Advanced Analysis

GEORGIOS PSARADAKIS

05/07/2024
Contents

1 Prerequisites 7
1.1 Subsets of the Euclidean space . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.1 Abstract sets and sequences of abstract sets . . . . . . . . . . . . . . 7
1.1.2 Subsets of the real line . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.3 Subsets of the n-dimensional Euclidean space . . . . . . . . . . . . . 9
1.2 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Volume of unit ball in Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 Measure 17
2.1 Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Measurable sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Regularity of measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4 Metric measures and Carathéodory’s criterion . . . . . . . . . . . . . . . . . 29

3 Brunn-Minkowski and isoperimetric inequalities 33

4 Integral 39
4.1 Measurable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 Lusin and Egoroff theorems . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 Integration of measurable functions . . . . . . . . . . . . . . . . . . . . . . 45
4.4 L p spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5 The basic theorems of advanced analysis 53

5.1 Limit theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.2 Product measures and Fubini’s theorem . . . . . . . . . . . . . . . . . . . . 59
5.3 Some density results for L n -summable functions . . . . . . . . . . . . . . . 61

6 Some representation formulas 63

3
4 CONTENTS

7 The Fourier transform 69

7.1 The Fourier transform in L1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.2 The Fourier transform in L2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

8 More examples of Radon measures 75

8.1 The Riesz representation theorem . . . . . . . . . . . . . . . . . . . . . . . 76
8.2 Positive linear functionals and distributions . . . . . . . . . . . . . . . . . . 79
8.3 Weak derivatives of functions of bounded variation . . . . . . . . . . . . . . 81

9 Elementary convexity 83
9.1 Convex functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
9.2 Lipschitz functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
9.3 A characterization of convex functions . . . . . . . . . . . . . . . . . . . . . 91
Preface

This set of notes grew from the graduate course “Advanced Analysis” I taught at the Mathe-
matics Institute of the University of Mannheim (Spring 2017, 2018, 2019 and Fall 2020).
The aim of the course was to present the analytic techniques required to prove various
functional inequalities. The lectures started with basic knowledge of real analysis (mea-
sure and integral) and then it went into some advanced topics in analysis such as L p spaces,
symmetric-decreasing rearrangement of functions, the Fourier transform, distributions and
Sobolev spaces. In parallel, the proofs of inequalities such as Riesz’s rearrangement in-
equality, Young’s inequality, Hardy-Littlewood-Sobolev inequality and logarithmic Sobolev
inequality, were given. The skeleton of the course was the first half of the second edition of
Analysis by E. H. Lieb and M. Loss. The first chapter of this book includes a very brief intro-
duction to measure theory. For instance, outer/inner regularity of the Lebesgue measure and
Lusin’s theorem are included in the exercises of the first chapter. Having taught only some
basic facts about the Lebesgue integral (sometimes without proofs), students found difficult
to study the book itself and were restricted to the classroom notes.
The aim of these notes is to quickly cover possible measure theoretic gaps of the students
knowledge so that to appreciate the remarkable book by Lieb and Loss. In order not to deviate
from the proving-inequalities-purpose of the course, we present: (i) a special proof of the
Euclidean isoperimetric inequality due to H. Hadwiger and D. Ohmann, that essentially uses
only the outer/inner regularity of Lebesgue measure, (ii) an elementary version of the coarea
formula which together with the isoperimetric inequality leads to the L1 -Sobolev inequality.
I start with a section of material students are expected to know from undergraduate anal-
ysis. The main part of sections 2, 4 and 5 is a more explanatory presentation of particular
subsections from the first two chapters of Measure theory and fine properties of functions by
L. C. Evans and R. F. Gariepy. For the presentation and proof of the isoperimetric inequality
through the Brunn-Minkowski inequality of section 3, I followed the book Geometric mea-
sure theory by H. Federer. The coarea formula of section 6 is from Sobolev spaces by V. G.
Maz’ya. Section 7 is a brief account on convex and Lipschitzian functions. For these notes
and throughout the course I have consulted several times the books found at the end of these
notes.
I want to thank Veniamin Gvozdik and Paul Nikolaev for spotting several misprints. Paul

5
6 CONTENTS

has provided me with the nice proof of Lemma 9.2.8. Any further misprints or comments are
very welcome at
psaradakis at outlook dot com
Chapter 1

Prerequisites

1.1 Subsets of the Euclidean space

1.1.1 Abstract sets and sequences of abstract sets
Let X be a nonempty set and 2X be the set of all subsets of X.
1. Writing Ac for the complement of A ⊆ X, that is Ac := X \ A, there holds
 [ c \  \ c [
A = Ac and A = Ac ,
A∈F A∈F A∈F A∈F

for any family of sets F ⊆ 2X .

2. A sequence {Ak ⊆ X}k∈N is said to non-decrease (to k∈N Ak ) if Ak ⊆ Ak+1 for all k ∈ N,
S

and to non-increase (to k∈N Ak ) if Ak ⊇ Ak+1 for all k ∈ N.

T

3. For a sequence {Ak ⊆ X}k∈N and j ∈ N, let

[ \
M j := Ak and M j := Ak .
k≥ j k≥ j

Then {M j } j∈N is non-increasing, {M j } j∈N is non-decreasing and we set

\ [
lim sup Ak := M j and lim inf Ak := M j.
k→∞ k→∞
j∈N j∈N

1.1.2 Subsets of the real line

1. For a nonempty subset A of the extended real line R̄ = R ∪ {±∞}, let

7
8 CHAPTER 1. PREREQUISITES

(i) sup A to be the least upper bound of A,

(ii) inf A to be the greatest lower bound of A.

We have the following “variational” characterizations

n x ≤ M ∀x ∈ A, (upper bound)
(i) R ∋ M = sup A ⇔
∀ ε > 0, ∃ xε ∈ A such that xε > M − ε. (least one)
Taking ε = 1/k, k ∈ N, we obtain a sequence {xk ∈ A}k∈N such that xk → M as k → ∞.

n x ≥ m ∀x ∈ A, (lower bound)
(ii) R ∋ m = inf A ⇔
∀ ε > 0, ∃ xε ∈ A such that xε < m + ε. (greatest one)
Taking ε = 1/k, k ∈ N, we obtain a sequence {xk ∈ A}k∈N such that xk → m as k → ∞.

Remark 1.1.1. Initially we will refer to this characterization when we use it. After some
point we will use it without any comment.

2. For a sequence {ak ∈ R}k∈N and j ∈ N, let

m j := sup ak and m j := inf ak .

k≥ j k≥ j

Then {m j } j∈N is non-increasing, {m j } j∈N is non-decreasing and we set

lim sup ak := lim m j and lim inf ak := lim m j .

k→∞ j→∞ k→∞ j→∞

When finite, lim supk→∞ ak and lim infk→∞ ak correspond respectively to the largest and small-
est subsequential limit points of {ak }k∈N . Thus,

ak → a ∈ R̄ ⇐⇒ lim sup ak = lim inf ak = a.

k→∞ k→∞

Exercise 1.1.2. Prove lim infk→∞ (−ak ) = − lim supk→∞ ak .

Exercise 1.1.3. Suppose there exists K ∈ N such that ak ≥ bk for all k ∈ N with k ≥ K. Prove
lim infk→∞ ak ≥ lim infk→∞ bk .
1.1. SUBSETS OF THE EUCLIDEAN SPACE 9

1.1.3 Subsets of the n-dimensional Euclidean space

We assume the student is familiar with elementary metric space properties of Rn . Notions like
Cauchy sequence, density and separability will be reviewed when needed in the classroom.
For our purposes in these notes, we only need to say that for a point x = (x1 , ..., xn ) ∈ Rn we
use |x| to denote its Euclidean distance to the origin; that is,
s
n
|x| := ∑ xi2,
i=1

and recall the fundamental notion of a limit point of a sequence of points in Rn :

Definition 1.1.4. A point x ∈ Rn is called limit point of the sequence {xk ∈ Rn }k∈N , whenever
|xk − x| → 0 as k → ∞.

1. Diameter, distance between sets and open balls

Definition 1.1.5. (i) The diameter of a nonempty A ⊆ Rn is given by


diam(A) := sup |x − y| x, y ∈ A .

(ii) The distance between nonempty A1 , A2 ⊆ Rn is given by


dist(A1 , A2 ) := inf |x − y| x ∈ A1 , y ∈ A2 .

(iii) The open ball of radius r > 0 having centre at x ∈ Rn is denoted by Br (x); that is,

Br (x) := {y ∈ Rn |y − x| < r}.

2. Interior of a set, open/closed sets, limit point of a set, closure and boundary of a set
Definition 1.1.6. Let A be a nonempty subset of Rn . Then

(i) x ∈ A is called interior point of A if Br (x) ⊆ A for some r > 0,

(ii) the set of all interior points of A is called interior of A and is denoted by A◦ ,

(iii) A is called open if A◦ = A and closed if its complement is open,1

(iv) x ∈ Rn is called limit point of the set A if it is the limit point of a sequence in A,
10
/ is open by convention
10 CHAPTER 1. PREREQUISITES

(v) x ∈ A is called isolated point of A if it is not the limit of any nontrivial sequence in A,

(vi) the union of A and all its limit points is called the closure of A and is denoted by Ā,

(vii) the set Ā \ A◦ is called the boundary of A and is denoted by ∂ A.

Exercise 1.1.7. Prove we have diam(C̄) = diam(C) for any C ⊂ Rn .

Theorem 1.1.8. (i) B̄r (x) = {y ∈ Rn |y − x| ≤ r}.

(ii) A ⊂ Rn is closed if and only if A = Ā.

(iii) Given A ⊂ Rn we have that Ā is the smallest closed set containing A.

(iv) x ∈ A is an isolated point of A if and only if Br (x) ∩ A = {x} for some r > 0.

Theorem 1.1.9. (i) The union (intersection) of any number of open (closed) sets is open
(closed).

(ii) The intersection (union) of a finite number of open (closed) sets is open (closed).

3. Bounded and compact sets

Definition 1.1.10. E ⊂ Rn is called bounded if E ⊂ Br (0) for some r > 0.

Exercise 1.1.11. (i) Prove that for any x = (x1 , ..., xn ) ∈ Rn , there holds
n
|x j | ≤ |x| ≤ ∑ |xi |, j = 1, ..., n.
i=1

(ii) Write down a proof of the triangle inequality in Rn :

|x| − |y| ≤ |x + y| ≤ |x| + |y| ∀ x, y ∈ Rn .

(iii) Prove Br (x0 ) is an open subset of Rn . Prove it is also bounded.

(iv) Prove E ⊂ Rn is bounded if and only if for each i ∈ {1, ..., n}, the i-th coordinates of its
points form a bounded subset of R.

Definition 1.1.12. K ⊂ Rn is called compact if every open cover of K has a finite subcover.

Theorem 1.1.13. K ⊂ Rn is compact if and only if

(i) it is closed and bounded,

1.1. SUBSETS OF THE EUCLIDEAN SPACE 11

(ii) every sequence of points of K has a subsequence converging in K.

Exercise 1.1.14. Prove the distance of a compact set to a disjoint closed set is positive.
Definition 1.1.15. We say U1 is compactly contained in U2 and write this as U1 ⋐ U2 , when-
ever U1 and U2 are open subsets of Rn such that Ū1 ⊂ U2 and U1 is bounded.
Remark 1.1.16. From exercise 1.1.14 we deduce that given a compact subset K of an open
set U ⊆ Rn , then δ := dist{K,U c } ∈ (0, ∞]. For any ε ∈ (0, δ ) we define

Kε := {x ∈ Rn dist(x, K) ≤ ε}.

Then Kε is again a compact subset of U with K ⊂ (Kε )◦ .

As a consequence, given U1 ⋐ U2 , by taking K = Ū1 and U = U2 in the above remark, it
readily follows that
Proposition 1.1.17. Given U1 ⋐ U2 , there exists V such that U1 ⋐ V ⋐ U2 .

4. Connected and convex sets

Definition 1.1.18. A nonempty A ⊊ Rn is called connected if it cannot be included in the
disjoint union of two nonempty open sets, each having nonempty intersection with A.
Remark 1.1.19. If A is open, connectedness of A implies that any two points of A can be
connected by a polygonal curve which is contained in A.
Definition 1.1.20. A nonempty A ⊊ Rn is called convex if the line segment connecting any
two of its points is contained in A.

5. Intervals
Definition 1.1.21. For n ∈ N, an n-dimensional interval I is a subset of Rn of the form

I = x = (x1 , ..., xn ) αk ≤ xk ≤ βk , k = 1, ..., n
≡ [α1 , β1 ] × [α2 , β2 ] × ... × [αn , βn ],

where αk < βk for all k = 1, ..., n. If the edge lengths βk − αk are all equal, then I is called a
cube. For an open interval/cube take strict inequalities in the above form. Two intervals are
non-overlapping whenever their interiors are disjoint.
The volume of an interval equals the volume of its interior and is given by
n
Vol(I) = ∏ (βk − αk ).
k=1
12 CHAPTER 1. PREREQUISITES

Theorem 1.1.22. Every open set in R can be written as a countable union of disjoint open in-
tervals. Every open set in Rn , n ∈ N, can be written as a countable union of non-overlapping
intervals.
Exercise 1.1.23. For compact X ⊂ Rn , prove there exists a non-increasing sequence {Xk ⊂
Rn }k∈N with X = k∈N Xk and each Xk consists of a finite union of non-overlapping intervals.
T

Exercise 1.1.24. Show that an open set in Rn , n ∈ N\{1}, is not necessarily a countable union
of disjoint open sets by explaining why the open unit disk in the plane cannot be written as a
countable disjoint union of open squares.
Disjointness can be achieved in the following way:
Definition 1.1.25. Let k ∈ Z. The n-dimensional cubes of the form
h j j + 1 h j j + 1
1 1 n n
k
, k
× ... × k
, k , ( j1 , ..., jn ) ∈ Zn ,
2 2 2 2
are called (right) half-open cubes of side length 2−k , or simply dyadic cubes.
Theorem 1.1.26. Every open set in Rn , n ∈ N, can be written as a countable union of disjoint
dyadic cubes.

1.2 Functions
1. Pre-image properties
For any function f : X → Y and any A, {Ak }k∈N subsets of Y , there holds

c
f −1 (Ac ) = f −1 (A) and f −1 f −1 (Ak ) .
[
[

Ak =
k∈N k∈N

2. Continuity
We denote by E an arbitrary nonempty subset of Rn and by f : E → R̄ a given function.
For any limit point x0 of E we consider the punctured balls
B∗r (x0 ) := Br (x0 ) \ {x0 }, r > 0.
We set
Mr (x0 ) := sup f (x) and mr (x0 ) := inf f (x).
B∗r (x0 )∩E B∗r (x0 )∩E

As r decreases, {Mr (x0 )}r>0 is non-increasing, {mr (x0 )}r>0 is non-decreasing and we set
lim sup f (x) := lim Mr (x0 ) and lim inf f (x) := lim mr (x0 ).
E∋x→x0 r→0 E∋x→x0 r→0
1.2. FUNCTIONS 13

Definition 1.2.1. Let x0 ∈ E be a limit point of E.

(i) f is continuous at x0 if ∀ ε > 0, ∃ δ > 0 s.t. | f (x) − f (x0 )| < ε whenever |x − x0 | < δ ,
x ∈ E.
(ii) f is upper semi-continuous at x0 if lim supE∋x→x0 f (x) ≤ f (x0 ).
(iii) f is lower semi-continuous at x0 if lim infE∋x→x0 f (x) ≥ f (x0 ).
Theorem 1.2.2. f is continuous at x0 if and only if | f (x0 )| < ∞ and is both upper semi-
continuous and lower semi-continuous at x0 .
Theorem 1.2.3. f is continuous at a non-isolated point of x0 ∈ E if and only if it is se-
quentially continuous at x0 ; that is, for any sequence {xk ∈ E}k∈N with xk → x we have
f (xk ) → f (x).
Definition 1.2.4. (i) f is continuous in E if f is continuous at any limit point of E which
lies in E.
(ii) f is uniformly continuous in E if ∀ ε > 0, ∃ δ > 0 s.t. | f (x) − f (y)| < ε whenever
|x − y| < δ , x, y ∈ E.
Theorem 1.2.5. Let E be compact and f be continuous in E. Then the following are true:
(i) f is bounded in E.
(ii) f attains its supremum and infimum in E.
(iii) f is uniformly continuous in E.
Lemma 1.2.6. (Urysohn’s lemma - elementary form). Let U ⊆ Rn be open and K ⊂ U be
compact. There exists a real valued function g ∈ Cc (Rn ); that is, g is continuous in Rn and
vanishes outside some compact subset of Rn , such that
(i) 0 ≤ g(x) ≤ 1 for all x ∈ Rn ,
(ii) g(x) = 1 for all x ∈ K,

(iii) sprt(g) := {x ∈ Rn g(x) ̸= 0} ⊂ U.

The above lemma is essentially a corollary of proposition 1.1.17. Indeed, the hypotheses
K ◦ ⋐ U implies the existence of W such that K ◦ ⋐ W ⋐ U. The function defined by
dist(x,W c )
g(x) := , x ∈ Rn ,
dist(x, K) + dist(x,W c )
readily satisfies (i), (ii) while sprt(g) = W̄ ⊂ U. The only unclear at the moment requirement
is the continuity of g. However, this follows from the fact that the distance to any nonempty
closed subset of Rn is a Lipschitz function (see example 9.2.3-(i) and the remark before that).
14 CHAPTER 1. PREREQUISITES

3. Differentiability
We denote by E an arbitrary nonempty subset of Rn and by f : E → Rm a given function.

Definition 1.2.7. Let x0 ∈ E be an interior point of E. f is differentiable at x0 if there exist a

matrix Mx0 = [αi j ]m×n such that
| f (x0 + h) − f (x0 ) − Mx0 h|
lim = 0.
R n
∋h→0 |h|
The matrix Mx0 is called the derivative of f at x0 and is denoted by D f (x0 ). In case m = 1 we
write ∇ f (x0 ) instead of D f (x0 ) and call this the gradient vector of f at x0 .
Remark 1.2.8. Using mollification we can strengthen the continuity of g in lemma 1.2.6 to
read g ∈ Cc∞ (Rn ); that is, g is infinitely many times differentiable with respect to any variable
and has compact support. Simply, for any ε ∈ (0, 12 dist{K,U c }), consider the function
Z
g(x) = ηε ∗ χKε (x) = ηε (x − y)dy x ∈ Uε ,
Kε

where
Kε := {x ∈ Rn dist(x, K) ≤ ε}, Uε := {x ∈ Rn dist(x,U c ) > ε},
and ηε : Rn → [0, ∞) is the standard mollifier given by ηε (x) := ε −n η(x/ε), x ∈ Rn , where
c exp{(|x|2 − 1)−1 }
(
|x| < 1,
η(x) := (1.2.1)
0 |x| ≥ 1,
R
with the constant c > 0 being such that Rn η(x) dx = 1.
Theorem 1.2.9. The gradient vector of a C1 (Ω) scalar
function f at a point x0 of any non
empty level surface Sλ = {x ∈ Ω f (x) = λ }, λ ∈ R, is perpendicular to the tangent at x0 of
any C1 -curve that lies on Sλ and goes through x0 . In other words, ∇ f (x0 ) is perpendicular
to the surface Sλ at x0 ∈ Sλ .
proof. Suppose ⃗γ(t) ∈ Sλ for all t ∈ I := (α, β ) ⊆ R and t0 ∈ I is such that ⃗γ(t0 ) = x0 ∈ Sλ .
Then

f ⃗γ(t) = λ for all t ∈ I.
Differentiating with respect to t we readily get by the chain rule that

d⃗γ
∇ f ⃗γ(t) · (t) = 0 for all t ∈ I.
dt
For t = t0 this gives ∇ f (x0 ) · d⃗
γ
dt (t0 ) = 0, which implies the result since
d⃗γ
dt (t0 ) is the tangent
vector to the curve at its point ⃗γ(t0 ).
1.3. VOLUME OF UNIT BALL IN Rn 15

4. The Riemann integral

Let I be an interval in Rn and f : I → R be a bounded function. Decomposing I into a finite
collection I of non-overlapping intervals {Ik }k , k = 1, ..., N, we define

∥I ∥ := max diam(Ik ).
k∈{1,...,N}

Selecting xk ∈ Ik , k = 1, ..., N, we define

N
RI ( f ) := ∑ f (xk ) Vol(Ik ).
k=1

We further define
N N
UI ( f ) := ∑ sup f (x) Vol(Ik ) and LI ( f ) := inf f (x) Vol(Ik ).
∑ x∈I
k=1 x∈Ik k=1 k

We say the Riemann integral of f on I exists whenever the limit lim∥I ∥→0 RI exists. If the
Riemann integral of f on I exists we write
Z
f = lim RI ( f ).
I ∥I ∥→0

Theorem 1.2.10. (i) The Riemann integral of f on I exists if and only if

inf UI ( f ) = sup UI ( f ).
I I

(ii) If f is continuous in I then the Riemann integral of f on I exists.

Exercise 1.2.11. Let f : [0, 1] → R be a continuous function. Prove

Z 1 Z 1
−nx
f (0) = lim ne f (x)dx and f (1) = lim nxn f (x)dx.
n→∞ 0 n→∞ 0

1.3 Volume of unit ball in Rn

The gamma function is given by
Z ∞
Γ(s) := e−x xs−1 dx, s > 0.
0
16 CHAPTER 1. PREREQUISITES

Through the gamma function we define the number

π s/2
ωs := , s ≥ 0.
Γ(1 + s/2)

If s ∈ N then ωs is precisely the volume of the unit ball of Rs . Also the volume of a ball of
radius r > 0 is in this case given by

Vol(Br (x0 )) = ωs rs .

Exercise 1.3.1. Prove the following

2 √
(i) 0∞ e−t dt = π/2.
R

√
(ii) Γ(1/2) = π.

(iii) Γ(s + 1) = sΓ(s) for all s > 0.

√
(iv) For any s ∈ N there holds Γ(s) = (s − 1)! and Γ(s + 1/2) = 1 · 3 · ... · (2s − 1) · π/2s .
R ∞ −α/2 −r2 /(4s)
(v) For all α > 2 and r > 0, there holds 0 s e ds = (2/r)α−2 Γ(α/2 − 1).
Chapter 2

Measure

Throughout, X stands for a nonempty set.

2.1 Measures
Definition 2.1.1 (measure). A mapping µ : 2X → [0, ∞] is called measure1 provided

/ = 0, and
(i) µ(0)

(ii) µ(A) ≤ ∑k∈N µ(Ak ) whenever A ⊆

S
k∈N Ak .

Remark 2.1.2. The two properties readily imply the monotonicity of µ:

A ⊆ B =⇒ µ(A) ≤ µ(B),

as well as the subadditivity of µ:

[
µ( Ak ) ≤ ∑ µ(Ak ).
k∈N k∈N

It is easy to see that one can replace the second requirement in the above definition with
monotonicity plus subadditivity.

