Faculty of Mathematics

Institute for Analysis

JProf. Dr. Xian Liao
Englerstrasse 2, Geb. 20.30
76131 Karlsruhe
Telefon: `49 (0) 721/608-42616
Fax: `49 (0) 721/608-46530
E-Mail: [email protected]
Web: www.math.kit.edu/iana2/˜liao
Datum: February 17, 2023

Lecture notes for the lecture “Functional Analysis”

1 2
(Lecture 0104800, Winter Semester 2022/2023)

Instructor:
JProf. Xian Liao ([email protected])

Teaching assistant:
Ms. Rebekka Zimmermann ([email protected])

Time (weekly hours 4+2):

• Monday & Friday 11:30-13:00 (lecture)
• Wednesday 08:00-09:30 (problem class)
Place:
Seminar Room 1.067 (Mathematical Building 20.30)

Exam:
Writing exam (28.02.2023, 11am-13pm, Gaede Hörsaal)
1
Comments are welcome to be sent to me by email.
2
Updated version can be found here.

1 [February 17, 2023]

Contents
1 Introduction 4
1.1 Normed spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Sequence spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Continuous linear maps . . . . . . . . . . . . . . . . . . . . . . 9

2 Hilbert spaces 13
2.1 Definition and first properties . . . . . . . . . . . . . . . . . . 13
2.1.1 Inner products and (pre-)Hilbert spaces . . . . . . . . . 13
2.1.2 Parallelogram identity . . . . . . . . . . . . . . . . . . 15
2.1.3 Projection and orthogonality . . . . . . . . . . . . . . . 16
2.2 Orthonormality . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.1 Orthonormal system and Bessel’s inequality . . . . . . 19
2.2.2 Orthonormal basis and Parseval’s equality . . . . . . . 21
2.3 The Riesz representation theorem . . . . . . . . . . . . . . . . 24

3 Lebesgue spaces 26
3.1 Review of measure spaces . . . . . . . . . . . . . . . . . . . . 26
3.1.1 Measure spaces . . . . . . . . . . . . . . . . . . . . . . 26
3.1.2 Measurable and integrable functions . . . . . . . . . . 28
3.2 Preliminary inequalities . . . . . . . . . . . . . . . . . . . . . 31
3.2.1 Jensen’s inequality . . . . . . . . . . . . . . . . . . . . 32
3.2.2 Hölder’s inequality . . . . . . . . . . . . . . . . . . . . 33
3.2.3 Minkowski’s inequality . . . . . . . . . . . . . . . . . . 34
3.2.4 Hanner’s inquality . . . . . . . . . . . . . . . . . . . . 37
3.3 Lebesgue spaces Lp pµq . . . . . . . . . . . . . . . . . . . . . . 39
3.3.1 Fischer-Riesz theorem . . . . . . . . . . . . . . . . . . 39
3.3.2 Projection and duality . . . . . . . . . . . . . . . . . . 40
3.3.3 Young’s inequality on Rd . . . . . . . . . . . . . . . . . 42
3.3.4 Schur’s lemma . . . . . . . . . . . . . . . . . . . . . . . 45
3.4 Radon measures on metric spaces . . . . . . . . . . . . . . . . 47
3.4.1 Radon measure . . . . . . . . . . . . . . . . . . . . . . 48
3.4.2 Density of Lip.-functions with comp. supp. in Lp pµq . . 50
3.4.3 Arzela-Ascoli theorem and separability of Lp pµq . . . . 52
3.4.4 Compact sets in Lp pRd q . . . . . . . . . . . . . . . . . 53
3.4.5 Riesz representation theorem for Cc pXq . . . . . . . . . 55
3.4.6 Radon-Nikodym Theorem . . . . . . . . . . . . . . . . 61

4 Distributions and Sobolev spaces 64

2 [February 17, 2023]

 4.1 Baire category theorem, Banach-Steinhaus theorem and open
mapping theorem . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.2.2 Products, derivatives and supports . . . . . . . . . . . 71
4.2.3 Convolution in Rd and density of DpU q in D1 pU q . . . . 74
4.3 Schwartz functions and tempered distributions . . . . . . . . . 75
4.3.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3.2 Fourier transformation . . . . . . . . . . . . . . . . . . 77
4.4 Sobolev spaces W k,p . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.2 Extension and traces . . . . . . . . . . . . . . . . . . . 82
4.4.3 Poincaré inequality and Sobolev inequality . . . . . . . 86

5 Hahn-Banach Theorem, Reflexivity and Weak Topology 88

5.1 Hahn-Banach Theorems . . . . . . . . . . . . . . . . . . . . . 89
5.1.1 Hahn-Banach theorems . . . . . . . . . . . . . . . . . . 89
5.1.2 Consequences of Hahn-Banach theorem . . . . . . . . . 93
5.2 Reflexivity, weak and weak-˚ topology . . . . . . . . . . . . . 95
5.2.1 Reflexive spaces . . . . . . . . . . . . . . . . . . . . . . 95
5.2.2 Separation theorems . . . . . . . . . . . . . . . . . . . 97
5.2.3 Weak topology . . . . . . . . . . . . . . . . . . . . . . 98
5.2.4 Weak-* topology . . . . . . . . . . . . . . . . . . . . . 100
5.3 Introduction to spectral theory . . . . . . . . . . . . . . . . . 101
5.3.1 Compact operators . . . . . . . . . . . . . . . . . . . . 102
5.3.2 Spectrum of compact operators . . . . . . . . . . . . . 103

3 [February 17, 2023]

Main references:
• H. W. Alt, Linear functional analysis, Springer, 2016.

• H. Koch, PDE and functional analysis, lecture notes, WS2016/2017.

• P. Kunstmann, Functional analysis, lecture notes, WS2018.

• R. Schnaubelt, Functional analysis, lecture notes, WS2014.

• D. Werner, Funktionalanalysis, Springer, 2011.

[24.10.2022]

1 Introduction
In the Analysis courses we have dealt with real/complex-valued functions
f “ f pxq : Rn Ą Ω ÞÑ K, with K “ R or C, which are continuous in the sense
n
that for any x P Ω and ε ą 0, there exists δ ą 0 such that BδR pxq Ă Ω and
n
f pBδR pxqq Ă BεK pf pxqq.

Here the (open) ball Br pyq, r ą 0, y P Rn denotes the set of points whose
distance with y are smaller than r, either in Rn or in K. From now on we
take an abstract viewpoint and view functions f as points in a functional
space X:

f P X “ tf pxq | f satisfies some specific propertiesu.

In the lecture of Functional Analysis we are going to define and study var-
ious functional spaces, as well as the linear operators between them. In
this introduction part we will introduce Banach spaces and give some in-
tuitive examples, and then we will discuss Hilbert spaces, Lebesgue spaces,
distribution spaces, Sobolev spaces, etc. Functional analysis is built on the
structure of analysis and linear algebra, and it finds its applications to vast
areas of mathematics including probability, partial differential equations and
numerical analysis.
Before introducing Banach spaces, let us recall/introduce some abstract con-
cepts.
1. Topological space. A system τ of subsets of a set X defines a topology
in X if

4 [February 17, 2023]

 • H P τ, X P τ;
• If τ 1 Ă τ , then
Ť
U P τ;
U Pτ 1
• If U1 , U2 P τ , then U1 U2 P τ .
Ş

The sets in τ are called the open sets of the topological space pX, τ q.
We will often omit τ and simply refer to X as a topological space.
A topological space is called Hausdorf space, if in addition the following
separation axiom holds
• For any two points x1 , x2 P X with x1 ‰ x2 , there exist two open
sets U1 , U2 P τ such that x1 P U1 , x2 P U2 with U1 X U2 “ H.
A function f : X ÑÞ Y defined from a topological space X to another
topological space Y is called continuous if the inverse image of every
open set is open.
2. Metric space. A set X is called a metric space if there is a (distance)
function d : X ˆ X ÞÑ r0, 8q such that
• dpx, yq ě 0 and dpx, yq “ 0 iff x “ y;
• dpx, yq “ dpy, xq, @x, y P X;
• dpx, zq ď dpx, yq ` dpy, zq, @x, y, z P X (triangle inequality).
A metric space is called complete if each Cauchy sequence in pX, dq
converges in X. A sequence pxk qkPN is called a Cauchy sequence in
pX, dq if dpxk , xl q Ñ 0 for k, l Ñ 8. A sequence pxk qkPN is said to
converge to x in pX, dq if dpxk , xq Ñ 0 for k Ñ 8.
3. Induced topology by a metric. Let X be a metric space and let Br pxq “
ty P X | dpx, yq ă ru denote the ball centered at x P X with radius
r ą 0. For A Ă X we define its interior as

Å “ tx P X | Bε pxq Ă A for some ε ą 0upĂ Aq

and its closure as

Ā “ tx P X | Bε pxq X A ‰ H for all ε ą 0upĄ Aq.

Then pX, τ q, τ “ tA Ă X | Å “ Au is a topology space, and this

topology is induced by the metric. In particular all the balls of form
Br pxq are open sets, and pX, τ q is a Hausdorff space, since for all x, y P
X with x ‰ y and d :“ dpx, yq ą 0, the two open sets B d pxq and B d pyq
2 2
are disjoint.

