Ṙ is integrable.
Remark 1.25. If f : X → Ṙ is integrable, then f (x) ∈ R µ-a.e. We define f ∗ : X → R,
(
f (x), if f (x) ∈ R ,
f ∗ (x) =
0, if f (x) ∈ {−∞, +∞} .
Then f ∗ is integrable and Z Z
∗
f dµ = f dµ .
X X
Threfore we may often assume that the functions in question are real valued.
If f, g ∈ L1 , a, b ∈ R, then af + bg ∈ L1 . Hence L1 is a vector space with real coefficients.
We denote Z Z
kf k1 = kf k1,X = |f |dµ = |f |dµ .
X
Theorem 1.26. k·k satisfies:
(1) kf k1 ≥ 0,
(2) kλf k1 = |λ|kf k1 , ∀λ ∈ R,
(3) kf + gk1 ≤ kf k1 + kgk1 ,
(4) kf k1 = 0 ⇐⇒ f = 0 a.e.
Proof. Clear.

Theorem 1.26 ⇒ f 7→ kf k1 is a semi norm. It is not a norm because: kf k1 = 0 6⇒ f = 0.

Example.
X = R, µ = m, f = χQ , ⇒ kf k1 = 0, but f 6= 0 .
Definition 1.27. Functions f, g ∈ L1 ae equivalent, denoted by f ∼ g, if f = g a.e.
We denote
[f ] = f˜ = {g ∈ L1 : g ∼ f } = the equivalence class of f ,
L̃1 = {f˜: f ∈ L1 } .

L̃1 is a vector space:

[af + bg] = a[f ] + b[g] .
We define
kf˜k1 = kf k1 (well defined, does not depend on the representative f ) .
L̃1 is a normed space since in addition to the cases (1)–(3) in Theorem 1.26 we also have:
2017 11

(4’) kf˜k1 = 0 ⇐⇒ f˜ = 0,

where 0 = 0̃ = {f ∈ L1 : f = 0 a.e.}. In what follows we will get rid of the notation L̃1 and we say:
a normed space L1 . We also talk about (L1 -)functions rather than equivalence classes, that is, we
identify functions that agree almost everywhere.

1.28 The space L∞

Let (X, Γ, µ) be a complete measure space and f : X → Ṙ measurable.
We write

kf k∞ = inf {α ≥ 0 : µ {x ∈ X : |f (x)| > α} = 0} .
| {z }
=S

If S = ∅, we set kf k∞ = ∞.
R
{x ∈ X : |f (x)| > α}
α

−α

If kf k∞ < ∞, that is S 6= ∅, then we have:

[
{x ∈ X : |f (x)| > kf k∞ } ⊂ {x ∈ X : |f (x)| > kf k∞ + 1/j}
j∈N

X

⇒ µ {x ∈ X : |f (x)| > kf k∞ } ≤ µ {x ∈ X : |f (x)| > kf k∞ + 1/j} = 0
j∈N
| {z }
=0
⇒ kf k∞ ∈ S and hence |f | ≤ kf k∞ a.e.

Therefore we often denote kf k∞ = ess sup|f | (”essential supremum”). Denote

L∞ (X) = L∞ = L∞ (µ) = {f : X → Ṙ | f measurable and kf k∞ < ∞} .

Again we identify functions f, g ∈ L∞ if f = g a.e. Equivalence classes are denoted by f etc.:

L∞ = the set of equivalence classes but nevertheless we talk about functions.

Theorem 1.29. The space L∞ is a normed space equipped with the norm k·k∞ .

Proof. Clearly:

(i) L∞ is a vector space (see (iv)).

(ii) kf k∞ ≥ 0 and kf k∞ = 0 ⇐⇒ f = 0 a.e. (notice the equivalence class).

(iii) kλf k∞ = |λ|kf k∞ ∀λ ∈ R.

Moreover:
12 Real Analysis I

(iv) the triangle inequality holds:


|f | ≤ kf k∞ a.e.
⇒ |f + g| ≤ |f | + |g| ≤ kf k∞ + kgk∞ a.e.
|g| ≤ kgk∞ a.e.

Hence

µ {x ∈ X : |f (x) + g(x)| > kf k∞ + kgk∞ } = 0,
and so kf + gk∞ ≤ kf k∞ + kgk∞ and f + g ∈ L∞ .

Example 1.30. Let X = R, µ = m, and f : R → R continuous.

Claim: f ∈ L∞ ⇐⇒ f is bounded.
Proof: ⇐ clear.
⇒:
Assume on the contrary: f is not bounded, and hence ∀M > 0 ∃ x0 ∈ R s.t. |f (x0 )| > M.
Since f is continuous, we have |f (x)| > M ∀x ∈ ]x0 − δ, x0 + δ[ = J for some δ > 0.
m(J) > 0 ⇒ kf k∞ ≥ M.
M > 0 arbitrary ⇒ kf k∞ = ∞. Contradiction

1.31 The space Lp , 1 ≤ p < ∞

Let (X, Γ, µ) be a measure space, µ complete, and 1 ≤ p < ∞. Let us define
Z
p p p
L (X) = L = L (µ) = {f : X → Ṙ | f measurable and |f |p dµ < ∞} .
X
Denote Z 1/p
p
kf kp = |f | dµ .
X
We identify functions that agree a.e. as earlier. The exponent p has a great influence on which
functions belong to Lp .
Example 1.32. Let X =]0, 1[ and µ = m|]0, 1[ be the (restriction of) Lebesgue measure. If
f is measurable and bounded, then f ∈ Lp ∀p ≥ 1. (Reason: |f |p measurable and bounded,
µ(X) < ∞ ⇒ |f |p integrable, hence f ∈ Lp ).
Let
1
f (x) = √ .
x
Then

1 . 1 1−p/2 1
lim x = p < ∞, if p < 2,

 p


 a→0+ 1 − 2 a 1 − 2




Z Z 1 

p −p/2 1 . 1 1−p/2
|f | dµ = lim x dx = lim x = ∞, if p > 2,
X a→0+ a a→0+ 1 − p2 a






 .1


 lim log x = ∞, if p = 2.
a→0+ a

Hence
f ∈ Lp ⇐⇒ 1 ≤ p < 2 .
2017 13

Theorem 1.33. If µ(X) < ∞ and 1 ≤ q ≤ p, then Lp (µ) ⊂ Lq (µ).

Proof. Exercise.

We will prove next: Lp is a normed space. We need some ”tools”.

Lemma 1.34 (Young’s inequality). If a, b ≥ 0, α, β > 0 and α + β = 1, then

aα bβ ≤ αa + βb .

Proof. The case a = 0 or b = 0 is trivial, therefore we may assume that a, b > 0.

The function x 7→ log x, x > 0, is concave, and therefore


log aα bβ = α log a + β log b ≤ log(αa + βb) .

Since log is increasing, we obtain the claim.

log

log b

log(αa + βb)

α log a + β log b

αa + βb b
log a

Next we will prove a very important inequality.

1 1
Theorem 1.35 (Hölder’s inequality). If p, q > 1, p + q = 1, f ∈ Lp , and g ∈ Lq , then

f g ∈ L1 and kf gk1 ≤ kf kp kgkq , i.e.

Z Z 1/p Z 1/q
p q
|f g|dµ ≤ |f | dµ |g| dµ .
X X X

Proof. If kf kp = 0, then f = 0 a.e.,and consequently kf gk1 = 0 and the claim follows. Similarly, if
kgkq = 0. Hence we may assume that

kf kp , kgkq > 0 .

Furthermore, we may assume that f (x), g(x) ∈ R ∀x (see Remark 1.25). By applying Young’s
inequality with
|f (x)|p |g(x)|q 1 1
a= p , b = q , α= , β= ,
kf kp kgkq p q
14 Real Analysis I

we obtain (aα bβ ≤ αa + βb)

|f (x)| |g(x)| 1 |f (x)|p 1 |g(x)|q

≤ + .
kf kp kgkq p kf kpp q kgkqq

By integrating over X:n (notice that the functions are measurable) we get

kf gk1 1 kf kpp 1 kgkqq 1 1

≤ p + q = + = 1.
kf kp kgkq p kf kp q kgkq p q

Remark 1.36. The constants p, q > 1, for which p1 + 1q = 1, are called Hölder conjugates (of each
p
other). Often we denote q = p′ = p−1 . The exponent 2 is the only constant that is conjugate with
itself.

Corollary 1.37 (Schwarz inequality).

kf gk1 ≤ kf k2 kgk2 .

Example 1.38. Let X = {1, 2, . . . , n}, µ : P(X) → [0, ∞[, µ(A) = card A = the number of
elements in A.If 1p + 1q = 1 and a1 , b1 , a2 , b2 , . . . , an , bn ∈ R, then

n n
!1/p n
!1/q
X X X
|ai bi | ≤ |ai |p |bi |q .
i=1 i=1 i=1

Proof: Choose X X
f= ai χ{i} , g= bi χ{i} .
i i

In general:
X = {xi : i = 1, 2, . . .} , Γ = P(X) , and µ(A) = card A .
If f : X → R, then
Z 1/p ∞
!1/p
X
p
kf kp = |f | dµ = |f (xi )|p .
X i=1

(Note that every function f is measurable.) Denote

Lp (X) = ℓp (X) .
1 1
If f ∈ ℓp (X), g ∈ ℓq (X), p + q = 1, and ai = f (xi ), bi = g(xi ), then the Hölder inequality takes
the form
∞ ∞
!1/p ∞
!1/q
X X X
|ai bi | ≤ |ai |p |bi |q .
i=1 i=1 i=1

Theorem 1.39 (Minkowski’s inequality). If f, g ∈ Lp , then f + g ∈ Lp and

kf + gkp ≤ kf kp + kgkp .
2017 15

p 1 1
Proof. The case p = 1 is already proven. Let p > 1 and q = p−1 so that p + q = 1 (Hölder
conjugates).
If a, b ≥ 0, then
p
p
(a + b)p ≤ 2 max(a, b) = 2p max(a, b) ≤ 2p (ap + bp ) .
We may assume that f (x), g(x) ∈ R ∀x (Remark 1.25). Then


|f + g|p ≤ (|f | + |g|)p ≤ 2p |f |p + |g|p ⇒ f + g ∈ Lp .

On the other hand,

|f + g|p = |f + g||f + g|p−1 ≤ |f ||f + g|p−1 + |g||f + g|p−1

and
q f +g∈Lp
|f + g|p−1 = |f + g|p =⇒ |f + g|p−1 ∈ Lq .
By Hölder’s inequality we get
Z Z Z
p p p−1
kf + gkp = |f + g| ≤ |f | |f + g| + |g| |f + g|p−1
|{z} | {z } |{z} | {z }
∈Lp ∈Lq ∈Lp ∈Lq
Z 1/q Z 1/q


p−1 q


p−1 q
≤ kf kp |f + g| + kgkp |f + g|
Z 1/q Z 1/q
= kf kp |f + g|p + kgkp |f + g|p


= kf kp + kgkp kf + gkp/q
p

p/q=p−1
=⇒ kf + gkp ≤ kf kp + kgkp .

We have proved:

Theorem 1.40. Lp is a normed space when equipped with the norm k·kp .

Example 1.41. 1. If µ(X) = ∞, it may happen that Lp 6⊂ Lq , q < p: Let X = R and µ = m.

1
f (x) = f ∈ Lp , if p > 1 ,
1 + |x|
f 6∈ L1 .

2. In general Lp 6⊂ Lq , p 6= q. Above is the case q < p. Earlier: X =]0, 1[ , µ = m and

1
f (x) = √
x
p
f ∈ L , if 1 ≤ p < 2
f 6∈ Lq , if q ≥ 2 .
16 Real Analysis I

1.42 Completeness of Lp spaces

In this section we will prove that normed spaces Lp , 1 ≤ p ≤ ∞, are Banach spaces, that is
complete normed spaces.
Some terminology: Let (Y, d) be a metric space. We say that a sequence (xj ), xj ∈ Y, is a
Cauchy sequence in Y if for every ε > 0 there exists iε ∈ N s.t. d(xi , xj ) < ε for every i, j ≥ iε .
A metric space (Y, d) is complete if every Cauchy sequence in Y converges towards a point in Y .
Recall that a sequence (xj ) converges to x ∈ Y if d(xj , x) → 0 as i → ∞.
Let (V, k·k) be a normed space. It is also a metric space equipped with the natural metric

d(x, y) = kx − yk .

A normed space (V, k·k) is a Banach space if, for every Cauchy sequence (xj ) in Y , there exists
x ∈ V s.t.
j→∞
kxj − xk −−−→ 0 .

Let (X, Γ, µ) be a complete measure space.

We say that (fj ) converges to f in Lp , denoted by fj → f in Lp , if fj , f ∈ Lp and kfj − f kp → 0
as j → ∞.

Theorem 1.43. If (fj ) is a Cauchy sequence in Lp , 1 ≤ p < ∞, then there exists a subsequence
(fjk ) that converges (pointwise) almost everywhere.

Proof. For every k ∈ N we choose jk s.t.

1
(1) kfi − fj kp < 2k
whenever i, j ≥ jk ,

(2) j1 < j2 < · · · .

Note that by Remark 1.25 we may assume that all functions (above) are real valued. We define

gk = |fj1 | + |fj2 − fj1 | + · · · + |fjk+1 − fjk | .

(gk ) increasing sequence ⇒ ∃ g = lim gk .

k→∞

By Minkowski’s inequality,

k
X

kgk kp = |fj1 | + |fjν+1 − fjν | p
ν=1
k
Minkowski X
≤ kfj1 kp + kfjν+1 − fjν kp
ν=1
k
X 1
≤kfj1 kp + ν
≤ kfj1 kp + 1 ∀k
ν=1
2
2017 17

The Monotone convergence theorem (MCT) then implies that

Z Z
MCT
p
gp = lim gkp = lim kgk kpp ≤ kfj1 kp + 1 < ∞
k→∞ k→∞
⇒ g(x) < ∞ a.e.