Exercise 2.1.3. Let µ be a measure on X and suppose {Ak ⊆ X}k∈N satisfies ∑k∈N µ(Ak ) < ∞.
Prove that µ(lim supk→∞ Ak ) = 0.

Example 2.1.4. Important examples of measures are:

1 outer measure in other texts

17
18 CHAPTER 2. MEASURE

(i) Dirac’s delta measure concentrated to a given x ∈ X; that is the measure on X defined
by
n 1 if x ∈ A,

δx (A) := = χA (x) , A ⊆ X.
0 if x ∈
/ A.
Proof. Let A ⊆ k∈N Ak . If x ∈ / k∈N Ak then δx (A) ≤ δx ( k∈N Ak ) = 0 = ∑k∈N δx (Ak ).
S S S

If x ∈ k∈N Ak then x is in Ak0 for some k0 ∈ N, which implies δx (A) ≤ δx ( k∈N Ak ) =

S S

1 = δx (Ak0 ) ≤ ∑k∈N δx (Ak ).

(ii) The Lebesgue measure on Rn (denoted throughout by L n ), defined for any A ⊆ Rn by
n [ o
L n (A) := inf ∑ Vol(Ik ) {Ik }k∈N are intervals and A ⊆ Ik .
k∈N k∈N

Proof. Let A ⊆ k∈N Ak with L n (Ak ) < ∞ for all k ∈ N and fix ε > 0. Given k ∈ N, by
S

definition of L n (Ak ) and the variational characterization of inf, we can choose intervals
(k) (k)
{I j } j∈N such that Ak ⊆ j∈N I j and
S

(k) ε
∑ Vol(I j ) < L n (Ak ) + .
j∈N 2k
(k)
Since A ⊆ the definition of L n (A) implies
S S
k∈N j∈N I j ,
(k)
L n (A) ≤ ∑ ∑ Vol(I j ).
k∈N j∈N

Combining the two estimates, L n (A) < ∑k∈N L n (Ak ) + ε, which implies the subaddi-
tivity since ε > 0 is arbitrary.
(iii) The s-dimensional Hausdorff measure on Rn (denoted throughout by H s ), defined for
any s ≥ 0 and any A ⊆ Rn by
H s (A) := lim Hδs (A),
δ →0+
where (see §1.3 for the definition of ωs )
n  diamC s o
j
Hδ (A) := inf ∑ ωs
[
s
A⊆ C j , diam(C j ) ≤ δ .
j∈N 2 j∈N

Remark 2.1.5. Observe that for δ1 < δ2 , all coverings of A by sets with diameters no
more than δ1 are included in the set of coverings of A by sets with diameters no more
than δ2 . Thus Hδs increases as δ decreases. This justifies writing “lim” in the above
definition and also
H s (A) = sup Hδs (A).
δ >0
2.1. MEASURES 19

Remark 2.1.6. For s = 0 we deduce that H 0 is just the counting measure; that is,
H 0 (A) equals the number of elements in A when A is
finite, and infinity otherwise.
This follows from the fact that ω0 = 1, hence H 0 {a} = 1 for any a ∈ Rn .

Proof. We show first that Hδs is a measure. Let A ⊆ k∈N Ak with Hδs (Ak ) < ∞ for all
S

k ∈ N and fix ε > 0. Given k ∈ N, by definition of Hδs (Ak ) and the variational char-
(k) (k) (k)
acterization of inf, we can choose {C j } j∈N such that Ak ⊆ j∈N C j , diam(C j ) ≤ δ
S

for all j ∈ N and

 diamC(k) s ε
j
∑ ωs < Hδs (Ak ) + k .
j∈N 2 2
(k) (k)
Since A ⊆ and diam(C j ) ≤ δ for all k, j ∈ N, the definition of Hδs (A)
S S
k∈N j∈N C j
implies
 diamC(k) s
j
Hδs (A) ≤ ∑ ∑ ωs .
k∈N j∈N 2
Combining the two estimates, Hδs (A) < ∑k∈N Hδs (Ak ) + ε, which implies the subad-
ditivity since ε > 0 is arbitrary. □
We show next that H s is a measure. Since Hδs is a measure, if A ⊆
S
k∈N Ak then
Hδs (A) ≤ ∑ Hδs(Ak ).
k∈N

By remark 2.1.5 we have Hδs (Ak ) ≤ H s (Ak ) for all k ∈ N, thus

Hδs (A) ≤ ∑ H s(Ak ),
k∈N

and the subadditivity follows by taking the limit δ → 0.

(iv) The restriction µ|A of a measure µ on X to an arbitrary A ⊂ X; that is,
µ|A (B) := µ(B ∩ A), B ⊆ X.

Proof. Let B ⊆ with µ|A (Bk ) < ∞ for all k ∈ N and note
S
k∈N Bk
[
[
Bk ∩ A = (Bk ∩ A). (2.1.1)
k∈N k∈N
Using the monotonicity of µ, then (2.1.1) and finally the subadditivity of µ,
 [

µ|A (B) ≤ µ Bk ∩ A ≤ ∑ µ(Bk ∩ A) = ∑ µ|A (Bk ).
k∈N k∈N k∈N

Thus µ|A is also subadditive.

20 CHAPTER 2. MEASURE

Exercise 2.1.7. Prove for each one of the measures given in the above example, that its value
on the empty set is zero.

Exercise 2.1.8. Prove L n {a} = 0 for a ∈ Rn . Prove that for any countable set E in Rn we

have L n (E) = 0.

Exercise 2.1.9. Prove that the Lebesgue measure of an interval equals its volume.

Exercise 2.1.10. Prove that the Lebesgue measure of a hyperplane is zero.

Exercise 2.1.11. We set λ X := {λ x x ∈ X} whenever λ > 0 and X ⊆ Rn . Prove that

L n (λ X) = λ n L n (X).

2.2 Measurable sets

Let µ be a measure on X.

Definition 2.2.1 (µ-measurable set). A ⊆ X is called µ-measurable if

µ(B) ≥ µ(B ∩ A) + µ(B \ A) ∀ B ⊆ X, µ(B) < ∞.

Remark 2.2.2. The above inequality is trivial when µ(B) = ∞. Also, since B = (B ∩ A) ∪
(B \ A) we get by the subadditivity of µ that µ(B) ≤ µ(B ∩ A) + µ(B \ A). In particular,
µ-measurability of A ⊆ X is equivalent with

µ(B) = µ(A ∩ B) + µ(B \ A) ∀ B ⊆ X.

Exercise 2.2.3. Let X ̸= 0/ and define the following function on subsets A of X

n 1 if A ̸= 0,
/
µ(A) =
0 if A = 0.
/

Prove that µ is a measure on X and find all µ-measurable subsets of X.

Theorem 2.2.4. (i) A is µ-measurable ⇔ Ac is µ-measurable.

(ii) 0,
/ X and sets of µ-measure 0 are µ-measurable.

(iii) Any µ-measurable set is a also µ|A -measurable for any A ⊂ X.

(iv) A finite union (or intersection) of µ-measurable sets is µ-measurable.

2.2. MEASURABLE SETS 21

Proof. The first assertion readily follows from the definition of µ-measurability and so is
the µ-measurability of 0/ and X. Let µ(A) = 0 and B ⊆ X with µ(B) < ∞. Since B ∩ A ⊆ A
and B \ A ⊆ B, we get from the monotonicity of µ that µ(B ∩ A) = 0 and µ(B \ A) ≤ µ(B),
respectively. Add these to get µ(B ∩ A) + µ(B \ A) ≤ µ(B) as required. To prove the third
assertion let C ⊆ X be µ-measurable and B ⊆ X with µ|A (B) < ∞. It is enough to show
µ|A (B
∩C) + µ|A (B \C)
≤ µ|A (B). Since (B \C) ∩ A = (B ∩ A) \C, this is written as µ (B ∩
A) ∩C + µ (B ∩ A) \C ≤ µ(B ∩ A), which is implied by the µ-measurability of C. For the
final assertion, let B ⊆ X with µ(B) < ∞ and pick A1 , A2 ⊆ X. If A1 is µ-measurable then

µ(B) = µ(B ∩ A1 ) + µ(B \ A1 ).

If A2 is µ-measurable then



µ(B \ A1 ) = µ (B \ A1 ) ∩ A2 + µ (B \ A1 ) \ A2 .

Thus if they are both µ-measurable we have

µ(B) = µ(B ∩ A1 ) + µ (B \ A1 ) ∩ A2 + µ (B \ A1 ) \ A2 .


Noting B ∩ (A1 ∪ A2 ) = (B ∩ A1 ) ∪ (B \ A1 ) ∩ A2 , the subadditivity of µ gives



µ B ∩ (A1 ∪ A2 ) ≤ µ(B ∩ A1 ) + µ (B \ A1 ) ∩ A2 ,

and inserting this in the previous equality,

µ(B) ≥ µ B ∩ (A1 ∪ A2 ) + µ (B \ A1 ) \ A2 .

Observing (B \ A1 ) \ A2 = B \ (A1 ∪ A2 ), this last inequality says A1 ∪ A2 is µ-measurable.

By induction, the union of finitely many µ-measurable sets is µ-measurable. Taking com-
plements we get using (i) that the intersection of finitely many µ-measurable sets is µ-
measurable.

Theorem 2.2.5. Let {Ak ⊂ X}k∈N be a sequence of µ-measurable sets.

(i) If {Ak }k∈N are disjoint, then µ k∈N Ak = ∑k∈N µ(Ak ).
S

(ii) If {Ak }k∈N is non-decreasing, then limk→∞ µ(Ak ) = µ k∈N Ak .
S

(iii) If {Ak }k∈N is non-increasing and µ(A1 ) < ∞, then limk→∞ µ(Ak ) = µ
T
k∈N Ak .
S T
(iv) The sets k∈N Ak and k∈N Ak are µ-measurable.
22 CHAPTER 2. MEASURE

Proof. (i) Let {Ak }k∈N be a sequence of µ-measurable disjoint sets. By the µ-measurability
of A j+1 , j ∈ N,
j+1
[  j+1
[   j+1
[ 




µ Ak = µ Ak ∩ A j+1 + µ Ak \ A j+1 (2.2.1)
k=1 k=1 k=1
j
[

= µ(A j+1 ) + µ Ak .
k=1
Repeating this formula in the last summand, we arrive at
j+1
[ j−1
[ j+1



µ Ak = µ(A j+1 ) + µ(A j ) + µ Ak = ... = ∑ µ(Ak ) ∀ j ∈ N.
k=1 k=1 k=1

This readily implies by the monotonicity of µ that

j+1 [

∑ µ(Ak ) ≤ µ Ak ∀ j ∈ N.
k=1 k∈N

The proof is completed by taking the limit as j → ∞. Note µ( k∈N Ak ) ≤ ∑k∈N µ(Ak )
S
is also
true by the subadditivity of µ. □
(ii) Let {Ak }k∈N be a non-decreasing sequence of µ-measurable sets. For any k ∈ N write
Ak as a disjoint union of a finite number of µ-measurable sets (because of Theorem 2.2.4-(i)
and (iv)) as follows
k
[

Ak = A1 ∪ (A j+1 \ A j ) .
j=1
Using (i),
k
µ(Ak ) = µ(A1 ) + ∑ µ(A j+1 \ A j ),
j=1
and taking the limit as k → ∞,
∞
lim µ(Ak ) = µ(A1 ) + ∑ µ(A j+1 \ A j ).
k→∞ j=1

But again because of (i), µ(A1 ) + ∑∞j=1 µ(A j+1 \ A j ) = µ(

S
k∈N Ak ). □
(iii) Let {Ak }k∈N be a non-increasing sequence of µ-measurable sets. Then {A1 \ Ak }k∈N
is a non-decreasing sequence of µ-measurable sets (because of Theorem 2.2.4-(i) and (iv))
and from (ii) we get [

lim µ(A1 \ Ak ) = µ (A1 \ Ak ) . (2.2.2)
k→∞
k∈N
2.2. MEASURABLE SETS 23

Because of (i), µ(A1 ) = µ(A1 \ Ak ) + µ(Ak ), k ∈ N, and since µ(A1 ) < ∞ we write µ(A1 \
Ak ) = µ(A1 ) − µ(Ak ), k ∈ N. Taking the limit,
lim µ(A1 \ Ak ) = µ(A1 ) − lim µ(Ak ). (2.2.3)
k→∞ k→∞


\ \ \ ∪
S T T
On the other

hand, noting k∈N (A1 A k ) = A1 k∈N A k and since A

1 = A 1 k∈N

2(A k )
k∈N Ak , the subadditivity of µ gives µ(A1 ) ≤ µ k∈N (A1 \ Ak ) + µ
T S T
k∈N Ak . Since
µ(A1 ) < ∞ we arrive at
[
\

µ (A1 \ Ak ) ≥ µ(A1 ) − µ Ak . (2.2.4)
k∈N k∈N

Inserting (2.2.3) and (2.2.4) in (2.2.2) we end up with

\

lim µ(Ak ) ≤ µ Ak .
k→∞
k∈N


≤ µ(Ak ) for all k ∈ N.
T
The reverse inequality is also true because µ k∈N Ak □
(iv) To prove next that k∈N Ak is µ-measurable, let B ⊆ X with µ(B) < ∞. By definition
S

of the restriction measure we have

 [
  [
 [
\

µ B∩ Ak + µ B \ Ak = µ|B Ak + µ|B Ack . (2.2.5)
k∈N k∈N k∈N k∈N
Sj
Defining the sequence of sets {B j } j∈N by B j := k=1 Ak , j ∈ N, it is trivial that k∈N Ak =
S
c c
k∈N Bk and thus also k∈N Ak = k∈N Bk . The advantage is {B j } j∈N is a non-decreasing
S T T

sequence of sets, and thus by (ii) and (iii) we get respectively

[
\

µ|B Ak = lim µ|B (Bk ) and µ|B Ack = lim µ|B (Bck ),
k→∞ k→∞
k∈N k∈N

provided B j , j ∈ N, are µ|B -measurable. By theorem 2.2.4-(iii) it is enough to establish B j ,

j ∈ N, are µ-measurable which is however true since each B j is a finite union of µ-measurable
sets (see Theorem 2.2.4-(iv)). Taking into account that both limits exist (µ(B) < ∞), equation
(2.2.5) is the same as
 [
  [

Ak = lim µ|B (Bk ) + µ|B (Bck ) .
 
µ B∩ Ak + µ B \
k→∞
k∈N k∈N

Since Bk ∩ Bck = 0,
/ we get from (i) that µ|B (Bk ) + µ|B (Bck ) = µ|B (X) = µ(B). Taking com-
plements we see the intersection of countably many µ-measurable sets is also µ-measurable.

2 wecannot use equality in place of inequality here (see (i)) because we don’t know yet that
T
is
k∈N Ak
µ-measurable
24 CHAPTER 2. MEASURE

Definition 2.2.6 (σ -algebra). A collection of subsets A ⊆ 2X is called a σ -algebra provided

/ X ∈A,
(i) 0,

(ii) A ∈ A ⇒ X \ A ∈ A ,

(iii) {Ak ∈ A }k∈N ⇒ ∈A.

S
k∈N Ak

Example 2.2.7. (i) the collection of all µ-measurable sets,

(ii) the Borel σ -algebra of Rn ; that is the smallest σ -algebra of Rn containing the open
subsets of Rn .3 A set in the Borel σ -algebra will be called a Borel set.

2.3 Regularity of measures

Definition 2.3.1 (regular measure). A measure µ on X is called regular if for each C ⊆ X
there exists µ-measurable set B such that C ⊆ B and µ(B) ≤ µ(C) (so µ(C) = µ(B) by the
monotonicity of measures).

if {Ak }k∈N is a nondecreasing sequence of subsets

Exercise 2.3.2. For regular µ prove that
S
of X, then limk→∞ µ(Ak ) = µ k∈N Ak .

Definition 2.3.3 (Borel, Borel regular and Radon measures). Let µ be a measure on Rn .

(i) µ is called Borel if each Borel set is µ-measurable,

(ii) µ is Borel regular if: (a) µ is Borel and (b) for each C ⊆ Rn there exists Borel set B ⊇ C
with µ(B) ≤ µ(C) (hence µ(C) = µ(B) by the monotonicity of measures),

(iii) µ is called Radon if µ is Borel regular and µ(K) < ∞ for all compact K ⊂ Rn .

Exercise 2.3.4. Prove first that

L n (A) = inf{L n (G) open G ⊇ A} ∀ A ⊆ Rn .

Show next that for each C ⊆ Rn , there exists Borel set B ⊇ C with L n (B) ≤ L n (C) (this does
not imply L n is Borel regular because we don’t know yet it is Borel).

Theorem 2.3.5. Let µ be a measure on Rn .

(i) If µ is Borel then µ|A is Borel.

3 one can more generally consider any topological space X in place of Rn and the same applies for definition
2.3.3
2.3. REGULARITY OF MEASURES 25

(ii) If µ is Borel regular and A ⊂ Rn is a Borel set, then µ|A is Borel regular.

(iii) If µ is Borel regular and A ⊂ Rn is a µ-measurable set with µ(A) < ∞, then µ|A is
Radon.
Proof. (i) By definition of µ being Borel, any Borel set is µ-measurable. From theorem
2.2.4-(iii) we know that any µ-measurable set is µ|A -measurable. Thus any Borel set is
µ|A -measurable and hence µ|A is Borel. □
n
(ii) From (i) we readily have that µ|A is a Borel measure. Now let C ⊆ R . We want to
find a Borel set D such that C ⊆ D and µ|A (D) ≤ µ|A (C). By definition of µ being Borel
regular, given C ⊆ Rn , there exists Borel set B such that C ∩ A ⊆ B and µ|A (C) = µ(B). Thus,
it suffices to find a Borel set D with C ⊆ D and µ|A (D) ≤ µ(B). The set D := B ∪ Ac is clearly
Borel and also C ⊆ D. Furthermore, D ∩ A = B ∩ A and we get µ|A (D) = µ(B ∩ A) ≤ µ(B).
□
(iii) By definition of µ being Borel regular, there exists Borel set A′ such that A ⊆ A′ and
µ(A′ ) = µ(A). Since A is µ-measurable we have µ(A′ ) = µ(A) + µ(A′ \ A).
Since µ(A) < ∞
′ n ′
we conclude µ(A \ A) = 0. Now given C ⊆ R we have µ (C ∩ A ) \ A = 0 and using the
µ-measurability of A once more,

µ|A′ (C) = µ (C ∩ A′ ) ∩ A + µ (C ∩ A′ ) \ A = µ|A (C).

This shows µ|A′ and µ|A agree on any set and since A′ is a Borel set we conclude from (ii)
that µ|A is Borel regular. Finally note µ|A (Rn ) = µ(A) < ∞; that is µ is a finite measure. In
particular µ|A (K) < ∞ for any compact K ⊂ Rn .
Lemma 2.3.6. Let µ be a Borel measure on Rn and let B be a Borel set.
(i) If µ(B) < ∞, then ∀ ε > 0, ∃ closed C ⊆ B with µ(B \C) < ε.

(ii) If µ is a Radon measure, then ∀ ε > 0, ∃ open U ⊇ B with µ(U \ B) < ε.

Proof. (i) It is enough to show B ∈ F with

F := A ⊆ Rn A is µ-measurable and ∀ ε > 0, ∃ closed C ⊆ A with µ|B (A \C) < ε .



Indeed, that would imply the existence of a closed set C with C ⊆ B and µ|B (B \ C) < ε,
which in turn implies (i) because µ|B (B \C) = µ(B \C). We next prove B ∈ F by showing

G := {A ∈ F Ac ∈ F }

is a σ -algebra containing all open sets. This would imply it contains all Borel sets, hence B
in particular.
We start by listing the required properties of F .
26 CHAPTER 2. MEASURE

(a) F contains all closed sets (this follows readily from the definition of F ).

(b) If {Ak ∈ F }k∈N then A := ∈ F.

T
k∈N Ak
Proof. Let ε > 0. Since Ak ∈ F for all k ∈ N, there exist closed sets {Ck }k∈N with
Ck ⊆ Ak and µ|B (Ak \ Ck ) < ε2−k , k ∈ N. By theorem 1.1.9 we have C := k∈N Ck is
T

closed and since A \C ⊆ k∈N (Ak \Ck )4 ,

S

µ|B (A \C) ≤ ∑ µ|B(Ak \Ck ) < ε.

k∈N

Thus A ∈ F . □

(c) If {Ak ∈ F }k∈N then A := ∈ F.

S
k∈N Ak
Proof. Let ε and Ck as above. Setting C := k∈N Ck we have A \ C ⊆ 5
k∈N (Ak \ Ck )
S S

which gives
µ|B (A \C) ≤ ∑ µ|B (Ak \Ck ) < ε.
k∈N
Sm
The sequence A \ ( k=1 Ck ) is non-increasing and so the


above inequality together
Sm
with theorem 2.2.5-(iii) imply lim µ|
m→∞ B A \ C
k=1 k < ε. In particular µ|B A \
Sm0
< ε for some m0 ∈ N. By theorem 1.1.9 we get m


S 0
k=1 Ck k=1 Ck is also closed and
thus A ∈ F . □

(d) F contains all open sets.

Proof. According to theorem 1.1.22, any open set in Rn is a (countable) union of
intervals, which are closed by definition. But as observed in (a), F contains all closed
sets and thus (c) implies F contains all open sets. □

We may now proceed to check the collection of subsets G is a σ -algebra containing all
open sets. Since F contains all open and all closed sets, G contains all open sets. Conse-
quently it contains 0/ and Rn and by it’s definition, it contains the complement of each of it’s
elements. Let {Ak ∈ G }k∈N . Then Ak ∈ F , k ∈ N, and also Ack ∈ F
c, k ∈ N, which imply
because of (c) and (b), respectively, that k∈N Ak ∈ F and k∈N Ak = k∈N Ak ∈ F . □
c
S S T

(ii) The sequence {Bk (0) \ B}k∈N is comprised of Borel sets with

µ(Bk (0) \ B) ≤ µ(B̄k (0)) < ∞ ∀ k ∈ N.

4 Clearly, A \ Ck ⊆ Ak \ Ck for all k ∈ N, hence k∈N (A \ Ck ) ⊆ k∈N (Ak \ Ck ). k∈N (A \
S S S
Now observe that
Ck ) = A \C.
5 Clearly, A \C ⊆ A \C for all k ∈ N, hence S
k∈N (Ak \C) ⊆ k∈N (Ak \Ck ). Now observe that k∈N (Ak \
S S
k k k
C) = A \C.
2.3. REGULARITY OF MEASURES 27

Applying (i) we get closed sets {Ck ⊆ Bk (0) \ B}k∈N such that µ (Bk (0) \ B) \ Ck < ε2−k ,

k ∈ N. Set [
U := (Bk (0) \Ck ).
k∈N

Since B ⊆ Ckc 6 for all k ∈ N, we have B = k∈N (Bk (0) ∩ B) ⊆ k∈N (Bk (0) ∩ Ckc ) = U. Fur-
S S

thermore, observing
[
[
[

U \B = (Bk (0) \Ck ) \ B = (Bk (0) \Ck ) \ B = (Bk (0) \ B) \Ck ,
k∈N k∈N k∈N

we have

µ(U \ B) ≤ ∑µ (Bk (0) \ B) \Ck < ε.
k∈N
Finally notice U is open by it’s definition.