5 [February 17, 2023]

 A map f : X ÞÑ Y between two metric spaces is continuous if for
any ball BεY pf pxqq, there exists a ball BδX pxq such that f pBδX pxqq Ă
BεY pf pxqq.

1.1 Normed spaces

We will focus on normed spaces, the most important class of topological
vector spaces.
Definition 1.1 (Normed spaces). Let X be a K vector space. A map }.} :
X Ñ r0, 8q is called norm if
• }x} “ 0 ðñ x “ 0;
• }x ` y} ď }x} ` }y}, @x, y P X (triangle inequality);
• }λx} “ |λ|}x}, @λ P K, x P X.
It is called normed space.
If pX, dq is a complete metric space, with the metric defined by dpx, yq “
}x ´ y}, then X is called a Banach space.
Remark 1.2. It is straightforward to check that dpx, yq “ }x´y} is a distance
function defined in X. But not all the metrics are induced by the norms, e.g.
the non-vector metric spaces.
Example 1.3 (Normed spaces). Exercise.
1. Rn and Cn equipped with the Euclidean norm are real resp. complex
Banach spaces.
2. Let X be a set. The space of bounded functions BpXq equipped with the
supremum norm is a Banach space.
3. Let pX, dq be a metric space. The space of bounded continuous functions
Cb pXq equipped with the supremum norm is a Banach space, or more
precisely a closed sub vector space of BpXq.
4. The space C0 pRn q Ă Cb pRn q of functions converging to 0 at 8 is a Ba-
nach space. Similarly the space c0 of sequences converging to 0 equipped
with the sup norm is a Banach space.
5. Let U Ă Rd be open. Cbk pU q is the vector space of k times differentiable
functions on U which are together with there derivatives bounded. The
norm
}u}C k pU q “ max |α|ďk }B α u}BpU q
turns Cbk into a Banach space.

6 [February 17, 2023]

 6. Let U Ă C be open. The space of bounded holomorphic functions
H 8 pU q is a Banach space when equipped with the supremum norm.

A closed subvector space of a Banach space is also a Banach space (Exer-

cise):

Lemma 1.4. Let X be a Banach space, and U Ă X be a closed subvector

space. Then U is a Banach space.
Furthermore, X{U is a vector space,

}x̃}X{U “ inf }y ´ x}
yPU

defines a norm (here x̃ is the equivalence class of x) and X{U is a Banach

space.

1.2 Sequence spaces

We study now sequence spaces as examples.

Definition 1.5 (Sequence spaces). Let 1 ď p ď 8. A K sequence pxj qjPN is

called p summable if
}pxj q}lp ă 8,
where the expression }pxj q}lp is given by
˜ ¸1{p
8
ÿ
}pxj q}lp :“ |xj |p if p ă 8, }pxj q}l8 :“ sup jPN |xj | if p “ 8.
j“1

The set of all p summable sequences is denoted by lp pNq “ lp .

Lemma 1.6 (Hölder’s inequality). If p1 ` 1q “ 1, 1 ď p, q ď 8 and pxj q P lp ,

pyj q P lq , then pxj yj q is summable and Hölder’s inequality holds:
ˇ ˇ
ˇÿ 8 ˇ ÿ 8
ˇ xj y j ˇ ď |xj yj | ď }pxj q}lp }pyj q}lq .
ˇ ˇ
ˇj“1 ˇ j“1

Proof. It holds obviously if p “ 1 or q “ 1.

For p1 ` 1q “ 1, 1 ă p, q ă 8, we recall Young’s inequality

1 1
|xy| ď |x|p ` |y|q
p q

7 [February 17, 2023]

for all x, y P R. Without loss of generality we assume x, y ą 0 and this can
be proven by searching the maximum of
1
x Ñ xy ´ xp
p
for y ą 0 which is attained at x0 “ y 1{pp´1q :
1 p ´ 1 p´1
p 1
x0 y ´ xp0 “ y “ yq .
p p q
As a consequence
ÿ ÿ1 1 1 1
|xj yj | ď |xj |p ` |yj |q “ }pxj q}plp ` }pyj q}qlq
j j
p q p q

and we obtain Hölder’s inequality

ÿ ÿ
|xj yj | “}pxk q}lp }pyk q}lq |xj |}pxk q}´1 ´1
lp |yj |}pyk q}lq
j j
1 1
ď}pxk q}lp }pyk q}lq p ` q
p q
“}pxj q}lp }pyj q}lq .

[24.10.2022]
[28.10.2022]

Theorem 1.7. For all 1 ď p ď 8, plp , } ¨ }lp q are Banach spaces.

Remark 1.8. We may replace N by Z, by a finite set, or even an arbi-
trary set. Then l8 pXq “ BpXq. The triangle inequality is called Minkowski
inequality.
Proof. l8 pNq “ BpNq is showed to be Banach space in the exercise.
We show first the triangle inequality. It is obvious if p “ 1. For 1 ă p ă 8
8
ÿ 8
ÿ
p
|xj ` yj | “ |xj ` yj |p´1 |xj ` yj |
j“1 j“1
ÿ8 8
ÿ
p´1
ď |xj ` yj | |xj | ` |xj ` yj |p´1 |yj |
j“1 j“1

p´1
´ ¯ 1 1
ď}p|xj ` yj | q}lq }pxj q}lp ` }pyj q}lp , ` “1
p q
´ ¯
p´1
“}pxj ` yj q}lp }pxqj }lp ` }pyj q}lp ,

8 [February 17, 2023]

and hence
}pxj ` yj q}lp ď }pxj q}lp ` }pyj q}lp ,
provided we can divide both sides by }pxj ` yj q}lp . There is nothing to show
if this quantity is 0. It is finite whenever we sum over a finite number of
indices, and hence a limit argument completes the triangle inequality.
It is easy to check that lp is a vector space. One easily sees that }pxj q}lp “ 0
implies pxj q “ 0 and
}pλxj q}lp “ |λ|}pxj q}lp .
Thus the spaces lp are normed vector spaces.
Now suppose that xn “ pxn,j q is a Cauchy sequence in lp : }pxn,j ´xm,j q}lp Ñ 0
as n, m Ñ 8. Then for every j, n Ñ xn,j is a Cauchy sequence in K. Let
yj “ limnÑ8 xn,j and y “ pyj q. Then, for every m ą 1 (assuming p ă 8)
N
ÿ
}y ´ xm }plp “ lim |yj ´ xm,j |p
N Ñ8
j“1
N
ÿ
“ lim lim |xn,j ´ xm,j |p
N Ñ8 nÑ8
j“1

ď lim }xn ´ xm }plp

nÑ8
Ñ 0 as m Ñ 8.

y “ pyj q is the limit of the Cauchy sequence pxn,j q in lp .

Lemma 1.9. Let X and Y be normed spaces. Their direct sum X ‘ Y p“

X ˆ Y q is a vector space. If 1 ď p ď 8 then

}px, yq}p “ }p|x|X , |y|Y q}lp

defines a norm with which X ‘ Y becomes a Banach space.

Proof. Exercise.

1.3 Continuous linear maps

We can define continuous maps between two topological spaces or two metric
spaces. Between two normed spaces we can define continuous linear maps.

Definition 1.10 (Continuous linear maps). Let X and Y be normed spaces.

We define LpX, Y q as the set of all continuous linear maps from X to Y .

9 [February 17, 2023]

Remark 1.11. Let X be a normed space, then addition, scalar multiplication
and the map to the norm are continuous. But the map to the norm is not
linear. (Exercise)
For two normed spaces X, Y , LpX, yq is a vector space with the obvious ad-
dition and multiplication. (Exercise)
Theorem 1.12. Let X, Y be two normed spaces.
For any continuous linear map T : X Ñ Y ,

}T }XÑY :“ sup }x}X ď1 }T pxq}Y ă 8

and }.}XÑY defines a norm on LpX, Y q.

For any linear operator S : X Ñ Y , S is continuous if and only if }S}XÑY
is finite.
If Y is a Banach space, then LpX, Y q is a Banach space.
Remark 1.13. For the linear map T : X ÞÑ Y ,
• }T x}Y ď }T }XÞÑY }x}X for all x P X;

• if }T x}Y ď a}x}X for all x P X and some a ă 8, then }T }XÞÑY ď a.

Proof. Let T : X Ñ Y be a continuous linear map. Then T 0 “ 0. For any
ball B1 p0q in Y , there exists δ ą 0 so that

}T x}Y ă 1 if }x}X ă δ.
› 1 ›
› p 2 δqx ›
In particular, if x P X, x ‰ 0, then › }x} X
› “ 12 δ ă δ and
X

}x}X ›› 12 δx ›› }x}X
}T x}Y “ 1 ›T › ă 1
2
δ }x}X Y 2
δ

and thus
}T }XÑY ď 2δ ´1 .
It is straightforward to check that }}XÞÑY defines a norm in LpX, Y q.
Vice versa: Let S : X Ñ Y be linear so that }S}XÑY ă 8. If }S}XÞÑY “ 0,
then S “ 0 and it is obviously continuous. If }S}XÞÑY ą 0, then for any ε ą 0
we choose δ “ ε{}S}XÑY . Then for any y P BδX pxq, we have Sy P BεY pSxq:

}Sx ´ Sy}Y “ }Spx ´ yq}Y ď }S}XÑY }x ´ y}X ă ε.