Hence the series

∞
X

fj1 (x) + fjν+1 (x) − fjν (x)
ν=1

converges a.e. Denote its sum by f (x),

∞
X

f (x) = fj1 (x) + fjν+1 (x) − fjν (x) .
ν=1

We obtain
k
X

fjk+1 = fj1 + fjν+1 − fjν → f a.e.
ν=1

Remark 1.44. The condition fj → f in Lp does not (in general) imply that (the whole sequence)
fj → f a.e.
Example: Let Ik be the closed subinterval of I = [0, 1] as in the picture.
I1

I2 I3

I4 I5 I6 I7

jne

Let fk = χIk : I → R. Then fk ∈ Lp , ∀p ∈ [1, ∞) and

Z 1/p
k→∞
kfk − 0kp = χIk dm = m(Ik )1/p −−−→ 0 .
I

Hence fk → 0 in Lp .
Claim: fk (x) 6→ 0 as k → ∞ for any x ∈ I.
Proof: Let x ∈ I and k0 ∈ N be arbitrary.
[
Ik = I ⇒ ∃ k1 > k0 s.t. x ∈ Ik1 and fk1 (x) = 1 .
k>k0

Theorem 1.45. Lp is a Banach avaruus for every 1 ≤ p ≤ ∞.

Remark 1.46. The case 1 ≤ p < ∞ is so-called Riesz-Fischer theorem (1906).

18 Real Analysis I

Proof. (a) 1 ≤ p < ∞: Let (fj ) be a Cauchy sequence in Lp . Theorem 1.43 implies that there
exists a subsequence (fjk ) s.t. fjk → f a.e.
Claim: f ∈ Lp and fjk → f in Lp .
Proof: Let ε > 0. Then there exists j0 ∈ N s.t.

kfi − fj kp < ε as i, j ≥ j0 .

If j ≥ j0 , we have
Z Z Z
Fatou
p p
|fj − f | dµ = lim |fj − fjk | dµ ≤ lim inf |fj − fjk |p dµ
k→∞ k→∞
= lim inf kfj − fjk kpp ≤ εp
k→∞

 
p
 fj − f ∈ L
  f = fj − (fj − f ) ∈ Lp
⇒ ⇒
 j→∞ 
 kf − f k −
j p −−→ 0 fj → f in Lp .

(b) p = ∞: Let (fj ) be a Cauchy sequence in L∞ . Denote

Aj = {x : |fj (x)| > kfj k∞ }

Aj,k = {x : |fj (x) − fk (x)| > kfj − fk k∞ } .

Then µ(Aj ) = 0 = µ(Aj,k ) (this follows directly from the definition of k·k∞ ). Denote
[ [
A= Aj ∪ Aj,k , hence µ(A) = 0 .
j j,k

If x ∈ Ac , then

(1.47) |fj (x) − fk (x)| ≤ kfj − fk k∞ .

Thus fj (x) is a Cauchy sequence in R, and therefore the sequence fj (x) converges. Define

f (x) = lim fj (x) .

j→∞

The Cauchy criterion for the uniform convergence and (1.47) imply that

fj → f uniformly in Ac .

For x ∈ A, we set f (x) = 0. By [Ho, L. 2.29] f is measurable.

Claim: f ∈ L∞ and kfj − f k∞ → 0.
Proof: Let j0 ∈ N s.t.
kfj − fk k∞ < 1 , if j, k ≥ j0 .
Since k·k∞ is a norm, we have

kfj k∞ ≤ kfj0 k∞ + kfj − fj0 k∞ ≤ kfj0 k∞ + 1 =: M

2017 19

for j ≥ j0 . If x ∈ Ac , we have

|fj (x)| ≤ kfj k∞ , and so

|f (x)| = lim |fj (x)| ≤ M
j→∞

⇒ kf k∞ ≤ M , because µ(A) = 0 .
⇒ f ∈ L∞ .

(In fact, |f (x)| ≤ M ∀x.)

Since fj → f uniformly in Ac and µ(A) = 0, we have kfj − f k∞ → 0.

Remark 1.48. The theory of Lp spaces generalizes to mappings f : X → Rm , f = (f1 , . . . , fm ).

Then the norm is Z 1/p
p
kf kp = |f | dµ ,
X

1/2
where |f (x)| = f1 (x)2 + · · ·+ fm(x)2 is the Euclidean norm. Similarly, for functions f : X → C.

2 Approximation in Lp spaces
2.1 Absolute continuity of measures
Let (X, Γ, µ) be a measure space and let σ : Γ → [0, +∞] be another measure.

Definition 2.2. A measure σ is absolutely continuous with respect to µ (denoted by σ ≪ µ) if

σ(E) = 0 for every E ∈ Γ with µ(E) = 0.

Example 2.3. 1. Let f : X → [0, +∞] be Γ-measurable. Define

Z
σ(A) = f dµ , A ∈ Γ .
A

By properties of an integral ([Ho, Lause 3.32]) σ is a measure and

µ(A) = 0 ⇒ σ(A) = 0 .

Hence σ ≪ µ. (See the next Remark for the converse direction.)

2. Let X = R and σ : Leb R → [0, +∞] be the counting measure. Then the Lebesgue measure
m({0}) = 0 but σ({0}) = 1, and therefore σ 6≪ m.

3. Let X = R, x ∈ R, and δx : Leb R → [0, +∞] be the Dirac measure at x (or, in fact, the
restriction of the Dirac measure to Leb R). Then m({x}) = 0 but δ({x}) = 1, and therefore
δ 6≪ m.

Remark 2.4. Let (X, Γ, µ) be a measure space and ϕ : Γ → Ṙ. We say that

(a) ϕ is countably additive if

(i) ϕ(∅) = 0,
20 Real Analysis I

P
(ii) if A1 , A2 , . . . ∈ Γ are pairwise disjoint, then i ϕ(Ai ) is defined and
X

ϕ(Ai ) = ϕ ∪i Ai .
i

(b) ϕ is absolutely continuous with respect to µ if µ(A) = 0 ⇒ ϕ(A) = 0. Then we denote ϕ ≪ µ.

(c) ϕ is σ-finite if
X = ∪j Aj , Aj ∈ Γ, |ϕ(Aj )| < ∞.

Unfortunately, we shall not prove the following Radon-Nikodym theorem in this course: If (X, Γ, µ)
is σ-finite measure space and if ϕ : Γ → Ṙ is countably additive, σ-finite, and ϕ ≪ µ, then there
exists a measurable function f : X → R s.t.
Z
ϕ(E) = f dµ, ∀E ∈ Γ.
E

If, in addition, g is another function satisfying the equation above, then f = g a.e.
Theorem 2.5. Let (X, Γ, µ) be a measure space σ : Γ → [0, +∞) a measure s.t. σ(X) < ∞. Then

σ ≪ µ ⇐⇒
(2.6) ∀ε > 0 ∃ δ > 0 s.t. µ(A) < δ ⇒ σ(A) < ε .

Proof. ⇒ Suppose that σ ≪ µ.

Assume on the contrary that there exists ε > 0 and a sequence E1 , E2 , . . . ∈ Γ s.t.

σ(Ei ) ≥ ε and µ(Ei ) < 2−i .

Denote
[ ∞
\
Ak = Ei , A= Ak .
i≥k k=1
Ak ⊃ Ek ⇒ σ(Ak ) ≥ σ(Ek ) ≥ ε ∀k .


A1 ⊃ A2 ⊃ · · ·
⇒ σ(A) = lim σ(Ak ) ≥ ε .
σ(A1 ) ≤ σ(X) < ∞ k→∞

∞
X 1 1
µ(A) ≤ µ(Ak ) ≤ i
= k−1 ∀k
2 2
i=k

σ≪µ
⇒ µ(A) = 0 =⇒ σ(A) = 0 . Contradiction

⇐ If the condition (2.6) holds, then σ ≪ µ trivially (µ(A) = 0 ⇒ σ(A) < ε ∀ε > 0 ⇒ σ(A) =
0).

Corollary 2.7. Let f ∈ L1 . Then ∀ε > 0 ∃ δ > 0 s.t.

Z
µ(E) < δ ⇒ |f |dµ < ε .
E
2017 21

Proof. We can apply Theorem 2.5 to the measure

Z
σ(E) = |f |dµ E∈Γ
E

since σ ≪ µ and σ(X) < ∞ (because f ∈ L1 ).

2.8 Egorov’s Theorem and Lusin’s Theorem

Let (X, Γ, µ) be a measure space. In ”Mitta and integraali”-course we proved ([Ho, Lause 3.14]):
Theorem 2.9. If f : X → [0, ∞] is measurable, there exists an increasing sequence 0 ≤ f1 ≤ f2 ≤
· · · of simple functions s.t.
f = lim fj .
j→∞

Remark 2.10. 1. g : X → [0, ∞) is simple if

k
X
g= ai χAi , ai ≥ 0, Ai ∈ Γ disjoint.
i=1

2. The proof of Theorem 2.9 is the same as the one in the case of (Rn , Leb Rn , m).

3. If the function f in Theorem 2.9 is bounded, then fj → f uniformly in X, i.e. ∀ε > 0 ∃ iε ∈ N

s.t. |fj (x) − f (x)| < ε for every x ∈ X (the index iε is independent of x).

4. f : X → [0, ∞] is measurable ⇐⇒ f = limj→∞ fj , where (fj ) is an increasing sequence of

simple functions fj : X → [0, ∞).
In general: The convergence is uniform in a large portion of X as the following theorem reveals.
Theorem 2.11 (Egorov’s Theorem). Let µ be complete, µ(X) < ∞, and let functions fk : X →
R, k = 1, 2, . . . , be measurable s.t. fk → f a.e., where f : X → R. Then
1. ∀ε > 0 ∃ a measurable F ⊂ X s.t. µ(X \ F ) < ε and fk |F → f |F uniformly;

2. if X ⊂ Rn and µ = m = is the Lebesgue measure, the set F can be chosen to be compact.

Lemma 2.12. Let A ⊂ Rn be measurable and ε > 0. Then
1. ∃ open G ⊃ A s.t. m(G \ A) < ε;

2. ∃ closed F ⊂ A s.t. m(A \ F ) < ε.

3. If, in addition, m(A) < ∞, then ∃ compact F ⊂ A s.t. m(A \ F ) < ε.

Proof. (Extra) exercise.

Proof of Egorov’s Theorem. 1. µ complete ⇒ f measurable ([Ho, Lause 2.29]). Denote

∞
\ 1
Ek,l = {x ∈ X : |fm (x) − f (x)| < } , k, l ∈ N ,
k}
m=l | {z
measurable

H = {x ∈ X : lim fm (x) = f (x)} .

m→∞
22 Real Analysis I

1
If x ∈ H and k ∈ N, then ∃ lk ∈ N s.t. |fm (x) − f (x)| < k ∀m ≥ lk ⇒ x ∈ Ek,lk . Hence
∞
[
H⊂ Ek,l ∀k .
l=1

The sets H and Ek,l are measurable and µ(H) = µ(X) since fm → f a.e.

Ek,l ⊂ Ek,l+1 ⇒
∞
[

(2.13) µ(X) ≥ lim µ(Ek,l ) = µ Ek,l ≥ µ(H) = µ(X)
l→∞
l=1

µ(X) < ∞, (2.13) ⇒

lim µ(X \ Ek,l ) = µ(X) − lim µ(Ek,l ) = 0 ∀k .
l→∞ l→∞

Let ε > 0. Then ∀k ∃ lk ∈ N s.t.

ε
µ(X \ Ek,lk ) < .
2k
Claim: The
∞
\
F = Ek,lk
k=1

satisfies the desired conditions.

Proof. F measurable and
∞ ∞ ∞
[
X X ε
µ(X \ F ) = µ (X \ Ek,lk ) ≤ µ(X \ Ek,lk ) < = ε.
2k
k=1 k=1 k=1

Furthermore F ⊂ Ek,lk ∀k and

1
|fm (x) − f (x)| < , if x ∈ Ek,lk and m ≥ lk .
k

1
⇒ |fm (x) − f (x)| < , if x ∈ F and m ≥ lk .
k
⇒ fm |F → f |F uniformly (the index lk is independent of x ∈ F ).

2. Suppose now that µ = m = is the Lebesgue measure. Lemma 2.12 ⇒ ∃ compact F0 ⊂ F

s.t. µ(F \ F0 ) < ε. Thus
µ(X \ F0 ) ≤ µ(X \ F ) + µ(F \ F0 ) < 2ε.

A measurable function is continuous in a large portion of its domain:

Theorem 2.14 (Lusin’s Theorem). Let A ⊂ Rn be measurable, m(A) < ∞, and let f : A → R be
measurable. Then

∀ε > 0 ∃ compact F ⊂ A s.t. m(A \ F ) < ε and f |F is continuous.

2017 23

Proof. Let ε > 0. (a): Suppose that f is simple

k
X
f= ai χ A i .
i=1

Lemma 2.12 ⇒ ∃ compact sets Fi ⊂ Ai s.t. m(Ai \ Fi ) < ε/k. Then F = F1 ∪ · · · ∪ Fk is compact
and Fi ’s are disjoint. If x ∈ F, then ∃ r > 0 s.t. B(x, r) ∩ Fi 6= ∅ for exactly one i.
Fj compact
(Reason: x ∈ Fi ⇒ dist (x, Fj ) = inf |x − y| = min |x − y| > 0 ∀j 6= i .)
y∈Fj y∈Fj

Hence
f (y) = f (x) = ai ∀y ∈ B(x, r) ∩ F ,
and therefore f |F is locally constant and hence f |F is continuous. Also
k
[ k

disjoint X
m(A \ F ) = m (Ai \ Fi ) = m(Ai \ Fi ) < ε .
i=1 i=1

(b): Suppose then that f ≥ 0 is measurable. L. 2.9 ⇒ ∃ simple functions fj s.t. fj ր f. The part
(a) of the proof ⇒ ∃ compact sets Fj ⊂ A s.t.
ε
fj |Fj continuous and m(A \ Fj ) < j .
2
T
Let F0 = j Fj , then
[
X X ε
m(A \ F0 ) = m (A \ Fj ) ≤ m(A \ Fj ) < = ε.
2j
j j j

Egorov’s Theorem ⇒
∃ compact F ⊂ F0 s.t. fj |F → f |F uniformly and m(F0 \ F ) < ε .
Now

fj |F compact
⇒ f |F continuous,
fj |F → f |F uniformly

moreover m(A \ F ) = m(A \ F0 ) + m(F0 \ F ) < 2ε .