Theorem 2.3.7 (outer regularity). If µ is a Radon measure on Rn and A ⊆ Rn , then

µ(A) = inf{µ(U) open U ⊇ A}.

Proof. If µ(A) = ∞, by the monotonicity of µ we have µ(U) = ∞ for any open U ⊇ A

and so the above infimum is also ∞. Assume next µ(A) < ∞ and let ε > 0. If in addition
A is a Borel set, then by the second part of the previous lemma we find open U ⊇ A with
µ(U \ A) < ε. The µ-measurability of A together with µ(A) < ∞ imply µ(U) − µ(U ∩ A) < ε
and so µ(U) < µ(A) + ε. For arbitrary A we know there exists a Borel set B such that A ⊆ B,
µ(B) = µ(A) and, by the above argument, µ(B) = inf{µ(U) open U ⊇ B}. Altogether,

µ(A) = inf{µ(U) open U ⊇ B} ≥ inf{µ(U) open U ⊇ A}.

The reverse inequality holds true as well by the monotonicity of µ.

Theorem 2.3.8 (inner regularity). If µ is a Radon measure on Rn and A ⊆ Rn is µ-measurable,

then
µ(A) = sup{µ(K) compact K ⊆ A}.

Proof. Assume first µ(A) < ∞ and let ε > 0. By theorem 2.3.5-(iii) we know µ|A is a Radon
measure. Applying theorem 2.3.7 for µ|A on the set Ac (observe µ|A (Ac ) = 0) we find open
U ⊇ Ac with µ|A (U) < ε. Then C := U c is closed with C ⊆ A. Since also U ∩ A = A \C, there
holds
ε > µ(A \C) = µ(A) − µ(C),
6 Taking complements in Ck ⊆ Bk (0) \ B, we get Bck (0) ∪ B ⊆ Ckc . In particular, B ⊆ Ckc .
28 CHAPTER 2. MEASURE

by the µ-measurability of C. Hence,

µ(A) = sup{µ(C) closed C ⊆ A}. (2.3.1)

Next we show (2.3.1) is valid also when µ(A) = ∞. Define Dk := Bk (0) \ Bk−1 (0), k ∈ N, and
write A as a disjoint union of sets: A = k∈N (Dk ∩ A). The subadditivity of µ implies
S

∑ µ(Dk ∩ A) ≥ µ(A) = ∞. (2.3.2)

k∈N

Since µ is Radon, each Dk ∩ A is of finite µ-measure7 and we may apply the above argument
to get closed Ck , k ∈ N, such that Ck ⊆ Dk ∩ A with µ(Ck ) ≥ µ(Dk ∩ A) − 2−k , k ∈ N. Clearly,
k∈N Ck ⊆ A. Since {Ck }k∈N is a sequence of disjoint µ-measurable sets, using theorem
S

2.2.5-(i) we have

Ck = ∑ µ(Ck ) ≥ ∑ µ(Dk ∩ A) − 2−k .

[


µ
k∈N k∈N k∈N

The last sum is ∞ by (2.3.2). Hence by theorem 2.2.5-(ii) we conclude limm→∞ µ m

S

k=1 Ck =
Sm
∞. But theorem 1.1.9 says k=1 Ck m∈N is comprised of closed sets and thus (2.3.1) is true
also if µ(A) = ∞. To complete the proof of the theorem, it suffices to show

sup{µ(C) closed C ⊆ A} ≤ sup{µ(K) compact K ⊆ A},

the reverse inequality being trivially true. Let C be a closed subset of A. By theorem 2.2.5-(ii)
we get

µ(C) = lim µ C ∩ B̄m (0) ≤ sup{µ(K) compact K ⊆ A},
m→∞

where in the last inequality we used C ∩ B̄m (0), m ∈ N, are compact subsets of A.

Exercise 2.3.9. Let 0 < λ < 1 and suppose X ⊂ R satisfies 0 < L 1 (X) < ∞. Prove there
exists an open interval I such that L 1 (X ∩ I) ≥ λ L 1 (I).

Lemma 2.3.10. Let {Et ⊂ Rn }t∈I be a family of disjoint Borel sets and µ be a Radon
measure on Rn . Then µ(Et ) > 0 for at most countably many t ∈ I .

Proof. This is self-evident in case I is at most countable. Suppose next I is uncountable.

We write (why?)
t ∈ I µ(Et ) > 0 =
 [
Ik , (2.3.3)
k∈N
7 µ(D
k ∩ A) ≤ µ(B̄k (0)) < ∞
2.4. METRIC MEASURES AND CARATHÉODORY’S CRITERION 29

where n
1o
Ik := t ∈ I µ Bk (0) ∩ Et > .
k
Then for any finite set J ⊂ Ik , k ∈ N, we have

 [ 
µ Bk (0) ≥ µ Bk (0) ∩ Et
t∈J


= ∑ µ Bk (0) ∩ Et
t∈J
H 0 (J)
≥ .
k
Hence, for any k ∈ N the finite number kµ Bk (0) bounds H 0 (J) independently of the choice

 J ⊂ Ik . This implies for any k ∈ N, the set Ik is itself finite. Conse-

of the finite set of indices
quently, (2.3.3) implies t ∈ I µ(Et ) > 0 is at most countable.

2.4 Metric measures and Carathéodory’s criterion

Definition 2.4.1 (metric measure). A measure µ on Rn is called a metric measure if
µ(A ∪ B) ≥ µ(A) + µ(B) whenever A, B ⊂ Rn with dist(A, B) > 0 and µ(A ∪ B) < ∞.
Remark 2.4.2. The above inequality is trivial when µ(A ∪ B) < ∞. Also, by the subadditivity
property of measures, the reverse inequality is true as well. Hence we have the following
equivalent condition that a measure µ has to satisfy in order to be a metric measure:
µ(A ∪ B) = µ(A) + µ(B) whenever A, B ⊂ Rn with dist(A, B) > 0.
Example 2.4.3. (i) The Lebesgue measure on Rn .
Proof. Let A, B ⊂ Rn with dist(A, B) > 0 and L n (A ∪ B) < ∞. Let ε > 0 and pick a
covering of A ∪ B by intervals {Ik }k∈N , such that

∑ Vol(Ik ) < L n(A ∪ B) + ε.

k∈N

By possibly dividing Ik we may further assume diam(Ik ) < dist(A, B) for all k ∈ N.
Hence the covering {Ik }k∈N splits in two subsequences {IkA }k∈N , {IkB }k∈N , the first of
which covers A and the second covers B. We have
L n (A) + L n (B) ≤ ∑ Vol(IkA) + ∑ Vol(IkB) = ∑ Vol(Ik ).
k∈N k∈N k∈N

Coupling these two inequalities, L n (A) + L n (B) < L n (A ∪ B) + ε, which implies the
claim since ε > 0 is arbitrary.
30 CHAPTER 2. MEASURE

(ii) The s-dimensional Hausdorff measure on Rn .

Proof. Let A, B ⊂ Rn with dist(A, B) > 0 and H s (A ∪ B) < ∞. Let ε > 0. By remark
2.1.5 we know Hδs (A ∪ B) ↑ H s (A ∪ B) as δ ↓ 0. We pick a covering of A ∪ B by sets
{Ck }k∈N with diam(Ck ) ≤ δ < dist(A, B), such that
 diamC s
k
∑ ωs < Hδs (A ∪ B) + ε ≤ H s (A ∪ B) + ε.
k∈N 2

Let A be the set of all sets of {Ck }k∈N having nonempty intersection with A; that is,

A := {Ck Ck ∩ A ̸= 0}.
/

Correspondingly set
B := {Ck Ck ∩ B ̸= 0}.
/
Clearly, A ⊆ Ck ∈A Ck and B ⊆ Ck ∈B Ck , while the assumption δ < dist(A, B) gives
S S

that any set in A is disjoint with any set from B. We have

 diamC s  diamC s
k k
Hδs (A) + Hδs (B) ≤ ∑ ωs + ∑ ωs
Ck ∈A 2 Ck ∈B 2
 diamC s
k
= ∑ ωs .
k∈N 2

Coupling these two inequalities, Hδs (A) + Hδs (B) < H s (A ∪ B) + ε. Letting δ → 0,
H s (A) + H s (B) < H s (A ∪ B) + ε and the claim follows since ε > 0 is arbitrary.

Theorem 2.4.4 (Carathéodory). If µ is a metric measure on Rn , then µ is a Borel measure.

Proof. It is enough to prove any closed set is µ-measurable. Let C ⊆ Rn be closed. It is

enough to prove µ(A) ≥ µ(A ∩C) + µ(A \C) for any A ⊂ Rn with µ(A) < ∞. To this end we
set
Cm := {x ∈ Rn dist(x,C) ≤ 1/m}, m ∈ N.
Let A ⊆ Rn with µ(A) < ∞. Since dist(A \ Cm , A ∩ C) ≥ 1/m > 0 for all m ∈ N, by the
hypotheses of the theorem we have


µ(A \Cm ) + µ(A ∩C) = µ (A \Cm ) ∪ (A ∩C) ≤ µ(A).

It remains to show limm→∞ µ(A \Cm ) = µ(A \C). Set

Rk := {x ∈ A 1/(k + 1) < dist(x,C) ≤ 1/k}, k ∈ N.

2.4. METRIC MEASURES AND CARATHÉODORY’S CRITERION 31
S∞

Then A \C = (A \Cm ) ∪ k=m Rk and by the subadditivity of µ,
∞
µ(A \Cm ) ≤ µ(A \C) ≤ µ(A \Cm ) + ∑ µ(Rk ).
k=m

In turn, for limm→∞ ∑∞

k=m µ(Rk ) = 0 to hold true, it is enough that the series ∑k∈N µ(Rk )
converges. But dist(Ri , R j ) > 0 for j ≥ i + 2 and using successively the hypotheses of the
theorem,
m m
[

∑ µ(R2k ) = µ R2k ≤ µ(A) ∀ m ∈ N,
k=1 k=1
and also
m m
[

∑ µ(R2k+1) = µ R2k+1 ≤ µ(A) ∀ m ∈ N.
k=0 k=0
These readily imply ∑k∈N µ(Rk ) ≤ 2µ(A).

Remark 2.4.5. By exercise 2.3.4 and by example 2.4.3-(i) we obtain L n is Borel regular.
Since it is clearly finite on compact sets, we conclude L n is Radon.

Exercise 2.4.6. Prove H s , s ≥ 0, is Borel regular.

Remark 2.4.7. H s is not a Radon measure for 0 ≤ s < n. Using the Brunn-Minkowski
inequality of the next section one can prove the isodiametric inequality; that is, among all
bounded sets of Rn having the same fixed diameter, it is the ball that maximizes the volume,
or what is the same, for all X ⊆ Rn there holds

L n (X) ≤ L n (BX ).

where BX is defined to be any ball with radius equal to half the diameter of X. Using the
isodiametric inequality one can show that H n = L n . These facts will be included in a
possible future addendum to this notes.
32 CHAPTER 2. MEASURE
Chapter 3

Brunn-Minkowski and isoperimetric

inequalities

Given 0/ ̸= X ⊆ Rn and ε > 0 we define the set

X ε := x ∈ Rn dist(x, X) < ε ,


with X ε := Rn in case X = Rn . We start with a version of the isoperimetric inequality not

involving the notion of perimeter.

Theorem 3.0.1 (isoperimetric inequality). Given any L n -measurable X ⊂ Rn satisfying

0 < L n (X) < ∞, consider any ball BrX with the same Lebesgue measure; that is,

L n (X) = ωn rXn . (3.0.1)

Then for any ε > 0 there holds

L n (X ε ) ≥ L n (BεrX ).

Definition 3.0.2 (Minkowski sum). The Minkowski sum of nonempty X,Y ⊆ Rn is given by

X +Y := {x + y x ∈ X, y ∈ Y }.

Remark 3.0.3. Throughout we use the simple fact that

X1 ⊆ Y1 and X2 ⊆ Y2 =⇒ X1 + X2 ⊆ Y1 +Y2 .

33
34 CHAPTER 3. BRUNN-MINKOWSKI AND ISOPERIMETRIC INEQUALITIES

Observing X ε ⊇ X + Bε (0)1 and BεrX = BrX +ε 2 , the isoperimetric inequality would be readily
implied by the following inequality

1/n 1/n
L n X + Bε (0) ≥ ωn (rX + ε) (3.0.2)

1/n  n
1/n
= L n (X) + L Bε (0) .

This shows the isoperimetric inequality is a special case of the following inequality

Theorem 3.0.4 (Brunn-Minkowski inequality). For any non-empty and L n -measurable

X,Y ⊆ Rn , there holds

1/n
1/n
1/n
L n (X +Y ) ≥ L n (X) + L n (Y ) . (3.0.3)

Now define the n − 1 dimensional lower Minkowski content of a bounded X ⊂ Rn by3

L n X + Bε (0) − L n (X)


Per(X) := lim inf .
ε↓0 ε

If X ⊂ Rn is bounded, noting BrX +ε ⊇ Brx + Bε (0) (combine the last two footnotes taking
X = BrX ), we further obtain from (3.0.2) and (3.0.1)

L n X + Bε (0) − L n (X) L n BrX + Bε (0) − L n (BrX )

≥ .
ε ε
Passing to the lim inf (see Exercise 1.1.3) we conclude

Per(X) ≥ Per(BrX ), (3.0.4)

which asserts

among all bounded sets of Rn having the same fixed volume, it is the ball that minimizes
the perimeter,
1 Given y ∈ X + Bε (0) we know there exists x ∈ X and z ∈ Bε (0) such that y = x + z. Hence, dist(y, X) ≤
|y − x| = |z| < ε ⇒ y ∈ X ε . As an exercise, show these two sets are in fact equal whenever X is closed.
2 It is enough to prove Bε (0) = B


r r+ε (0). Given x ∈ B ε (0) we know dist x, B (0) < ε. For any z ∈ B (0)
r
r r
we have |x| ≤ |x − z| + |z| < |x − z| + r. Hence |x| ≤ dist x, Br (0) + r < ε + r ⇒ x ∈ Br+ε (0). Conversely,
given x ∈ Br+ε (0) we know |x| < r + ε, so if x ∈ Br+ε (0) \ Br (0) then dist(x, Br (0)) = |x| − r < ε. Hence
Br+ε (0) \ Br (0) ⊂ Bεr (0). But also Br (0) ⊂ Bεr (0) and so Br+ε (0) ⊂ Bεr (0).
3 The smoother the boundary of the set X is, the n − 1 dimensional lower Minkowski content coincides with

its perimeter (in a weak notion which is out of the scope of these notes), H n−1 (∂ X) or even the surface area of
∂ X from calculus. This is why here (by abuse of notation) we denote it by Per.
 35

or, equivalently,

among all bounded sets of Rn having the same fixed perimeter, it is the ball that maxi-
mizes the volume.

This last formulation is the classical isoperimetric statement. Finally we notice the perimeter
of BrX can be explicitly computed. Indeed,

ωn (rX + ε)n − ωn rXn

Per(BrX ) = lim inf = nωn rXn−1 ,
ε↓0 ε

and using once more (3.0.1) on the right, we deduce after coupling with (3.0.4),

1/n
1−1/n
Per(X) ≥ nωn L n (X) .

This is the classical expression of the isoperimetric inequality involving the notion of perime-
ter.

Proof of the Brunn-Minkowski inequality. The proof splits in three steps.

Step 1 - The simplest case. Suppose first X and Y are intervals with sides lengths {x1 , ..., xn },
{y1 , ..., yn } respectively. Then X +Y is an interval with sides lengths {x1 + y1 , ..., xn + yn } and
the Brunn-Minkowski inequality reads
 n 1/n  n 1/n  n 1/n
∏(xi + yi) ≥ ∏ xi + ∏ yi ,
i=1 i=1 i=1

where we have used exercise 2.1.9. Writing this as

n
 xi 1/n  n yi 1/n
1≥ ∏ + ∏ ,
i=1 xi + yi i=1 xi + yi

we see it is a direct consequence of the arithmetic-geometric mean inequality, which we

recall next. □

Lemma 3.0.5 (arithmetic-geometric mean inequality). 4 For positive real numbers a1 , ..., an ,
 n 1/n 1 n
∏ ai ≤ ∑ ai .
i=1 n i=1
4 it’s proof will be an easy exercise (see remark 9.1.4) as soon as we learn about convex functions
36 CHAPTER 3. BRUNN-MINKOWSKI AND ISOPERIMETRIC INEQUALITIES

Step 2 - Reductions. From remark 2.4.5 we know L n is Radon and this implies the inner
regularity (theorem 2.3.8) of L n . Suppose X and Y are L n -measurable and pick compact
sets KX , KY such that KX ⊆ X and KY ⊆ Y . Then KX + KY ⊆ X +Y and assuming the Brunn-
Minkowski inequality holds true for all compact sets, we get from the monotonicity of L n

1/n
1/n
1/n
L n (X +Y ) ≥ L n (KX ) + L n (KY ) .

Taking the supremum over all compact KX ⊆ X and over all KY ⊆ Y , we deduce (3.0.3). Thus
(3.0.3) will be true for all L n -measurable X and Y , if it is true for all compact X and Y . Let
X and Y be compact. Recalling exercise 1.1.23, we construct two non-increasing sequences
{Xk ⊇ X}k∈N and {Yk ⊇ Y }k∈N with X = k∈N Xk , Y = k∈N Yk and such that each Xk and Yk ,
T T

k ∈ N, consists of a finite union of non-overlapping intervals. Suppose

1/n
1/n
1/n
L n (Xk +Yk ) ≥ L n (Xk ) + L n (Yk ) .

Since Xk ⊇ X and Yk ⊇ Y , from the monotonicity of L n we arrive at

1/n
1/n
1/n
L n (Xk +Yk ) ≥ L n (X) + L n (Y ) .

Taking the limit k → ∞ we use theorem 2.2.5-(iii) to arrive to (3.0.3). Indeed, noting {Xk +
Yk ⊇ X +Y }k∈N is a non-increasing sequence of sets with X +Y = k∈N (Xk +Yk ),5 we deduce
T

L n (X +Y ) = limk→∞ L n (Xk +Yk ). □

Step 3 - The case where X and Y are finite unions of non-overlapping intervals. Let X and
Y be finite unions of non-overlapping intervals. The proof in this case follows by induction
on the total number of intervals N(X,Y ) that constitute X and Y (we will denote simply by
N(X) the case where X = Y ). If N(X,Y ) = 2 then (3.0.3) has been proved in the first step.
Let N(X,Y ) = m ≥ 3 and suppose that (3.0.3) is true when N(X,Y ) ≤ m − 1. Since m ≥ 3
we assume that X consists of at least two non-overlapping intervals I1 , I2 . Then there exists
i ∈ {1, ..., n} and s ∈ R such that the hyperplane {x = (x1 , ..., xn ) ∈ Rn xi = s} separates
them. From now on, for simplicity, we denote a hyperplane of this form by {xi = s} and by
{xi ≥ s}, {xi ≤ s} the two “half spaces” produced by it. We define the sets

X + := X ∩ {xi ≥ s} and X − := X ∩ {xi ≤ s}.

Then X + and X − are non-overlapping sets comprised of non-overlapping intervals and the
number of intervals in each one is strictly less than m − 1 (it can be at most m − 2 in case Y is
5 If
α ∈ k∈N (Xk + Yk ), then ∀ k ∈ N, ∃ xk ∈ Xk and ∃ yk ∈ Yk such that xk + yk = α. But {Xk }k∈N is non-
T

increasing, hence xk ∈ X1 for all k ∈ N and since X1 is compact, a subsequence of xk converges to x ∈ X1 . Since
xk ∈ Xk for all k ∈ N and {Xk }k∈N is non-increasing to X, we deduce x ∈ Xk for all k ∈ N, that is x ∈ X. In the
same way, we get a a subsequence of yk converging to y ∈ Y . Taking the subsequential limit in xk + yk = α, we
get x + y = α. Thus α ∈ X +Y . The reverse inclusion is immediate since X +Y ⊆ Xk +Yk for all k ∈ N.
 37

comprised of only one interval). Set

L n (X + )
λ := .
L n (X)

Then λ ∈ (0, 1) and since X = X + ∪ X − , we have L n (X) = L n (X + ) + L n (X − ) by exercise

2.1.10. Thus
L n (X − ) = (1 − λ )L n (X). (3.0.5)
Pick a parallel hyperplane {xi = t} such that if Y + := Y ∩ {xi ≥ t} and Y − := Y ∩ {xi ≤ t},
then
L n (Y + )
= λ .6 (3.0.6)
L n (Y )
Since X ± ⊂ X and Y ± ⊂ Y , we get X ± +Y ± ⊂ X +Y which gives (X + +Y + ) ∪ (X − +Y − ) ⊂
X +Y . Thus
L n (X +Y ) ≥ L n (X + +Y + ) ∪ (X − +Y − ) .

But X + +Y + ⊆ {xi ≥ s+t} and X − +Y − ⊆ {xi ≤ s+t}, which implies X + +Y + and X − +Y −

are non-overlapping. Using exercise 2.1.10 the above inequality becomes

L n (X +Y ) ≥ L n (X + +Y + ) + L n (X − +Y − ).

Since N(X + +Y + ), N(X − +Y − ) ≤ m − 1, the induction hypothesis further gives

1/n
1/n n 
1/n
1/n n
L n (X +Y ) ≥ L n (X + ) + L n (Y + ) + L n (X − ) + L n (Y − )

.

As we worked to get (3.0.5), we may use (3.0.6) to get L n (Y − ) = (1 − λ )L n (Y ). This

together with (3.0.6), (3.0.5) and the definition of λ when inserted in the last inequality imply

1/n
1/n n
1/n
1/n n
L n (X +Y ) ≥ λ L n (X) + L n (Y ) + (1 − λ ) L n (X) + L n (Y )
 

1/n
1/n n
= L n (X) + L n (Y )

,

which in turn gives (3.0.3).

6 Note that the quotient on the left is a continuous function of t ∈ R, that takes all values of [0, 1].
38 CHAPTER 3. BRUNN-MINKOWSKI AND ISOPERIMETRIC INEQUALITIES
Chapter 4

Integral

4.1 Measurable functions

Unless otherwise stated, µ stands for a measure on a set X.

Definition 4.1.1 (µ-measurable function). 1 f : X → Rm is called µ-measurable if

U ⊆ Rm is open ⇒ f −1 (U) is µ-measurable.

Definition 4.1.2 (Borel measurable function). 2 Let µ be a measure on Rn . A function

f : Rn → Rm is called Borel measurable if

U ⊆ Rm is open ⇒ f −1 (U) is a Borel set.

Example 4.1.3. If f : Rn → Rm is continuous then it is Borel measurable. This is because

the pre-image f −1 (U) of an open set U ⊆ Rm through a continuous function f is again open.

Theorem 4.1.4. (i) If f : X → Rm is µ-measurable, then f −1 (B) is µ-measurable for each

Borel set B ⊆ Rm .