In particular S is uniformly continuous.

Now assume that Y is a Banach space. The idea of the proof of the com-
pleteness of LpX, Y q is the same as in the proof of Theorem 1.7 above. Let

10 [February 17, 2023]

Tn P LpX, Y q be a Cauchy sequence. For any x, Tn x is a Cauchy sequence
in Y since
}Tn x ´ Tm x}Y ď }Tn ´ Tm }XÑY }x}X .
Let
T x :“ lim Tn x.
nÑ8

The convergence is uniform on bounded sets, and hence the limit T is con-
tinuous and in LpX, Y q. Moreover

}T ´ Tn }XÑY “ sup }x}X ď1 }pT ´ Tn qx}Y “sup }x}X ď1 lim sup }pTm ´ Tn qx}Y
mÑ8
ď lim sup }Tm ´ Tn }XÑY Ñ 0
mÑ8

as n Ñ 8. Here we used continuity of addition and the map to the norm.

[28.10.2022]
[31.10.2022]

Definition 1.14. Two norms }.} and |.| on a normed space X are called
equivalent, if there exits C ě 1 so that for all x P X

C ´1 }x} ď |x| ď C}x}.

Remark 1.15. If the two norms }}1 and }}2 as well as the two norms }}2
and }}3 are equivalent on a normed space X, then the two norms }}1 and }}3
are also equivalent.

Theorem 1.16. All norms on finite dimensional normed spaces are equiva-
lent. Finite dimensional normed spaces are Banach spaces.

Proof. We first prove the case X “ Kd . Let |.| be the Euclidean norm on Kd
and }.} a second norm. Let tej uj“1,¨¨¨ ,d be the standard basis. Then
ˇ ˇ
›ÿ d › ÿ d ? ˇÿ d ˇ
aj e j › ď |aj |max k }ek } ď p dmax k }ek }q ˇ aj ej ˇ .
› › ˇ ˇ
›
j“1 j“1
ˇ j“1
ˇ

Thus v Ñ }.} is continuous with respect to |.|. On the Euclidean unit sphere
(which is compact), the continuous function } ¨ } : Kd ÞÑ p0, 8q attains its
infimum, which is positive and we denote it by λ´1 . Then

|v| “ |v||v{|v|| ď λ|v|}v{|v|} “ λ}v}.

11 [February 17, 2023]

The two inequalities imply the equivalence of the norms }.} and |.| by choosing
␣? (
C “ max dmax }ek }, λ, 1 .

Thus every norm on Kd is equivalent to the Euclidean norm, and any two
norms are equivalent. A Cauchy sequence vm “ pvm,j q with respect to }.}
is also a Cauchy sequence with respect to the Euclidean norm, hence it
converges to a vector v with respect to | ¨ |, and hence also respect to } ¨ }.
This proves the claim for Kd .
Now let X be a K-vector spaces of dimension d. Then there is a basisřd of d
d
vectors tvj uj“1,¨¨¨ ,d , and a bijective linear map ϕ : K ÞÑ X maps j“1 aj ej
to dj“1 aj vj . If }.}X is a norm on X then x Ñ }ϕpxq}X is a norm on Kd .
ř
Thus the first part follows. Since ϕpxn q is a Cauchy sequence with respect to
}.}X iff pxn q is a Cauchy sequence in Kd with respect to the second metric,
and one converges iff the second converges.This completes the proof.
Remark 1.17. The norms on infinite dimensional normed spaces are not
necessarily equivalent. For example, the norm }}BpU q and the norm }}C 1 pU q
are not equivalent on Cb1 pU q.
Definition 1.18 (Dual space). Let X be a normed space. We define the dual
(Banach) space as X ˚ “ LpX, Kq.
Example 1.19. Let X “ Rd with the Euclidean norm. For any y P Rd , we
can define a continuous linear map
d
ÿ
d
Ty : R ÞÑ R via x ÞÑ xj y j .
j“1

The linear map

ϕ : Rd Q y Ñ Ty P pRd q˚
is
• isometric: |y| “ }Ty }Rd ÞÑR ;

• injective: ϕ´1 pt0uq “ t0u;

• surjective: For any T : Rd ÞÑ R, there exists y “ pT e1 , ¨ ¨ ¨ , T ed q P Rd

such that ϕpyq “ T .
It allows to identify Rd and pRd q˚ .
Question: Can we identify lq and plp q˚ , p1 ` 1q “ 1? In next chapter we will
answer it when p “ q “ 2, and we will answer the general case later.

12 [February 17, 2023]

2 Hilbert spaces
We introduce in this chapter Hilbert spaces, which are equipped with inner
products and are Banach spaces with the induced norms. With the structure
of inner products, we can introduce the notion of orthogonality between
two elements. The Hilbert projection theorem implies the existence of the
projection operator from a Hilbert space to its closed subspace. Thanks to
this fact, the celebrated Riesz representation theorem identifies a Hilbert
space with its dual space. As a variant, the Lax-Milgram theorem becomes
powerful in showing the existence of weak solutions of some partial differential
equations of elliptic type.

2.1 Definition and first properties

2.1.1 Inner products and (pre-)Hilbert spaces
Definition 2.1. Let X be a K vector space. A map x., .y : X ˆ X Ñ K is
called inner product if
• xλ1 x1 ` λ2 x2 , yy “ λ1 xx1 , yy ` λ2 xx2 , yy, @x1 , x2 , y P X, @λ1 , λ2 P K;
• xx, yy “ xy, xy @x, y P X;
• xx, xy ě 0, @x P X and xx, xy “ 0 iff x “ 0.
The vector space with inner product pX, x¨, ¨yq is called pre-Hilbert space.
Remark 2.2. It is straightforward to derive from the first and second axioms
that
• xx, λyy “ λ̄xx, yy;
• xx ` y, x ` yy “ xx, xy ` 2Re xx, yy ` xy, yy.
Example 2.3. 1. Euclidean vector spaces over K, Euclidean inner prod-
uct.
Real and complex square summable sequences space l2 pNq with xpxj q, pyj qy “
2. ř
xj y j .
´
3. Let U Ă Rn be an open set, xf, gy “ U f gdx defines an inner product
in Cb pU q.
Lemma 2.4 (Cauchy-Schwarz inequality). Let X be a pre-Hilbert space.
Then 1
|xx, yy| ď pxx, xyxy, yyq 2
for all x, y P X.

13 [February 17, 2023]

Proof. Let x, y P X and λ P K. Then
0 ďxx ´ λy, x ´ λyy
“xx, xy ´ λxy, xy ´ λ̄xx, yy ` |λ|2 xy, yy.
xx,yy
If y “ 0 there is nothing to show, so we assume y ‰ 0 and define λ “ xy,yy
.
Then
|xx, yy|2
0 ď xx, xy ´
xy, yy
which implies the Cauchy-Schwarz inequality.
Lemma 2.5. Let X be a pre-Hilbert space. The map
a
x Ñ }x} :“ xx, xy

defines a norm on X.
With this notation the Cauchy-Schwarz inequality becomes

|xx, yy| ď }x}}y}.

Proof. Clearly }x} ě 0, }x} “ 0 iff x “ 0. Moreover

}λx}2 “ |λ|2 }x}2

and by the Cauchy-Schwarz inequality

}x ` y}2 “}x}2 ` xx, yy ` xy, xy ` }y}2
ď}x}2 ` 2}x}}y} ` }y}2
“p}x} ` }y}q2 .

[31.10.2022]
[04.11.2022]

Lemma 2.6. Let X be a pre-Hilbert space. Then the inner product defines
a continuous map from X ˆ X to K.
Proof. Exercise.
Definition 2.7. A pre-Hilbert space is a Hilbert space if it is a Banach space
with respect to the metric-induced norm.
Example: Rd , Cd , l2 pNq are all Hilbert space. But Cb pU q is not a Hilbert
space, although it is Banach space with respect to the sup-norm.

14 [February 17, 2023]

2.1.2 Parallelogram identity
It is not hard to verify that the induced norm by the inner product x¨, ¨y
satisfies the parallelogram identity

}x ` y}2 ` }x ´ y}2 “ 2}x}2 ` 2}y}2 (2.1)

for x, y P X some pre-Hilbert space. On the other side, if the norm in a

normed space satisfies the parallelogram identity, we can define the inner
product correspondingly.

Theorem 2.8 (Jordan von Neumann). Let X be a normed K vector space

with norm }.} satisfying the parallelogram identity (2.1). Then
1´ ¯
xx, yy :“ }x ` y}2 ´ }x ´ y}2 (2.2)
4
if K “ R or
3
1ÿ k 1´ ¯
xx, yy :“ i }x ` ik y}2 “ }x ` y}2 ´ }x ´ y}2 ` i}x ` iy}2 ´ i}x ´ iy}2
4 k“0 4
(2.3)
if K “ C defines an inner product on X, such that the norm }.} is the norm
of the pre-Hilbert space. Vice versa: The norm of a pre-Hilbert space satisfies
the parallelogram identity.