(c): Suppose finally that f is measurable and write f = f + − f − . The part (b) of the proof ⇒ ∃
compact sets F1 , F2 ⊂ A s.t. f + |F1 , f − |F2 are continuous and m(A \ Fi ) < ε/2, i = 1, 2. The set
F = F1 ∩ F2 fulfils the requirements.
Remark 2.15. 1. The assumption that f is real valued is essential in Egorov’s and Lusin’s
theorems, that is the values ±∞ are not allowed.
2. The assumption µ(X) < ∞ is essential in Egorov’s theorem: Example: X = R, fj = χ[j,∞[.
Then fj (x) → 0 ∀x ∈ R. Denote f = 0. If F ⊂ R s.t. fj |F → f |F uniformly, then ∃j0 s.t.
1
|fj (x) − f (x)| = |fj (x)| <, if j ≥ j0 , x ∈ F
2
⇒ F ∩ [j, ∞[= ∅ ∀j ≥ j0
⇒ [j, ∞[ ⊂ R \ F ⇒ m(R \ F ) = ∞ .
24 Real Analysis I

3. Lusin’s theorem holds also in the case m(A) = ∞ if we just require that the set F is closed.
(Exerc.)

2.16 Convolution in Rn
In this section we consider the measure space (Rn , Leb Rn , m).
Let f, g ∈ L1 (Rn ). The mapping ϕ : Rn → Rn , ϕ(y) = x − y, where x ∈ Rn is a constant,
satisfies:
ϕ(A) measurable ⇐⇒ A measurable.
Hence

y 7→ f (x − y) is measurable
⇒ y 7→ f (x − y)g(y) is measurable.

Therefore the integral Z

h(x) = f (x − y)g(y)dm(y)
Rn
is defined if f, g ≥ 0.
Questions: When h(x) < ∞? Can h be defined without assumptions f, g ≥ 0?

Theorem 2.17. Suppose that f, g ∈ L1 (Rn ). Then

Z
(2.18) |f (x − y)||g(y)|dm(y) < ∞ a.e. x ∈ Rn .
Rn

For these x we denote

Z
(2.19) h(x) = f (x − y)g(y)dm(y) .
Rn

Then h ∈ L1 (Rn ) and

(2.20) khk1 ≤ kf k1 kgk1 .

The function h is called the convolution of f and g and it is denoted by h = f ∗ g.

Proof. We are going to apply Fubini’s theorem to the function F : Rn × Rn → R, (x, y) 7→ f (x −

y)g(y), and therefore we must prove that it is measurable. We start with:
Claim: ∃ Borel functions f0 , g0 : Rn → R s.t.

f0 = f a.e.
g0 = g a.e.

(i.e. f0−1 U, g0−1 U ∈ Bor Rn ∀ open U ⊂ R.)

Proof: f + measurable ⇒ ∃ a sequence of simple functions 0 ≤ f1 ≤ f2 ≤ · · · s.t.

fj ր f + ,
k
X
fj = ai χ A i .
i=1
2017 25

Choose1 Borel sets Bi ⊂ Ai s.t. m(Ai \ Bi ) = 0. Then

k
X
ϕj = ai χBi is a Borel function, 0 ≤ ϕj ≤ fj and ϕj = fj a.e.;
i=1
ϕ+ = lim inf ϕj Borel function and ϕ+ = f + a.e.
j→∞

(Note that (ϕj ) need not be increasing ⇒ limj→∞ ϕj need not exist.)
Similarly ∃ a Borel function ϕ− = f − a.e.
Now f0 = ϕ+ − ϕ− is a Borel function and f0 = f a.e.
Similarly for g.
The values of integrals (2.18) and (2.19) do not change if f and g are replaced by f0 and g0 . ⇒
We mays assume: f, g are Borel functions.
Claim: F : Rn × Rn → R, F (x, y) = f (x − y)g(y), is a Borel function.
Proof: The mappings u : Rn × Rn →
Rn , u(x, y) n n n

= x − y, and v : R × R → R , v(x, y) = y, are
continuous. Now F (x, y) = f u(x, y) g v(x, y) , eli

F = (f ◦ u)(g ◦ v).

Let V ⊂ R be open. Since f is a Borel function, we have f −1 V ∈ Bor Rn . Furthermore,

(f ◦ u)−1 V = u−1 ( f −1 V ) ∈ Bor R2n

| {z }
∈Bor Rn

since u : R2n → Rn is continuous (ks. L. 1.7). Hence f ◦ u is Borel. Similarly, we see that g ◦ v is
Borel, and therefore is a Borel function as a product of two Borel functions. By Exercise 1/4, the
”original” F agrees with the ”new” F a.e. in R2n , and therefore the original F is measurable.
We may thus apply Fubini’s theorems. Fubini 1. ⇒
Z Z  Z Z 
|F (x, y)|dy dx = |F (x, y)|dx dy
Rn Rn Rn Rn
Z  Z 
= |g(y)| |f (x − y)|dx dy
Rn Rn
= kf k1 kgk1 < ∞ ,

because
Z
(2.21) |f (x − y)|dx = kf k1 .
Rn

Hence (2.18) holds.

Fubini 2. ⇒ h ∈ L1 (Rn ) and
Z Z Z
 
khk1 = |h(x)|dx =  f (x − y)g(y)dy dx
n Rn n
ZR Z R
≤ |F (x, y)|dy dx = kf k1 kgk1 .
Rn Rn

1
Lemma 2.12 ⇒ ∃ Fσ -set Bi ⊂ Ai s.t. m(Ai \ Bi ) = 0.
26 Real Analysis I

Remark 2.22. The equation (2.21) holds by the translatoin invariance of the Lebesgue measure:
If
Xk
f= aj χ A j
j=1

is simple and ϕ(x) = x − y is a translation, then ϕ−1 (x) = x + y,

k
X
f ◦ϕ= aj χϕ−1 Aj and m(ϕ−1 Aj ) = m(Aj )
j=1
Z k
X k
X Z
⇒ f ◦ϕ = aj m(ϕ−1 Aj ) = aj m(Aj ) = f.
j=1 j=1

In other words, (2.21) holds for simple functions. After this, the general case follows from the
definition of integral.

Question: Why we used Borel sets/functions and not just measurable sets/functions?
Reason: g : Rn → Rm measurable and E ∈ Leb Rm 6⇒ g −1 E ∈ Leb Rn .

2.23 Approximation by C ∞ -functions

First some notation:
Let A ⊂ Rn , f : A → R. We denote

spt f = A ∩ {x ∈ A : f (x) 6= 0} (support of f ),

0
f ∈ C(A) = C (A) ⇐⇒ f continuous.

Let U ⊂ Rn be open, k ∈ N, f : U → R.

f ∈ C k (U ) ⇐⇒ f has continuous partial derivatives of order k

( ⇐⇒ f is k times continuously differentiable),
f ∈ C ∞ (U ) ⇐⇒ f ∈ C k (U ) ∀k,
f ∈ C0k (U ) ⇐⇒ f ∈ C k (U ) and spt f ⊂ U compact,
f ∈ C0∞ (U ) ⇐⇒ f ∈ C ∞ (U ) and spt f ⊂ U compact.

Denote also f ∈ C k etc.

Theorem 2.24. If f ∈ L1 (Rn ) and g ∈ C0 (Rn ), then f ∗ g ∈ C(Rn ).

Proof. Let’s make a simple change of variables x − y 7→ y. The value of the integral does not change
(cf. Remark 2.22), and therefore
Z
f ∗ g(x) = f (y)g(x − y)dy
Rn
2017 27

that is defined for all x, because

Z Z
|f (y)g(x − y)|dy ≤ M |f (y)|dy = M kf k1 < ∞ ∀x .
Rn Rn

Here M = max|g| (the maximum exists since g is continuous and compactly supported).
Z 


f ∗ g(x + h) − f ∗ g(x) = f (y) g(x − y + h) − g(x − y) dy  

Rn ⇒ ∀ε > 0 ∃ δ > 0 s.e.


g uniformly continuous in Rn

Z
|f ∗ g(x + h) − f ∗ g(x)| ≤ |f (y)| |g(x − y + h) − g(x − y)| dy
Rn | {z }
<ε
n
< ε kf k1 , for |h| < δ and x ∈ R
| {z }
<∞
⇒ f ∗ g continuous at x.

Remark 2.25. The proof above implies that f ∗ g is uniformly continouos in Rn .

Theorem 2.26. If f ∈ L1 (Rn ) and g ∈ C0k (Rn ), then f ∗ g ∈ C k (Rn ).

Proof. For (z, t) ∈ Rn × (R \ {0}), we set

g(z + tei ) − g(z)

ϕ(z, t) = − Di g(z) , where e1 , . . . , en is the standard basis of Rn .
t
The mean value theorem ⇒

(2.27) ϕ(z, t) = Di g(z + ϑtei ) − Di g(z) , for some 0 < ϑ < 1 .

(2.27)
g ∈ C0k (Rn ) ⇒ Di g uniformly continuous in Rn =⇒ ϕ(z, t) → 0 uniformly in Rn as t → 0, hence
t→0
σ(t) = sup |ϕ(z, t)| −−→ 0 .
z∈Rn

Let x ∈ Rn . Then
 f ∗ g(x + te ) − f ∗ g(x) 
 i 
lim  − f ∗ Di g(x)
t→0
 t  
Z g(x + tei − y) − g(x − y)
 
= lim  f (y) − Di g(x − y) dy 
t→0 n t
Z R
≤ lim |f (y)| |ϕ(x − y, t)| dy
t→0 Rn | {z }
≤σ(t)

≤ lim σ(t)kf k1 = 0 ,
t→0

and therefore

Di f ∗ g (x) = f ∗ Di g(x) .
28 Real Analysis I

2.24
Di g ∈ C0 (Rn ) =⇒ Di (f ∗ g) ∈ C(Rn ). Repeating the above we get


D f ∗ g = f ∗ Dg ,

where D is any partial derivative of order p ≤ k.

Remark 2.28. Theorems 2.24 and 2.26 hold also for functions f ∈ Lp (Rn ), 1 ≤ p ≤ ∞. Reason:
in the proofs we need not integrate over the whole Rn , it suffices to integrate over a sufficiently
large compact set K since g is compactly supperted. Indeed, by Hölder’s inequality
Z Z
p−1  1/p p−1
|f | ≤ m(K) p |f |p ≤ m(K) p kf kp
K K

p−1
if p > 1 (we interprete above p = 1 if p = ∞).

Our next goal is to use convolution in approximating Lp functions (1 ≤ p < ∞) by C0∞ func-
tions.
For this purpose:

Theorem 2.29. If f ∈ Lp (Rn ), 1 ≤ p < ∞, then

Z
lim |f (x + h) − f (x)|p dx = 0 .
h→0 Rn

Proof. Let ε > 0. We need to show: ∃ η > 0 s.t.

Z
|h| < η ⇒ |f (x + h) − f (x)|p dx < ε .
Rn

Denote
Z
Ih (A) = |f (x + h) − f (x)|p dx , A ∈ Leb Rn ,
A
Bk = B(0, k) = {x ∈ Rn : |x| < k} , k > 1.

Let h ∈ B(0, 1). (Then |x| ≥ k ⇒ |x + h| ≥ k − 1. )

Now
Z Z


|f (x + h) − f (x)|p dx ≤ 2p |f (x + h)|p + |f (x)|p dx
Rn \Bk Rn \Bk
Z Z !
DCT
≤ 2p |f |p + |f |p −−−→ 0, kun k → ∞ .
Rn \Bk−1 Rn \Bk

⇒ ∃ k s.t.

(2.30) Ih (Rn \ Bk ) < ε/4 .

Similarly,
Z Z 
p p p
Ih (A) ≤ 2 |f | + |f | , if A ∈ Leb Rn and A + h = {a + h : a ∈ A}.
A+h A
2017 29

The integral is absolutely continuous with respect to the Lebesgue measure, and therefore Theorem
2.5 implies ∃ δ > 0 s.t.

(2.31) Ih (A) < ε/4 , if m(A) < δ .

Lusin’s theorem ⇒ ∃ a compact F ⊂ Bk+1 s.t.

m(Bk+1 \ F ) < δ and f |F continuous.

F compact ⇒
f |F uniformly continuous.
⇒ ∃ η ∈ ]0, 1[ s.t.
ε
(2.32) |f (x + h) − f (x)|p < , if |h| < η and x, x + h ∈ F .
4m(F )
Let h ∈ B(0, η) be arbitrary. Denote

A1 = {x ∈ F : x + h ∈ F },
A2 = {x : x + h ∈ Bk+1 \ F },
A3 = Bk+1 \ F .

(We observe: A2 = A3 − h, and so m(A2 ) = m(A3 ) = m(Bk+1 \ F ) < δ.)

Then
x ∈ Bk ⇒ x + h ∈ Bk+1 ,
and therefore

Bk ∩ F ⊂ {x ∈ F : x + h ∈ Bk+1 } ⊂ {x ∈ F : x + h ∈ F } ∪ {x ∈ F : x + h ∈ Bk+1 \ F }
| {z } | {z }
=A1 ⊂A2
⊂ A1 ∪ A2

⇒ Bk = (Bk \ F ) ∪ (Bk ∩ F ) ⊂ A1 ∪ A2 ∪ A3
| {z } | {z }
⊂A3 ⊂A1 ∪A2

Now Rn = A1 ∪ A2 ∪ A3 ∪ (Rn \ Bk ) and

Ih (Rn ) ≤ Ih (A1 ) + Ih (A2 ) + Ih (A3 ) + Ih (Rn \ Bk ) .

Let us estimate the terms on the right hand side:

Z 
(2.32) ⇒ Ih (A1 ) = |f (x + h) − f (x)|p dx < ε/4 



A1 | {z } A1 ⊂F 

ε 
< 4m(F )
⇒
(2.31) ⇒ Ih (A2 ) < ε/4 

(2.31) ⇒ Ih (A3 ) < ε/4 



(2.30) : Ih (Rn \ Bk ) < ε/4


Z
n
Ih (R ) = |f (x + h) − f (x)|p dx < ε .
Rn
30 Real Analysis I

Define η : R → [0, ∞[ ,
( 1
e t2 −1 , if |t| < 1,
η(t) =
0, if |t| ≥ 1.

Let |t| < 1. Then

1

(k) e t2 −1 · P3k (t)

η (t) = ; P3k = polynomial of order 3k,
(t2 − 1)2k
⇒ η (k) (t) → 0, as t → 1 or t → −1 .
⇒ η ∈ C0∞ (R) .

We want a function ϕk : Rn → [0, ∞[ , k ∈ N, s.t.

(a) ϕk ∈ C0∞ (Rn ) ,

(b) spt ϕk ⊂ B̄1/k = B̄(0, 1/k) ,

Z
(c) ϕk = 1 .
Rn

We may choose

(2.33) ϕk (x) = ak η(k|x|) ,

where the constant ak is chosen such that (c) holds.