(ii) f : X → R̄ is µ-measurable if and only if f −1 [−∞, α) is µ-measurable for each

α ∈ R.

(iii) If f : X → Rn and g : X → Rm are µ-measurable, then ( f , g) : X → Rn+m is µ-

measurable.
1 you can replace Rm in this definition by any topological space Y
2 you can replace Rn , Rm in this definition by any topological spaces X,Y

39
40 CHAPTER 4. INTEGRAL

Proof. (i) By assumption the set

A := {A ⊆ Rm f −1 (A) is µ-measurable},

contains all open sets of Rm . If we show it is a σ -algebra then by definition of the Borel
σ -algebra it will contain also all Borel sets. To this end notice first that 0,
/ R m ∈ A . Also,
c
if A ∈ A then by a property of the pre-image we have f −1 (Ac ) = f −1 (A) , which is µ-

S if {Ak ∈ A }k∈N , then again by a property of the pre-image we have

measurable. Finally,
f −1 k∈N Ak = k∈N f −1 (Ak ) which is µ-measurable.
S

(ii) Similarly, the set

A := {A ⊆ R̄ f −1 (A) is µ-measurable},

is a σ -algebra containing [−∞, α) for each α ∈ R.

(iii) One proves the set

A := {A ⊆ Rn+m h−1 (A) is µ-measurable}, with h := ( f , g),

is a σ -algebra containing U ×V whenever U ⊆ Rn and V ⊆ Rm are open.

Exercise 4.1.5. Let f , g : Rn → Rn be Borel measurable. Prove f ◦ g is Borel measurable.

Theorem 4.1.6. (i) If f , g : X → R̄ are µ-measurable, then so are

f + g, f g, | f |, min{ f , g}, max{ f , g} and f /g (provided g ̸= 0 in X).

(ii) If { fk : X → R̄}k∈N are µ-measurable, then so are

inf fk , sup fk , lim inf fk and lim sup fk .

k∈N k∈N k→∞ k→∞

Exercise 4.1.7. Prove the above theorem.

Exercise 4.1.8. Let E ⊂ Rn be L n -measurable with L n (E) < ∞ and let f : E → R be an

L n -measurable function. Define µE, f : R → [0, ∞) by

µE, f (s) := L n {x ∈ E f (x) > s} .

Prove the following assertions

(i) µE, f is bounded, non-increasing, µE, f (−∞) = L n (E) and µE, f (∞) = 0.

(ii) µE, f (s) − µE, f (t) = L n {x ∈ E s < f (x) ≤ t} whenever s < t.

4.1. MEASURABLE FUNCTIONS 41

(iii) lims→t + µE, f (s) = µE, f (t) (thus µE, f is right-continuous).

(iv) lims→t − µE, f (s) = L n {x ∈ E f (x) ≥ t} = µE, f (t) + L n {x ∈
E f (x) = t} (thus

µE, f is continuous at t ∈ R if and only if L n {x ∈ E f (x) = t} = 0).

(v) µE, f is constant in the interval (s,t) if and only if L n {x ∈ E s < f (x) < t} = 0.

Theorem 4.1.9. If f : X → [0, ∞] is µ-measurable, there exist µ-measurable sets {Ak ⊂ X}k∈N
such that
1
f (x) = ∑ χAk (x) ∀ x ∈ X.
k∈N k
Proof. We define the sets
n 1 k−1 1 o
A1 := {x ∈ X f (x) ≥ 1} and Ak := x ∈ X f (x) ≥ + ∑ χA j (x) , k ∈ N \ {1}.
k j=1 j
First we observe
f (x) ≥ χA1 (x) ∀ x ∈ X.
Indeed, this is self evident if x ∈ A1 and trivially true by the nonnegativity of f in case x ∈
X \ A1 . Next we claim
1
f (x) ≥ χA1 (x) + χA2 (x) ∀ x ∈ X. (4.1.1)
2
Indeed, noting
A2 = {x ∈ X f (x) ≥ 3/2 if x ∈ A1 and f (x) ≥ 1/2 if x ∈ X \ A1 },
we see again that (4.1.1) is self evident if x ∈ A1 ∩ A2 and trivially true by the nonnegativity
of f in case x ∈ X \ (A1 ∩ A2 ). If x ∈ A2 \ A1 then 1/2 ≤ f (x) < 1 and this gives f (x) ≥ 1/2
which is (4.1.1) in this case. Finally if x ∈ A1 \ A2 then 1 ≤ f (x) < 3/2 and this gives f (x) ≥ 1
which is (4.1.1) in this case. In the same fashion we can prove inductively that for any m ∈ N
m
1
f (x) ≥ ∑ χA j (x) ∀ x ∈ X.
j=1 j

Letting m → ∞ we obtain
1
f (x) ≥ ∑ k χAk (x) ∀ x ∈ X. (4.1.2)
k∈N
To show the equality we observe first that in case f (x) = ∞ we have x ∈ Ak for all k ∈ N.
Thus ∑k∈N (1/k)χAk (x) = ∑k∈N (1/k) = ∞ = f (x) in this case. If f (x) < ∞, then there exists
kx ∈ N such that x ∈
/ Ak for all k ≥ kx . This means
1 k−1 1
f (x) < + ∑ χA j (x) ∀ k ≥ kx ,
k j=1 j
42 CHAPTER 4. INTEGRAL

which coupled with (4.1.2) gives

k−1
1 1
0 ≤ f (x) − ∑ χA j (x) < ∀ k ≥ kx . (4.1.3)
j=1 j k

The proof follows by letting k → ∞.

4.2 Lusin and Egoroff theorems

Theorem 4.2.1 (Lusin). Let µ be a Borel regular measure on Rn and f : Rn → Rm be µ-
measurable. Assume A ⊂ Rn is µ-measurable with µ(A) < ∞. Then

∀ ε > 0, ∃ compact K ⊂ A such that µ(A \ K) < ε and f |K is continuous.

Proof. By theorem 1.1.26, given i ∈ N, we may consider a partition of Rm comprised of

dyadic cubes having diameter less than 1/i. Actually, given i ∈ N, any countable family
{Bi, j } j∈N comprised of disjoint Borel sets such that Rm = j∈N Bi, j and diam(Bi, j ) < 1/i
S

will work in what follows. So pick such a partition and set

Ai, j := A ∩ f −1 (Bi, j ), i, j ∈ N.

These sets enjoy the following properties

(P1 ) {Ai, j }i, j∈N are µ-measurable. Indeed, since {Bi, j }i, j∈N are Borel and f is µ-measurable,
theorem 4.1.4-(i) implies { f −1 (Bi, j )}i, j∈N are µ-measurable. Since A is µ-measurable
we conclude that {Ai, j }i, j∈N are µ-measurable.

for any i ∈ N. By definition and standard properties of the pre-image

S
(P2 ) A = j∈N Ai, j

f −1 (Bi, j ) = A ∩ f −1 Bi, j = A ∩ f −1 (Rm ) = A ∩ Rn = A.

[ [
[

Ai, j = A ∩
j∈N j∈N j∈N

(P3 ) {Ai, j } j∈N are disjoint for any i ∈ N. Indeed, x ∈ Ai, j if and only if x ∈ A and there exists
y ∈ Bi, j such that y = f (x). Assuming x ∈ Ai, j ∩ Ai,k with j ̸= k, we get y ∈ Bi, j and
y′ ∈ Bi,k such that y = f (x) = y′ . This contradicts the disjointness of {Bi, j } j∈N .

Hence, out of a partition {Bi, j } j∈N of the target space Rm , we have constructed a partition
{Ai, j } j∈N of A ⊆ Rn comprised of µ-measurable sets.
Now let ε > 0. Theorem 2.3.5-(iii) assures that µ|A is Radon and because of (P1 ) we may
apply inner regularity (theorem 2.3.8) to get compact sets {Ki, j ⊆ Ai, j }i, j∈N with µ|A (Ai, j \
4.2. LUSIN AND EGOROFF THEOREMS 43

Ki, j ) < ε2−i− j . Using (P2 ) we observe A \

Ai, j \ ⊆ j∈N (Ai, j \
S S S S
j∈N Ki, j = j∈N j∈N Ki, j
Ki, j ) and so
[
[

µ A\ Ki, j = µ|A A \ Ki, j
j∈N j∈N

(Ai, j \ Ki, j ) < ε2−i ,

[

≤ µ|A
j∈N

by the subadditivity of measures. Noting {A \ Nj=1 Ki, j }N∈N non-increases to A \

S S
j∈N Ki, j ,
we use theorem 2.2.5-(ii) to deduce further from the above estimate
N
Ki, j < ε2−i
[

lim µ A \ ∀ i ∈ N.
N→∞
j=1

This implies for each i ∈ N there exists Ni ∈ N such that

Ni
Ki, j < ε2−i .
[

µ A\ (4.2.1)
j=1

Next we define the sets

Ni
[
Di := Ki, j , i ∈ N.
j=1

Being a finite union of compact sets, each Di is compact. For each i ∈ N we pick {bi, j ∈
Bi, j } j∈N and define functions gi : Di → Rm by

Ni
gi (x) := ∑ bi, j χKi, j (x), x ∈ Di .
j=1

(P3 ) together with {Ki, j ⊆ Ai, j }i, j∈N imply {Ki, j }i, j∈N are disjoint and thus, each being com-
pact, of positive distance apart (use exercise 1.1.14). This means each gi is continuous. Also,
if x ∈ Di , then x ∈ Ki, j ⊆ Ai, j for some j ∈ {1, ..., Ni }. Thus f (x) ∈ Bi, j for some j ∈ {1, ..., Ni }.
Since gi (x) = bi, j ∈ Bi, j for this j, we obtain

1
| f (x) − gi (x)| ≤ diam(Bi, j ) < ∀ x ∈ Di . (4.2.2)
i
Finally we define the set \
K := Di .
i∈N
44 CHAPTER 4. INTEGRAL

Then K is a compact subset of A and A \ K = A ∩ i∈N Dci = i∈N (A ∩ Dci ). Subadditivity

S
S

of measures gives
µ(A \ K) ≤ ∑ µ(A \ Di ) < ε,
i∈N
by the definition of Di and (4.2.1). Also, if x ∈ K then x ∈ Di for all i ∈ N and because of
(4.2.2)
1
sup | f (x) − gi (x)| < .
x∈K i
m
Thus {gi : Di → R }i∈N is a sequence of continuous functions which converges uniformly to
f on K = i∈N Di . Standard analysis implies the limit function f is continuous in K.3
T

Notation 4.2.2. Let µ be a measure on a set X. The expression “µ-a.e. in A” where A ⊆ X

means “almost everywhere in A with respect to the measure µ”; that is, “for all x ∈ A \ N
where µ(N) = 0”.
Theorem 4.2.3 (Egoroff). Let µ be a measure on X and suppose { fk X → Rm }k∈N are
µ-measurable. Let A ⊂ X be a µ-measurable set such that µ(A) < ∞ and fk → f µ-a.e. in
A. Then
∀ ε > 0, ∃ µ-measurable B ⊂ A such that µ(A \ B) < ε and fk → f uniformly in B.
Proof. Let N ⊆ A be the set of µ-measure 0 out of which fk → f . For each i ∈ N we define
x ∈ X | fk (x) − f (x)| > 2−i , j ∈ N.
[
Ci, j :=
k≥ j

Then {A ∩Ci, j } j∈N are µ-measurable (by theorem 4.1.6-(ii)) and non-increase to j∈N A ∩
T



Ci, j . But fk → f µ-a.e. in A implies j∈N A ∩Ci, j ⊆ N. Taking into account the fact that
T

µ(A ∩Ci,1 ) ≤ µ(A) < ∞, we apply theorem 2.2.5-(iii) to get

lim µ(A ∩Ci, j ) ≤ µ(N) = 0 ∀ i ∈ N.
j→∞

Hence, for each i ∈ N there exists Ni ∈ N such that µ(A ∩Ci,Ni ) < ε2−i .
We define next the set [
B := A \ Ci,Ni .
i∈N
c



S
Observing that A \ B = A ∩ A ∪ i∈N Ci,Ni = A ∩ i∈N Ci,Ni = i∈N (A ∩Ci,Ni )
S S
we get
µ(A \ B) ≤ ∑ µ(A ∩Ci,Ni ) < ε.
i∈N
Moreover, since for each i ∈ N we have x ∈ B ⇒ x ∈ / A ∩Ci,Ni , we get by the definition of the
Ci, j ’s that for each i ∈ N, any x ∈ B and all k ≥ Ni , there holds | fk (x) − f (x)| ≤ 2−i .
3 | f (x) −
f (x0 )| ≤ | f (x) − gi (x)| + |gi (x) − gi (x0 )| + |gi (x0 ) − f (x0 )| ≤ |gi (x) − gi (x0 )| + 2 maxx∈K | f (x) −
gi (x)| ∀x, x0 ∈ K
4.3. INTEGRATION OF MEASURABLE FUNCTIONS 45

4.3 Integration of measurable functions

Notation 4.3.1. For a function g : X → R̄ we set

g+ := max{0, g} and g− := max{0, −g}.

Observe these are nonnegative functions and that the following decompositions are valid

g = g+ − g− and |g| = g+ + g− .

Definition 4.3.2 (simple function). A function g : X → R̄ is called simple function if the

image of g is countable.

Remark 4.3.3. The above definition implies that if g : X → R̄ is simple, then there exist
disjoint {Ak ⊆ X}k∈N and {αk ∈ R̄}k∈N such that
[
g= ∑ αk χAk and Ak = X.
k∈N k∈N

Consider now µ to be a measure on the set X. The sets {Ak }k∈N in the above expression can
be taken µ-measurable if g is known to be µ-measurable. Indeed, for each k ∈ N we have

Ak = g−1 {αk } = g−1 [−∞, αk ] ∩ g−1 [αk , ∞)

c
g−1 [−∞, αk + 1/m) ∩ g−1 [−∞, αk ) ,
\

=
m∈N

and thus by theorem 4.1.4-(ii) Ak is µ-measurable. Consequently, we have the useful fact4

Proposition 4.3.4. Given a µ-measurable simple function g : X → [0, ∞], then there exist
disjoint µ-measurable sets {Ak }k∈N and {αk ∈ (0, ∞]}k∈N such that
[
g= ∑ αk χAk and Ak ⊆ X.
k∈N k∈N

Definition 4.3.5. (i) If g : X → [0, ∞] is a simple, µ-measurable function, we define its

integral on X by Z
g dµ := ∑ yµ g−1 {y} .


X y∈[0,∞]
R
Note that applying proposition 4.3.4 on g we get Xg dµ = ∑k∈N αk µ(Ak ).
4 compare with theorem 4.1.9
46 CHAPTER 4. INTEGRAL
R R
(ii) If g : X → R̄ is a simple µ-measurable function and either X g+ dµ < ∞ or X g− dµ <
∞, we call g a µ-integrable simple function and define its integral on X by
Z Z Z
g dµ := g+ dµ − g− dµ.
X X X

Definition 4.3.6. Let f : X → R̄.

(i) The upper integral of f on X is defined by
Z nZ o
f dµ := inf g dµ g µ-integrable, simple, g ≥ f µ-a.e. in X .
X X

(ii) The lower integral of f on X is defined by

Z nZ o
f dµ := sup g dµ g µ-integrable, simple, g ≤ f µ-a.e. in X .
X X

Theorem 4.3.7. Let f , h : X → R̄.

R R
(i) X
f dµ = − X (− f )dµ.
R
(ii) If f ≥ 0 µ-a.e. in X, then X f dµ ≥ 0.
R R
(iii) If α > 0 then Xα f dµ = α X f dµ.
R R R R R
(iv) If X f dµ + Xh dµ < ∞, then X( f + h)dµ ≤ X f dµ + Xh dµ.
R R
(v) X
f dµ ≤ X f dµ.
Exercise 4.3.8. Prove the above theorem.
Definition 4.3.9
R
(µ-integrable
R
function). A µ-measurable function
R
f :X →
R
R̄ is called µ-
integrable if X f dµ ≤ X f dµ. By 4.3.7-(v) we have then X f dµ = X f dµ and we
write Z Z Z
f dµ = f dµ = f dµ.
X X X

Remark 4.3.10. For a µ-integrable simple function g : X → R̄ we have

Z
g−1 {y} .


g dµ = ∑ yµ
X
y∈R̄
R
Theorem 4.3.11. Any µ-measurable f : X → [0, ∞] is µ-integrable with X f dµ ∈ [0, ∞].
4.3. INTEGRATION OF MEASURABLE FUNCTIONS 47


Proof. Suppose first that µ {x ∈ XR| f (x) = ∞} > 0. For t > 0 we use the simple function
tχ{x∈X | f (x)=∞} in the definition of X f dµ to obtain
Z

f dµ ≥ tµ {x ∈ X | f (x) = ∞} for any t > 0.
X
R R
Thus X f dµ = ∞. Because of theorem 4.3.7-(v) we also get X f dµ = ∞. Hence f is
R

µ-integrable and X f dµ = ∞. Suppose now µ {x ∈ X | f (x) = ∞} = 0. Then f (x) < ∞
µ-a.e. in X and for t > 1 we consider the disjoint µ-measurable sets

Ek := {x ∈ X t k ≤ f (x) < t k+1 }, k ∈ Z,

as well as the simple µ-measurable function

g := ∑ t k χEk .
k∈Z

Then X \{ f = 0} = k∈Z Ek and assigning the value 0 to g in { f = 0}, we have g(x) ≤ f (x) ≤
S

tg(x) µ-a.e. x ∈ X. From the definitions of upper and lower integrals of f we deduce
Z Z Z Z
f dµ ≤ tg(x) dµ = t g(x) dµ ≤ t f (x) dµ,
X X X X
R
where
R
we have used 4.3.7-(iii) to get the equality. This holds for any t > 1 and so X f dµ ≤
X
f (x) dµ.

Theorem 4.3.12. Let f , h : X → R̄.

R R
(i) If α ̸= 0 then Xα f dµ = α X f dµ.
R R R R R
(ii) If X f dµ + Xh dµ ∈ R̄, then X( f + h) dµ = X f dµ + Xh dµ.
R R
(iii) If f is µ-integrable | X f dµ| ≤ X |f| dµ.

Exercise 4.3.13. Prove the above theorem.

Proposition 4.3.14. Let µ be a Radon measure on Rn and suppose A ⊂ Rn is µ-measurable

with µ(A) < ∞. Then, given any p > 0 we have
Z
∀ ε > 0, ∃ g ∈ Cc (Rn ) such that |χA − g| p dµ < ε.
Rn
48 CHAPTER 4. INTEGRAL

Proof. Given ε > 0 we employ the outer and inner regularity properties of Radon measures
(theorems 2.3.7 and 2.3.8), to get open U ⊃ A and compact K ⊂ A such that

µ(U) − ε/2 < µ(A) < µ(K) + ε/2.

Hence µ(U \ K) < ε. Further, from lemma 1.2.6 we get a continuous function g : Rn → [0, 1]
with g = 1 on K and sprt(g) ⊂ U. It follows that |χA − g| = 0 in (U \ K)c and |χA − g| ≤ 1 on
U \ K. Altogether,
Z Z
p
|χA − g| dµ = |χA − g| p dµ ≤ µ(U \ K) < ε.
Rn U\K

Definition 4.3.15
R
(µ-summable function). f : X → R̄ is called µ-summable if it is µ-
integrable and X | f | dµ < ∞.

Exercise 4.3.16. Prove that if f : X → R̄ is µ-summable then

1
Z
µ {x ∈ X | f (x)| ≥ s} ≤ | f | dµ whenever s > 0.
s X
R
Use this to prove that X |f| dµ = 0 if and only if f = 0 µ-a.e. in X.

4.4 L p spaces
Let µ be a measure on a set X ̸= 0.
/

Notation 4.4.1. From now on we write {g > α} for {x ∈ X | g(x) > α}, etc.

Definition 4.4.2. For a µ-measurable g : X → R̄ we set


 0
 if µ(X) = 0,


ess supX g := ∞ if µ {g > α} > 0 ∀α ∈ R,


inf{α ∈ R | µ {g > α} = 0} otherwise.


Remark

4.4.3. Suppose µ {g > α} > 0 for all α ∈ R. Then ∄ α ∈ R such that µ {g >
α} = 0; that is, ∄ α ∈ R such that g(x) ≤ α for µ-a.e. x ∈ X; in other words, g = ∞ on a
set of positive µ-measure. This justifies

the middle definition of ess sup. On the other hand,
if ∃ α ∈ R such that µ {g > α} = 0, then g(x) ≤ α for µ-a.e. x ∈ X; that is α is an upper
bound for g (except possibly on a subset of X of µ-measure 0). Hence it is natural to define
ess sup as the least upper bound in this case.
4.4. L p SPACES 49

Definition 4.4.4. For p ∈ (0, ∞):

L p ≡ L p (X, µ) := {all µ-measurable functions f : X → R̄ such that p

R
X |f| dµ is finite}.

Also,

L∞ ≡ L∞ (X, µ) := {all µ-measurable functions f : X → R̄ such that ess supX | f | is finite}.

Remark 4.4.5. From Definition 4.3.15 we notice that L1 (X, µ) is the collection of all µ-
summable functions.

Theorem 4.4.6. Suppose µ(X) < ∞. Then

(i) 0 < p1 < p2 ≤ ∞ implies L p2 ⊂ L p1 ,

R 1/p
(ii) lim p→∞ X | f | p dµ = ess supX | f |.

Proof. If f ∈ L p2 and p2 < ∞ then write

Z Z Z
| f | p1 dµ = | f | p1 dµ + | f | p1 dµ
X {| f |≤1} {| f |>1}
Z Z
p2
≤ µ({| f | ≤ 1}) + |f| dµ ≤ µ(X) + | f | p2 dµ < ∞.
{| f |>1} X

If f ∈ L p2 and p2 = ∞ then write

Z
p
| f | p1 dµ ≤ ess supX | f | 1 µ(X) < ∞.
X

this
proves (i). For (ii), let µ(X) > 0 and assume first that ∃ α ∈ [0, ∞) such that µ {| f | >
α} = 0. This implies (see the remark following the definition of ess sup) that


ess supX | f | = inf{α ∈ R | µ {| f | > α} = 0} =: M ∈ [0, ∞).

If α = 0 or M = 0 then f = 0 µ-a.e. in X and the result true. If M > 0, observe that

Z 1/p
1/p Z 1/p
| f | p dµ ≤ M µ(X) ⇒ lim sup | f | p dµ ≤ M. (I)
X p→∞ X

On the other hand, given M ′ < M we have µ {| f | > M ′ } > 0, hence

Z 1/p Z 1/p
1/p

p
| f | dµ ≥ | f | p dµ ≥ M ′ µ({| f | > M ′ }) ⇒
X {| f |>M ′ }
50 CHAPTER 4. INTEGRAL
Z 1/p
lim inf p
| f | dµ ≥ M′.
p→∞ X
But this holds true for any M ′ < M, hence
Z 1/p
lim inf | f | p dµ ≥ M. (II)
p→∞ X


Inequalities (I) and (II) readily imply the result. In the case where µ {| f | > α} > 0 for
all α ∈ [0, ∞), then by definition ess supX | f | = ∞. By the remark following the definition of
ess sup we know | f | = ∞ on a set of positive measure, hence X | f | p dµ = ∞ and the result
R

follows. Finally, the case where µ(X) = 0 is also clear since ess supX | f | = 0 by its definition,
p
R
and also X | f | dµ = 0.
Theorem 4.4.7 (Hölder inequality). Let p, q ∈ [1, ∞] satisfying 1/p + 1/q = 1. If f ∈ L p ,
g ∈ Lq then

R
ess supX |g| X | f | dµ if p = 1,



Z
1/p R 1/q
   
| f g| dµ ≤ p q
R
X | f | dµ X |g| dµ if 1 < p < ∞,
X 


R
ess supX | f | X |g| dµ if p = ∞.