Proof. We begin with a real normed spaces whose norm satisfies the paral-
lelogram identity. We define the inner product by (2.2), then

xx, yy “ xy, xy.

Since by the parallelogram identity

}x ` y ` z}2 “ 2}x ` z}2 ` 2}y}2 ´ }x ´ y ` z}2

“ 2}y ` z}2 ` 2}x}2 ´ }y ´ x ` z}2 ,

we derive
1 1
}x ` y ` z}2 “ }x}2 ` }y}2 ` }x ` z}2 ` }y ` z}2 ´ }x ´ y ` z}2 ´ }y ´ x ` z}2 ,
2 2
and hence by changing z into ´z,
1 1
}x ` y ´ z}2 “ }x}2 ` }y}2 ` }x ´ z}2 ` }y ´ z}2 ´ }x ´ y ´ z}2 ´ }y ´ x ´ z}2 .
2 2

15 [February 17, 2023]

We arrive at
1
xx ` y, zy “ p}x ` y ` z}2 ´ }x ` y ´ z}2 q
4
1 1
“ p}x ` z}2 ´ }x ´ z}2 q ` p}y ` z}2 ´ }y ´ z}2 q
4 4
“xx, zy ` xy, zy.

We claim
xλx, yy “ λxx, yy, @x, y P X, @λ P R .
It obviously holds for λ “ ˘1, and hence for all λ P Z by the previous step.
Thus it holds for all rational λ and, by continuity of the map λ Ñ }λx ˘ y},
for λ P R.
We complete the proof for complex Hilbert spaces: We define
3
1ÿ k
xx, yy “ i }x ` ik y}2
4 k“0

and observe that xix, yy “ ixx, yy, xx, yy “ xy, xy by definition, Rexx, yy is the
previous real inner product and Imxx, yy “ Rexx, iyy. The above argument
implies xx ` y, zy “ xx, zy ` xy, zy.

Remark 2.9. We could define a Hilbert spaces as a Banach space whose

norm satisfies the paralellogram identity. By an abuse of notation we call a
normed space pre-Hilbert space if it satisfies the parallelogram identity.

Corollary 2.10. A normed space is a pre-Hilbert space if and only if all two
dimensional subspaces are pre-Hilbert spaces.

Proof. It is a pre-Hilbert space if and only if its norm satisfies the parallelo-
gram identity which holds if and only if the parallelogram identity holds for
all two dimensional subspaces.

2.1.3 Projection and orthogonality

Recall that a set A in a metric space is called closed if Ā “ A, and is called
compact if any sequence in A has a subsequence which converges to a limit
in A 3 . A set B is called convex if it holds tx ` p1 ´ tqy P B for all x, y P B
and all t P r0, 1s.
3
Equivalently, A is compact if any open cover of A contains a finite subcover, or if A
is totally bounded and complete. In particular, totally boundedness implies boundedness
and completeness implies closedness.

16 [February 17, 2023]

Lemma 2.11. Let H be a Hilbert space, K Ă H compact, and C Ă H closed
and convex, such that C and K disjoint. Then there exist x P K and y P C
so that
}x ´ y} “ dpC, Kq,
where the distance between two sets is defined as dpA, Bq “ inftdpx, yq : x P
A, y P Bu.
Proof. Let pxj q Ă K and pyj q Ă C be minimizing sequences. Since K is
compact there is a subsequence which we denote again by pxj q and x P K so
that xj Ñ x. We also denote the corresponding subsequence of pyj q again by
pyj q. By the triangle inequality
}x ´ yj } Ñ dpC, Kq.
Then
}yn ´ ym }2 “}px ´ yn q ´ px ´ ym q}2
“2}x ´ yn }2 ` 2}x ´ ym }2 ´ }2x ´ pyn ` ym q}2
ď2p}x ´ yn }2 ` }x ´ ym }2 q ´ 4d2 pC, Kq
Ñ0 as n, m Ñ 8
since by convexity 21 pyn ` ym q P C. Thus pyn q is a Cauchy sequence with
limit y P C. Moreover
dpC, Kq “ lim }x ´ yn } “ }x ´ y}.
nÑ8

Remark 2.12 (Definition of ppxq for x R C). In particular, if K “ txu in a

Hilbert space H, then the closest point in the close and convex subset C to x
is unique and we denote it by ppxq.
The uniqueness is a consequence of the proof of Lemma 2.11: If there are
two closest points y1 , y2 , then by denoting d :“ dpx, Cq,
}y1 ´ y2 }2 “ 2}x ´ y1 }2 ` 2}x ´ y2 }2 ´ }2x ´ py1 ` y2 q}2 ď 2d2 ` 2d2 ´ 4d2 “ 0.
Definition 2.13. Let H be a pre-Hilbert space.
We call two elements x, y P H orthogonal if xx, yy “ 0.
We call two subsets M, N in H orthogonal if xx, yy “ 0 for all x P M , y P N .
Remark 2.14. (Exercise) For any subset A in a pre-Hilbert space H, we
can define AK as
AK “ ty P H | xx, yy “ 0 @x P Au.
Then AK is a closed subspace of H.
If B is a closed subspace in a Hilbert space H, then pB K qK “ B.

17 [February 17, 2023]

 [04.11.2022]
[07.11.2022]

Theorem 2.15 (Hilbert projection theorem). Let H be a Hilbert space and

x P H.

1. Suppose that C is a closed and convex subset of H such that x R C.

Then for y P C,

y “ ppxq
ô Rexx ´ y, z ´ yy ď 0, @z P C. (2.4)

2. Suppose that C is a closed subspace such that x R C. Then for y P C,

y “ ppxq
ô xx ´ y, zy “ 0, @z P C. (2.5)

Moreover, in this case

}x}2 “ }x ´ ppxq}2 ` }ppxq}2 . (2.6)

Proof. If C is a closed and convex subset, then by the convexity of C,

y ` tpz ´ yq P C, @y, z P C, @t P r0, 1s.

If y “ ppxq, then

}x ´ y}2 ď }x ´ y ´ tpz ´ yq}2 “ }x ´ y}2 ´ 2t Rexx ´ y, z ´ yy ` t2 }z ´ y}2 ,

which implies (2.4). If conversely (2.4) holds, then for all z P C it holds

}x ´ z}2 “ }x ´ y ´ pz ´ yq}2 “ }x ´ y}2 ´ 2 Rexx ´ y, z ´ yy ` }z ´ y}2 ě }x ´ y}2 ,

and y “ ppxq is the unique closest point in C to x (by Remark 2.12).

If C is a closed subspace, then we claim that (2.4) and (2.5) are equivalent.
Obviously (2.5) implies (2.4). Conversely, in the case that C is a closed
subspace, then (2.4) implies Re xx ´ y, zy ď 0 for all z P C, and hence
Re xx ´ y, zy “ 0 for all z P C, and thus Im xx ´ y, zy “ 0 for all z P C, such
that the orthogonality relation (2.5) holds. In particular, (2.6) follows from
(2.5) with y “ z “ ppxq.
An intuitive example of Theorem 2.15 is H “ R2 with xx, yy “ x1 y1 ` x2 y2 :

18 [February 17, 2023]

 1. Suppose that C is a closed convex subset and y P C. We can take y
as the origin, and take the coordinate of x to be p0, x2 q with x2 ą 0.
Then

y “ ppxq ô z “ pz1 , z2 q, z2 ď 0, @z P C
ô Re xx ´ y, z ´ yy “ x2 z2 ď 0, @z P C.

2. If C is a closed subspace and is not t0u or R2 , then without loss of

generality we can take it to be the horizontal (real) axis. Then for
y “ py1 , 0q,

y “ ppxq ô x1 “ y1 ô xx ´ y, zy “ 0, @z “ pz1 , z2 q with z2 “ 0.

Remark 2.16 (Projection operator). If C is a closed subspace in a Hilbert

space H, then the operator
"
ppxq if x R C,
P : H ÞÑ H via P x “
x if x P C,

is an projection operator, i.e. P 2 “ P .

This projection operator P is linear and bounded, with the norm }P }HÞÑH “ 1
if C ‰ t0u. Indeed, the linearity of P follows from
yPC
(2.5) ô y “ P x.

We calculate

}P x}2 “ xP x, P xy “ xx, P xy ´ xx ´ P x, P xy “ xx, P xy ď }x}}P x},

which implies }P }HÞÑH ď 1. Obviously P z “ z on C, and hence }P } “ 1 if

C ‰ t0u. If C “ t0u, then P “ 0 is the zero operator in LpH, Hq.

2.2 Orthonormality
2.2.1 Orthonormal system and Bessel’s inequality
Definition 2.17. Let H be a pre-Hilbert space. We call a subset A Ă H
an orthogonal set if any two different vectors in A are orthogonal. If, in
addition, every vector x in A is of unit norm: }x} “ 1, this orthogonal set A
is called an orthonormal set.