We notice: If f ∈ Lp (Rn ) and g ∈ C0 (Rn ) (i.e. g continuous and spt g compact), then

y 7→ f (x − y)g(y) is integrable ∀x since

Z Z Z
|f (x − y)| |g(y)| dy ≤ M |f (x − y)|dy = M |f (y)|dy < ∞ ,
Rn | {z } spt g A
≤M <∞

where A = x − spt g = {x − z : z ∈ spt g}, and it holds:

f ∈ Lp (A), m(A) < ∞ ⇒ f ∈ L1 (A) .

Hence the convolution f ∗ g(x) is defined ∀x ∈ Rn .

We apply this to the following:

gk : Rn → [0, ∞[ continuous,
spt gk ⊂ B̄1/k ,
Z
gk = 1 .
Rn

Theorem 2.34. Let f ∈ Lp (Rn ), 1 ≤ p < ∞, and gk as above. Then

lim kf − f ∗ gk kp = 0 .
k→∞
2017 31

Proof. For the case p = 1 we have

Z Z
f (x) − f ∗ gk (x) = f (x) gk (y)dy − f (x − y)gk (y)dy
Z Rn Rn

= (f (x) − f (x − y)) gk (y)dy

n
ZR
⇒ |f (x) − f ∗ gk (x)| ≤ |f (x) − f (x − y)|gk (y)dy .
Rn

If p > 1,
Z p
|f (x) − f ∗ gk (x)|p ≤ |f (x) − f (x − y)|gk (y)dy
R n
Z p p
1/p 1/q
= |f (x) − f (x − y)| gk (y) gk (y) dy (where q = )
Rn | {z } p−1
=gk (y)
Z Z  q p/q
p
≤ |f (x) − f (x − y)| gk (y)dy gk (y)1/q dy
Hölder Rn n
|R {z }
=1
Z
p
= |f (x) − f (x − y)| gk (y)dy .
Rn

Hence ∀p ≥ 1:
Z
kf − f ∗ gk kpp = |f (x) − f ∗ gk (x)|p dx
Rn
Z Z 
≤ |f (x) − f (x − y)|p gk (y)dy dx
n Rn
ZR Z 
Fubini 1. p
= gk (y) |f (x) − f (x − y)| dx dy .
Rn Rn

Since spt gk ⊂ B̄1/k , we may assume that |y| ≤ 1/k in the inner integration.
Z 
Thm. 2.29 ⇒ |f (x) − f (x − y)|p dx → 0, 


Rn 


as k → ∞ and |y| ≤ 1/k
⇒ claim.
Z 




gk = 1 
Rn

Clearly the spaces C0 (Rn ) and C0k (Rn ), k = 1, 2, . . . , ∞, vector subspaces of Lp (Rn ) and, when
equipped with the norm k·kp , they are also normed subspaces of Lp (Rn ).

Definition 2.35. If W is a subspace of a normed space (V, k·k), we say that W is dense in V if
∀v ∈ V ∃ a sequence w1 , w2 , . . . ∈ W s.t. kwi − vk → 0 as i → ∞. (That is, W̄ = V .)

Theorem 2.36. C0∞ (Rn ) is a dense subspace of Lp (Rn ) if 1 ≤ p < ∞.

32 Real Analysis I

Proof. Let f ∈ Lp (Rn ). We need to show: ∃ ψ1 , ψ2 , . . . ∈ C0∞ (Rn ) s.t.

kf − ψk kp → 0 as k → ∞.

(a): Suppose that spt f is compact. Then f ∈ L1 . Choose functions ϕk as inn (2.33). Thm. 2.26
⇒
f ∗ ϕk ∈ C ∞ (Rn ).
If d(x, spt f ) > 1/k and y ∈ spt ϕk (⊂ B̄(0, 1/k)), then x − y 6∈ spt f. and therefore
Z
f ∗ ϕk (x) = f (x − y)ϕk (y)dy = 0 ⇒ spt(f ∗ ϕk ) kompakti.
Rn
k→∞
Thm. 2.34 ⇒ kf − f ∗ ϕk kp −−−→ 0.

Hence we may choose ψk = f ∗ ϕk . (Note: Above spt(f ∗ ϕk ) is compact since it is both closed and
bounded.)
(b): General case: spt f not necessarily compact. Let ε > 0. Denote

fj = f χBj , Bj = B(0, j).

Then fj ∈ Lp (Rn ), spt fj is compact. Furthermore, there exists j0 s.t.

kf − fj kp < ε/2 ∀j ≥ j0 .

Part (a) ⇒ ∃ ψ1j , ψ2j , . . . ∈ C0∞ (Rn ) s.t.

kfj − ψkj kp < ε/2 as k ≥ kj .

Then
kf − ψkj j kp < ε ∀j ≥ j0 .

3 Derivative
In this section we study differentiability of integrals, in particular, the question when a function
f : [a, b] → R can be recovered by integrating its derivative f ′ ?

Example 3.1. 1. (Let g : [a, b] → R be continuous and

Z x
G(x) = g(t)dt, x ∈ [a, b].
a

Then G is differentiable and G′ (x) = g(x), x ∈ [a, b].

2. There is a more general result (Lebesgue’s theorem): Let g : [a, b] → R be integrable. Then
the function G : [a, b] → R, Z x
G(x) = g(t)dt,
a
is differentiable a.e. and
G′ (x) = g(x) a.e. x ∈ [a, b].
2017 33

3. Converse to the case 1 (starting from the function):

Z x
1
f ∈ C ([a, b]) ⇒ f ′ (t)dt = f (x) − f (a).
a

4. Let f : [0, 1] → [0, 1] be the Cantor 1/3 function. Then f is a continuous increasing surjection
and f ′ (t) = 0 for a.e. t ∈ [0, 1] (f ′ measurable), but
Z 1 Z 1
′
f (t)dt = 0dt = 0 6= 1 = f (1) − f (0).
0 0

We will study these questions by using ”covering theorems” as a tool. Here we try to ”almost”
cover a given subset of Rn by closed, pairwise disjoint balls. The disjointness of the balls is required
in order to apply the countable additivity of the Lebesgue measure.

3.2 Covering theorems

We denote kB = B(x, kr) if B = B(x, r) and k > 0 (or, respectively, kB = B̄(x, kr) if B = B̄(x, r)).
Note: We assume that the radius r of a closed ball B̄(x, r) is positive (r > 0).
({y ∈ Rn : |x − y| < 0} = ∅, {y ∈ Rn : |x − y| ≤ 0} = {x})

Theorem 3.3 (Basic covering theorem). Let F be an arbitrary family of balls in Rn s.t.

D = sup{d(B) : B ∈ F} < ∞,

where d(B) = sup{|x − y| : x, y ∈ B} is the diameter of B. Then there exists a countable (possibly
finite) family G ⊂ F s.t.

Bi ∩ Bj = ∅ ∀Bi , Bj ∈ G, Bi 6= Bj , i.e. the balls in G are pairwise disjoint; and

[ [
B⊂ 5B .
B∈F B∈G

Proof. 1. Denote
Fj = {B ∈ F : D/2j < d(B) ≤ D/2j−1 }, j ∈ N,
S
hence F = ∞ j=1 Fj.
Define inductively families Gj ⊂ Fj :

(a) Let G1 be a maximal family of pairwise disjoint balls in F1 , i.e.

B ∈ F1 ⇒ ∃B ′ ∈ G1 s.t. B ∩ B ′ 6= ∅.

(Thus: We can not add to G1 any ball of F1 without destroying the pairwise disjointness.)

(b) Suppose that families G1 , . . . , Gk−1 are chosen. Let Gk be any maximal collection of pairwise
disjoint balls of Fk s.t.
k−1
[
B ∩ B ′ = ∅ ∀B ′ ∈ Gj .
j=1
34 Real Analysis I

Denote
∞
[
G= Gj ,
j=1

then G (⊂ F) is a family of pairwise disjoint balls.

2. Claim: G is countable.
Proof: It is enough to show: Gj countable ∀j. Write
∞
[
Gj = Gj,i , where Gj,i = {B ∈ Gj : B ⊂ B̄(0, i)},
i=1

and let us prove that Gj,i is finite (possibly empty), hence Gj ’s and therefore G is countable.

B ∈ Gj,i ⇒ d(B) > D/2j and B ⊂ B̄(0, i) (∗)
⇒ Gj,i finite.
B̄(0, i) compact

Reason for (∗): assume on the contrary: there are infinitely many disjoint balls in Gj,i ⇒ ∃a
sequence xk ∈ B̄(0, i), k ∈ N, s.t.

(3.4) |xk − xl | ≥ D/2j ∀k 6= l.

(for example, every xk is a center of some ball in G  j,i .) B̄(0, i) compact ⇒ ∃ a convergent
subsequence of (xk ) that is a contradiction with (3.4).
3. Claim: For all B ∈ F there exists B ′ ∈ G s.t. B ∩ B ′ 6= ∅ and B ⊂ 5B ′ . In particular, then
[ [
B⊂ 5B .
B∈F B∈G

Sk
Proof: If B ∈ F, then B ∈ Fk for some k ∈ N. Since Gk is maximal, there exists B ′ ∈ j=1 Gj s.t.
B ∩ B ′ 6= ∅. On the other hand,

d(B ′ ) > D/2k and d(B) ≤ D/2k−1 ⇒ d(B) < 2d(B ′ )
⇒ B ⊂ 5B ′ .
B ∩ B ′ 6= ∅

5B ′
B B′

Remark 3.5. 1. The part 2 in the proof follows also from the disjointness of the balls in G and
the separability of Rn (Qn is a countable dense subset of Rn ).

2. The basic covering theorem (stated as above) holds for some metric spaces (X, d) provided cer-
tain additional assumptions hold. Read the proof again and think of what kind of properties
(X, d) should have in order to the proof work.
2017 35

Definition 3.6. Let V be a family of balls in Rn . We say that V is a Vitali covering of a set
E ⊂ Rn if
∀x ∈ E and ∀ε > 0 ∃B ∈ V s.t. x ∈ B and d(B) < ε .
Such a family V is a closed Vitali covering (respectively, open) if every ball B ∈ V is closed (resp.
open).

Remark 3.7. If V is a Vitali covering of E and R > 0, then

{B ∈ V : d(B) < R}

is also a Vitali covering of E.

From the proof of Theorem 3.3 we obtain:

Corollary 3.8. Let V be a closed Vitali covering of E ⊂ Rn s.t. d(B) < R ∀B ∈ V. Then there
exists a countable family G ⊂ V of pairwise disjoint balls such that for every finite G ∗ ⊂ G we have:
[ [
E\ B⊂ 5B .
B∈G ∗ B∈G\G ∗

Proof. Let G ⊂ V be as in the proof of Theorem 3.3 and let G ∗ = {B1 , B2 , . . . , Bm } ⊂ G be arbitrary.
If
[m
E⊂ Bi , we are done.
i=1
Sm Sm
Otherwise, let x ∈ E \ i=1 Bi . Then i=1 Bi is compact, and therefore
m
[ m
[ m
[


d x, Bi = inf{|x − y| : y ∈ Bi } = min{|x − y| : y ∈ Bi } > 0 .
i=1 i=1 i=1
Sm

Since V is a Vitali covering of E, ∃ B ∈ V s.t. x ∈ B and B ∩ i=1 Bi = ∅ (d(B) small enough).

E
Bi

x
B∈V

It follows from the part 3 of the proof of Theorem S 3.3 that
∃ B ′ ∈ G s.t. B ∩ B ′ 6= ∅ and B ⊂ 5B ′ .
m
In particular x ∈ 5B ′ . Now B ′ 6∈ G ∗ , because B ∩ ′ ∗
i=1 Bi = ∅. Thus B ∈ G \ G , and so
[ [
E\ B⊂ 5B .
B∈G ∗ B∈G\G ∗
36 Real Analysis I

Theorem 3.9 (Vitali’s covering theorem). Let E ⊂ Rn (not necessarily measurable) and let V be
a closed Vitali covering of E. Then there exists a countable subfamily G ⊂ V of pairwise disjoint
balls s.t. [

m E\ B = 0.
B∈G

Proof. 1. Suppose fist that E is bounded. Then we may assume that there exists a bounded open
H ⊂ Rn s.t. B ⊂ H ∀B ∈ V. Let G be as in the proof of Theorem 3.3 (and of Corollary 3.8). Let
ε > 0. We will show: [

m∗ E \ B <ε
B∈G

that implies the claim (since ε > 0 arbitrary). The balls in G are pairwise disjoint and ⊂ H, and
hence X [

m(B) = m B ≤ m(H) < ∞ .
B∈G B∈G

⇒ ∃ a finite G∗ ⊂ G s.t. X
m(B) < ε/5n .
B∈G\G ∗

Corollary 3.8 ⇒
[
[
[

m∗ E \ B ≤ m∗ E \ B ≤m 5B
B∈G B∈G ∗ B∈G\G ∗
| {z } | {z }
⊂E\∪B∈G ∗ B ⊂ ∪B∈G\G ∗ 5B
X X
≤ m(5B) = 5n m(B) < ε .
B∈G\G ∗ B∈G\G ∗

ε > 0 arbitrary ⇒ [

m E\ B = 0.
B∈G

2. General case: E not necessarily bounded. Denote

A1 = B(0, 1) and Ai = B(0, i) \ B̄(0, i − 1), i ≥ 2 .

Then the sets Ai are open and disjoint, and

∞
[ ∞
[



(3.10) m Rn \ Ai = m S(0, i) = 0 , S(0, i) = {x ∈ Rn : |x| = i} .2
i=1 i=1

Idea: Applying the part 1 to the sets E ∩ Ai we obtain subfamilies Gi ⊂ V. We must take care
that the balls in Gi and Gj intersect each other. This can be dealt with as follows: Ai open,
x ∈ Ai ⇒ ∃ rx > 0 s.t. B(x, r) ⊂ Ai ∀r ≤ rx ⇒ Vi = {B ∈ V : B ⊂ Ai } is a Vitali covering of
E ∩ Ai .
Ai ’s disjoint ⇒

(3.11) if B ∈ Vi and B ′ ∈ Vj , i 6= j, then B ∩ B ′ = ∅.

2


Find a simple argument that shows that mn S(0, i) = 0.
2017 37

1. part ⇒ ∃ a countable family Gi ⊂ Vi of disjoint balls s.t.

[

(3.12) m (E ∩ Ai ) \ B = 0.
B∈Gi
S
Then G = ∞ i=1 Gi satisfies the requirements. Clearly G is countable and the balls in G are disjoint
(see (3.11)). Moreover,
 ∞   ∞
[ [
[ [
[

E\ B= E\ B \ Ai ∪ E \ B ∩ Ai
B∈G B∈G i=1 B∈G i=1
| {z } | {z }
of measure 0 (see (3.10)) S∞ S

= i=1 (E∩Ai )\ B∈G B
i

[ ∞

X [
(3.12)
⇒ m E\ B ≤ m (E ∩ Ai ) \ B = 0.
B∈G i=1 B∈Gi

Remark 3.13. Also the covering theorem holds for certain ”metric measure spaces” (X, d, Γ, µ).
Read the proof again and think of what kind of requirements (X, d) and the measure µ should
fulfil.