Proof. Let 1 < p < ∞ (the cases p = 1 and p = ∞ are trivial).R We know Rthe convexity
inequality |ab| ≤ 1p |a| p + q1 |b|q for all a, b ∈ R (prove it!). Hence if X | f | p dµ = X |g|q dµ = 1
we get
Z
1
Z
1
Z
1 1  Z 1/p  Z 1/q
p q p q
| f g|dµ ≤ | f | dµ + |g| dµ = + = 1 = | f | dµ |g| dµ .
X p X q X p q X X
p q
R R
If X |f| dµ, X |g| dµ > 0, normalize f , g as follows
f g
f˜ := R
1/p , g̃ := R
1/q .
p q
X | f | dµ X |g| dµ

Then X | f˜| p dµ = q ˜ ≤ 1. Substituting f˜ and g̃

R R R
X |g̃| dµ = 1 and as before we have X | f g̃|dµ
gives the result.
Theorem 4.4.8 (Minkowski inequality). If f , g ∈ L p with p ∈ [1, ∞), then
Z 1/p Z 1/p Z 1/p
p p p
| f + g| dµ ≤ | f | dµ + |g| dµ .
X X X

If f , g ∈ L∞ then ess supX | f + g| ≤ ess supX | f | + ess supX |g|.

4.4. L p SPACES 51

Proof. For 1 < p < ∞ use the triangle inequality to get

Z Z Z
p p−1
| f + g| dµ ≤ | f || f + g| dµ + |g|| f + g| p−1 dµ.
X X X

The proof follows by applying Hölder’s inequality with exponents p and p/(p − 1) on both
terms of the right hand side and then rearranging terms in the resulting inequality.

Hence, if p ∈ [1, ∞], the function ∥ · ∥ p : L p → [0, ∞] given by

( R 1/p
p
∥ f ∥ p ≡ ∥ f ∥L p ≡ ∥ f ∥L p (X,µ) := X | f | dµ if p ∈ [1, ∞)
, f ∈ L p,
ess supX | f | if p = ∞

defines a norm on the linear space L p .

Theorem 4.4.9. Let 1 ≤ p ≤ ∞ and suppose fk : X → R̄, k ∈ N, is a Cauchy sequence in L p .

There exists then a subsequence { flk }k∈N such that

(i) | flk | ≤ F for all k ∈ N, µ-a.e. in X, and some nonnegative F ∈ L p ,

(ii) flk → f as k → ∞, µ-a.e. in X, and some f : X → R̄.

In particular, applying Fatou’s lemma to the sequence gk := | flk | p gives f ∈ L p . Applying then
the dominated convergence theorem for the sequence hk := | flk − f | p , we deduce ∥ flk − f ∥ p →
0, as k → ∞. This, together with the fact that { fk }k∈N is Cauchy in L p , imply ∥ fk − f ∥ p → 0,
as k → ∞.
52 CHAPTER 4. INTEGRAL
Chapter 5

The basic theorems of advanced analysis

5.1 Limit theorems

Theorem 5.1.1 (Fatou). Let fk : X → [0, ∞], k ∈ N be µ-measurable. Then
Z Z
lim inf fk dµ ≤ lim inf fk dµ.
X k→∞ k→∞ X
R
Proof. We assume lim infk→∞ X fk dµ < ∞, otherwise the thesis is immediate. Theorem
4.1.6-(ii) says lim infk→∞ fk : X → [0, ∞] is µ-measurable. Its nonnegativity implies through
theorem 4.3.11 that it is also µ-integrable. In particular
Z Z
lim inf fk dµ = sup g dµ, (5.1.1)
X k→∞ g∈A

where
A := g µ-integrable, simple, g ≤ lim inf fk µ-a.e. in X .

k→∞

Let g ∈ A . By proposition 4.3.4 we may assume g = ∑ j∈N α j χA j , for some {α j > 0} j∈N and
disjoint µ-measurable {A j ⊆ X} j∈N with j∈N A j ⊆ X.
S

Claim. Let t ∈ (0, 1). Then for each j ∈ N we have A j =

S
k∈N B j,k , where

B j,k := A j ∩ {x ∈ X fℓ (x) > tα j ∀ ℓ ≥ k}.

Proof of claim. Let j ∈ N. We only prove A j ⊆ k∈N B j,k , the reverse inclusion being self
S

evident. Since g ∈ A we know α j ≤ lim infk→∞ fk µ-a.e. in A j . Hence

tα j < lim inf fk µ-a.e. in A j .

l→∞ k≥ℓ

53
54 CHAPTER 5. THE BASIC THEOREMS OF ADVANCED ANALYSIS

This implies the existence of k0 ∈ N such that tα j < infk≥ℓ fk µ-a.e. in A j , for all ℓ ≥ k0 ,
which in turn gives tα j < fℓ µ-a.e. in A j , for all ℓ ≥ k0 . □
Using {A j } j∈N are disjoint with ⊆ X and then A j ⊇ B j,k for all k ∈ N, we get for
S
j∈N A j
any m ∈ N
Z m Z
fk dµ ≥ ∑ fk dµ
X j=1 A j
m Z m
≥ ∑ fk dµ ≥ t ∑ α j µ(B j,k ),
j=1 B j,k j=1

the last inequality coming straight from the definition of B j,k . Observing next B j,k+1 ⊇ B j,k
for all k ∈ N and taking the lim infk→∞ , we arrive at
Z m
lim inf fk dµ ≥ t lim µ(B j,k )
∑ α j k→∞
k→∞ X j=1
m m
=t ∑ α j µ(∪k∈NB j,k ) = t ∑ α j µ(A j ),
j=1 j=1

the penultimate equality coming from theorem 2.2.5-(ii). Consequently, letting m → ∞,

Z Z
lim inf fk dµ ≥ t g dµ ∀ t ∈ (0, 1), ∀ g ∈ A ,
k→∞ X X

which implies Z Z
lim inf fk dµ ≥ sup g dµ.
k→∞ X g∈A X

Coupling this with (5.1.1) gives the desired estimate.

Theorem 5.1.2 (monotone convergence - Beppo Levi). Let fk : X → [0, ∞], k ∈ N be µ-
measurable such that fk ≤ fk+1 µ-a.e. in X, for all k ∈ N. Then
Z Z
lim fk dµ = lim fk dµ.
X k→∞ k→∞ X

Proof. Fatou’s lemma readily gives

Z Z
lim fk dµ ≤ lim inf fk dµ. (5.1.2)
X k→∞ k→∞ X

On the other hand, fk ≤ limk→∞ fk µ-a.e. in X for all k ∈ N. This implies

Z Z
fk dµ ≤ lim fk dµ ∀ k ∈ N,
X X k→∞
5.1. LIMIT THEOREMS 55

and so Z Z
lim sup fk dµ ≤ lim fk dµ. (5.1.3)
k→∞ X X k→∞

The proof follows now from (5.1.2) and (5.1.3).

Exercise 5.1.3. Let f , fk : X → R̄, k ∈ N, be µ-summable and satisfy

Z
lim | fk − f | dµ = 0.
k→∞ X

Prove there exists a subsequence { flk }k∈N converging to f µ-a.e. in X.

R
[Hint: Setting αk := X | fk − f |dµ, extract a subsequence {αℓk }k∈N such that ∑k∈N αℓk ≤ 1.
Apply then the monotone convergence theorem to the sequence gk := ∑kj=1 | fℓ j − f |.]

Theorem 5.1.4 (absolute continuity of integrals - Vitali). If f : X → R̄ is µ-summable then

Z
∀ ε > 0, ∃ δ > 0 such that if A ⊆ X is µ-measurable with µ(A) < δ , then | f | dµ < ε.
A

Proof. We define a sequence of functions gk : X → [0, ∞], k ∈ N, by truncating | f | as follows

| f (x)| if | f (x)| < k,


gk (x) :=
k if | f (x)| ≥ k.

The sequence {gk }k∈N enjoys the following properties

(i) gk ≤ min{k, | f |} µ-a.e. in X, for all k ∈ N,

(ii) gk → | f | as k → ∞, µ-a.e. in X,

(iii) gk ≤ gk+1 µ-a.e. in X, for all k ∈ N.

Indeed, for (i) observe that gk ≤ k and gk ≤ | f | both µ-a.e. in X. (ii) is obvious. For
(iii) notice in case | f (x)| < k that gk (x) = | f (x)| = gk+1 (x), in case k ≤ | f (x)| < k + 1 that
gk (x) = k ≤ | f (x)| = gk+1 (x) and in case | f (x)| ≥ k + 1 that gk (x) = k < k + 1 = gk+1 (x).
Because of (ii) and (iii) we get through the monotone convergence theorem
Z Z
lim gk dµ = | f | dµ;
k→∞ X X

that is, given ε > 0, there exists kε ∈ N such that

Z Z
gk dµ − | f | dµ < ε for all k ≥ kε .
X X
56 CHAPTER 5. THE BASIC THEOREMS OF ADVANCED ANALYSIS

But (i) implies the difference in the absolute value is non-positive and so
Z Z
| f | dµ < gkε dµ + ε.
X X

Let δ := ε/kε . For any µ-measurable A ⊆ X with µ(A) < δ we rewrite the above inequality
Z Z Z
| f | dµ < gkε dµ + (gkε − | f |) dµ + ε.
A A Ac

Property (i) implies the second integral on the right is non-positive and also the first integral
does not exceed kε µ(A). Hence
Z
| f | dµ < kε δ + ε = 2ε.
A
Rescaling ε the proof is complete.
Theorem 5.1.5 (dominated convergence - Lebesgue). Let f , fk : X → R̄, k ∈ N, be µ-
measurable, g : X → [0, ∞] be µ-summable, satisfying
(i) fk → f as k → ∞, µ-a.e. in X,
(ii) | fk | ≤ g µ-a.e. in X, for all k ∈ N.
Then Z
lim | fk − f | dµ = 0. (5.1.4)
k→∞ X

Proof. First notice that (ii) and (i) imply | f | ≤ g µ-a.e. in X. This and (ii) again give through
the triangle inequality
2g − | fk − f | ≥ 2g − | fk | − | f | ≥ 0 µ-a.e. in X ∀ k ∈ N.
Applying Fatou’s lemma to the sequence defined by the left hand side gives
Z
Z

lim inf 2g − | fk − f | dµ ≤ lim inf 2g − | fk − f | dµ,
X k→∞ k→∞ X

and using (i) once more together with the summability of g and exercise 1.1.2
Z Z Z
2g dµ ≤ 2g dµ − lim sup | fk − f | dµ.
X X k→∞ X

Cancelling the common integral on both sides we get

Z
lim sup | fk − f | dµ ≤ 0,
k→∞ X

which yields (5.1.4).

5.1. LIMIT THEOREMS 57

Exercise 5.1.6. Prove Z 1 

2 /k
lim 1 − e−x x−1/2 dx = 0.
k→∞ 0

Exercise 5.1.7. Let f , fk : X → R̄, k ∈ N, be µ-measurable and g, gk : X → [0, ∞] be µ-

summable, satisfying
(i) fk → f as k → ∞, µ-a.e. in X,
(ii) gk → g as k → ∞, µ-a.e. in X,
R R
(iii) limk→∞ X gk dµ = Xg dµ, and
(iv) | fk | ≤ gk µ-a.e. in X, for all k ∈ N.
R
Prove limk→∞ X | fk − f | dµ = 0.
Exercise 5.1.8. Compute the following limit
sin(x/k)
Z ∞
lim dx.
k→∞ 0 (1 + x/k)k
Theorem 5.1.9 (missing term in Fatou’s lemma - Brezis and Lieb). Let f , fk : X → R̄,
k ∈ N, be µ-measurable satisfying
(i) fk → f as k → ∞, µ-a.e. in X, and
p
R
(ii) M := supk∈N X | fk | dµ < ∞ for some p > 0.
Then Z
lim | fk | p − | fk − f | p − | f | p dµ = 0.
k→∞ X
Remark 5.1.10. Assumption (i) implies through the triangle inequality that | fk | → | f | as
k → ∞, µ-a.e. in X. So we further deduce | fk | p → | f | p as k → ∞, µ-a.e. in X. Applying now
Fatou’s lemma to | fk | p gives Z
| f | p dµ ≤ M. (5.1.5)
X
Remark 5.1.11. The conclusion of the theorem implies (use theorem 4.3.12-(iv))
Z
| fk | p − | fk − f | p − | f | p dµ = 0.


lim
k→∞ X

This together with the above remark imply

Z Z
| f | p dµ = lim | fk | p − | fk − f | p dµ,


(5.1.6)
X k→∞ X
p p f | p dµ = 0.
R R R
but we don’t know if limk→∞ X | fk | dµ = X |f| dµ or limk→∞ X | fk −
58 CHAPTER 5. THE BASIC THEOREMS OF ADVANCED ANALYSIS

Remark 5.1.12. Rewriting (5.1.6) as

Z Z Z 
| f | p dµ = lim inf | fk | p dµ − | fk − f | p dµ ,
X k→∞ X X

and comparing this with Fatou’s lemma

Z Z
| f | p dµ ≤ lim inf | fk | p dµ,
X k→∞ X

justifies the “missing term in Fatou’s lemma”.

Proof. It is enough to prove the following statement: Let gk : X → R̄, k ∈ N, be µ-measurable
satisfying
(I) gk → 0 as k → ∞, µ-a.e. in X, and
p
R
(II) M̃ := supk∈N X |gk | dµ < ∞ for some p > 0.
p dµ
R
Then for any µ-measurable g : X → R̄ such that X |g| < ∞, there holds
Z
lim |gk + g| p − |gk | p − |g| p dµ = 0. (5.1.7)
k→∞ X

Indeed, let fk , f satisfy the hypotheses of the theorem and take gk := fk − f , g := f in the
above statement (the choice

p for g is eligible

because of remark 5.1.10). Then the elementary
1 p p
inequality |a| + |b| ≤ 2 |a| + |b| gives p

Z Z
p
|gk | p dµ ≤ | fk | + | f | dµ
X X Z Z
 
p p
≤2 | fk | dµ + | f | dµ ≤ 2 p+1 M,
p
X X

by (ii) and (5.1.5). This establishes (II) with M̃ = 2 p+1 M.

Now let ε > 0. We define {Gε,k : X → R̄}k∈N , by
Gε,k := |gk + g| p − |gk | p − |g| p − ε|gk | p .

Claim.2 ∀ ε > 0, ∃ Cε > 0 such that |a + b| p − |a| p ≤ ε|a| p +Cε |b| p ∀ a, b ∈ R.

Accepting this claim for the moment, we readily have
|gk + g| p − |gk | p − |g| p ≤ |gk + g| p − |gk | p + |g| p
≤ ε|gk | p +Cε |g| p + |g| p .
1

p
p
|a| + |b| ≤ 2 max{|a|, |b|} = 2 p max{|a| p , |b| p } ≤ 2 p |a| p + |b| p


2 it’s proof will be an easy exercise (see remark 9.1.5) as soon as we learn about convex functions
5.2. PRODUCT MEASURES AND FUBINI’S THEOREM 59

Thus Gε,k ≤ (Cε + 1)|g| p and taking positive parts

Gε,k + ≤ (Cε + 1)|g| p .

The right hand side of this is µ-summable by assumption. In addition, Gε,k → 0 as k → ∞,

µ-a.e. in X and so

Gε,k + → 0 as k → ∞, µ-a.e. in X.


Therefore the dominated convergence theorem applies to Gε,k + to deduce
Z

lim Gε,k +
dµ = 0. (5.1.8)
k→∞ X


Integrating on X with respect to µ the obvious inequality Gε,k ≤ Gε,k +
, we have
Z Z Z
p p p p


|gk + g| − |gk | − |g| dµ ≤ ε |gk | dµ + Gε,k +
dµ (5.1.9)
X X Z X


≤ ε M̃ + Gε,k +
dµ,
X

where we used (II) to get to the last inequality. Taking the limit and using (5.1.8)
Z
lim sup |gk + g| p − |gk | p − |g| p dµ ≤ ε M̃.
k→∞ X

Since ε > 0 is arbitrary we deduce (5.1.7).

5.2 Product measures and Fubini’s theorem

Let µ be a measure on a set X, and ν be measure on a set Y .

Definition 5.2.1 (product measure). The product measure of µ and ν is the measure µ × ν :
2X×Y → [0, ∞] given by

(µ × ν)(S) := inf ∑ µ(Ak )ν(Bk ) S ⊆ X ×Y,

k∈N

where the infimum is taken over all collections of µ-measurable sets Ak ⊆ X and ν-measurable
sets Bk ⊆ Y , k ∈ N, such that [
S⊆ (Ak × Bk ).
k∈N
60 CHAPTER 5. THE BASIC THEOREMS OF ADVANCED ANALYSIS

Definition 5.2.2 (σ -finite set). The set A ⊆ X is called σ -finite with respect to µ if
[
A= Ak , where Ak are µ-measurable and µ(Ak ) < ∞ ∀ k ∈ N.
k∈N

Exercise 5.2.3. Prove that if A ⊆ X is σ -finite with respect to µ, then there exist disjoint
µ-measurable sets Bk , k ∈ N, such that A = k∈N Bk and µ(Bk ) < ∞ for all k ∈ N.
S

Definition 5.2.4 (σ -finite function). The function f : X → R̄ is called σ -finite with respect
to µ if f is µ-measurable and the set {x ∈ X f (x) ̸= 0} is σ -finite with respect to µ.
Theorem 5.2.5 (Fubini). (i) µ × ν is a regular measure on X ×Y .
(ii) If A ⊆ X is µ-measurable and B ⊆ Y is ν-measurable, then A × B is µ × ν-measurable
and (µ × ν)(A × B) = µ(A)ν(B).
(iii) If S ⊆ X × Y is σ -finite with respect to µ × ν, then Sy := {x ∈ X (x, y) ∈ S} is µ-
measurable for ν-a.e. y ∈ Y , Sx := {y ∈ Y (x, y) ∈ S} is ν-measurable for µ-a.e.
x ∈ X, µ(Sy ) is ν-integrable and ν(Sx ) is µ-integrable. Moreover,
Z Z
(µ × ν)(S) = µ(Sy ) dν(y) = ν(Sx ) dµ(x).
Y X

(iv) If f : X × Y → R̄ Ris µ × ν-integrable and also σ -finite with respect R

to µ × ν, then
the mapping y 7→ X f (x, y)dµ(x) is ν-integrable, the mapping x 7→ Y f (x, y)dν(y) is
µ-integrable, and
Z Z hZ i Z hZ i
f d(µ × ν) = f (x, y) dµ(x) dν(y) = f (x, y) dν(y) dµ(x).
X×Y Y X X Y
Proof. to be added
Exercise 5.2.6. Let α > 0. Given an L 1 -summable function f : (0, α) → R̄, define g :
(0, α) → R̄ by
f (t)
Z α
g(x) := dt.
x t
Prove g is also L 1 -summable and satisfies
Z α Z α
g(x) dx = f (x) dx.
0 0
Exercise 5.2.7. The convolution of two L n -measurable functions f , g is defined by
Z
( f ∗ g)(x) := f (x − y)g(y) dL n (y),
Rn
for any x ∈ Rn such that the integral exists. Use Fubini’s theorem and Hölder’s inequality to
prove the following version of Young’s inequality:
5.3. SOME DENSITY RESULTS FOR L n -SUMMABLE FUNCTIONS 61

If f ∈ L1 (Rn , L n ) and g ∈ L p (Rn , L n ), p ∈ [1, ∞], then f ∗ g ∈ L p (Rn , L n ) with

∥ f ∗ g∥L p (Rn ,L n ) ≤ ∥ f ∥L1 (Rn ,L n ) ∥g∥L p (Rn ,L n ) .

Lemma 5.2.8 (continuous version of the Minkowski inequality). Let p ∈ [1, ∞). For any
f ∈ L p (Rn × Rn , L n ) we have
Z Z p 1/p Z Z 1/p
nn p n
| f (x, y)| dL (y) dL (x) ≤ | f (x, y)| dL (x) dL n (y).
Rn Rn Rn Rn

p
p−1
dL n (y). Then, writing F(x) = F(x) F(x)
R
Proof. Set F(x) := Rn | f (x, y)| , we get
Z
p Z Z 
p−1
n
I := F(x) dL (x) = | f (x, y)| dL n (y) F(x) dL n (x)
Rn R n
Z Z Rn

p−1 
= | f (x, y)| F(x) dL n (x) dL n (y),
Rn Rn

from Fubini’s theorem. Applying now Hölder’s inequality on the inner integral we deduce
Z Z 1/p  Z
p (p−1)/p
p n n
I≤ | f (x, y)| dL (x) F(x) dL (x) dL n (y)
Rn Rn Rn
Z Z 1/p
= I(p−1)/p | f (x, y)| p dL n (x) dL n (y).
Rn Rn

1/p
Hence I1/p ≤ p dL n (x) dL n (y) which is the desired inequality.
R R
Rn Rn | f (x, y)|

5.3 Some density results for L n-summable functions

Theorem 5.3.1 (Separability of the space of L n -summable functions).

...
62 CHAPTER 5. THE BASIC THEOREMS OF ADVANCED ANALYSIS
Chapter 6

Some representation formulas

Theorem 6.0.1 (layer cake

representation formula). Let ν be a Radon measure on [0, ∞)
and set φ (t) := ν [0,t) . Let µ be a regular measure on X and suppose that X is σ -finite
with respect to µ. For any µ-measurable function f : X → [0, ∞] we have1
Z
Z ∞

φ f (x) dµ(x) = µ { f > t} dν(t). (6.0.1)
X 0



If in addition the measure ν satisfies ν [0,t) = ν [0,t] for any t ≥ 0, then also
Z
Z ∞

φ f (x) dµ(x) = µ { f ≥ t} dν(t). (6.0.2)
X 0

In particular, if φ : [0, ∞] → [0, ∞] is a continuously differentiable non-decreasing function

with φ (0) = 0, we get
Z Z ∞
φ ′ (t)µ { f > t} dL 1 (t).