19 [February 17, 2023]

Let pxj qj“1,¨¨¨ ,n be a linearly independent vectors in a pre-Hilbert space X.
Then by the Gram-Schmidt procedure we obtain an orthonormal system
pyj qj“1,¨¨¨ ,n such that xyk , yl y “ δkl , k, l “ 1, ¨ ¨ ¨ , n. More precisely, one can
perform the following procedure recursively
1 1
y1 “ x1 , ỹ2 “ x2 ´ xx2 , y1 yy1 , y2 “ ỹ2 .
}x1 } }ỹ2 }

We can take n “ 8.
[07.11.2022]
[11.11.2022]

Lemma 2.18 (Bessel inequality). Let pxj qjPN be an orthonormal system in

a pre-Hilbert space X. Then for any finite n and x P X,
n
ÿ n
ÿ
2
} xx, xj yxj } “ |xx, xj y|2 ď }x}2 ,
j“1 j“1

and n n
› ÿ ›2 ÿ
›x ´ xx, xj yxj › “ }x}2 ´ |xx, xj y|2 ě 0.
› ›
j“1 j“1
řn
is equivalent to }x}2 “ nj“1 |xx, xj y|2 .
ř
In particular, x “ j“1 xx, xj yxj

Proof. The first equality } nj“1 xx, xj yxj }2 “ nj“1 |xx, xj y|2 follows from the
ř ř
orthonormality.
For any λ1 , ¨ ¨ ¨ , λn P K, we have
› n
ÿ ›2 n
ÿ n
ÿ n
ÿ
2
›x ´ λj xj › “ }x} ´ xx, xj yλ̄j ´ xxj , xyλj ` |λj |2
› ›
j“1 j“1 j“1 j“1
ÿn n
ÿ
“ }x}2 ´ |xx, xj y|2 ` |xx, xj y ´ λj |2 .
j“1 j“1

The minimum is obtained at pλj q “ pxx, xj yq:

› n
ÿ ›2 n
ÿ
2
›x ´ xx, xj yxj › “ }x} ´ |xx, xj y|2 ě 0.
› ›
j“1 j“1

20 [February 17, 2023]

Example 2.19 (Legendre polynomials).
´1 Let X “ Cpr´1, 1s; Rq be equipped
with the inner product xf, gy “ ´1 f gdx. All the monomials xn in the vector
space X are linearly independent. We define the Legendre polynomials by the
Rodrigues’ formula
1 dn
Pn pxq “ rpx2 ´ 1qn s.
2n n! dxn
Show
´ 1 mthat Pn is the unique polynomial of degree n such that Pn p1q “ 1 and
´1
x Pn pxqdx “ 0 for all 0 ď m ă n. Show that the family of functions
´c 2n ` 1 ¯
Pn pxq
2 ně0

is an orthonormal system in X. Exercise.

2.2.2 Orthonormal basis and Parseval’s equality

Recall that a subset A of a metrix space X is called dense if its closure is X.
A metric space X is called separable, if there is a countable dense subset.
Example 2.20 (Countable sets and separable metric spaces). 1. N, Z and
QN are countable.
2. Subsets of countable sets are countable; the product of countable sets is
countable; If Xj are countable sets then there union is countable.
3. RN is separable since QN is countable and dense.
4. Cpr0, 1sq is separable by Weierstrass approximation theorem.
5. lp pNq, 1 ď p ă 8 is separable. Exercise.
Definition 2.21. An orthonormal set pxj qjPN of a Hilbert space is called
orthonormal basis if

xx, xj y “ 0, @j P N implies x “ 0. (2.7)

Theorem 2.22. The following properties are equivalent for a Hilbert space
H which is not finite dimensional.
(1) The space H is separable.
(2) There exists an orthonormal basis pxj qjPN such that the Parseval’s equal-
ity holds ÿ
}x}2 “ |xx, xj y|2 , @x P H.
jPN

21 [February 17, 2023]

(3) The exists a linear surjective isometry l2 Ñ H.

Proof. “p1q ñ p2q” Suppose that H is separable. Let pyn qnPN be a dense
sequence and let XN be the span of py1 , ¨ ¨ ¨ , yN q. The dimension of XN is at
most N , and we can use the Gram-Schmidt procedure to find an orthonormal
basis pxn q of XN . We do this recursively by increasing N . This leads to a
countable orthonormal sequence pxn qnPN so that its span is dense in H.
For any x P H, the map
N
ÿ N
ÿ
2 2
φ : N Ñ }x ´ xx, xj yxj } “ }x} ´ |xx, xj y|2
j“1 j“1

is monotonically decreasing. We observe that ifřx P CN :“ span ptxn unďN q,

then x “ n“1 xx, xn yxn , and if x R CN , then N
řN
n“1 xx, xn yxn “ pN x is the
closed point in CN to x (by the proof of Bessel inequality). Since the span
of pxn qnPN is dense, φpN q converges to 0:
N
ÿ
0 Ð distpx, span tpxj q1ďjďN uq2 “ }x ´ xx, xj yxj }2 “ φpN q,
j“1

that is,
N
ÿ
xx, xn yxn Ñ x in H as N Ñ 8.
n“1
ř8
This in turn implies x “ n“1 xx, xn yxn and
8
ÿ
2
}x} “ |xx, xn y|2 .
n“1

Hence pxn q is an orthonormal basis.

“p2q ñ p3q” Now suppose that pxn q is an orthonormal basis such that the
Parseval’s equality holds. We claim that the following map is well-defined
8
ÿ
2
ϕ : l Q pan q Ñ an xn P H.
n“1

Indeed, for M ą N , due to the orthonormality, the norm of

M
ÿ N
ÿ M
ÿ
an x n ´ an x n “ an x n
n“1 n“1 n“N `1

22 [February 17, 2023]

 bř
M 2
is given by n“N `1 |aj | . This implies that the partial sums are a Cauchy
sequence in H and the map ϕ : l2 ÞÑ H is well-defined, and is isometric (by
an approximation argument):
8
ÿ 8
ÿ 8
ÿ
xϕppan qq, ϕppbn qqyH “ x an x n , bn , xn y “ an bn “ xpan q, pbn qyl2 pNq .
n“1 n“1 n“1

The map is clearly linear. The surjectivity follows from Parseval’s identity:
For any x P H we can take an “ xx, xn y such that ϕppan qq “ x.
“p3q ñ p1q” It follows from the fact that l2 pNq is separable.
[11.11.2022]
[14.11.2022]

Remark 2.23. • An isometric (anti-)linear map ϕ between two normed

spaces X, Y is injective since if ϕpxq “ 0, then
}x}X “ }ϕpxq}Y “ 0 ñ x “ 0.

• A Hilbert space is either isomorphic (there exists an isometric surjective

linear map) to Rd resp. Cd , to l2 pNq, or it is not separable.
• One can show that if pxn q is an orthonormal basis, then the Parseval’s
identity holds. Vice versa: An orthonormal set such that Parseval’s
identity holds is an orthonormal basis. Exercise.
Example 2.24. Exercise.
• The space l2 pRq with inner product
ÿ
xf, gy “ f pxqgpxq
xPR

is not separable.
• Let Cpr0, 2πs; Cq be equipped with the inner product
ˆ 2π
1
xf, gy “ f gdx.
2π 0
Then the system peinx qnPZ is an orthonormal set, such that Parseval’s
identity holds on Cpr0, 2πs; Cq.
2
(The Lebesgue spacea L pr0, 2πs; Cqis the completeness inx of Cpr0, 2πs; Cq
under the norm x¨, ¨y, which is a Hilbert space and pe qnPZ is an or-
thonormal basis on it.ř In particular, fˆpnq :“ xf, einx y is called Fourier
coefficient of f , and nPZ fˆpnqeinx is called Fourier series of f .)

23 [February 17, 2023]

2.3 The Riesz representation theorem
By use of Hilbert projection theorem, we can prove the Riesz representation
theorem for Hilbert spaces. Recall the definition of dual spaces in Definition
1.18.
Theorem 2.25 (Riesz representation theorem). Let H be a Hilbert space.
Then
J : H Q x Ñ py Ñ xy, xyq P H ˚
is an antilinear isometric bijective map (isomorphism).
Proof. By the Cauchy-Schwarz inequality

}Jpxq}H ˚ “ sup }y}ď1 |xy, xy| ď }x}H

and the map is well defined and antilinear.

Since
}x}H }Jpxq}H ˚ ě |pJpxqqpxq| “ xx, xy “ }x}2H
we see that
}Jpxq}H ˚ ě }x}H .
Thus J is an isometry: }Jpxq}H ˚ “ }x}H . In particular J is injective.
To show that J is surjective we assume that x˚ P H ˚ and try to find x1 so
that x˚ “ Jpx1 q. Let
N “ ty P H : x˚ pyq “ 0u.
Then N is a closed subspace. Let p be the orthogonal projection to N as in
Theorem 2.15. If x˚ ‰ 0, then there exists y1 ‰ 0 in H such that x˚ py1 q ‰ 0.
We take y0 “ x˚ypy1 1 q P H such that x˚ py0 q “ 1 and define

x0 “ y0 ´ ppy0 q.

Then x˚ px0 q “ 1 and for all y P N by (2.5) xy, x0 y “ 0. Moreover, for any
x P H, x˚ px ´ x˚ pxqx0 q “ 0, and hence x ´ x˚ pxqx0 P N and by (2.5)

xx ´ x˚ pxqx0 , x0 y “ 0.