3.14 Maximal function

Let us start with the following useful result:
Let (X, Γ, µ) be a measure space and g : X → [0, +∞] measurable. The function [0, +∞) →
[0, +∞],


t 7→ µ {x : g(x) > t} ,
is called the distribution function of g. It is decreasing and hence measurable.

Lemma 3.15. Let f : X → [0, +∞] be measurable and 0 < p < ∞. Then
Z Z ∞
p


(3.16) f dµ = p tp−1 µ {x : f (x) > t} dt .
X 0

Proof. Idea: (i) Suppose first that f is simple and prove that (by direct computation) that (3.16)
holds. (ii) In the general case, choose a sequence of simple functions fk ր f and apply the monotone
convergence theorem. Details are left as an exercise.
R

f
A
t

{x ∈ X : f (x) > t} X
38 Real Analysis I

R R∞

In the picture above X f dµ = 0 µ {x ∈ X : f (x) > t} dt can be interpreted as the ”product
measure” (µ × m1 )(A) of the shaded area A.
Denote f ∈ L1loc (Rn ) if f : Rn → Ṙ is measurable and
Z
|f | < ∞ ∀ compact K ⊂ Rn .
K

We say: f is locally integrable (or locally in L1 ).

Remark 3.17. 1. f ∈ C(Rn ) ⇒ f ∈ L1loc (Rn ).

2. f ∈ L1 (Rn ) ⇒ f ∈ L1loc (Rn ). The converse ”⇐” does not hold: for instance f (x) ≡ 1.
Our aim is to prove:
Z
1
f∈ L1loc (Rn ) ⇒ lim
f (y)dy = f (x) a. e. x ∈ Rn .
r→0+ m B(x, r) B(x,r)

For A ∈ Leb Rn and m(A) > 0 we denote

Z Z
1
f (y)dy = f (y)dy,
m(A) A
A

the ”integral mean” of f over A.

Definition 3.18. If f ∈ L1loc (Rn ) and x ∈ Rn , we set
Z
M f (x) = sup |f (y)|dy,
B∋x
B

where B is an (arbitrary) open ball that contains x. Then function M f : Rn → [0, ∞] is the
(Hardy-Littlewood) maximal function of f .
Note: There are different kind of maximal functions in the literature. For instance, the supre-
mum can be taken over balls B(x, r) centered at x. Then we obtain a ”centered” maximal function
of f ∈ L1loc (Rn ), denoted by M̃ f ,
Z
M̃ f (x) = sup |f (y)|dy.
r>0
B(x,r)

It holds: M̃ f (x) ≤ M f (x) ≤ 2n M̃ f (x) ∀x ∈ Rn .

Lemma 3.19. The maximal function M f : Rn → [0, ∞] is measurable.
Proof. Write Et = {x : M f (x) > t}. We prove a stronger results that Et is open3 ∀t ∈ R (and
hence, in particular, measurable). Let x ∈ Et . Then there exists an open ball B ∋ x s.t.
Z
|f (y)|dy > t.
B
3
A function u : X → Ṙ of a topological space X is lower semicontinuous if {x ∈ X : u(x) > t} is open ∀t ∈ R.
Upper semicontinuity is defined similarly. Thus u is continuous ⇐⇒ u is both lower and upper semicontinuous.
Proof of Lemma 3.19 ⇒ M f lower semicontinuous.
2017 39

Then ∀y ∈ B we have: Z
sup
M f (y) ≥ |f (y)|dy > t ⇒ y ∈ Et .
B
Hence B ⊂ Et and therefore Et is open.

Remark 3.20. 1. What can be said about integrability of M f ?

Answer: M f is ”vey seldom” integrable.

More precisely: M f ∈ L1 (Rn ) ⇒ f = 0 a.e.
Reason: Recall that n
R m B(x, r) = cn r , where cn is a constant depending on n. Assume on
the contrary that Rn |f | > 0, hence ∃R > 0 s.t.
Z
|f (y)|dy > 0.
B(0,R)
| {z }
def.
=I

0 2|x|
x
R

If x ∈ Rn \ B(0, R), then B(0, R) ⊂ B(x, 2|x|) (see the picture) and
Z Z
1 1
M f (x) ≥
|f (y)|dy ≥ |f (y)|dy .
m B(x, 2|x|) B(x,2|x|) cn (2|x|)n B(0,R)
| {z }
=I>0

Z Z Z
⇒ M f (x)dx ≥ M f (x)dx ≥ I c−1
n (2|x|)
−n
dx = ∞
Rn Rn \B(0,R) Rn \B(0,R)

since
Z ∞ Z
X
c−1
n (2|x|)
−n
dx = c−1 (2|x|)−n dx
Rn \B(0,R) B(0,2i+1 R)\B(0,2i R) |n {z }
i=0
≥c−1
n (2
i+2 R)−n

∞
X

≥ m B(0, 2i+1 R) \ B(0, 2i R) c−1
n (2
i+2
R)−n
i=0
| {z }
=cn ((2i+1 R)n −(2i R)n )
∞ ∞
X 2n − 1 X
= = c = ∞.
4n
i=0 | {z } i=0
=c>0

2. Chebyshev’s inequality: f ∈ L1 (Rn ) ⇒ an estimate

c
(3.21) m {x ∈ Rn : |f (x)| > t} ≤ ∀t > 0
t
holds with the constant c = kf k1 . The converse does not hold: a measurable function f for
which (3.21) holds ∀t > 0 with some constant c need not be integrable.
40 Real Analysis I

3. We say that a measurable function f : Rn → Ṙ belongs to the weak L1 -space, weak-L1 (Rn ), if
there is a constant c = cf < ∞ such that (3.21) holds ∀t > 0. Thus: L1 (Rn ) ⊂ weak-L1 (Rn ),
but weak-L1 (Rn ) 6⊂ L1 (Rn ).
It turns out that the maximal function M f of an integrable function f ∈ L1 (Rn ) satisfies
(3.21). This is one of the most important properties of M f . The proof is based on the Basic
covering theorem 3.3.
Theorem 3.22 (Hardy-Littlewood). If f ∈ L1 (Rn ), then

5n kf k1
(3.23) m {x ∈ Rn : M f (x) > t} ≤ ∀t > 0.
t
Proof. Fix t > 0 and write Mt = {x ∈ Rn : M f (x) > t}. Then ∀x ∈ Mt ∃ an open ball Bx ∋ x (not
necessarily centered at x) s.t. Z
|f (y)|dy > t.
Bx

In other words,
Z
1 kf k1
(3.24) m(Bx ) ≤ |f (y)|dy (≤ ).
t Bx t

Let F = {Bx : x ∈ Mt }, hence (trivially)

[
Mt ⊂ B.
B∈F

Since m(Bx ) = cn (d(Bx )/2)n , then

 1/n
kf k1
(3.24) ⇒ sup{d(Bx ) : Bx ∈ F} ≤ 2 < ∞.
cn t

We may apply the Basic covering theorem 3.3 ⇒ ∃ countable subfamily G = {B1 , B2 , . . .} ⊂ F of
disjoint open balls s.t. [ [
Mt ⊂ B⊂ 5Bi .
B∈F Bi ∈G

Hence

countable union X X
m(Mt ) ≤ m ∪i 5Bi ≤ m(5Bi ) = 5n m(Bi )
i i

(3.24) X1Z Bi ’s disjoint 5n

Z
n
≤ 5 |f (y)|dy = |f (y)|dy
t Bi t ∪i Bi
i

Z
5n
≤ |f (y)|dy.
t Rn

By Remark 3.20, M f ∈ L1 (Rn ) ⇒ f = 0 a.e. The situation is completely different if p > 1.

2017 41

Theorem 3.25. Let 1 < p < ∞ and f ∈ Lp (Rn ). Then M f ∈ Lp (Rn ) and there is a constant
c = c(p, n) s.t.
kM f kp ≤ ckf kp .
Proof. The proof uses Lemma 3.15, the Hardy-Littlewood theorem and Fubini’s theorem.
We omit the details.

3.26 Lebesgue differentiation theorem

Let A ⊂ Rn , h : A → Ṙ and x0 an accumulation point of A (i.e. B(x0 , r) ∩ (A \ {x0 }) 6= ∅ ∀r > 0).
Define
lim sup h(x) = lim sup{h(x) : x ∈ B(x0 , r) ∩ (A \ {x0 })}.
x→x0 r→0+

x
x0
A
A
x
x0 B(x0 , r) ∩ (A \ {x0 })

Observation: 0 < r1 < r2 ⇒

sup{h(x) : x ∈ B(x0 , r1 ) ∩ (A \ {x0 })} ≤ sup{h(x) : x ∈ B(x0 , r2 ) ∩ (A \ {x0 })},
and therefore the limit exists (±∞ allowed). Similarly we can define lim inf.
Motivation: If f : Rn → R is continuous, then
Z
(3.27) lim |f (y) − f (x)|dy = 0
r→0+
B(x,r)

∀x ∈ Rn (see the proof of part 3 below). On the other hand, Lusin’s theorem says that a measurable
function is ”almost continuous” (f measurable, ε > 0 ⇒ ∃ closed F ⊂ Rn s.t. m(Rn \ F ) < ε and
f |F continuous), hence a question arises: in what sense (3.27) holds for locally integrable functions.
Theorem 3.28 (Lebesgue differentiation theorem). Let f ∈ L1loc (Rn ) be a locally integrable func-
tion. Then
Z
(3.29) lim |f (y) − f (x)|dy = 0 a.e. x ∈ Rn .
r→0+
B(x,r)

In particular,
Z
(3.30) lim f (y)dy = f (x) a.e. x ∈ Rn .
r→0+
B(x,r)

Proof. If f ∈ L1loc (Rn ) and x ∈ Rn , let us define

Z
Λf (x) = lim sup |f (y) − f (x)|dy.
r→0
B(x,r)

(Note: Λf (x) is defined ∀x ∈ Rn and the value depends on f (x), thus in particular, on the chosen
representative of the equivalence class.)
Then Λf satisfies:
42 Real Analysis I

1. Λf (x) ≥ 0 ∀x ∈ Rn (clear).
2. Λ is sub linear, i.e.
Λ(f + g) ≤ Λf + Λg, ∀f, g ∈ L1loc (Rn ).
Reason:
Z
Λ(f + g)(x) = lim sup |f (y) + g(y) − f (x) − g(x)|dy
r→0
B(x,r)
△-ineq.  Z Z 
≤ lim sup |f (y) − f (x)|dy + |g(y) − g(x)|dy
r→0
B(x,r) B(x,r)
Z Z
≤ lim sup |f (y) − f (x)|dy + lim sup |g(y) − g(x)|dy
r→0 r→0
B(x,r) B(x,r)
n
= Λf (x) + Λg(x), ∀x ∈ R .

3. g ∈ C(Rn ) ⇒ Λg(x) = 0 ∀x ∈ Rn .
Reason: Fix x ∈ Rn and ε > 0.
g continuous at x ⇒ ∃ δ > 0 s.t. |g(y) − g(x)| < ε ∀y ∈ B(x, δ).
Hence ∀ 0 < s ≤ δ we have:
Z Z

1 εm B(x, s)
|g(y) − g(x)| dy <
εdy =
=ε
| {z } m B(x, s) B(x,s) m B(x, s)
B(x,s) <ε
Z !
⇒ Λg(x) = lim sup |g(y) − g(x)|dy ≤ε
r→0 0<s<r
B(x,s)
| {z }
<ε, if 0<s≤δ

⇒ Λg(x) = 0 , sinc ε > 0 arbitrary.

4. Λf ≤ M f + |f |.
Reason:
Z △-ey.
Z Z
|f (y) − f (x)|dy ≤ |f (y)|dy + |f (x)|dy ≤ M f (x) + |f (x)|.
B(x,r) B(x,r) B(x,r)
| {z } | {z }
≤M f (x) =|f (x)|

(Note |f (x)| is constant in the last integral.)

Let then f ∈ L1loc (Rn ) be given and t > 0. It suffices to show that ∀k ∈ N (3.29) holds a.e.
x ∈ B(0, k). The validity of (3.29) in B(0, k) is independend of the values of f in Rn \ B(0, 2k), and
therefore we may assume that f = 0 in Rn \ B(0, 2k) and hence f ∈ L1 (Rn ).
If g ∈ C(Rn ), then
2. 3. 4.
Λf = Λ(f − g + g) ≤ Λ(f − g) + Λg = Λ(f − g) ≤ M (f − g) + |f − g|.
Hence at least one of the values M (f − g)(x) or |f (x) − g(x)| is at least Λf (x)/2, and therefore
{x ∈ Rn : Λf (x) > t} ⊂ {x ∈ Rn : M (f − g)(x) > t/2} ∪ {x ∈ Rn : |f (x) − g(x)| > t/2}.
2017 43

Hence




m∗ {x : Λf (x) > t} ≤ m {x : M (f − g)(x) > t/2} + m {x : |f (x) − g(x)| > t/2}
| {z } | {z }
H.-L. Cheb.
≤ 2·5n kf −gk1 /t ≤ 2kf −gk1 /t

2(5n + 1)kf − gk1

≤ .
t
Theorem 2.36 ⇒ continuous functions are dense in L1 ⇒ ∀ε > 0 ∃ g ∈ C(Rn ) s.t. kf − gk1 < ε
ε>0 arbitr.