φ f (x) dµ(x) = (6.0.3)
X 0

Proof. Starting from the left hand side and using the definition of ν we have
Z
Z 


φ f (x) dµ(x) = ν 0, f (x) dµ(x) (6.0.4)
X ZX Z ∞
= χ
(t) dν(t) dµ(x).
X 0 0, f (x)

Fubini’s theorem applies to give

Z
Z ∞Z
φ f (x) dµ(x) = χ
(t) dµ(x) dν(t).
X 0 X 0, f (x)

1 Throughout we write for instance { f > t} in place of {y ∈ X f (y) > t}

63
64 CHAPTER 6. SOME REPRESENTATION FORMULAS

Since χ
(t) = χ
{ f >t} (x) for all t ∈ [0, ∞), x ∈ X, we conclude
0, f (x)
Z
Z ∞Z
φ f (x) dµ(x) = χ{ f >t} (x) dµ(x) dν(t)
X Z0∞ X

= µ { f > t} dν(t).
0



For the formula (6.0.2), the assumption ν [0,t) = ν [0,t] for any t ≥ 0 when applied to
(6.0.4) gives Z Z

 

φ f (x) dµ(x) = ν 0, f (x) dµ(x).
X X
We conclude with the proof by following the same steps as before. Finally, to get formula
(6.0.3) notice that
Z Z Z f (x)
φ ′ (t) dt dµ(x)


φ f (x) dµ(x) =
X ZX Z0 ∞
= χ
(t)φ ′ (t) dL 1 (t) dµ(x).
X 0 0, f (x)

Now proceed as above from the application of Fubini’s theorem and on.

Remark 6.0.2. Under the assumptions for (6.0.2) to hold true, since { f > t} ⊆ { f ≥ t},
monotonicity of measures, formulas (6.0.1), (6.0.2) and exercise 4.3.16 readily imply µ({ f >
t}) = µ({ f ≥ t}) for ν-a.e. t ∈ [0, ∞); that is,

µ({ f = t}) = 0 for ν-a.e. t ∈ [0, ∞).

Observe that in case µ is a Radon measure on Rn and f is Borel measurable then this follows
from lemma 2.3.10. Indeed, the sets Et := { f = t}, t ∈ I := [0, ∞) form a family of disjoint
Borel sets. Hence lemma 2.3.10 asserts µ(Et ) > 0 for at most countable many t ∈ I ; that is,

µ({ f = t}) = 0 for all but countably many t ∈ [0, ∞).

But the condition ν [0,t) = ν [0,t] for any t ≥ 0 implies that singletons are of ν-measure
0 and consequently so is any countable subset of [0, ∞).
Example 6.0.3. Some further useful cases are:
(i) Take X = Rn and µ to be Dirac’s delta measure δx , x ∈ Rn . For ν = L 1 |[0,∞) , the
formulas above read: For any L n -measurable function f : Rn → [0, ∞] we have
Z ∞ Z ∞
f (x) = χ{ f >t} (x) dL 1 (t) = χ{ f ≥t} (x) dL 1 (t) L n -a.e. in Rn . (6.0.5)
0 0
 65

(ii) Take X = Rn and µ = L n . If u : Rn → R is L n -measurable, then f := |u| is L n -

measurable by Theorem 4.1.6. Hence, with these choices for X, µ, f and taking
ν [0,t) = t p , p ≥ 1, we get from (6.0.3) the following representation formula for

the integral of the absolute value to the p-th power

Z Z ∞
p n
t p−1 L n {|u| > t} dL 1 (t).


|u| dL = p (6.0.6)
Rn 0

Definition 6.0.4. Let p, q > 0. The Lorentz space L p,q is the collection of all measurable
functions f defined on Rn , such that [ f ]L p,q < ∞, where
!1/q
q
Z ∞
q
[ f ]L p,q := f ∗ (t) t q/p−1 dt .
p 0

Here, f ∗ denotes the decreasing rearrangement of f

f ∗ (t) := inf{s ≥ 0 s.t. µ f (s) ≤ t},

where µ f is the distribution function of f

µ f (s) := L n {x ∈ Rn s.t. | f (x)| > s} .

Exercise 6.0.5. Prove the following properties of Lorenz spaces:

p dL n .
R
(i) We have [ f ]L p,p = Rn | f |

(ii) If χω is the characteristic function of a set ω with finite Lebesgue measure, then
[χω ]L p,q = [L n (ω)]1/p for all p, q > 0.

(iii) Let p > 0 and q2 ≥ q1 > 0. There holds [ f ]L p,q1 ≥ [ f ]L p,q2 .

We prove next an analogous to (6.0.6) formula for f = |∇u|, where u ∈ Cc1 (Rn ).2 To establish
it we will need the following proposition and its consequence, the divergence or Gauss-Green
theorem. The interested student can find the proof in [EG]-§3.3.

Proposition 6.0.6. The measure H n−1 of a sufficiently smooth bounded surface agrees with
the surface’s area as taught in vector calculus.

A consequence is the following3

2 This a very simple case of the coarea formula; see also §3.4.2 of [EvG], or Theorem 3.2.11 of [H].
3A nice proof of this fundamental calculus fact can be found in the Appendix of [W].
66 CHAPTER 6. SOME REPRESENTATION FORMULAS

Theorem 6.0.7 (divergence theorem). Let Ω ⊂ Rn be a bounded domain with smooth bound-
ary and ⃗F = (F1 , ..., Fn ), Fi ∈ C1 (Ω̄) for any i = 1, ..., n, be a vector field. Then
Z Z
div ⃗F(x) dx = ⃗F(x) ·⃗ν∂ Ω (x) dH n−1 (x),
Ω ∂Ω

where ⃗ν∂ Ω (x) is the outwards pointing unit normal at the point x ∈ ∂ Ω.

Theorem 6.0.8 (representation formula for the integral of the modulus of the gradient).
For any u ∈ Cc1 (Rn ) we have
Z Z ∞
H n−1 {|u| = t} dL 1 (t),


|∇u(x)| dx = (6.0.7)
Rn 0

where H n−1 is the n − 1-dimensional Hausdorff measure.

⃗ ε : Rn → Rn by
Proof. Let ε > 0 and u ∈ Cc1 (Rn ). Define also U

⃗ ε (x) := p ∇u(x)
U .
|∇u(x)|2 + ε

Noting that for any t ≥ 0 we have {u+ > t} = {u > t}, formula (6.0.5) implies
Z ∞
+
u (x) = χ{u>t} (x) dL 1 (t) L n -a.e. in Rn .
0

Also, since for any t ≥ 0 we have {u− ≥ t} = {u ≤ −t}, the same formula gives
Z ∞
−
u (x) = χ{u≤−t} (x) dL 1 (t)
0
Z 0
= χ{u≤t} (x) dL 1 (t)
−∞
Z 0
χ{u>t} (x) − 1 dL 1 (t) L n -a.e. in Rn .


=−
−∞

We then have
Z Z Z ∞ 
+ ⃗ ε (x) dL (x) =
u (x) div U n
χ{u>t} (x) dL 1 (t) div U
⃗ ε (x) dL n (x)
Rn n
ZR∞ Z 0
= ⃗ ε (x) dx dL 1 (t),
div U
0 {u>t}
 67

by the Fubini theorem. Correspondingly,

Z Z Z 0 
− ⃗ ε (x) dL n (x) = 1 ⃗ ε (x) dL n (x)


− u (x) div U χ{u>t} (x) − 1 dL (t) div U
Rn Rn −∞
Z 0 Z
⃗ ε (x) dL n (x) dL 1 (t)


= χ{u>t} (x) − 1 div U
−∞ Rn
Z 0 Z Z 
= ⃗ ε (x) dx −
div U ⃗ ε (x) dx dL 1 (t)
div U
−∞ {u>t} Rn
Z 0 Z
= ⃗ ε (x) dx dL 1 (t),
div U
−∞ {u>t}

where we have applied Fubini’s theorem in the second equality and the divergence theorem
⃗ ε (x)dx = 0). Adding the last two equalities
R
in the last one (to show that Rn div U
Z Z ∞Z
⃗ ε (x) dx =
u(x) div U ⃗ ε (x) dx dL 1 (t),
div U
Rn −∞ {u>t}

from which, after an integration by parts on the left hand side and applying the divergence
theorem on the right hand side4 , we conclude
Z Z ∞Z
− ⃗ ε (x) dx =
∇u(x) · U ⃗ ε (x) ·⃗ν(x) dH n−1 (x) dL 1 (t),
U
Rn −∞ {u=t}

where ⃗ν(x) is the unit normal at a point x of the surface {u = t} pointing in the direction
where u ≤ t. From Theorem 1.2.9 we know ∇u(x) ·⃗ν(x) = −|∇u(x)| on {u = t} and so

⃗ ε (x) ·⃗ν(x) = − p |∇u(x)|

U on {u = t}.
|∇u(x)|2 + ε
Hence
|∇u(x)|2 |∇u(x)|
Z Z ∞Z
p dx = p dH n−1 (x) dL 1 (t).
Rn |∇u(x)|2 + ε −∞ {u=t} |∇u(x)|2 + ε

The results follows letting ε ↓ 0 through the monotone convergence theorem applied in both
sides.

Combining the above representation theorem with the isoperimetric inequality we get the
sharp L1 -Sobolev inequality, or sharp Gagliardo-Nirenberg inequality; that is,
4 we
need to know that the level set {u = t} is a smooth surface to apply it. This is a consequence of Sard’s
lemma which is going to be added in the next update of the notes (see the Appendix of [W])
68 CHAPTER 6. SOME REPRESENTATION FORMULAS

Theorem 6.0.9. For any u ∈ Cc∞ (Rn ) we have

Z Z (n−1)/n
1/n n/(n−1)
|∇u(x)| dx ≥ nωn |u(x)| dx . (6.0.8)
Rn Rn

Proof. From the isoperimetric inequality we have

1/n 
(n−1)/n
H n−1 {|u| = t} ≥ nωn L n {|u| > t}


.

Combining this with the representation formula of the last theorem,

Z Z ∞
(n−1)/n
1/n
|∇u(x)| dx ≥ nωn L n {|u| > t} dL 1 (t).
Rn 0

It suffices to prove
Z ∞
(n−1)/n
L n {|u| > t} dL 1 (t) ≥ I (n−1)/n , (6.0.9)
0

where I := Rn |u(x)|n/(n−1) dx. To this end using (6.0.5) with f = |u| and then Fubini’s
R

theorem we write
Z
I= |u(x)||u(x)|1/(n−1) dx
Z Rn Z ∞
= χ{|u|>t} (x) dL 1 (t)|u(x)|1/(n−1) dL n (x)
Rn
Z ∞ Z0
= χ{|u|>t} (x)|u(x)|1/(n−1) dL n (x) dL 1 (t).
0 Rn

Applying next Hölder’s inequality with the conjugate exponents n/(n − 1) and n,
Z ∞ Z
n/(n−1) (n−1)/n
I≤ χ{|u|>t} (x) n
dL (x) I 1/n dL 1 (t)
0 n
Z ∞R
(n−1)/n
= I 1/n L n {|u| > t} dL 1 (t),
0

from which (6.0.9) readily follows.

Chapter 7

The Fourier transform

7.1 The Fourier transform in L1

Definition 7.1.1 (Fourier transform in L1 ). For f ∈ L1 (Rn ), the Fourier transform fˆ of f is
given by Z
fˆ(κ) := e−2πiκ·x f (x) dL n (x), k ∈ Rn .
Rn

Proposition 7.1.2 (properties of fˆ). The Fourier transform enjoys the following properties:

(i) For f ∈ L1 (Rn ) we have fˆ ∈ L∞ (Rn ) with ∥ fˆ∥∞ ≤ ∥ f ∥1 . In particular, since for non-
negative f ∈ L1 (Rn ) we get fˆ(0) = ∥ f ∥1 , we deduce ∥ fˆ∥∞ = ∥ f ∥1 in this case.

f + β g = α fˆ + β ĝ for all f , g ∈ L1 (Rn ) and all α, β ∈ C.

(ii) α \

∗ g(κ) = fˆ(κ)ĝ(κ) for all f , g ∈ L1 (Rn ), where f ∗ g(x) := f (x − y)g(y) dL n (y).

R
(iii) fd Rn

fˆg dL n = f ĝ dL n for all f , g ∈ L1 (Rn ).

R R
(iv) Rn Rn

−2πiκ·h fˆ(κ) for all h ∈ Rn , where τ f (x) := f (x − h), f ∈ L1 (Rn ).

h f (κ) = e
(v) τc h

n ˆ 1 n


λ f (κ) = λ f (λ κ) for all λ ∈ (0, ∞), where δλ f (x) := f (x/λ ), f ∈ L (R ).
(vi) δd

g(y)(κ) = τx ĝ(κ) for all x ∈ Rn and all g ∈ L1 (Rn ).

(vii) e2πiy·x
\

(viii) fˆ ∈ C(Rn ).
(ix) lim|κ|→∞ fˆ(κ) = 0.
2 /λ
(x) e−πλ |x| (κ) = λ −n/2 e−π|κ|
\2
for all λ ∈ (0, ∞).

69
70 CHAPTER 7. THE FOURIER TRANSFORM

Proof. Properties (i) and (ii) follow from Theorem 4.3.12. Properties (iii) and (iv) are con-
sequences of the Fubini theorem. Each of the property (v) and (vi) follows from the obvious
change of variables. Property (vii) is elementary. For a given κ0 ∈ Rn we write
Z
ˆf (κ) − fˆ(κ0 ) = gκ (x) dL n (x) where gκ (x) := e−2πiκ·x − e−2πiκ0 ·x f (x).


Rn

Clearly, limκ→κ0 gκ (x) = 0 and |gκ (x)| ≤ 2| f (x)|, both for L n -a.e. x ∈ Rn . Since f ∈ L1 (Rn ),
the dominated convergence theorem applies to establish (viii). @@@@ To prove (x), we
compute
Z Z √ √
−2πik·x−πλ |x|2 −π|k|2 /λ 2
e−|i π/λ k+ πλ x| dx
\
−πλ |x|2
e (k) = e dx = e
Rn Rn
Z Z n
−n/2 −π|k|2 /λ −|y|2 −n/2 −π|k|2 /λ 2
= (πλ ) e e dy = (πλ ) e e−t dt .
Rn R
R −t 2
√
Finally notice that Re dt = π (see Exercise 1.3.1).

Lemma 7.1.3 (continuity of the translation operator in L p ). Let p ∈ [1, ∞). For any f ∈
L p (Rn , L n ) we have limh→0 ∥τh f − f ∥ p = 0.
Proposition 7.1.4 (invertibility of the Fourier transform). If f , fˆ ∈ L1 (Rn ) then
Z
e2πik·x fˆ(k) dk = fˆˆ(−x) , for a.e. x ∈ Rn .


f (x) =
Rn

Proof. For any ε > 0 and any x ∈ Rn we set

Z
2
Jε (x) := e2πik·x−επ|k| fˆ(k) dk.
Rn
2
Using first property (iv), then property (vii) with g(y) := e−επ|k| and finally property (x)
with λ = ε, we deduce
Z
−n/2 2 /ε
Jε (x) = ε e−π|k−x| f (k) dL n (k).
Rn

Because of Fubini’s theorem and the continuity of the translation operator in L1 , we can see
that Jε (x) converges to f (x) in L1 (Rn ) (see below, in the proof of “Approximation in L p theo-
rem”, for the details). Hence, for some subsequence δ (ε) we know Jδ (ε) (x) → f (x) for L n -
a.e. x ∈ Rn , as ε → 0. On the other hand, the integrand of Jδ (ε) (x) converges to e2πik·x fˆ(k)
as ε → 0, for all x ∈ Rn . Moreover, this same integrand is dominated by | fˆ(k)| which is (by
assumption) an L1 (Rn , dL n (k)) function. So the dominated convergence theorem applies to
end the proof.
7.1. THE FOURIER TRANSFORM IN L1 71

Remark 7.1.5. It turns out from the regularity properties (i), (viii) and (ix) of Proposition
7.1.2 that, given f ∈ L1 (Rn ) we have f ∈ L∞ (Rn ) ∩ C0 (Rn ), but note carefully that it is not
necessarily an L1 (Rn ) function. Hence the invertibility formula is not valid in general (con-
sider for instance f = χ(α,β ) in R).
Using the last lemma (continuity of the translation operator in L p ), as well as the con-
tinuous version of Minkowski’s inequality (see Lemma 5.2.8), one has the following useful
approximation theorem, already used with p = 1 in the “Invertibility of the Fourier trans-
form”.

Theorem 7.1.6 (approximation in L p ). Let f ∈ L p (Rn ), p ∈ [1, ∞). Then,

Z
2 /ε
lim ε −n/2 e−π|k−x| f (k) dL n (k) − f (x)
= 0.
ε→0 Rn L p Rn ,dL n (x)

Proof. Let
Z
−n/2 2 /ε
Iε := ε e−π|k−x| f (k) dL n (k) − f (x)
.
Rn L p Rn ,dL n (x)

Since ε −n/2 −π|k−x|2 /ε

R
Rn e dk = 1, we have
Z
−n/2 2
e−|y| f (x +
p
ε/πy) − f (x) dL n (y)


Iε = π
,
Rn L p Rn ,dL n (x)
p
where we have also changed variables by k = x + ε/πy. By Lemma 5.2.8 , or just by the
Fubini theorem in case p = 1, we arrive at
Z Z 1/p
−n/2 −|y|2 p
p n
Iε ≤ π e f (x + ε/πy) − f (x) dL (x) dL n (y). (⋆)
Rn Rn
p
But h(ε) := ε/πy → 0 as ε → 0. Hence from the continuity of the translation operator
in L p we get that limε→0 ∥τh(ε) f − f ∥ p = 0; that is, the integrand on the right term of (⋆)
vanishes as ε → 0. Using Minkowski’s inequality we easily see that this same integrand is
2
dominated by 2∥ f ∥ p e−|y| which is an L1 Rn , dL n (y) function. So the dominated conver-

gence theorem applies to end the proof.

We use the above theorem with p = 2 in the proof of the following fundamental result:
Theorem 7.1.7 (Plancherel’s formula). If f ∈ L1 (Rn )∩L2 (Rn ), then fˆ ∈ L2 (Rn ) with ∥ fˆ∥2 =
∥ f ∥2 .
...
72 CHAPTER 7. THE FOURIER TRANSFORM

7.2 The Fourier transform in L2

Definition 7.2.1 (Fourier transform in L2 ). Given f ∈ L2 (Rn ), think of a sequence { f j ∈
L1 (Rn ) ∩ L2 (Rn )} j∈N such that f j → f in L2 . By Plancherel’s formula we get then

∥ fˆj − fˆl ∥2 = ∥ f j − fl ∥2 ∀ j, l ∈ N;

that is, { fˆj } j∈N is a Cauchy sequence in L2 . But L2 is complete and thus { fˆj } j∈N converges
to a function of L2 (Rn ) which we call the Fourier transform of f and denote it by fˆ.
Remark 7.2.2. Given f ∈ L2 (Rn ) we can always find sequences { f j ∈ L1 (Rn ) ∩ L2 (Rn )} j∈N
such that f j → f in L2 . For example, taking f j := η1/ j ∗ f , where ηε for ε > 0 is the standard
mollifier, we have { f j ∈ Cc∞ (Rn )} j∈N such that f j → f in L2 . Another example is { f j :=
f χB j } j∈N . Since by Hölder’s inequality ∥ f j ∥1 ≤ ∥ f ∥2 [L n (B j )]1/2 for all j ∈ N and also
∥ f j ∥2 ≤ ∥ f ∥2 for all j ∈ N, we have f j ∈ L1 (Rn ) ∩ L2 (Rn ) for all j ∈ N. Moreover,
Z
∥fj − f ∥22 = g j dL n , where g j := (1 − χB j )| f |2 .
Rn

Since lim j→∞ g j → 0 L n -a.e. in Rn and g j ≤ | f |2 ∈ L1 (Rn ), the dominated convergence

theorem readily gives f j → f in L2 . This last example provides us with a fairly simple
sequence of functions whose L2 -limit defines fˆ:
nZ o
e−2πik·x f (x) dL n (x) .
Bj j∈N

Remark 7.2.3. Although there are many sequences such that f j → f in L2 , fˆ is independent
of the one we choose. Indeed, suppose { f j ∈ L1 (Rn ) ∩ L2 (Rn )} j∈N satisfies lim j→∞ ∥ f j −
f ∥2 = 0. Then fˆ is defined as the L2 -limit of fˆj , hence lim j→∞ ∥ fˆj − fˆ∥2 = 0. Let {g j ∈
L1 (Rn ) ∩ L2 (Rn )} j∈N be one more sequence such that lim j→∞ ∥g j − f ∥2 = 0. Then

∥ĝ j − fˆ∥2 ≤ ∥ĝ j − fˆj ∥2 + ∥ fˆj − fˆ∥2

= ∥g j − f j ∥2 + ∥ fˆj − fˆ∥2
≤ ∥g j − f ∥2 + ∥ f j − f ∥2 + ∥ fˆj − fˆ∥2 ,

where we have used Plancherel’s formula to get the middle equality. Thus, lim j→∞ ∥ĝ j −
fˆ∥2 = 0 which says fˆ is the L2 -limit of {ĝ j } j∈N too. □

Proposition 7.2.4 (Properties of fˆ). (i) (isometry) If f ∈ L2 (Rn ), then ∥ fˆ∥2 = ∥ f ∥2 .

f + β g = α fˆ + β ĝ.
(ii) (linearity) If f , g ∈ L2 (Rn ) and α, β ∈ C, then α \
7.2. THE FOURIER TRANSFORM IN L2 73

(iii) (Parseval’s formula) If f , g ∈ L2 (Rn ), then ⟨ f , g⟩ = ⟨ fˆ, ĝ⟩, where

Z
⟨ f , g⟩ := f¯g dL n .
Rn

(iv) (invertibility) If f ∈ L2 (Rn ), then f (x) = fˆˆ(−x).