This implies
xx, x0 y “ xx˚ pxqx0 , x0 y “ x˚ pxq}x0 }2H ,
and hence
˚
A x0 E x0 x0
x pxq “ x, 2
“ Jp qpxq, i.e. x˚ “ Jpx1 q with x1 “ .
}x0 } }x0 }2 }x0 }2

24 [February 17, 2023]

Since l2 pNq is a Hilbert space, Riesz representation theorem ensures that we
can identify pl2 pNqq˚ and l2 pNq in the sense that there exists an isomorphism
between them.
The Lax-Milgram theorem can be viewed as a variant of the Riesz represen-
tation theorem.
Theorem 2.26 (Lax-Milgram). Let H be a Hilbert space and

Q : H ˆ H Q px, yq Ñ Qpx, yq P K

• linear in x, antilinear in y;

• bounded in the sense that there exists C ą 0 such that

|Qpx, yq| ď C}x}|y};

• coercive in the sense that there exists δ ą 0 so that

Re Qpx, xq ě δ}x}2 .

Then there exists a unique continuous linear map A : H Ñ H with continuous

inverse A´1 so that
Qpx, yq “ xAx, yy.
Moreover
}A}HÑH ď C, }A´1 }HÑH ď δ ´1 .
Proof. Exercise.

A little on operators between Hilbert spaces By use of Riesz repre-

sentation theorem, the following holds true (Exercise):
For any linear continuous map T P LpH1 , H2 q between two Hilbert spaces,
there exists a unique T ˚ P LpH2 , H1 q such that

xT x, yyH2 “ xx, T ˚ yyH1 .

The operator T ˚ is called the adjoint of T , and it satisfies

}T }H1 ÞÑH2 “ }T ˚ }H2 ÞÑH1 .

It is easy to see that T ˚˚ “ T . The situation will be more delicate when

the unbounded operators are concerned. Spectral theory is devoted to the
relevant study.
A map U P LpH1 , H2 q between two Hilbert spaces is called unitary, if

25 [February 17, 2023]

 • It is invertible: There exists V P LpH2 , H1 q (denoted by U ´1 ) such that
V U “ 1H1 and U V “ 1H2 ;

• It preserves the inner product: xU x, U yyH2 “ xx, yyH1 , @x, y P H1 .

Obviously the inverse of a unitary operator U P LpH1 , H2 q is also its adjoint

U ´1 “ U ˚ , since

xU x, yyH2 “ xU x, U U ´1 yyH2 “ xx, U ´1 yyH1 .

The composition of unitary operators is unitary, and in particular the set

of all unitary operator LpHq :“ LpH, Hq on the Hilbert space H is a group
(called unitary group U pHq).
[14.11.2022]
[18.11.2022]

3 Lebesgue spaces
3.1 Review of measure spaces
3.1.1 Measure spaces
Definition 3.1 (Measure space). Let X be a set. A family of subset A is
called a σ´algebra if

1. tu P A

2. A P A implies XzA P A
8
Ť
3. An P A implies An P A.
n“1

A map µ : A Ñ r0, 8s is called a measure if whenever An P A are pairwise

disjoint then
´ď 8 ¯ ÿ 8
µ An “ µpAn q.
n“1 n“1

The triple pX, A, µq is called a measure space.

A measure space pX, A, µq is called sigma finite if there exists a sequence of
8
Ť
measurable sets An of finite measure so that X “ An .
n“1
A measure µ for a measure space pX, A, µq is called complete, if the σ algebra
contains every subset of a set of measure zero.

26 [February 17, 2023]

Example 3.2. 1. X a set, A “ 2X the set of all subsets, and µpAq the
number of elements.

2. X a set, "A “ 2X the set of all subsets, x P X a fixed element, and

1 x P A,
δx pAq “ for A Ă X. δx is called the point measure or
0 x R A,
Dirac measure.

3. If pX, dq is a metric space (or pX, τ q is a topological space), then there

is a smallest σ algebra containing all open sets. It is called the Borel σ
algebra of X, and a Borel measure is a measure on the Borel sets.

4. X “ Rn , A the Borel sets, µ the Lebesgue measure restricted to the

Borel sets.

5. X “ Rn , A the Lebesgue sets, µ the Lebesgue measure.

Recall the construction of Lebesgue measures m “ mn on Rn :

Step 1. We define the measure of a coordinate rectangle as the product of

the sidelengths;

Step 2. We define the measure of a union of countably many pairwise disjoint

coordinate rectangles as the sum over the measures of the rectangles;

Step 3. We define the outer measure m˚ of a general set as the infimum of

all measures of coverings by unions of disjoint coordinate rectangles;

Step 4. All the sets A satisfying Caratheodory’s criterium: m˚ pBq “ m˚ pAX

Bq ` m˚ pB X Ac q, @B Ă Rn form a σ algebra M;

Step 5. The triple pRn , M, mq is a measure space, where mpAq “ m˚ pAq is

the Lebesgue measure, and A P M is a Lebesgue measurable set.

Measures are often constructed by restricting outer measures on Caratheodory

measurable sets, by the same procedure above.

Definition 3.3. • Let X be a set. A map µ : X ÞÑ r0, 8s is called an

outer measure if

1. µpt uq “ 0.
2. A Ă B implies µpAq ď µpBq.
8
´Ť ¯ ř
3. µ Aj ď 8 j“1 µpAj q, @Aj Ă X.
j“1

27 [February 17, 2023]

 • Let µ be an outer measure on X. We call a subset A Ă X Caratheodory
measurable if for all B Ă X
µpBq “ µpB X Aq ` µpB X pXzAqq.

• Let pX, dq be a metric space. We call µ an outer metric measure if it

is an outer measure which satisfies
µpA Y Bq “ µpAq ` µpBq, @A, B Ă X with distpA, Bq ą 0.
We state Caratheodory’s theorem which constructs a measure on the Caratheodory
measurable sets, without proof.
Theorem 3.4 (Caratheodory). Let µ be an outer measure on the set X.
Then the Caratheodory measurable sets C are a σ algebra and pX, C, µ|C q is
a measure space. Moreover C contains all sets of exterior measure 0.
If X is a metric space and µ is a metric outer measure, then C contains all
open sets and hence all Borel sets.
Example 3.5 (Hausdorff measure). Let pX, dq be a metric space and s ě 0.
We define for any r ą 0 and A Ă X,
8 8
!ÿ π s{2 ˇ ď )
Hrs pAq “ inf ´s
2 pdiamAn qsˇ
A n Ą A with diampA n q ă r .
Γp 2s ` 1q
ˇ
n“1 n“1

Hrs pAq is nonincreasing in r and we define

Hs pAq “ lim Hrs pAq.
rÑ0`

Then Hs is a metric outer measure, and by Caratheodory theorem its restric-

tion on Caratheodory measurable sets is a measure, which is called Hausdorff
measure.

3.1.2 Measurable and integrable functions

Definition 3.6. Let pX, A, µq be a measure space.
• A map f : X Ñ R Yt´8, 8u is called measurable if
f ´1 ppt, 8sq P A, @t P R .

• For a measurable function f : X Ñ r0, 8s, we define its Lebesgue

integral by the Riemann integral
ˆ ˆ 8
f dµ “ µpf ´1 ppt, 8sqqdt P r0, 8s,
X 0

28 [February 17, 2023]

 ´
and we call it integrable if X
f dµ ă 8.
We call a measurable function f : X ÞÑ R Yt´8, 8u integrable if |f |
is integrable, and we define
ˆ ˆ ˆ
`
f dµ “ f dµ ´ f ´ dµ ,
X X X

where f “ max tf, 0u denotes its positive part and f ´ “ max t´f, 0u
`

denotes its negative part.

We call a function f : X Ñ C integrable if the real and imaginary parts
are both integrable, and we define
ˆ ˆ ˆ
f dµ “ pRe f q dµ ` i pIm f q dµ .
X X X

• Let 1 ď p ă 8. We call a measurable function f p integrable if |f |p is

integrable and denote
ˆˆ ˙1{p
p
}f }Lp “ |f | dµ .
X

• We call a measurable function 8 integrable or essentially bounded if

there is a constant C so that
µptx : |f pxq| ą Cuq “ 0.
The best constant is denoted by }f }L8 .
• We call two measurable functions equivalent (denoted by f „ g), if they
are the same almost everywhere (i.e. µptx P X|f pxq ‰ gpxquq “ 0).
We define Lp pµq as the set of equivalence classes of p integrable func-
tions.
[18.11.2022]
[21.11.2022]

Remark 3.7. We notice the following facts from the definitions:

• For measurable functions
´ f, g :´X ÞÑ R such that f ď g almost ev-
erywhere, it holds X f dµ ď X g dµ . It follows straightforward if
f, g : X ÞÑ r0, 8s,
ˆ ˆ 8 ˆ 8 ˆ
f dµ “ µptf ą tuqdt ď µptg ą tuqdt “ g dµ .
X 0 0 X
`
For general f, g : X ÞÑ R with f ď g, one notices f ď g ` and
f ´ ě g´.