=⇒ m∗ {x ∈ Rn : Λf (x) > t} = 0 ∀t > 0
∞

X

⇒ m∗ {x ∈ Rn : Λf (x) > 0} ≤ m∗ {x ∈ Rn : Λf (x) > 1/k} = 0
| S {z } | {z }
k=1
⊂ k {x : Λf (x)>1/k} =0
n
⇒ Λf (x) = 0 a.e. x ∈ R
Z Z
⇒ 0 ≤ lim inf |f (y) − f (x)|dy ≤ lim sup |f (y) − f (x)|dy = 0 a.e. x ∈ Rn
r→0 r→0
B(x,r) B(x,r)
Z
⇒ lim |f (y) − f (x)|dy = 0 a.e. x ∈ Rn .
r→0+
B(x,r)

Finally,
 Z   Z Z   Z

    
 f (y)dy − f (x) =  f (y)dy − f (x)dy  =  f (y) − f (x) dy 
B(x,r) B(x,r) B(x,r) B(x,r)
| {z }
=f (x)
Z
r→0+
≤ |f (y) − f (x)|dy −−−−→ 0 a.e. x ∈ Rn .
B(x,r)

Definition 3.31. A point x ∈ Rn is density point of a set E ∈ Leb Rn if

m E ∩ B(x, r)
lim
= 1.
r→0+ m B(x, r)

Remark 3.32. If x ∈ Rn is a density point of E, it need not belong to E.For instance, 0 is a

density point of Rn \ {0}.
Corollary 3.33. Let E ∈ Leb Rn . Then almost every x ∈ E is a density point of E, i.e.


m E ∩ B(x, r)
lim
= 1 a.e. x ∈ E.
r→0+ m B(x, r)

Furthermore,

m E ∩ B(x, r)
lim
= 0 a.e. x ∈ Rn \ E.
r→0+ m B(x, r)
44 Real Analysis I

Proof. E ∈ Leb Rn ⇒ χE ∈ L1loc (Rn ) since χE is measurable and

Z
χE = m(E ∩ K) < ∞ ∀ compact K ⊂ Rn .
K

Furthermore,

Z
m E ∩ B(x, r)
χE (y)dy =
∀x ∈ Rn , r > 0.
m B(x, r)
B(x,r)

Lebesgue diff. theorem 3.28 ⇒

m E ∩ B(x, r)
lim
= χE (x) a.e. x ∈ Rn .
r→0+ m B(x, r)

A point x ∈ Rn is a Lebesgue point of a function f ∈ L1loc (Rn ) if

Z
lim |f (y) − f (x)|dy = 0.
r→0+
B(x,r)

This property depends on the value f (x) and hence on the choice of the representative f ∈ L1loc (Rn ).
To get rid of this dependence we say that a point x ∈ Rn b elongs to the Lebesgue set, Leb(f ), of
f if ∃ A = A(x) ∈ R s.t.
Z
(3.34) lim |f (y) − A|dy = 0.
r→0+
B(x,r)

If such A exists, it is unique since

Z
lim f (y)dy = A.
r→0+
B(x,r)

Remark 3.35. 1. f = g a.e. in Rn ⇒ Leb(f ) = Leb(g). In particular, the Lebesgue set

Leb(f ) is well defined in the whole equivalence class f ∈ L1loc (Rn ), i.e. does not depend on
the choice of the representative.

2. Lebesgue diff. theorem 3.28: If f ∈ L1loc (Rn ), then almost every x ∈ Rn belongs to Leb(f ).
Furthermore, A = f (x) for a.e. x ∈ Rn . Thus by modifying f in a set of measure 0 (by setting
f (x) = A(x)) we may assume in what follows that
Z
lim |f (y) − f (x)|dy = 0 ∀x ∈ Leb(f ).
r→0+
B(x,r)

Let x ∈ Rn . We say the a sequence of measurable sets Ej ⊂ Rn , j ∈ N, shrinks nice to x if ∃ a

constant c > 0 and a sequence rj > 0 s.t.

Ej ⊂ B(x, rj ) ∀j and lim rj = 0,

j→∞


(3.36) m B(x, rj ) ≤ c m(Ej ) ∀j.
2017 45

rj
Ej
x

Theorem 3.37. Let f ∈ L1loc (Rn ) and x ∈ Leb(f ). If a sequence Ej shrinks nicely to x, then
Z
lim f (y)dy = f (x).
j→∞
Ej

Proof.
Z  Z

Z
  
 f (y)dy − f (x) =  f (y)dy − f (x)  ≤ |f (y)dy − f (x)|dy
Ej Ej Ej

Z
m B(x, rj ) rj →0
≤ |f (y) − f (x)|dy −−−→ 0 since x ∈ Leb(f ).
m(Ej )
| {z } B(x,rj )
≤ c
(3.36)

Rx
Now we can study differentiability of the function F (x) = a f (t)dt if f : [a, b] → R is integrable
(see Example 3.1.2).
Theorem 3.38. Let f : [a, b] → R be integrable and
Z x
F (x) = f (t)dt, x ∈ [a, b].
a

YThen F is differentiable a.e. and F ′ (x) = f (x) for a.e. x ∈ [a, b].
Proof. Extend f to a function f : R → R by setting f (t) = 0 ∀t ∈ R \ [a, b], hence f ∈ L1 (R).
It suffices to show:
F ′ (x) = f (x) ∀x ∈ Leb(f ) ∩ [a, b]
since almost every x ∈ [a, b] belongs to Leb(f ) (Lebesgue diff. theorem 3.28).
Let x ∈ Leb(f )∩[a, b] and let rj > 0, j ∈ N, be an arbitrary

sequence rj → 0. Denote Ej =]x, x+rj [,,
and so Ej ⊂]x − rj , x + rj [= B(x, rj ) and m B(x, rj ) = 2m(Ej ). Thus the sequence Ej shrinks
nicely to x. Theorem 3.37 ⇒
Z
F (x + rj ) − F (x) 1 x+rj j→∞
= f (t)dt −−−→ f (x).
rj rj x
Similarly, we see that
Z x
F (x − rj ) − F (x) 1 j→∞
= f (t)dt −−−→ f (x)
−rj rj x−rj

by using sets Ej′ =]x − rj , x[. Since (rj ) is an arbitrary sequence, we get
F (x + h) − F (x)
lim = f (x).
h→0 h
46 Real Analysis I

3.39 Monotonic functions in R

In this section we will study differentiability of monotonic functions by using Vitali’s covering
theorem as a tool.
Let ∆ ⊂ R be an interval. A function f : ∆ → R is

• increasing if x1 , x2 ∈ ∆, x1 < x2 ⇒ f (x1 ) ≤ f (x2 );

• decreasing if x1 , x2 ∈ ∆, x1 < x2 ⇒ f (x1 ) ≥ f (x2 );

• monotonic if f is increasing or decreasing.

Example 3.40. Write Q = {qj : j ∈ N}. Set

X
f (x) = 2−j , x ∈ R.
j∈N
qj ≤x
P −j
Then f : R → R is strictly increasing and 0 < f (x) < j∈N 2 = 1 ∀x ∈ R. Furthermore, f is
continuous in a point x ∈ R ⇐⇒ x ∈ R \ Q.

Above the set of discontinuity points of f is countable. This holds for all monotonic functions:

Lemma 3.41. A monotonic function f : [a, b] → R has at most countably many points of discon-
tinuity.

Proof. We may assume that f is increasing (g decreasing ⇒ −g increasing). For x ∈ (a, b) we

define
H(x) = lim f (y) − lim f (y)
y→x+ y→x−

f increasing and bounded in a neighborhood of x ⇒ the limit exists and H(x) ≥ 0. We denote

Hk = {x ∈ (a, b) : H(x) > 1/k}, k ∈ N,

and prove that Hk is finite. Suppose that x1 , . . . , x2j ∈ Hk s.t. x1 < x2 < · · · < x2j .
For i = 2, 3, . . . , j choose arbitrary x and y s.t. x2i−2 < y < x2i−1 < x < x2i .
Since f is increasing,

f (x2i ) − f (x2i−2 ) ≥ f (x) − f (y) ≥ H(x2i−1 ) > 1/k.

j
X j

X
⇒ f (b) − f (a) ≥ f (x2i ) − f (x2i−2 ) ≥ H(x2i−1 ) > (j − 1)/k
i=2 i=2


⇒ j < k f (b) − f (a) + 1
⇒ the set Hk is finite.

Furthermore f is discontinuous at x ∈ (a, b) ⇐⇒ H(x) > 0. Thus

∞
[
{x ∈ (a, b) : f discontinuous at x} ⊂ Hk
k=1

which is countable.

Lebesgue theorem for monotonic functions.

2017 47

Theorem 3.42. Let f : [a, b] → R be a monotonic function. Then the derivative f ′ (x) exists for
a.e. x ∈ [a, b].

Proof. We may assume that f is increasing. If x ∈ [a, b], we define

 
f (z) − f (y)
Df (x) = lim sup : a ≤ y ≤ x ≤ z ≤ b, 0 < z − y < ε and
ε→0+ z−y

 
f (z) − f (y)
Df (x) = lim inf : a ≤ y ≤ x ≤ z ≤ b, 0 < z − y < ε .
ε→0+ z−y

f increasing ⇒ 0 ≤ Df (x) ≤ Df (x) ≤ ∞ ∀x ∈ [a, b].

We will show:

(3.43) Df (x) = Df (x) < ∞ a.e. x ∈ [a, b].

Indeed, if Df (x) = Df (x) < ∞, then the derivative f ′ (x) = Df (x) = Df (x) exists.
We split the proof of (3.43) in parts:
1. Write

Ek = {x ∈ [a, b] : Df (x) > k}, k ∈ N,

Fs,t = {x ∈ [a, b] : Df (x) < s < t < Df (x)}, 0 < s < t, s, t ∈ Q.

Clearly
\
{x ∈ [a, b] : Df (x) = ∞} = Ek
k∈N
[
{x ∈ [a, b] : Df (x) < Df (x)} = Fs,t . (Note: countable union.)
0<s<t
s,t∈Q

It’s enough to show:

c
m∗ (Ek ) ≤ , ∀k ∈ N, for some constant c > 0,
k
m∗ (Fs,t ) = 0 for all 0 < s < t, s, t ∈ Q,

because then

c

m∗ {x : Df (x) = ∞} ≤ ∀k ⇒ m∗ {x : Df (x) = ∞} = 0 

| {z } k 


⊂Ek ∀k 


X ⇒ (3.43).
∗ ∗ 
m {x : Df (x) < Df (x)} ≤ m (Fs,t ) = 0 


| {z } 

0<s<t =0

s,t∈Q

f (b)−f (a)
2. Claim: m∗ (Ek ) ≤ k , k ∈ N.4
4
Rd
Compare: g ∈ C 1 , g ′ (t) > k > 0 ∀t ∈ [c, d] ⇒ g(d) − g(c) = c
g ′ (t)dt > k(d − c) ⇒ m([c, d]) = d − c <
(g(d) − g(c))/k.
48 Real Analysis I

Proof: Let x ∈ Ek be arbitrary. Then

 
f (z) − f (y)
Df (x) = lim sup : a ≤ y ≤ x ≤ z ≤ b, 0 < z − y < ε > k
ε→0+ z−y

⇒ ∀ε > 0 ∃ closed interval Ix, ε = [y, z] ⊂ [a, b] s.t.

x ∈ [y, z], 0 < z − y = d([y, z]) < ε and

f (z) − f (y)

(3.44) > k, thus m (]f (y), f (z)[) > k m [y, z] .
z−y | {z } | {z }
=f (z)−f (y) =z−y

Hence such intervals Ix, ε = [y, z] form a closed Vitali cover of Ek . By Vitali’s covering theorem 3.9
there exists a countable subfamily of disjoint closed intervals Ij = [yj , zj ] ⊂ [a, b] s.t. (3.44) holds
and [

m Ek \ Ij = 0.
j∈N
Now
=0
[ z }| [ { [
X



m∗ (Ek ) ≤ m∗ Ek ∩ Ij + m∗ Ek \ Ij ≤ m Ij ≤ m(Ij )
j∈N j∈N j∈N j∈N
(3.44)
1X
≤ m (]f (yj ), f (zj )[) .
k
j∈N

f increasing, intevals Ij = [yj , zj ] disjoint ⇒ open intervals ]f (yj ), f (zj )[ are disjoint. Hence
1X disjoint 1
[ 
m∗ (Ek ) ≤ m (]f (yj ), f (zj )[) = m ]f (yj ), f (zj )[
k k
j∈N j∈N
| {z }
⊂[f (a),f (b)]
f (b) − f (a)
≤ .
k
3. Claim: m∗ (Fs,t ) = 0 ∀s, t ∈ Q, 0 < s < t.
Proof: We use Vitali’s covering theorem twice. It follows from the definition of outer measure that
∀ε > 0 ∃ an open G ⊃ Fs,t s.t.
m(G) < m∗ (Fs,t ) + ε.
3a. We apply Vitali’s covering theorem to the set Fs,t (arguing as in part 2).
It follows from definitions of Df (x) and Fs,t that ∀x ∈ Fs,t ∃ arbitrary small closed intervals
[y, z], y < z, s.t. x ∈ [y, z] ⊂ G ∩ [a, b] and


m ]f (y), f (z)[ f (z) − f (y)
(3.45)
= < s.
m [y, z] z−y
(Note: G ∋ x open.)
By Vitali’s covering theorem there exists disjoint closed intervals Ij = [yj , zj ] ⊂ G ∩ [a, b], j ∈ N,
s.t. (3.45) holds and [

m Fs,t \ Ij = 0.
j∈N
2017 49

Then
[  disjoint X
(3.45) X

m ]f (yj ), f (zj )[ = m ]f (yj ), f (zj )[ < s m [yj , zj ]
j∈N j∈N j∈N
disjoint
[ 
(3.46) = sm [yj , zj ] ≤ s m(G) < s(m∗ (Fs,t ) + ε).
j∈N
| {z }
⊂G

Furthermore,
[

(3.47) m Fs,t \ ]yj , zj [ = 0,
j∈N
S
because {yj , zj : j ∈ N} is of measure 0. Denote A = j∈N ]yj , zj [.
3b. We apply Vitali’s covering theorem to the set Fs,t ∩ A.
It follows from the definitions of Df (x) and Fs,t that ∀x ∈ Fs,t ∩ A there exists arbitrary small
closed intervals [u, v], u < v, s.t. x ∈ [u, v],


m ]f (u), f (v)[ f (v) − f (u)
(3.48)
= > t,
m [u, v] v−u
S
and [u, v] ⊂ ]yj , zj [ for some j ∈ N (this is possible because A = j ]yj , zj [ is a disjoint union of
open intervals).
By Vitali’s covering theorem there exists disjoint closed intervals Jk = [uk , vk ], k ∈ N, as above s.t.
every Jk ⊂ ]yj , zj [ for a suitable j = jk and
[

(3.49) m Fs,t ∩ A \ Jk = 0.
k∈N
S
Denote B = k Jk , hence B ⊂ A and

(3.47) ⇒ Fs,t = (Fs,t ∩ A) ∪ A0 , where A0 = Fs,t \ A of measure 0 
⇒

(3.49) ⇒ Fs,t ∩ A = (Fs,t ∩ A ∩ B) ∪ B0 , where B0 = Fs,t ∩ A \ B of measure 0

  B⊂A
Fs,t = (Fs,t ∩ A) ∪ A0 = (Fs,t ∩ A ∩ B) ∪ B0 ∪ A0 = (Fs,t ∩ B) ∪ A0 ∪ B0 ,
S
where A0 ∪ B0 is of measure 0. Thus m∗ (Fs,t ) ≤ m(B), B = k Jk . Then
[  disjoint X (3.48) 1 X

m∗ (Fs,t ) ≤ m Jk = m(Jk ) < m ]f (uk ), f (vk )[
t
k k∈N k∈N
disjoint 1
 [  1  [  (3.46) s

= m ]f (uk ), f (vk )[ ≤ m ]f (yj ), f (zj )[ < m∗ (Fs,t ) + ε .
t t t
k∈N j∈N
| S {z }
⊂ j ]f (yj ),f (zj )[

ε > 0 arbitrary ⇒
s 
m∗ (Fs,t ) ≤ m∗ (Fs,t ) 

t s 
0<s<t ⇒ <1 ⇒ m∗ (Fs,t ) = 0 ∀s, t ∈ Q, 0 < s < t.
∗

t 


m (Fs,t ) ≤ m [a, b] = b − a < ∞
50 Real Analysis I

Remark 3.50. The statement of Theorem 3.42 is the best possible. Indeed:
Let A ⊂ R be an arbitrary set of measure 0. Then there exists a continuous and increasing function
f : R → R for which
Df (x) = ∞ ∀x ∈ A.
Theorem 3.51. If f : [a, b] → R is increasing, the derivative f ′ is integrable and
Z b
(3.52) f ′ (x)dx ≤ f (b) − f (a).
a
Proof. Extend f y setting f (x) = f (b) ∀x > b. Theorem 3.42 ⇒ ∃ derivative f ′ (x) a.e. x ∈ [a, b],
i.e.