Proof. (i) Suppose { f j ∈ L1 (Rn ) ∩ L2 (Rn )} j∈N satisfies lim j→∞ ∥ f j − f ∥2 = 0, hence ∥ f ∥2 =
lim j→∞ ∥ f j ∥2 . But fˆ is defined as the L2 -limit of fˆj ; that is, lim j→∞ ∥ fˆj − fˆ∥2 = 0, hence
∥ fˆ∥2 = lim j→∞ ∥ fˆj ∥2 . Now the proof follows by Plancherel’s formula: ∥ fˆj ∥2 = ∥ f j ∥2 by
letting j → ∞.
(iv) Suppose { f j ∈ L1 (Rn ) ∩ L2 (Rn )} j∈N satisfies lim j→∞ ∥ f j − f ∥2 = 0. As in the proof
of the Proposition - Invertibility of the Fourier transform, we have
Z Z
n
2πik·x−επ|k|2 2 /ε
e fˆj (k) dk = ε − 2 e−π|k−x| f j (k) dL n (k) ∀ j ∈ N, ∀ ε > 0. (⋆)
Rn Rn

By Hölder’s inequality,
Z rZ
− n2 −π|k−x|2 /ε n
f j (k) − f (k) dL n (k) ≤ ε − 2
2 /ε
e−2π|k−x|


ε e dk ∥ f j − f ∥2 → 0,
Rn Rn

as j → ∞. Also fˆ is defined as the L2 -limit of fˆj ; that is, lim j→∞ ∥ fˆj − fˆ∥2 = 0. Hence,
Z rZ
2πik·x−επ|k|2 ˆ ˆ n 2
e−2επ|k| dk ∥ fˆj − fˆ∥2 → 0,


e f j (k) − f (k) dL (k) ≤
Rn Rn

as j → ∞. So taking the limit as j → ∞ in (⋆),

Z Z
n
2πik·x−επ|k|2 2 /ε
e fˆ(k) dL n (k) = ε − 2 e−π|k−x| f (k) dL n (k) ∀ ε > 0. (∗)
Rn Rn

From Theorem - Approximation in L p with p = 2, we know the rhs of (∗) converges to f (x) in
L2 . Hence there is a subsequence (that we don’t rename) such that the rhs of (∗) converges to
f (x) for L n -a.e. x ∈ Rn . To see that the lhs of (∗) converges (up to a subsequence) to fˆˆ(−x)
2
for L n -a.e. x ∈ Rn , set first gε (k) := e−επ|k| fˆ(k) and observe that gε ∈ L1 (Rn ) ∩ L2 (Rn ) for
all ε > 0. This allows to write lhs(∗) = ĝε (−x). Next we claim that gε converges to fˆ in L2 .
Indeed, we have
Z
2
2
∥gε − fˆ∥22 = 1 − e−επ|k| | fˆ(k)|2 dL n (k), ε > 0.
Rn

Clearly the integrand converges to 0 as ε → 0 for L n -a.e. k ∈ Rn , while it is also dominated

by 4| fˆ(k)|2 which is summable; hence the dominated convergence theorem applies to prove
74 CHAPTER 7. THE FOURIER TRANSFORM

the claim. Summarizing, we have {gε ∈ L1 (Rn )∩L2 (Rn )}ε>0 such that limε→0 ∥gε − fˆ∥2 = 0.
By definition of the Fourier transform in L2 we readily get fˆˆ is the L2 -limit of ĝε ; or, fˆˆ(−x)
is the L2 -limit of ĝε (−x) which equals the lhs of (∗). Passing to new subsequence we get that
lhs(∗) converges to fˆˆ(−x) for L n -a.e. x ∈ Rn .
Chapter 8

More examples of Radon measures

In this chapter, after proving in detail the Riesz representation theorem, we will combine
it with the Hahn-Banach theorem from functional analysis, to produce two more important
examples of Radon measures:

• positive distributions, and

• the variation measure of a function in BVloc .

We assume all students are familiar with the notions of a linear (or vector) space and also
linear subspace of a linear space.

Definition 8.0.1 (Minkowski functional, seminorm, norm). A Minkowski functional on the

linear space Λ is any mapping p : Λ → R with the properties

(i) p(λ1 + λ2 ) ≤ p(λ1 ) + p(λ2 ) for all λ1 , λ2 ∈ Λ,

(ii) p(tλ ) = t p(λ ) for all λ ∈ Λ, t > 0.

A seminorm on Λ is any Minkowski functional p on Λ which satisfies in addition

(iii) p(tλ ) = −t p(λ ) for all λ ∈ Λ, t ≤ 0.

A norm on Λ is any seminorm p on Λ such that

(iv) p(λ ) = 0 implies λ = 0Λ ,

where we have denoted by 0Λ the zero element of Λ.

For the proof of the following fundamental theorem of functional analysis, we refer to the
first two pages from the book of H. Brezis; [Br].

75
76 CHAPTER 8. MORE EXAMPLES OF RADON MEASURES

Theorem 8.0.2 (Hahn-Banach). Let Λ be a linear space and p be a Minkowski functional

on Λ. Suppose for some linear subspace Λ′ of Λ, a linear functional F : Λ′ → R satisfies

F(λ ) ≤ p(λ ) ∀ λ ∈ Λ′ .

There exists then a linear functional F : Λ → R such that

F (λ ) ≤ p(λ ) ∀ λ ∈ Λ,

which is an extension of F; that is F (λ ) = F(λ ) whenever λ ∈ Λ′ .

8.1 The Riesz representation theorem

Theorem 8.1.1. Let Ω ⊆ Rn be open and ℓ : Cc (Ω; Rm ) → R be a linear functional such that1

∀ compact K ⊂ Ω, ∃ C > 0 : |ℓ( f )| ≤ C∥| f |∥L∞ (Ω) ∀ f ∈ Cc (Ω; Rm ), sprt( f ) ⊆ K.

There exist then a Radon measure µ on Ω and a µ-measurable function σ : Ω → Rm such

that

(i) |σ (x)| = 1 for µ-a.e. x ∈ Ω,

f dµ for each f ∈ Cc (Ω; Rm ).

R
(ii) ℓ( f ) = Ω

Proof. For open V ⊆ Rn , set

µ(V ) := sup ℓ( f ) f ∈ Cc (Rn ; Rm ), | f | ≤ 1, sprt( f ) ⊂ V ,


(8.1.1)

then, for arbitrary A ⊆ Rn , set


µ(A) := inf µ(V ) open V ⊇ A .

Furthermore, denoting by Cc+ (Rn ) all nonnegative Cc (Rn ) functions, set

λ ( f ) := sup |ℓ(g)| g ∈ Cc (Rn ; Rm ), |g| ≤ f , f ∈ Cc+ (Rn ).



Accepting for the moment that µ is a Radon measure and that λ is positively homogeneous,
nondecreasing and additive, we show firstly that
Z
λ(f) = f dµ for all f ∈ Cc+ (Rn ). (8.1.2)
Rn
1 equivalently: sup ℓ( f ) f ∈ Cc (Ω; Rm ), | f | ≤ 1, sprt( f ) ⊆ K < ∞, for all compact K ⊂ Ω

8.1. THE RIESZ REPRESENTATION THEOREM 77

So pick f ∈ Cc+ (Rn ) and let ε > 0. Consider a partition Nj=1 [t j−1 ,t j ] of the interval [0, 2M],
S

where
M := max | f |,
Rn
and such that 0 < t j − t j−1 < ε. Since f −1 {t} ≡ {x ∈ Rn f (x) = t}, t ∈ [0, ∞), is a

family of disjoint


Borel sets and µ is a Radon measure, by lemma 2.3.10 we may assume that
µ f −1 {t j } = 0 for all j = 1, ..., N. Then by the subadditivity of µ we have
N
f −1 {ti } .
[

µ(Z) = 0, where Z := (8.1.3)
j=1

Set
U j := f −1 (t j−1 ,t j ) ≡ {x ∈ Rn t j−1 < f (x) < t j },


j = 1, ..., N.
Obviously:
N
[ 
the sets Z,U1 , ...,UN are disjoint and { f > 0} = Z ∪ Uj . (8.1.4)
j=1

Let j ∈ {1, ..., N}. Since f is continuous,

the set U j is open. Since f has compact support and
µ is Radon, µ(U j ) ≤ µ sprt( f ) < ∞. Consequently, (8.1.1) gives
(a) ∃ g j ∈ Cc (Rn ; Rm ), s.t. |g j | ≤ 1, sprt(g j ) ⊂ U j and ℓ(g j ) ≥ µ(U j ) − ε/N.
On the other hand, U j is also measurable and again because of µ(U j ) < ∞, theorem 2.3.8
implies
(b) ∃ compact K j ⊂ U j s.t. µ(U j \ K j ) = µ(U j ) − µ(K j ) < ε/N.
S
In addition, since K j sprt(g j ) is a compact subset of the open set U j , lemma 1.2.6 implies
(c) ∃ h j ∈ Cc+ (Rn ), s.t. sprt(h j ) ⊂ U j , h j = 1 on K j sprt(g j ), h j ≤ 1.
S

Since |g j | ≤ 1 = h j on sprt(g j ) we have λ (h j ) ≥ |ℓ(g j )| ≥ ℓ(g j ) and from (a) we obtain


λ (h j ) ≥ µ(U j ) − ε/N. On the other hand, since h j ≤ 1 and sprt(h j ) ⊂ U j we have g∈
n m n m

Cc (R ; R ), |g| ≤ h j ⊂ f ∈ Cc (R ; R ), | f | ≤ 1, sprt( f ) ⊂ U j , which implies in turn
λ (h j ) ≤ µ(U j ). Altogether,
µ(U j ) − ε/N ≤ λ (h j ) ≤ µ(U j ), j = 1, ..., N. (8.1.5)
Next we write A for the subset of Rn where both f and 1−h j are positive for all j ∈ {1, ..., N};
that is,
N n
\ o   N  
n n


A := x ∈ R f (x) 1 − h j (x) > 0 = x ∈ R f (x) 1 − ∑ h j (x) > 0 .
j=1 j=1
78 CHAPTER 8. MORE EXAMPLES OF RADON MEASURES

The last equality holds

Sbecause the sets sprt(h j ) ⊂ U j are disjoint and for this same reason
N


we can write A = Z ∪ j=1 U j \ {h j = 1} . Hence

N

µ(A) = ∑ µ U j \ {h j = 1} ( by (8.1.4) and (8.1.3) )
j=1
N

≤ ∑µ Uj \ Kj ( because K j ⊆ {h j = 1} )
j=1
≤ε ( from (b) ).
Also, f 1 − ∑Nj=1 h j ≤ MχA and we claim that ℓ f (1 − ∑Nj=1 h j ) ≤ Mµ(A). Indeed,

  N 
≤ sup |ℓ(g)| g ∈ Cc (Rn ; Rm ), |g| ≤ MχA

ℓ f 1− ∑ hj
j=1
= M sup |ℓ(g)| g ∈ Cc (Rn ; Rm ), |g| ≤ χA ,


because ℓ(cg) = cℓ(g) for any c ≥ 0. Next note that since A is open there holds
{g ∈ Cc (Rn ; Rm ), |g| ≤ χA } = {g ∈ Cc (Rn ; Rm ), |g| ≤ 1, sprt(g) ⊂ A}.
We deduce
  N 
ℓ f 1− ∑ hj ≤ Mµ(A).
j=1

Now we are ready to estimate λ ( f ). We write,

 N   N 
λ( f) = λ f − f ∑ hj +λ f ∑ hj ,
j=1 j=1

hence, taking into account all the underlined properties of λ and (8.1.5), we obtain
 N   N 
λ f ∑ h j ≤ λ ( f ) ≤ εM + λ f ∑ h j ⇒
j=1 j=1
N N
∑ λ ( f h j ) ≤ λ ( f ) ≤ εM + ∑ λ ( f h j ) ⇒
j=1 j=1
N
N
∑ t j−1 µ(U j ) − ε/N ≤ λ ( f ) ≤ εM + ∑ t j µ(U j ) ⇒
j=1 j=1
N N
∑ t j−1 µ(U j ) − 2εM ≤ λ ( f ) ≤ εM + ∑ t j µ(U j ), (8.1.6)
j=1 j=1
8.2. POSITIVE LINEAR FUNCTIONALS AND DISTRIBUTIONS 79

where in the last inequality we used t j−1 ≤ tN = 2M for all j = 1, ..., N. Next we estimate
R
Rn f dµ. Because of (8.1.4) we have
Z N Z
f dµ = ∑ f dµ ⇒
Rn j=1 U j
N Z N
∑ t j−1 µ(U j ) ≤ Rn
f dµ ≤ ∑ t j µ(U j ). (8.1.7)
j=1 j=1

Subtracting (8.1.6) and (8.1.7) we get

N Z N
− ∑ (t j − t j−1 )µ(U j ) − 2εM ≤ λ ( f ) − f dµ ≤ εM + ∑ (t j − t j−1 )µ(U j ) ⇒
j=1 Rn j=1
Z N
λ(f)− f dµ ≤ 2εM + ∑ (t j − t j−1 )µ(U j )
Rn j=1


≤ 2εM + ε µ sprt( f ) ,

from the fact that t j − t j−1 < ε and (8.1.4). This proves (8.1.2) since ε > 0 is arbitrary. □

...

8.2 Positive linear functionals and distributions

Let 0/ ̸= U ⊆ Rn , n ∈ N, be open. We start by introducing a topology on Cc∞ (U).

Definition 8.2.1. We denote by D(U) the space consisting all functions in Cc∞ (U) accom-
panied with the following notion of convergence: A sequence {φm ∈ Cc∞ (U)}m∈N is said to
converge to φ ∈ Cc∞ (U) in D(U) (and write φm → φ in D(U)) whenever

(i) the supports of the φm are contained in a fixed compact K ⊂ U, and

(ii) for any choice of ai ∈ N ∪ {0}, i = 1, ..., n, there holds

 ∂ a1  ∂ an  ∂ a1  ∂ an
... φm → ... φ uniformly in K.
∂ x1 ∂ xn ∂ x1 ∂ xn

Definition 8.2.2. We say T is a distribution in U and write T ∈ D ′ (U), if T is a member of

the dual space of D(U). Thus, T ∈ D ′ (U) if
80 CHAPTER 8. MORE EXAMPLES OF RADON MEASURES

(i) T (αφ + β ψ) = αT (φ ) + β T (ψ) ∀ φ , ψ ∈ Cc∞ (U), ∀ α, β ∈ R, and

(ii) T (φm ) → T (φ ), whenever φm → φ in D(U).
Example 8.2.3. (i) Dirac’s delta distribution is defined for a given x ∈ Rn by δx (φ ) = φ (x).
(ii) More generally,
R
any locally finite Borel measure µ on U defines a distribution Tµ by
Tµ (φ ) := U φ dµ.
The converse of (ii) is also true for nonnegative distributions. More generally, the values
of any nonnegative linear functional on Cc∞ (U) are represented by a Radon measure, accord-
ing to the following fundamental result
Theorem 8.2.4. Suppose the linear functional ℓ : Cc∞ (U) → R is nonnegative; that is: ℓ(φ ) ≥
0 for allR nonnegative φ ∈ Cc∞ (U). There exists then a Radon measure µ on U such that
ℓ(φ ) = U φ dµ for all φ ∈ Cc∞ (U).
Proof. Fix a compact K ⊂ U and use lemma 1.2.6 plus standard mollification (see remark
1.2.8) to select a function g ∈ Cc∞ (U) such that 0 ≤ g(x) ≤ 1 for all x ∈ U and g ≡ 1 on K. Now
given f ∈ Cc∞ (U) with sprt( f ) ⊆ K we can define φ ∈ Cc∞ (U) by φ = g maxU | f | − f . Clearly,
φ is nonnegative, hence from the hypothesis ℓ(φ ) ≥ 0 we deduce ℓ( f ) ≤ ℓ(g) maxU | f |. We
have just showed

ℓ( f ) ≤ ℓ(g) max | f | ∀ f ∈ Cc∞ (U), sprt( f ) ⊆ K.

U

Now we can apply the Hahn-Banach theorem with


Λ = f ∈ Cc (U) sprt( f ) ⊆ K ,
p(λ ) = ℓ(g) maxU |λ |, λ ∈ Λ,
Λ′ = f ∈ Cc∞ (U) sprt( f ) ⊆ K ,


F(λ ) = ℓ(λ ), λ ∈ Λ′ ,
to get that ℓ can be extended to a linear mapping (that we don’t rename) from Cc (U) into R,
so that
ℓ( f ) ≤ ℓ(g) max | f | ∀ f ∈ Cc (U), sprt( f ) ⊆ K.
U
In other words,

sup ℓ( f ) f ∈ Cc (U; R), | f | ≤ 1, sprt( f ) ⊆ K < ∞,

for any compact K ⊂ U. Hence, the Riesz representation theorem with m = 1 applies to get
a Radon measure µ on U and a µ-measurable function σ : U → R such that σ = ±1 µ-a.e.
8.3. WEAK DERIVATIVES OF FUNCTIONS OF BOUNDED VARIATION 81

and ℓ( f ) = U f σ dµ whenever f ∈ Cc (U). Finally, suppose that σ = −1 on a set A ⊆ Rn

R

of positive µ-measure and, given ε > 0, employ the outer and inner regularity properties of
Radon measures (theorems 2.3.7 and 2.3.8), to get open O ⊃ A and compact K ⊂ A such that

µ(O) − ε/2 < µ(A) < µ(K) + ε/2. (8.2.1)

Hence µ(O\K) < ε. Further, from lemma 1.2.6 plus standard mollification (see remark 1.2.8)
we get a Cc∞R(U) function f : U →R
[0, 1] with f R= 1 on K and sprt( f ) ⊂ O. The hypothesis
that ℓ( f ) = U f σ dµ ≥ 0, gives O\K f σ dµ + K f σ dµ ≥ 0. By the properties of f and σ ,
this readily implies µ(K) < ε. Since ε is arbitrary, (8.2.1) gives µ(A) = 0 and the proof is
complete.

8.3 Weak derivatives of functions of bounded variation

Let U ⊆ Rn , n ∈ N, be open. To motivate the definition of functions of bounded variation, we
show that if f ∈ C2 (U), then
Z nZ o
|∇ f (x)| dx = sup f (x) div g(x) dx g ∈ Cc1 (U; Rn ), |g| ≤ 1 in U . (8.3.1)
U U

To this end, let us write S for the supremum on the right hand side of (8.3.1). Given g ∈
Cc1 (U; Rn ) we may integrate by parts to get
Z Z Z
f (x) div g(x) dx = − ∇ f (x) · g(x) dx ≤ |∇ f (x) · g(x)| dx.
U U U

Further, provided that |g| ≤ 1 in U, we get

Z Z Z
f (x) div g(x) dx ≤ |∇ f (x)||g(x)| dx ≤ |∇ f (x)| dx;
U U U
R
therefore S ≤ U |∇ f (x)| dx. We prove next the reverse inequality. The regularity assumption
on f implies the set U ′ := {x ∈ U |∇ f (x)| ̸= 0} is open. Combining lemma 1.2.6 with
standard mollification (see remark 1.2.8), given a compact set K ⊂ U ′ we can pick a function
φ ∈ Cc∞ (U ′ ) such that 0 ≤ φ ≤ 1 in U ′ and φ (x) = 1 for all x ∈ K. Consider the function
g ∈ Cc1 (U ′ ; Rn ) defined by
∇ f (x)
g(x) = −φ (x) .
|∇ f (x)|
Since U ′ ⊂ U we have g ∈ Cc1 (U; Rn ). Noting also that and |g| = φ ≤ 1 in U we get
Z
S≥ f (x) div g(x) dx.
U
82 CHAPTER 8. MORE EXAMPLES OF RADON MEASURES

An integration by parts on the right implies

Z Z
S≥− ∇ f (x) · g(x) dx = φ (x)|∇ f (x)| dx.
U U

Hence in particular Z
S≥ |∇ f (x)| dx.
K
We have shown nZ o
′
S ≥ sup |∇ f (x)| dx K ⊂ U , K compact .
K
It remains to apply Theorem 2.3.8 to get
nZ o Z Z
′
sup |∇ f (x)| dx K ⊂ U , K compact = |∇ f (x)| dx = |∇ f (x)| dx,
K U′ U

the last equality being trivial.

Since the right hand side of (8.3.1) makes sense if f is merely locally L n -summable on
U, defining the quantity
Z nZ o
n 1 n
|D f | := sup f div g dL g ∈ Cc (V ; R ), |g| ≤ 1 in V , V ⊆ U,
V V

we introduce:

Definition 8.3.1. A function f which is locally L n -summable

R
on U, is said to be of locally
bounded variation in U, written f ∈ BVloc (U), provided V |D f | < ∞ for each V ⋐ U.

Definition 8.3.2. A function f which is RL n -summable on U, is said to be of bounded varia-

tion in U, written f ∈ BV (U), provided U |D f | < ∞.

Theorem 8.3.3. For any f ∈ BVloc (U), there exists a Radon measure µ on U and a µ-
measurable function σ : U → Rn such that

(i) |σ (x)| = 1 for µ-a.e. x ∈ U, and

f div g dL n = − dµ for each g ∈ Cc1 (U; Rn ).

R R
(ii) U U g·σ

Proof.
...
Chapter 9

Elementary convexity

9.1 Convex functions

Definition 9.1.1 (convex function). Let 0/ ̸= Ω ⊆ Rn be convex. A function f : Ω → R is
called convex in Ω if


f λ x + (1 − λ )y ≤ λ f (x) + (1 − λ ) f (y) ∀ λ ∈ [0, 1], ∀ x, y ∈ Ω. (9.1.1)

In case the reverse inequality holds true in (9.1.1), then f is called concave.
Remark 9.1.2. Let −∞ ≤ a < b ≤ ∞. The mean value theorem shows that if f : (a, b) → R
has nonnegative second derivative at any point of (a, b), then f is convex in (a, b).
Remark 9.1.3. Using induction we obtain for any k ∈ N the equivalent inequality
 k  k
f λ x ≤
∑ i i ∑ λi f (xi),
i=1 i=1

valid for all λi ∈ [0, 1], i = 1, ..., k, satisfying ∑ki=1 λi = 1, and all xi ∈ Ω, i = 1, ..., k. For
instance, for k = 3 we have
!
 λ λ 
2 3
f (λ1 x1 + λ2 x2 + λ3 x3 ) = f λ1 x1 + (λ2 + λ3 ) x2 + x3
λ2 + λ3 λ2 + λ3
 λ λ3 
2
≤ λ1 f (x1 ) + (λ2 + λ3 ) f x2 + x3
λ2 + λ3 λ2 + λ3
 λ λ3 
2
≤ λ1 f (x1 ) + (λ2 + λ3 ) f (x2 ) + f (x3 )
λ2 + λ3 λ2 + λ3
≤ λ1 f (x1 ) + λ2 f (x2 ) + λ3 f (x3 ).

83
84 CHAPTER 9. ELEMENTARY CONVEXITY

Remark 9.1.4. The arithmetic-geometric inequality of lemma 3.0.5 follows from the above
remark by taking k = n, λi = 1/n, xi = log ai and f (t) = et , t ∈ R.
Remark 9.1.5. To show the claim made in the proof of theorem 5.1.9; that is,

∀ ε > 0, ∃ Cε > 0 such that |a + b| p − |a| p ≤ ε|a| p +Cε |b| p ∀ a, b ∈ R,

we observe first the function (0, ∞) ∋ t 7→ t p is convex in case p > 1. Hence, for A, B ∈ R and
λ ∈ (0, 1),
 |A| |B|  p
p p
|A + B| ≤ (|A| + |B|) = λ + (1 − λ ) ≤ λ 1−p |A| p + (1 − λ )1−p |B| p .
λ 1−λ

Taking A = a, B = b and λ = (1 + ε)1/(1−p) we get

|a + b| p − |a| p ≤ ε|a| p +Cε |b| p .

On the other hand, if ε < 1 then taking A = a + b, B = −b and λ = (1 − ε)1/(p−1) we get

|a| p − |a + b| p ≤ ε|a| p +Cε |b| p . (9.1.2)

Since (9.1.2) is also true if ε ≥ 1, the claim follows. For 0 < p ≤ 1 observe instead

p
|a| + |b| ≤ |a| p + |b| p .

Exercise 9.1.6. Let A ⊂ Rn and denote its indicator function by IA ; that is

0 if x ∈ A,

IA (x) :=
∞ if else.

Prove this is a convex function if and only if A is a convex set.