29 [February 17, 2023]

• The scale multiplication commutes with the Lebesgue integral. It holds
obviously true for λ ą 0 and nonnegative integrable function h : X ÞÑ
r0, 8s by use of change of variables:
ˆ ˆ 8 ˆ 8 ˆ
1 1
λ hdµ “ λ µpth ą tuqdt “ µpth ą t {λuq dt “ λh dµ .
X 0 0 X

Other cases follow correspondingly.

• If µpXq ă 8, then for any integrable function f : X ÞÑ C and complex

number a,
ˆ ˆ
pf ` aq dµ “ f dµ ` aµpXq.
X X

Indeed, it is easy to check with f : X ÞÑ r0, 8s and a ě 0,

ˆ ˆ 8 ˆ 8
pf ` aq dµ “ µptf ` a ą tuqdt “ µptf ą t ´ auqdt
X 0 0
ˆ 8 ˆ 0 ˆ
“ µptf ą tuqdt ` µptf ą tuqdt “ f dµ ` aµpXq.
0 ´a X

Other cases follow correspondingly.

• For any measurable function f : X ÞÑ C, it holds

ˇˆ ˇ ˆ
ˇ f dµ ˇ ď |f | dµ .
ˇ ˇ
X X

It holds obviously for real-valued functions:

ˇˆ ˇ ˇˆ ˆ ˇ "ˆ ˆ * ˆ
` ´ ` ´
ˇ f dµ ˇ “ ˇ f dµ ´ f dµ ˇ ď max f dµ , f dµ ď |f | dµ .
ˇ ˇ ˇ ˇ
X X X X X X

For complex-valued function f : X ÞÑ C, there exists a complex number

λ with modulus 1 such that
ˇˆ ˇ ˆ ˆ ˆ ˆ
ˇ f dµ ˇ “ λ f dµ “ pλf q dµ “ Re pλf q dµ “ Re pλf q dµ
ˇ ˇ
X ˆ X X ˆ Xˆ X

ď |Re pλf q| dµ ď |λf | dµ “ |f | dµ .

X X X

It is easy to see from the above that the equality implies f “ λ|f | for
some complex number λ with modulus 1.

30 [February 17, 2023]

We recall the convergence theorems about the relation between the limit of
integrals, and the integral over limits.

Theorem 3.8 (Fatou’s lemma). Let pX, A, µq be a measure space, and fn :

X ÞÑ r0, 8s, n P N be measurable functions. Then
ˆ ˆ
lim inf fn dµ ď lim inf fn dµ .
X nÑ8 nÑ8 X

If furthermore pfn qnPN converges to a measurable function f : X ÞÑ r0, 8s

almost everywhere, then
ˆ ˆ ˆ
f dµ ď lim inf fn dµ ď sup nPN fn dµ .
X nÑ8 X X

Theorem 3.9 (Lebesgue’s dominated convergence theorem). Let pX, A, µq

be a measure space, and fn , f : X ÞÑ R Yt´8, 8u, n P N be measurable
functions, such that fn Ñ f almost everywhere. Let g : X ÞÑ r0, 8s be an
integrable function, such that |fn | ď g for all n P N almost everywhere. Then
ˆ ˆ
lim fn dµ “ f dµ .
nÑ8 X X

Theorem 3.10 (Beppo Levi’s monotone convergence theorem). Let pX, A, µq

be a measure space, and fn : X ÞÑ r0, 8s, n P N be measurable functions, such
that fn ď fn`1 almost everywhere. Then there exists a measurable function
f : X ÞÑ r0, 8s such that fn converges to f almost everywhere, and
ˆ ˆ
lim fn dµ “ f dµ .
nÑ8 X X

Theorem 3.11 (Fubini-Tonelli). Let pX, A, µq and pY, B, νq be σ-finite mea-

sure spaces, A ˆ B the product σ algebra and µ ˆ ν the product measure. Let
f be ´µ ˆ ν integrable. Then for almost of x P X, y Ñ f px, yq is ν integrable,
x Ñ Y f px, yqdνpyq is µ integrable and
ˆ ˆ ˆ
f px, yqdµ ˆ ν “ f px, yqdνpyqdµpxq.
XˆY X Y

3.2 Preliminary inequalities

In this section we review some preliminary inequalities for p-integrable func-
tions.

31 [February 17, 2023]

3.2.1 Jensen’s inequality
The convex functions F : R ÞÑ R have left and right derivatives, and hence
there are at most countably many points where F is not differentiable.

Lemma 3.12. Suppose that f : R Ñ R is convex. Then both one sided

derivatives exist and if x ă y then

df ` df ´ df `
pxq ď pyq ď pyq
dx dy dy
and for all z P R,
! df ` df ´ )
f pzq ě max f pxq ` pxqpz ´ xq, f pxq ` pxqpz ´ xq .
dx dx
Proof. If x0 ă x1 ă x2 , then

f px1 q ´ f px0 q f px2 q ´ f px0 q f px2 q ´ f px1 q

ď ď .
x1 ´ x0 x2 ´ x0 x2 ´ x 1
Hence
f px ` hq ´ f pxq
hÑ
h
is monotonically increasing. This implies the differentiability from the right,
and similarly from the left and the relation between the derivatives.
[21.11.2022]
[25.11.2022]

We have Jensen’s inequality for the composition of convex functions and

integrable functions.

Theorem 3.13 (Jensen’s inequality). Let pX, A, µq be a measure space,

µpXq “ 1, F : R Ñ R convex and f : X ÞÑ R integrable. Then F ˝ f
is measurable and ˆ ˆ
F˝ f dµ ď F ˝ f dµ .
X X

Proof. Since F : R Ñ R is convex, F is continuous and hence F ˝ f is

measurable.
´
Let t0 “ X f dµ. By Lemma 3.12:

dF `
F ptq ě F pt0 q ` pt0 qpt ´ t0 q,
dt

32 [February 17, 2023]

it holds
ˆ ˆ ´
dF ` ¯
F ˝ f dµ ě F pt0 q ` pt0 qpf ´ t0 q dµ
dt
X X
` ˆˆ ˙
dF
“F pt0 q ` pt0 q f dµ ´ t0 “ F pt0 q.
dt X

Remark 3.14. For a particular case where X “ tx, yu, µptxuq “ t, µptyuq “
1 ´ t, for some t P r0, 1s, Jensen’s inequality reads as
´ ¯
F tf pxq ` p1 ´ tqf pyq ď tF pf pxqq ` p1 ´ tqF pf pyqq,

which follows also immediately from the definition of convexity.

3.2.2 Hölder’s inequality

We prove the Hölder’s inequality for general p-integrable functions. Recall
Lemma 1.6 for the case of lp , where pX, A, µq “ pN, 2N , counting measureq.

Lemma 3.15 (Hölder’s inequality). Let pX, A, µq be a measure space. Let

1 ď p, q ď 8 and p1 ` 1q “ 1. If f P Lp pµq and g P Lq pµq, then f g is integrable
and ˇˆ ˇ
ˇ ˇ
ˇ f gdµˇ ď }f }Lp }g}Lq . (3.1)
ˇ ˇ
X

If 1 ă p ă 8, f ‰ 0 and g ‰ 0, then the equality holds if and only if

“ λ|f |p´2 f¯}f }Lp
g ´pp´1q
}g}Lq
almost everywhere for some λ P K with |λ| “ 1 (in
p´2 ¯
particular if g “ λ|f | f ).

Proof. We follow the proof for Lemma 1.6 in the sequence space. If }f }Lp “ 0,
then f “ 0 almost everywhere, and the inequality (3.1) is trivial. The same
holds for }g}Lq “ 0. In the following we assume f ‰ 0 and g ‰ 0.
If pp, qq “ p1, 8q, then |gpxq| ď }g}L8 “: a ą 0 almost everywhere and the
inequality (3.1) follows:
ˆ ˆ 8 ˆ 8
|f g|dµ “ µpt|f g| ą tuqdt ď µpt|f | ą t{auqdt
X 0
ˆ 8
0
ˆ
“a µpt|f | ą tuqdt “ a |f |dµ.
0 X

It is the same for pp, qq “ p8, 1q.

33 [February 17, 2023]

Now show the inequality for the case 1 ă p, q ă 8. It suffices to consider f
and g with }f }Lp “ 1 and }g}Lq “ 1, and prove
ˆ ˆ ´
1 p 1 q¯
|f ||g|dµ ď |f | ` |g| dµ “ 1.
X X p q
It follows immediately from Young’s inequality and the inequality is strict
unless
1 1
|f gpxq| “ |f pxq|p ` |gpxq|q
p q
almost everywhere, which implies |g| “ |f |p´1 almost everywhere. We have
already shown ˇˆ ˇ ˆ
ˇ ˇ
ˇ f gdµˇ ď |f g|dµ ď }f }Lp }g}Lq
ˇ ˇ
X X

and in the case of equality all inequalities must be equalities:

f g “ λ|f g|, |g| “ |f |p´1

almost everywhere for some complex number λ of modulus 1, by Remark

3.7 and the above consideration. This implies f g “ λ|f |p and hence g “
λ|f |p´2 f¯. Back to our consideration, the equality implies g̃ “ λ|f˜|p´2 f˜ for
f˜ “ }f }fLp and g̃ “ }g}gLq , that is, g “ λ|f |p´2 f¯}g}Lq }f }Lp
´pp´1q
. This is also
sufficient condition for the equality in (3.1).