′ f x + k1 − f (x)
(3.53) f (x) = lim a.e. x ∈ [a, b].
k→∞ 1/k
Furthermore, f is measurable as an increasing function, hence the function


f x + k1 − f (x)
x 7→
1/k
are measurable ∀k ∈ N, and consequently
1


f x+ k − f (x)
x 7→ lim sup
k→∞ 1/k
is measurable. Since



f x + k1 − f (x) f x + k1 − f (x)
lim sup = lim = f ′ (x) a.e. x ∈ [a, b],
k→∞ 1/k k→∞ 1/k
f ′ is measurable. We notice (by performing a change of variables x + 1/k 7→ x) that
Z b   Z b+1/k Z b
1

k f x+ − f (x) dx = k f (x)dx − k f (x)dx
a k a+1/k a

Z b+1/k Z a+1/k
=k f (x) dx − k f (x) dx
b |{z} a |{z}
=f (b) ≥f (a)
| {z }
=f (b)

≤ f (b) − f (a) ∀k ∈ N.
f increasing ⇒  
1

k f x+ − f (x) ≥ 0 ∀x ∈ [a, b], k ∈ N,
k
and so by Fatou’s lemma
Z b Z b   Z b  
′ (3.53) 1
Fatou 1

f (x)dx = lim k f x + − f (x) dx ≤ lim inf k f x+ − f (x) dx
a a k→∞ | k k→∞ k
{z } |a {z }
≥0 and measurable ≤f (b)−f (a)

≤ f (b) − f (a).

Next we will study when an equality holds in (3.52), cf. Example 3.1.4.
2017 51

3.54 Functions of bounded variation in R

Definition 3.55. The total variation of a function f : [a, b] → R on an interval [a, x], a ≤ x ≤ b, is

k
X
Vf (a, x) = sup |f (xi ) − f (xi−1 )|,
i=1

where the supremum is taken over all divisions a = x0 < x1 < x2 < · · · < xk = x of [a, x]. We say
that f is of bounded variation (or has bounded variation) on the interval [a, b] (abbr. f ∈ BV ) if
Vf (a, b) < ∞. A function g : R → R is of bounded variation if Vg (R) = supa<b Vg (a, b) < ∞.

A trivial observation: |f (b) − f (a)| ≤ Vf (a, b), since the points a, b form a division of [a, b].


Example 3.56. 1. f
∈ C 1 [a, b] ⇒ f ∈ BV.
Proof f ∈ C 1 [a, b] ⇒ f ′ : [a, b] → R continuous ⇒ ∃M = max{|f ′ (x)| : x ∈ [a, b]} < ∞.
Let a = x0 < x1 < x2 < · · · < xk = b be an arbitrary division of [a, b]. Then

k k
X VAL X
|f (xi ) − f (xi−1 )| ≤ M (xi − xi−1 ) = M (b − a)
i=1 i=1
sup
=⇒ Vf (a, b) ≤ M (b − a).

2. f ∈ C [a, b] 6⇒ f ∈ BV.
Let f : [0, 1] → R,
(
x sin x1 , jos 0 < x ≤ 1,
f (x) =
0, jos x = 0.

Then f is contiuous but not of bounded variation (exerc.).

We want to prove that every function of bounded variation can be expressed as a difference of
two increasing functions.

Lemma 3.57. Let f : [a, b] → R be monotonic. Then f is of bounded variation and

Vf (a, b) = |f (b) − f (a)|.

Proof. Suppose that f is increasing. Let a = x0 < x1 < x2 < · · · < xk = b be a division of [a, b].
Then
k
X k
f incr. X

|f (xi ) − f (xi−1 )| = f (xi ) − f (xi−1 ) = f (xk ) − f (x0 ) = f (b) − f (a)
i=1 i=1
sup
=⇒ Vf (a, b) = f (b) − f (a) < ∞.

If f is decresing, then −f is increasing, and therefore Vf (a, b) = V−f (a, b) = f (a) − f (b).

Lemma 3.58. Let f : [a, b] → R be of bounded variation. Then

Vf (a, b) = Vf (a, c) + Vf (c, b) ∀c ∈ ]a, b[.

52 Real Analysis I

Proof. Let c ∈ ]a, b[. Let

a = x0 < x1 < · · · < xk = c

c = y0 < y1 < · · · < yn = b

be arbitrary divisions of [a, c] and [c, b], respectively. Then a = x0 < x1 < · · · < xk = y0 < y1 <
· · · < yn = b is a division of [a, b], and therefore
k
X n
X
|f (xi ) − f (xi−1 )| + |f (yi ) − f (yi−1 )| ≤ Vf (a, b).
i=1 i=1

Taking sup over all divisions of [a, c] and [c, b], we obtain

Vf (a, c) + Vf (c, b) ≤ Vf (a, b).

Conversely: Let a = x0 < x1 < x2 < · · · < xn = b be a division of [a, b].

Write k = min{i : c ≤ xi }. Then {x0 , x1 , . . . , xk−1 , c} is a division of [a, c] and {c, xk , . . . , xn } is a
division of [c, b], and therefore
n
X k−1
X n
X
|f (xi ) − f (xi−1 )| = |f (xi ) − f (xi−1 )| + |f (xk ) − f (xk−1 )| + |f (xi ) − f (xi−1 )|
| {z }
i=1 i=1 i=k+1
≤|f (xk )−f (c)|+|f (c)−f (xk−1 )|

k−1
X n
X
≤ |f (xi ) − f (xi−1 )| + |f (c) − f (xk−1 )| + |f (xk ) − f (c)| + |f (xi ) − f (xi−1 )|
|i=1 {z } |
i=k+1
{z }
≤Vf (a,c) ≤Vf (c,b)

≤ Vf (a, c) + Vf (c, b)

sup
=⇒ Vf (a, b) ≤ Vf (a, c) + Vf (c, b).

Now we can easily prove an important characterization of functions of bounded variation.

Theorem 3.59. A function f : [a, b] → R is of bounded variation ⇐⇒ f = g − h, where g and h

are increasing on the interval [a, b].
3.57
Proof. ⇐ g, h increasing =⇒ g, h of bounded variation ⇒ f = g − h of bounded variation
(follows easily from the △-ineq.).
⇒ Suppose that f : [a, b] → R is of bounded variation. Then


f (x) = Vf (a, x) − Vf (a, x) − f (x) , x ∈ [a, b] (convention: Vf (a, a) = 0).

Claim: x 7→ Vf (a, x) and x 7→ Vf (a, x) − f (x) are increasing on the interval [a, b].
Proof: Let a ≤ x1 ≤ x2 ≤ b. Lemma 3.58 ⇒

Vf (a, x2 ) = Vf (a, x1 ) + Vf (x1 , x2 ) ≥ Vf (a, x1 )

| {z }
≥0
2017 53

and


Vf (a, x2 ) − f (x2 ) − Vf (a, x1 ) − f (x1 ) = Vf (a, x1 ) + Vf (x1 , x2 ) − f (x2 ) − Vf (a, x1 ) + f (x1 )


= Vf (x1 , x2 ) − f (x2 ) − f (x1 ) ≥ 0.
| {z }
≥|f (x2 )−f (x1 )|

Thus we may choose g = Vf (a, ·) and h = Vf (a, ·) − f.

Consequences:

Theorem 3.60. Let f : [a, b] → R be of bounded variation. Then

1. f has at mst countably many points of discontinuity,

2. ∃f ′ (x) for a.e. x ∈ [a, b].

3. f ′ on integrable.

Proof. Theorem 3.59 ⇒ f = g − h, where g and h are increasing.

Lemma 3.41 ⇒ g and h have at most countably many points of discontinuity, and therefore the
same holds for f .
Theorem 3.42 ⇒ ∃ g ′ (x), h′ (x) for a.e. x ∈ [a, b] ⇒ ∃f ′ (x) = g ′ (x) − h′ (x) for a.e. x ∈ [a, b].
Furthermore, |f ′ (x)| ≤ |g′ (x)| + |h′ (x)| for a.e. x ∈ [a, b], and |g′ | and |h′ | are integrable (L. 3.51),
hence f ′ is integrable.

The next result concerning ”integral functions” will be very useful in Section 3.67. We also get
more examples of functions of bounded variation.


Theorem 3.61. Let f : [a, b] → R be integrable, f ∈ L1 [a, b] , and
Z x
F (x) = f (t)dt, x ∈ [a, b].
a

Then F is of bounded variation on the interval [a, b], and the variation of F is
Z b
(3.62) VF (a, b) = |f (t)|dt = kf k1 .
a

Proof. Let a = x0 < x1 < x2 < · · · < xn = b be an arbitrary division of [a, b]. Then
n
X n Z
X xi Z xi−1  X n Z xi 
   
|F (xi ) − F (xi−1 )| =  f (t)dt − f (t)dt =  f (t)dt
i=1 i=1 a a i=1 xi−1

Xn Z xi Z b
≤ |f (t)|dt = |f (t)|dt.
i=1 xi−1 a

Taking the sup over all divisions of [a, b] we obtain

Z b
(3.63) VF (a, b) ≤ |f (t)|dt = kf k1 < ∞.
a
54 Real Analysis I

Hence F is of bounded variation.

Converse inequality: A. Let g : [a, b] → R be continuous (and thus integrable) and
Z x
G(x) = g(t)dt.
a

Let ε > 0 be arbitrary. Since g is uniformly continuous ([a, b] closed interval), ∃δ > 0 s.t.

|x − y| < δ
⇒ |g(x) − g(y)| < ε.
x, y ∈ [a, b]

Let a = x0 < x1 < x2 < · · · < xn = b be a division of [a, b] s.t. |xi − xi−1 | < δ ∀i = 1, 2, . . . , n.
Applying △-inequality (several times) we get ∀i = 1, 2, . . . , n
Z xi   Z xi

   
|G(xi ) − G(xi−1 )| =  g(t)dt =  g(t) − g(xi−1 ) dt + (xi − xi−1 )g(xi−1 )
xi−1 xi−1
Z xi


≥ (xi − xi−1 )|g(xi−1 )| −  g(t) − g(xi−1 ) dt
x
Z xi−1
i
≥ (xi − xi−1 )|g(xi−1 )| − |g(t) − g(xi−1 )| dt
xi−1 | {z }
<ε
Z xi
≥ |g(xi−1 )|dt − ε(xi − xi−1 )
xi−1
Z xi


≥ |g(t)| − |g(t) − g(xi−1 )| dt − ε(xi − xi−1 )
x
Z xi−1
i
Z xi
= |g(t)|dt − |g(t) − g(xi−1 )| dt − ε(xi − xi−1 )
xi−1 xi−1 | {z }
<ε
Z xi
≥ |g(t)|dt − 2ε(xi − xi−1 ).
xi−1

Taking the sum over i = 1, 2, . . . , n ⇒

n
sup
X n Z
X xi n
X
VG (a, b) ≥ |G(xi ) − G(xi−1 )| ≥ |g(t)|dt −2ε (xi − xi−1 )
i=1 xi−1
|i=1 {z } |i=1 {z }
Rb =b−a
= a |g(t)|dt

= kgk1 − 2ε(b − a).


ε > 0 mv. ⇒ VG (a, b) ≥ kgk1
(3.64) ⇒ VG (a, b) = kgk1
(3.63)

if g : [a, b] → R is continuous.


B. General case f ∈ L1 [a, b] : We set f (x) = 0 ∀x ∈ R \ [a, b], and so f ∈ L1 (R).
Theorem 2.36 implies that for all ε > 0 there exists g ∈ C(R) s.t. kf − gk1 < ε.
Denote g0 = g|[a, b], hence in particular
Z b
kf − g0 k1 = |f (t) − g(t)|dt ≤ kf − gk1 < ε.
a
2017 55

Define Z x
G0 (x) = g0 (t)dt, x ∈ [a, b].
a
Then
△-ineq.
VG0 (a, b) = VG0 −F +F (a, b) ≤ VG −F (a, b) +VF (a, b) < ε + VF (a, b)
| 0 {z }
≤ kg0 −f k1 <ε
(3.63)

(3.64)
⇒ VF (a, b) ≥ VG0 (a, b) − ε = kg0 k1 − ε = kg0 − f + f k1 − ε
Minkowski
≥ kf k1 − kg0 − f k1 −ε > kf k1 − 2ε.
| {z }
<ε

ε > 0 arbitr. ⇒ VF (a, b) ≥ kf k1
⇒ VF (a, b) = kf k1 .
(3.63)

Example 3.65. Let F : [0, 1] → R,

(
x sin x1 , if 0 < x ≤ 1,
F (x) =
0, if x = 0.

Then there exists no Lebesgue integrable function f : [0, 1] → R s.t.