Exercise 9.1.7. (i) Prove using the Brunn-Minkowski inequality of theorem 3.0.3, the

1/n Rn
function L n (·) : 2 → [0, ∞] is concave with respect to the Minkowski addition;
that is, for any λ ∈ [0, 1] and all X,Y ⊆ Rn there holds

1/n
1/n
1/n
L n λ X + (1 − λ )Y ≥ λ L n (X) + (1 − λ ) L n (Y ) .

1−λ
(ii) Prove L n λ X + (1 − λ )Y ≥ L n (X) L n (Y )


λ
.

(iii) Prove that if the inequality of (ii) is known to be true for all λ ∈ [0, 1] and all X,Y ⊆ Rn ,
then it implies the Brunn-Minkowski inequality.
9.1. CONVEX FUNCTIONS 85

Exercise 9.1.8. Let f : [0, ∞) → [0, ∞) be concave. Prove

(i) g : (0, ∞) → [0, ∞) defined by g(x) = f (x)/x is decreasing.

(ii) f is sub-additive; that is f (x + y) ≤ f (x) + f (y) for all x, y ≥ 0.

Example 9.1.9. Let Ω ⊊ Rn be nonempty and convex.

(i) If Ω is closed, then the distance function to the set Ω is convex in Rn .

Proof. Denote by dΩ the distance to the set Ω; that is,

dΩ (x) := dist(x, Ω) = inf |x − y|, x ∈ Rn .

y∈Ω

Let x, y ∈ Rn with x ̸= y and choose ξx , ξy ∈ Ω such that dΩ (x) = |x − ξx | and dΩ (y) =

|y − ξy |. Let λ ∈ (0, 1) and consider the point λ x + (1 − λ )y. The convexity of Ω
implies that λ ξx + (1 − λ )ξy ∈ Ω and we have



dΩ λ x + (1 − λ )y ≤ λ x + (1 − λ )y − λ ξx + (1 − λ )ξy
≤ λ |x − ξx | + (1 − λ )|y − ξy |
≤ λ dΩ (x) + (1 − λ )dΩ (y),

where we have used the triangle inequality to get to the middle line.

(ii) If Ω is open, then the distance function to ∂ Ω is concave in Ω.

Proof. Denote by d∂ Ω (x) the distance of x ∈ Ω to the set ∂ Ω; that is,

d∂ Ω (x) := dist(x, ∂ Ω) = inf |x − y|, x ∈ Ω.

y∈∂ Ω

Let x, y ∈ Ω with x ̸= y and choose ξx , ξy ∈ ∂ Ω such that d∂ Ω (x) = |x − ξx | =: rx and

d∂ Ω (y) = |y − ξy | =: ry . Let λ ∈ (0, 1).
Claim. The open ball with center zλ := λ x + (1 − λ )y and radius rλ := λ rx + (1 − λ )ry
is contained in Ω.
Proof of claim. Given z ∈ Brλ (zλ ) consider the points
z − zλ z − zλ
zx := x + rx and zy := y + ry .
rλ rλ

Then zx ∈ Brx (x), zy ∈ Bry (y) and also λ zx + (1 − λ )zy = z. Thus z is in the Minkowski
sum of the sets λ Brx (x) and (1 − λ )Bry (y). Since Brx (x), Bry (y) ⊂ Ω and Ω is convex,
we conclude z ∈ Ω. □
86 CHAPTER 9. ELEMENTARY CONVEXITY

Since Brλ (zλ ) ⊂ Ω we readily have

d∂ Ω (zλ ) ≥ d∂ Br (zλ ) (zλ ) = rλ .
λ


Thus d∂ Ω λ x + (1 − λ )y ≥ λ d∂ Ω (x) + (1 − λ )d∂ Ω (y) as required.
Lemma 9.1.10. Suppose that f : R → R is convex. Then
∀ x ∈ R, ∃ m ∈ R such that f (y) ≥ f (x) + m(y − x) ∀ y ∈ R.
Proof. Let 0 < λ < κ. Then
κ − λ λ  κ −λ λ
f (x + λ ) = f x + (x + κ) ≤ f (x) + f (x + κ).
κ κ κ κ
This implies
f (x + λ ) − f (x) f (x + κ) − f (x)
≤ , (9.1.3)
λ κ
and so
f (x + λ ) − f (x) f (y) − f (x)
f ′ (x+) := lim ≤ ∀ y > x. (9.1.4)
λ ↓0 λ y−x
Similarly,
κ − λ λ  κ −λ λ
f (x − λ ) = f x + (x − κ) ≤ f (x) + f (x − κ).
κ κ κ κ
This implies
f (x − λ ) − f (x) f (x) − f (x − κ)
≥ , (9.1.5)
−λ κ
and so
f (x − λ ) − f (x) f (x) − f (y)
f ′ (x−) := lim ≥ ∀ y < x. (9.1.6)
λ ↓0 −λ x−y
Also, for λ > 0 we have
x − λ x + λ  1 1
f (x) = f + ≤ f (x − λ ) + f (x + λ ).
2 2 2 2
This implies
f (x − λ ) − f (x) f (x + λ ) − f (x)
≤ ,
−λ λ
and so
f ′ (x−) ≤ f ′ (x+) ∀ x ∈ R. (9.1.7)
Because of (9.1.7), given x ∈ Rn we may pick m ∈ [ f ′ (x−), f ′ (x+)]. Apply then (9.1.4) and
(9.1.6) to obtain the inequality of the lemma.

A byproduct of estimates presented in the above proof is

9.1. CONVEX FUNCTIONS 87

Lemma 9.1.11. Suppose that f : R → R is convex. Then

f (x2 ) − f (x1 ) f (x3 ) − f (x1 ) f (x3 ) − f (x2 )

≤ ≤ ∀ x1 < x2 < x3 .
x2 − x1 x3 − x1 x3 − x2

Proof. From (9.1.3) setting x = x1 , x + λ = x2 and x + κ = x3 we deduce the inequality on

the left. From (9.1.5) setting x − κ = x1 , x − λ = x2 and x = x3 we deduce the inequality on
the right.

Theorem 9.1.12 (Jensen’s inequality). Suppose that f : R → R is convex. Let µ be a mea-

sure on a set X satisfying 0 < µ(X) < ∞. If u : X → R̄ is µ-summable, then
 1 Z  1
Z
f u dµ ≤ f (u) dµ.
µ(X) X µ(X) X

Proof. Taking
1
Z
x := u dµ, y := u(z), z ∈ X,
µ(X) X

in lemma 9.1.10, we get

 1 Z   1
Z 
f u(z) ≥ f u dµ + m u(z) − u dµ µ-a.e. z ∈ X.
µ(X) X µ(X) X

The proof now follows by taking mean values on X with respect to µ on both sides.

Exercise 9.1.13. Let µ be a measure on the set X and µ(X) = 1. Let u : X → [0, ∞] be
µ-summable.
R
(i) If Xu dµ = 1, prove
Z
u log u dµ ≥ 0.
X

(ii) Prove that for all p ≥ 1 there holds

1 Z p
R 1
≤ u dµ .
X up dµ X

(iii) Prove r Z 2 Z p Z
1+ u dµ ≤ 1 + u2 dµ ≤ 1 + u dµ.
X X X
88 CHAPTER 9. ELEMENTARY CONVEXITY

9.2 Lipschitz functions

Definition 9.2.1 (Lipschitz function). Let A ⊂ Rn . A function f : A → R is called Lipschitz
continuous, or simply Lipschitz, if there exists a constant L > 0 such that

| f (x) − f (y)| ≤ L|x − y| ∀ x, y ∈ A.

Remark 9.2.2. A Lipschitz function is uniformly continuous in A.

Example 9.2.3. Two examples of Lipschitz functions are
(i) the distance function dF to a nonempty closed subset F ⊊ Rn ; that is,

dF (x) := inf |x − y|, x ∈ Rn . (9.2.1)

y∈F

Proof. Let x, y ∈ Rn and choose ξx ∈ F so that dF (x) := |x − ξx |. Then

dF (y) − dF (x) ≤ |y − ξx | − |x − ξx | ≤ |y − x|.

Chose now ξy ∈ F so that dF (y) := |y − ξy |. Then in the same fashion

dF (x) − dF (y) ≤ |x − y|.

These estimates show dF is Lipschitz continuous with L = 1.

(ii) real functions in C1 (Ω) having uniformly bounded derivative in Ω, with Ω being any
open and convex subset of Rn .
Proof. Let f : Ω → R, f ∈ C1 (Ω), for which there exists M > 0 such that

|∇ f (x)| ≤ M ∀ x ∈ Ω.

Let x, y ∈ Ω and consider the straight line γ(t) := x + t(y − x), t ∈ [0, 1]. Applying

the
mean value theorem to the function u : [0, 1] → R defined by u(t) := f γ(t) , we get
t0 ∈ (0, 1) such that u′ (t0 ) = u(1) − u(0). By the chain rule we deduce


D f γ(t0 ) · (y − x) = f (y) − f (x).

Coupling this with the bound on the derivative of f we conclude that f is Lipschitz
with L = M.
Exercise 9.2.4. Let Ω ⊊ Rn be open and suppose that f : Ω̄ → R is Lipschitz with f = 0 on
∂ Ω. Prove that there exists a positive constant C such that

| f (x)| ≤ Cd∂ Ω (x) ∀ x ∈ Ω̄.

9.2. LIPSCHITZ FUNCTIONS 89

Definition 9.2.5 (locally Lipschitz function). Let A ⊂ Rn . A function f : A → R is called

locally Lipschitz continuous, or simply locally Lipschitz, if for any compact set K ⊆ A there
exists a constant CK > 0 such that

| f (x) − f (y)| ≤ CK |x − y| ∀ x, y ∈ K.

Remark 9.2.6. A locally Lipschitz function is continuous.

Example 9.2.7. Any convex function f : Ω → R, where Ω ⊆ Rn is open and convex, is locally
Lipschitz.

Proof. The proof is done by induction on the space dimension n ∈ N:

For n = 1, let K ̸= 0/ be a compact subset of Ω. By the convexity of Ω and proposition
1.1.17 we can choose connected X,Y ⊂ Ω such that K ◦ ⋐ X ⋐ Y ⋐ Ω. Since n = 1 we have
that X = (b, c) and Y = (a, d) where a < b < c < d. Thus for any x, y ∈ K with x < y we apply
lemma 9.1.11 two times to get

f (b) − f (a) f (y) − f (x) f (d) − f (c)

C1 := ≤ ≤ =: C2 .
b−a y−x d −c

This in turn implies

| f (y) − f (x)|
≤ max{|C1 |, |C2 |} for all x, y ∈ K, x ̸= y.
|y − x|

Now assume the result is true for n = k − 1, k ∈ N \ {1}. Let f : Ω → R be a convex

function, where Ω is an open and convex subset of Rk . Given a compact K ⊂ Ω we choose as
above connected X,Y ⊂ Ω such that K ◦ ⋐ X ⋐ Y ⋐ Ω. Recalling exercise 1.1.23 we replace
X and Y by X ′ ⊃ X and Y ′ ⊃ Y respectively, where each one of X̄ ′ and Y¯′ is a union of a finite
number of nonoverlapping intervals of Rk and such that K ◦ ⋐ X ′ ⋐ Y ′ ⋐ Ω. The line segment
connecting two given distinct points x, y ∈ K is extended in the direction of x and hits ∂ X ′
at b and ∂Y ′ at a. It is also extended in the direction of y and hits ∂ X ′ at c and ∂Y ′ at d.
The function f restricted on the part of this line which lies in Ω is a one dimensional convex
function and as before we have the estimate
| f (y) − f (x)| n | f (b) − f (a)| | f (d) − f (c)| o
≤ max , .
|y − x| |b − a| |d − c|

The right hand side is always finite under the above construction. Indeed, the function

| f (y) − f (x)|
F(x, y) := x ∈ ∂ X ′ , y ∈ ∂Y ′ ,
|y − x|
90 CHAPTER 9. ELEMENTARY CONVEXITY

is well defined because ∂ X ′ and ∂Y ′ are of positive distance. Furthermore, ∂ X ′ (resp. ∂Y ′ )

is a finite union of parts of the k − 1-dimensional faces of those nonoverlapping intervals of
X ′ (resp. Y ′ ) that contribute to ∂ X ′ (resp. ∂Y ′ ). Clearly, f is convex on any such face and
by the induction hypothesis and remark 9.2.6, f is continuous on any such face. Thus f is
continuous on ∂ X ′ and on ∂Y ′ which implies in turn that F is continuous on ∂ X ′ × ∂Y ′ . Since
∂ X ′ × ∂Y ′ is compact, F attains its maximum there (theorem 1.2.5).

For instructive reasons we provide a second proof for the statement in the above example.
It is based on the following fact1
Lemma 9.2.8. Let Q be an interval in Rn and y1 , . . . , y2n the vertices. Then every point x ∈ Q
can be written as a convex linear combination of the vertices, i.e. there exist λi ∈ [0, 1],
i = 1, . . . , 2n , such that
2n 2n
x = ∑ λi yi , ∑ λi = 1.
i=1 i=1
Proof. The proof is done by induction on the space dimension n ∈ N:
For n = 1 the statement is trivial. Suppose now the claim holds true for n − 1. Let
x = (x1 , ..., xn ) ∈ Q = [l11 , l12 ] × [l21 , l22 ] × ... × [ln1 , ln2 ]. Further, we denote by A the set of all
vertices having their last coordinate equal to ln1 and by B the set of all vertices having their
last coordinate equal to ln2 . Observe that A and B are disjoint, card(A) = card(B) = 2n−1 and
their union is the whole set of vertices of Q. Hence we can define a bijection
(
A→B
ϕ:
(z1 , z2 , ..., zn−1 , ln1 ) → (z1 , z2 , ..., zn−1 , ln2 ).

Naturally, we obtain for each ai ∈ A, i = 1, ..., 2n−1 , a curve γi (t) = tai + (1 − t)ϕ(ai ) for
t ∈ [0, 1]. Furthermore, let H = {z = (z1 , ..., zn ) ∈ Rn zn = xn } be a hyperplane. Since γi is
continuous for each i = 1, ..., 2n−1 and all curves are identical in the last coordinate, we obtain
a t ∗ such that the last coordinate of γi (t ∗ ) is equal to xn for all i. Next we observe the points
{γ1 (t ∗ ), ..., γ2n−1 (t ∗ )} lie in the hyperplane H and generate an interval in this hyperplane. By
our induction hypotheses we can write the point (x1 , ..., xn−1 ) as a convex linear combination
of the points {γ1 (t ∗ ), ..., γ2n−1 (t ∗ )} restricted to the hyperplane H. But since the linear com-
bination is convex and {γ1 (t ∗ ), ..., γ2n−1 (t ∗ )} lie in the hyperplane H, we are allowed to write
(x1 , ..., xn ) as a linear combination of {γ1 (t ∗ ), ..., γ2n−1 (t ∗ )} in Rn . Finally, observe that the
points {γ1 (t ∗ ), ..., γ2n−1 (t ∗ )} are convex linear combinations of our 2n vertices.

Second proof for the statement in example 9.2.7. Let Ql (y0 ) ⊆ Ω be a cube of side length
l > 0, center at y0 and vertices y1 , y2 , ..., y2n . Set V := {y1 , y2 , ..., y2n }. If x ∈ Ql (y0 ) then
1 this is a primitive version of the Krein-Milman theorem in Functional Analysis; see [Br, Theorem 1.13]
9.3. A CHARACTERIZATION OF CONVEX FUNCTIONS 91
n
from the previous lemma we know there exists λi ∈ [0, 1], i = 1, ..., 2n , with ∑2k=1 λk = 1 and
n
x = ∑2k=1 λk yk . The convexity of f implies through exercise 9.1.3 that

2n
f (x) ≤ ∑ λk f (yk ) ≤ max | f | < ∞.
V
k=1

Thus M := supQl f < ∞. Since y0 = (1/2)x + (1/2)(2y0 − x) and 2y0 − x ∈ Ql (y0 ) (why?),
applying the convexity property of f gives

1 1 1 1
f (y0 ) ≤ f (x) + f (2y0 − x) ≤ f (x) + M.
2 2 2 2
This gives infQl f ≥ 2 f (y0 ) − M. Hence f is bounded in Ql . It follows that f is locally
bounded in Ω (why?). Next let B̄3r (x0 ) ⊂ Ω and x, y be two distinct points in B̄r (x0 ). Select
µ > 0 and z ∈ ∂ B2r (x) so that
z − x = µ(y − x).
Then µ = 2r/|y − x| > 1 and y = (1/µ)z + (1 − 1/µ)x. Hence, the convexity of f gives

f (y) ≤ (1/µ) f (z) + (1 − 1/µ) f (x) (9.2.2)

≤ f (x) + (1/µ) | f (z)| + | f (x)|
| f (z)| + | f (x)|
= f (x) + |y − x|
2r
≤ f (x) + L|y − x|,

where L = (1/r) supB̄3r | f |. Interchanging the roles of x and y we also have

f (x) ≤ f (y) + L|y − x|.

Hence f is Lipschitz in B̄r . It follows that f is locally Lipschitz in Ω (why?).

9.3 A characterization of convex functions

Theorem 9.3.1. Let Ω ⊆ Rn be nonempty and convex. A function f : Ω → R is convex if and
only if it is continuous and satisfies the inequality in the definition of convexity for λ = 1/2
only 2 ; that is,
x + y 1 1
f ≤ f (x) + f (y) ∀ x, y ∈ Ω. (9.3.1)
2 2 2
2 or any other value of λ in the interval (0, 1)
92 CHAPTER 9. ELEMENTARY CONVEXITY

Proof. If f is convex then (9.3.1) is true. Also, by the above example and remark we get f is
continuous. To prove the other direction let x, y ∈ Ω and define γ : [0, 1] → Ω by
γ(λ ) := λ x + (1 − λ )y, λ ∈ [0, 1].
Define the sequence of sets
Dk := γ(λ ) λ ∈ {0/2k , 1/2k , ..., 2k /2k } , k ∈ N.



Claim. There holds f γ(λ ) ≤ λ f (x) + (1 − λ ) f (y) for all γ(λ ) ∈ Dk .
Accepting the claim for a moment, we can use k∈N Dk is dense in {γ(λ ) : λ ∈ 
S
[0, 1]}
and the continuity of u to deduce (9.1.1). Indeed, for any λ ∈ [0, 1], there exists λk ∈
{0/2k , 1/2k , ..., 2k /2k } k∈N such that limk→∞ λk = λ . Then limk→∞ γ(λk ) = γ(λ ) and the



continuity of f implies through theorem 1.2.3 that f γ(λ ) = limk→∞ f γ(λk ) . From the
claim we further obtain



f γ(λ ) ≤ lim λk f (x) + (1 − λk ) f (y) ,
k→∞

which implies (9.1.1).

Proof of claim. This is done by induction. For k = 1 we have
D1 = {γ(0) = x, γ(1/2) = (x + y)/2, γ(1) = y}.
The claim is thus trivially true for λ = 0, 1 and also true for λ = 1/2 by the hypothesis (9.3.1).
Assuming next that (9.3.1) is true for all γ(λ ) ∈ Dk , we prove it is true for all γ(λ ) ∈ Dk+1 .
To this end let γ(λ ) ∈ Dk+1 \ Dk . Setting
λ− = λ − 1/2k+1 and λ+ = λ + 1/2k+1 ,
we readily have γ(λ− ), γ(λ+ ) ∈ Dk . Therefore, by the induction hypothesis,


f γ(λ− ) ≤ λ− f (x) + (1 − λ− ) f (y),


f γ(λ+ ) ≤ λ+ f (x) + (1 − λ+ ) f (y).
Adding these we get
1
1

f γ(λ− ) + f γ(λ+ ) ≤ λ f (x) + (1 − λ ) f (y),
2 2
and because of (9.3.1), we conclude
 γ(λ ) + γ(λ ) 
− +
f ≤ λ f (x) + (1 − λ ) f (y).
2
The proof follows since γ(λ− ) + γ(λ+ ) = 2γ(λ ).
9.3. A CHARACTERIZATION OF CONVEX FUNCTIONS 93

Example 9.3.2. The function f : Rn → R given by f (x) = |x|2 − dF2 (x), where dF is the
distance function to a nonempty closed subset F ⊊ Rn , is convex.

Proof. From example 9.2.3-(i) we know f is continuous, hence because of the above theorem
we only have to prove f (x/2 + y/2) ≤ f (x)/2 + f (y)/2. To this end pick z ∈ F such that
x + y x+y
dF = z− .
2 2
Then by expanding the square
x + y
f = −|z|2 + z · (x + y).
2
On the other hand
f (x) + f (y) |x|2 + |y|2 − dF2 (x) − dF2 (y)
=
2 2
|x|2 + |y|2 − |x − z|2 − |y − z|2
≥
2
2
= −|z| + z · (x + y),

as required.
94 CHAPTER 9. ELEMENTARY CONVEXITY
Bibliography

[Br] Brezis, H. Functional analysis, Sobolev spaces and partial differential equations. Univer-
sitext. Springer 2011.

[DB] DiBenedetto, E. Real analysis. 2nd ed. Birkhäuser Adv. Texts. Basler Lehrbücher.
Birkhäuser 2016.

[EvG] Evans, L. C., Gariepy, R. F. Measure theory and fine properties of functions. Stud. Adv.
Math. CRC Press 1992.

[F] Federer, H. Geometric measure theory. Grundlehren Math. Wiss. 153, Springer 1969.

[Fr] Friedlander, F.G. Introduction to the theory of distributions. 2nd ed. with additional material
by M. Joshi. Cambridge Univ. Press 1998.

[KrP] Krantz, S., Parks H. The geometry of domains in space. Birkhäuser Adv. Texts. Basler
Lehrbücher. Birkhäuser 1999.

[LL] Lieb, E. H., Loss, M. Analysis. 2nd ed. Grad. Stud. Math. 14. Amer. Math. Soc. 2001.

[Mg] Maggi, F. Sets of finite perimeter and geometric variational problems. An introduction to
geometric measure theory. Cambridge Stud. Adv. Math. 135. Cambridge Univ. Press 2012.

[Mz] Maz’ya, V. G. Sobolev spaces. Translated from Russian by T. Shaposhnikova. Springer Ser.
Soviet Math., Springer 1985.

[Schn] Schneider, R. Convex Bodies: The Brunn-Minkowski theory. 2nd expanded ed. Encyclope-
dia Math. Appl. 151. Cambridge Univ. Press 2014.

[V] Villani, C. Topics in Optimal Transportation. 2nd ed. Grad. Stud. Math. 58. Amer. Math.
Soc. 2016.

[WhZ] Wheeden, R. L.; Zygmund, A. Measure and integral. An introduction to real analysis. Pure
Appl. Math. 43. Marcel Dekker 1977.

[W] Willem, M. Functional analysis. Fundamentals and applications. Cornerstones.

Birkhäuser/Springer 2013.

95
96 BIBLIOGRAPHY

[Z] Ziemer, W. P. Modern real analysis. 2nd ed. with contributions by Monica Torres. Grad.
Texts in Math. 278. Springer 2017.