Z x
(3.66) F (x) = f (t)dt, x ∈ [a, b],
a
1
i.e. x 7→ x sin is not an ”integral function”.
x
Reason: For instance, 3.56.2

⇒ F is not of bounded variation. On the other hand, if there were
a function f ∈ L1 [0, 1] for which (3.66) holds, then by Theorem 3.61 F would be of bounded
variation.

3.67 Absolutely continuous functions

If f : [a, b] → R is a differentiable function whose derivative f ′ is Riemann integrable over an interval
[a, b] (ex. if f ′ continuous), then
Z x
(3.68) f (x) = f (a) + f ′ (t)dt ∀x ∈ [a, b].
a

Question: more general version by using Lebesgue integral?

Let f : [a, b] → R be a function such that the derivative f ′ (x) exists for a.e. x ∈ [a, b] and f ′
integrable. In this section we study the question for which functions (3.68) holds. Example 3.1.4:
(3.68) does not hold for the Cantor 1/3-function, and so we need an extra condition.
Definition 3.69. A function f : [a, b] → R is absolutely continuous (on [a, b]) if for all ε > 0 there
exists δ > 0 s.t.
k
X
|f (bj ) − f (aj )| < ε
j=1
56 Real Analysis I

whenever ]a1 , b1 [, . . . , ]ak , bk [⊂ [a, b] are disjoint and

k
X k

X
ℓ [aj , bj ] = (bj − aj ) < δ.
j=1 j=1

(Note: the number of intervals (k) is arbitrary but always finite.)

Integral functions are absolutely continuous:

Lemma 3.70. Let f : [a, b] → R be integrable and

Z x
F (x) = f (t)dt, x ∈ [a, b].
a

Then F is absolutely continuous.

Proof. Corollary 2.7 (”absolute continuity of integrals”) implies that for all ve > 0 there exists
δ > 0 s.t.
Z


(3.71) E ∈ Leb [a, b] , m(E) < δ ⇒ |f (t)|dt < ε.
E

If ]a1 , b1 [, . . . , ]ak , bk [⊂ [a, b] are disjoint intervals s.t.

k
X k
[ 
disjoint
(bj − aj ) = m ]ai , bi [ < δ,
j=1
|i=1 {z }
=E

then
k
X Xk Z bi Z ai 
 
|F (bi ) − F (ai )| =  f (t)dt − f (t)dt
i=1 i=1 | a {z a }
Z bi
= f (t)dt
ai
k Z bi Z
X disjoint (3.71)
≤ |f (t)|dt = |f (t)|dt < ε.
i=1 ai E

Theorem 3.72. Let f : [a, b] → R be absolutely continuous. Then:

1. f is (uniformly) continuous,

2. f is of bounded variation,

3. the derivative f ′ (x) exists for a.e. x ∈ [a, b].

4. f ′ is integrable.

Proof. 1. Clear (take k = 1 in the definition).

2. Exerc.
2017 57

2. 3.60
3. f absolutely continuous ⇒ f of bounded variation ⇒ ∃f ′ (x) for a.e. x ∈ [a, b].
2. 3.60
4. f absolutely continuous ⇒ f of bounded variation ⇒ f ′ integrable.

Example 3.73. 1. f : [0, 1] → R,

(
x sin x1 , if 0 < x ≤ 1,
f (x) =
0, if x = 0,
3.72
is continuous but not of bounded variation (Exerc.) ⇒ f is not absolutely continuous.

2. Let f : [a, b] → R be Lipschitz-function, i.e. ∃L < ∞ s.t.

|f (x) − f (y)| ≤ L|x − y| ∀x, y ∈ [a, b].

Claim: f Lipschitz ⇒ f absolutely continuous.

Proof: Let ε > 0 and ]a1 , b1 [, . . . , ]ak , bk [⊂ [a, b] be disjoint intervals s.t.
k
X
(bi − ai ) < ε/L.
i=1

Then
k
X k
X
|f (bi ) − f (ai )| ≤ L (bi − ai ) < ε.
i=1 i=1

”Uniqueness theorem”
Theorem 3.74. If f : [a, b] → R is absolutely continuous and f ′ (x) = 0 for a.e. x ∈ [a, b], then f
is constant.
Proof. Let us show that f (c) = f (a) ∀c ∈]a, b[. Fix c ∈]a, b[ and denote

E = {x ∈]a, c[ : f ′ (x) = 0},

hence m(E) = c − a. Let ε > 0 be arbitrary and choose δ > 0 as in the definition of absolutely
continuity of f . For every x ∈ E (⊂]a, c[) there exist arbitrary short intervals [x, x+h] ⊂]a, c[, h > 0,
s.t.
|f (x + h) − f (x)|
< ε,
h
and therefore such intervals form a closed Vitali covering of E. The Vitali covering theorem ⇒ ∃
disjoint intervals Ij = [xj , yj ] ⊂]a, c[, j ∈ N, s.t.

(3.75) |f (yj ) − f (xj )| < ε(yj − xj )

and [

m E\ Ij = 0.
j∈N

The convergence of measures ⇒ ∃k ∈ N s.t.

k
[

m ]a, c[\ Ij < δ,
j=1
58 Real Analysis I

because
k
[ [
[



lim m ]a, c[\ Ij = m ]a, c[\ Ij = m E \ Ij = 0.
k→∞
j=1 j∈N j∈N

We may assume
a < x1 < y1 < x2 < y2 < · · · < xk < yk < c,
Sk
and hence ]a, c[\ j=1 Ij is a disjoint unionof open intervals ∆j =]pj , qj [, j = 1, . . . , k + 1, (see the
picture) and
k+1
X k
[

(qj − pj ) = m E \ Ij < δ.
j=1 j=1

a I1 I2 Ik c

∆1 ∆2 ∆k+1

We obtain
Xk k+1

X

|f (c) − f (a)| =  f (yj ) − f (xj ) + f (qj ) − f (pj ) 
j=1 j=1
k
X k+1
X
≤ |f (yj ) − f (xj )| + |f (qj ) − f (pj )|
| {z }
j=1 j=1
< ε(yj −xj ) | {z }
(3.75)
< ε
abs. cont.

k
X
<ε (yj − xj ) +ε
j=1
| {z }
≤c−a

≤ ε(c − a + 1).

Letting ε → 0 we get f (c) = f (a).

Example 3.76. Let f : [0, 1] → [0, 1] be the Cantor 1/3-function. L. 3.59 ⇒ f of bounded variation
(f increasing). On the other hand, f ′ (x) = 0 a.e., but f is not a constant, hence f is not absolutely
continuous.

Theorem 3.77. Let f : [a, b] → R. Then TFAE 5

1. f is absolutely continuous,

2. ∃f ′ (x) for a.e. x ∈ [a, b], f ′ is integrable and

Z x
f (x) = f (a) + f ′ (t)dt ∀x ∈ [a, b],
a

3. ∃ integrable g : [a, b] → R s.t.

Z x
f (x) = f (a) + g(t)dt ∀x ∈ [a, b].
a
5
TFAE = ”the following are equivalent”
2017 59

Proof. 1. ⇒ 2. Suppose that f is absolutely continuous.

3.72 ⇒ ∃f ′ (x) a.e. x ∈ [a, b], and f ′ is integrable.
Let Z x
F (x) = f ′ (t)dt, x ∈ [a, b].
a
f′ intva
Theorem 3.38 =⇒ F ′ (x) = f ′ (x) for a.e. x ∈ [a, b] ⇒ (F − f )′ (x) = 0 for a.e. x ∈ [a, b].

Lemma 3.70 ⇒ F abs. continuous easily
=⇒ F − f abs. continuous.
assumption: f abs. continuous
”Uniqueness theorem” 3.74 ⇒ F − f constant, and so
F (x) − f (x) = F (a) −f (a) ∀x ∈ [a, b]
| {z }
=0
Z x
⇒ f (x) = f (a) + F (x) = f (a) + f ′ (t)dt ∀x ∈ [a, b].
a

2. ⇒ 3. Choose g = f ′ .
3. ⇒ 1. Suppose that ∃ integrable g : [a, b] → R s.t.
Z x
f (x) = f (a) + g(t)dt ∀x ∈ [a, b].
a

Z x 

Denote G(x) = g(t)dt, x ∈ [a, b] 3.70
a =⇒ G abs. continuous
assumption: g integrable 

⇒ f = f (a) +G abs. continuous.

|{z}
vakio

Example 3.78. Let α > 0 and

(
xα sin x1 , if 0 < x ≤ 1,
f (x) =
0, if x = 0.
Claim: f is absolutely continuous ⇐⇒ α > 1.
Reason: Example 3.73 ⇒ f is not absolutely continuous for α = 1. Similarly we may verify that f
is not absolutely continuous for 0 < α < 1.
Suppose α > 1 : Then:
1 1
f ′ (x) = αxα−1 sin − xα−2 cos , 0 < x < 1.
x x
Now
1
lim xα−1 sin = 0 (since α > 1),
x→0+ x
and therefore x 7→ xα−1 sin x1
integrable on the interval [0, 1]. Furthermore, h(x) = xα−2 cos x1 is
integrable on the interval [0, 1] since
Z 1 Z 1
α−2 1 1
x |cos | dx ≤ xα−2 dx = .
0 x
| {z } 0 α − 1
≤1
60 Real Analysis I

Hence f ′ is integrable on the interval [0, 1] and

Z x
f (x) = f ′ (t)dt, 0 ≤ x ≤ 1.
0

Theorem 3.70 ⇒ f absolutely continuous for α > 1.

Theorem 3.79. Let f : [a, b] → R be increasing. Then

Z b
f ′ (t)dt = f (b) − f (a)
a

if and only if f is absolutely continuous on the interval [a, b].

Proof. Suppose that f is absolutely continuous: Theorem 3.77 ⇒

Z x
f ′ (t)dt = f (x) − f (a) ∀x ∈ [a, b].
a

Choosing x = b ⇒ claim.
Conversely: Suppose that
Z b
(3.80) f ′ (t)dt = f (b) − f (a).
a

Claim: f is absolutely continuous.

Theorem 3.77 ⇒ it suffices to prove
Z x
f ′ (t)dt = f (x) − f (a) ∀x ∈ [a, b].
a

Suppose on the contrary: ∃c ∈]a, b[ s.t.

Z c
f ′ (t)dt < f (c) − f (a).
a

Since f is increasing on intervals [a, c] and [c, b], we have

Z b Z c Z b
3.51
=⇒ f ′ (t)dt = f ′ (t)dt + f ′ (t)dt < f (c) − f (a) + f (b) − f (c) = f (b) − f (a).
a
|a {z } | c {z }
<f (c)−f (a) ≤f (b)−f (c)

This is a contradiction with the assumption (3.80).

Hence: Z x
3.77
f ′ (t)dt = f (x) − f (a) ∀x ∈ [a, b] =⇒ f abs. continuous.
a

Definition 3.81. A function f : [a, b] → R of bounded variation is singular if f ′ (x) = 0 for a.e.
x ∈ [a, b].
(Note: f of bounded variation ⇒ ∃f ′ (x) for a.e. x ∈ [a, b].)

Example 3.82. 1. Constant functions are singular (and abs. continuous).

2017 61

2. Cantor 1/3-function is singular.

3. Let f : [a, b] → R be absolutely continuous. Then
f singular ⇐⇒ f constant function.
Reason: 
f ′ (x) = 0 a.e. 3.74
=⇒ f (x) ≡ c constant.
f abs. continuous
It turns out that every function of bounded variation can be decomposed into an absolutely
continuous and a singular part. (Recall: f abs. continuous ⇒ f of bounded variation but not
conversely.)
Theorem 3.83 (Lebesgue decomposition). If f : [a, b] → R is of bounded variation, then
f = g + h,
where g is absolutely continuous and h is singular. Furthermore, the decomposition is unique up to
an additive constant.
Proof. By Theorem 3.60, f ′ (x) exists for a.e. x ∈ [a, b] and, moreover, f ′ is integrable. Define
Z x
g(x) = f ′ (t)dt.
a
Theorem 3.70 ⇒ g abs. continuous. Let h = f − g. Then h is of bounded variation and
3.38
h′ (x) = (f − g)′ (x) = f ′ (x) − g ′ (x) = f ′ (x) − f ′ (x) = 0 a.e. x ∈ [a, b]
⇒ h singular.
Hence f = g + h is a desired decomposition.
Uniqueness: Suppose that f = g1 + h1 = g2 + h2 , where g1 , g2 are absolutely continuous and h1 , h2
are singular. Now

w = h2 − h1 = g1 − g2 abs. continuous, 3.74
=⇒ w(x) ≡ c constant function
w′ (x) = h′2 (x) − h′1 (x) = 0 a.e. x ∈ [a, b]

Some additional facts:

1. Let f : [a, b] → R be absolutely continuous. Then f satisfies Lusin’s condition (N ), that is,
E ⊂ [a, b], m(E) = 0 ⇒ m(f E) = 0.

2. Another characterization for absolutely continuous functions (see, for instance, [GZ, 7.45]).
Let f : [a, b] → R. Then f is absolutely continuous ⇐⇒
(a) f continuous,
(b) f is of bounded variation,
(c) E ⊂ [a, b], m(E) = 0 ⇒ m(f E) = 0 ”condition (N )”.
3. Let f : [a, b] → R be a function s.t. ∃f ′ (x) ∀x ∈ [a, b] and f ′ is integrable. Then f s absolutely
continuous (see, for instance, [GZ, 7.47]).

THE END
62 Real Analysis I

References
[EG] Evans, Lawrence and Gariepy Ronald. Measure theory and fine properties of functions, CRC
Press, 1992.

[Fr] Friedman, Avner. Foundations of modern analysis, Dover Publications Inc., 1982.

[GZ] Gariepy, Ronald and Ziemer, William. Modern real analysis, PWS Publishing Company,
1994.

[HS] Hewitt, Edwin and Stromberg, Karl. Real and abstract analysis, Springer-Verlag, 1975.

[Ho] Holopainen, Ilkka. Mitta and integraali, Kevätlk. 2002.

http://www.helsinki.fi/˜iholopai/MitInt02.ps

[Jo] Jones, Frank. Lebesgue integration on Euclidean space, Jones and Bartlett Publishers, 1993.

[Mar] Martio, Olli. Reaalianalyysi I, kevät 1999.

[Mat] Mattila, Pertti. Geometry of sets and measures in Euclidean spaces, Cambridge University
Press, 1995.

[MW] McDonald, John N. and Weiss, Neil A. A course in real analysis, Academic Press Inc., 1999.

[Ro] Royden, H. L. Real analysis, Macmillan Publishing Company, 1988.

[Ru] Rudin, Walter. Real and complex analysis, McGraw-Hill Book Co., 1987.