3.2.3 Minkowski’s inequality

By use of Hölder’s inequality one can show Minkowski’s inequality.

Theorem 3.16 (Minkowski’s inequality). Let 1 ď p ă 8 and let pX, A, µq

and pY, B, νq be σ finite measurable spaces. Let f be µ ˆ ν measurable. Then
ˆ ˆ ˇˆ ˇp ˙1{p ˆ ˆˆ ˙1{p
ˇ ˇ p
ˇ |f px, yq|dνpyqˇ dµpxq ď |f px, yq| dµpxq dνpyq.
ˇ ˇ
X Y Y X
(3.2)
If 1 ă p ă 8, the integrals above are finite, and
ˆ ˆ ˇˆ ˇp ˙1{p ˆ ˆˆ ˙1{p
ˇ ˇ p
ˇ f px, yqdνpyqˇ dµpxq “ |f px, yq| dµpxq dνpyq,
ˇ ˇ
X Y Y X
(3.3)
then there exist a µ-measurable function α and a ν-measurable nonnegative
function β so that
f px, yq “ αpxqβpyq

34 [February 17, 2023]

almost everywhere.
A special case is the triangle inequality (which assumes Y “ t0, 1u, ν counting
measure and no σ finiteness on X)
}f ` g}Lp pµq ď }f }Lp pµq ` }g}Lp pµq , (3.4)
whenever f and g are p-integrable, with equality for p ą 1 iff f and g are
positively linearly dependent, i.e. f “ λg for some λ ě 0 or g “ 0.
Proof. To show the inequality we assume first that f is nonnegative and omit
the absolute value. The case p “ 1 is trivial. We claim that
ˆ ˆ
p
Gpyq :“ f px, yqdµpxq and Hpxq :“ f px, yqdνpyq
X Y

are ν and µ measurable functions respectively. This follows from Fubini’s

Theorem 3.11 if f p resp f are µ ˆ ν integrable, and by an approximation
argument in the general case. More precisely, we may assume without loss
of generality X “ Yj Aj , Y “ Yk Bk with pAj q resp. pBk q pairwise disjoint
and of finite measure. Let fn “ f 1tf ďnu be monotone increasing sequence,
such that f “ limnÑ8 fn , and fn is integrable in the product measure space
´pAj ˆ Bk , pA ˆ Bqjk , pµ ˆ νqjk q restricted on Aj ˆ Bk , for any j, k. Hence
f px, yqdνpyq is µj -measurable on Aj and hence µ-measurable on X. Thus
´Bk n
f px, yqdνpyq is µ-measurable and hence Hpxq as the limit is µ-measurable.
Y n
[25.11.2022]
[28.11.2022]

Then if G1{p P L1 pνq and H P Lp pµq such that f H p´1 is µ ˆ ν integrable, the
Minkowski’s inequality (3.2) follows:
ˆ ˆ ˆ
p
H pxqdµpxq “ f px, yqdνpyqH p´1 pxqdµpxq
X ˆX ˆ Y
“ f px, yqH p´1 pxqdµpxqdνpyq
Y X
ˆ ˆˆ ˙1{p ˆˆ ˙ p´1
p
p p
ď f px, yqdµpxq H dµpxq dνpyq
Y X X
ˆ ˆˆ ˙ p´1
p
1{p p
“ pGpyqq dνpyq H dµpxq
´ˆ ˆ
Y X
¯1{p
ñ H p pxqdµpxq ď pGpyqq1{p dνpyq,
X Y
p
where we used Hölder’s inequality with q “ p´1
. An approximation argument
as above implies the general case.

35 [February 17, 2023]

Now assume that p ą 1, f is complex valued, such that the integrals in (3.2)
are finite. In order to have equality in the application of Hölder’s inequality
it holds for almost all x and y
ˆ
Hpxqp´1 ´ |f px, yq| ¯p´1
“ , where Hpxq “ |f px, yq|dνpyq,
}Hpxqp´1 } p´1p }f px, yq}Lpx Y
L

which implies
|f px, yq| “ α̃pxqβpyq
for some measurable nonnegative functions α̃ and β. In order to have f
instead of |f | on the left hand side of the equality (3.3), there exists λpxq
with |λ| “ 1 such that f px, yq “ λpxq|f px, yq| almost everywhere.
For the triangle inequality we apply the first part with the counting measure
on Y “ t0, 1u. The product measure is defined in the obvious fashion, even
without assuming that µ is σ finite. If f is p integrable then by the definition
of the integral
µptx : |f pxq| ą tuq ď t´p }f }pLp .
Indeed, for p “ 1, it follows from the definition that
ˆ ˆ 8 ˆ t0
|f | dµ “ µptx : |f pxq| ą tuqdt ě µptx : |f pxq| ą tuqdt
X
ˆ t0
0 0

ě µptx : |f pxq| ą t0 uqdt “ t0 µptx : |f pxq| ą t0 uq.

0

For general 1 ď p ă 8, we replace |f | by |f |p . Let

8 !
ď 1)
A“ x : |f pxq| ` |gpxq| ą
j“1
j

which is a countable union of sets of finite measure. We replace X by A,

take as σ algebra the sets in A which are contained in A, and µ restricted to
this σ algebra as measure. This is σ finite measure space.
Remark 3.17. If p “ 8, then the inequality (3.2) holds trivially since
ˆ ˆ
|f px, yq|dνpyq ď }f px, yq}L8
x pµq
dνpyq, µ ´ a.e.
Y Y

The Minkowski’s inequality (3.2) can be easily generalized to

› › › ›
› L pνq › q ,
› › › ›
}f } q
› L pνq › p ď }f } p
L pµq L pµq

whenever 1 ď q ď p ď 8. Indeed, we apply (3.2) with f , p replaced by |f |q ,

p{q ě 1 respectively.

36 [February 17, 2023]

3.2.4 Hanner’s inquality
There is an improvement of the triangle inequality.

Theorem 3.18 (Hanner’s inequality). Let pX, A, µq be a measure space and

f , g be p-integrable functions, 1 ď p ă 8. If 1 ď p ď 2 then
ˇ ˇp
}f ` g}pLp ` }f ´ g}pLp ě p}f }Lp ` }g}Lp qp ` ˇ}f }Lp ´ }g}Lp ˇ , (3.5)
ˇ ˇp
p}f ` g}Lp ` }f ´ g}Lp qp ` ˇ}f ` g}Lp ´ }f ´ g}Lp ˇ ď 2p p}f }pLp ` }g}pLp q. (3.6)
If 2 ď p ă 8 all inequalities are reversed.

Remark 3.19. The inequalities reduce to the parallelogram identity if p “ 2,

(recalling (2.1)).
Both (3.5) and (3.6) are equivalent: The second is obtained from the first by
replacing f by f ` g and g by f ´ g. It suffices to prove the first inequality.
If p “ 1, then (3.6) reads as

2max t}f ` g}L1 , }f ´ g}L1 u ď 2p}f }L1 ` }g}L1 q,

which follows from the triangle inequality (3.4).

Proof. As (3.6) follows from the triangle inequality (3.4), we restrict ourselves
to the case 1 ă p ă 8. We may assume that }g}Lp ď }f }Lp (otherwise we
exchange the two) and }f }Lp “ 1 (otherwise we multiply f and g by the
inverse of the norm }f }´1
Lp ).
The first inequality follows from the following pointwise inequality: Let

αprq “ p1 ` rqp´1 ` p1 ´ rqp´1 , βprq “ rp1 ` rqp´1 ´ p1 ´ rqp´1 sr1´p .

We claim that

αprq|f |p ` βprq|g|p ď |f ` g|p ` |f ´ g|p (3.7)

for 1 ă p ď 2, 0 ď r ď 1 and complex numbers f and g (and the reverse

inequality for 2 ď p ă 8). Indeed, (3.7) implies

αprq|f pxq|p ` βprq|gpxq|p ď |f pxq ` gpxq|p ` |f pxq ´ gpxq|p

and by integration

αprq}f }pLp ` βprq}g}pLp ď }f ` g}pLp ` }f ´ g}pLp .

37 [February 17, 2023]

We apply the inequality with r “ }g}Lp and recall that }f }Lp “ 1. The left
hand side becomes
” ı
p´1 p´1
p}f }L ` }g}L q
p p ` p}f }L ´ }g}L q
p p }f }Lp
” ı
` p}f }Lp ` }g}Lp qp´1 ´ p}f }Lp ´ }g}Lp qp´1 q }g}Lp
“p}f }Lp ` }g}Lp qp ` p}f }Lp ´ }g}Lp qp .
[28.11.2022]
[02.12.2022]

It remains to prove (3.7): If f, g are both real-valued,

αprq ` βprqRp ď |1 ` R|p ` |1 ´ R|p , @R ě 0,
if 1 ă p ď 2 (and the reverse inequality for 2 ď p ă 8). The proof is
elementary. Let for 0 ď R ď 1,
FR prq “ αprq ` βprqRp .
We claim that it attains its maximum at r “ R if 1 ď p ă 2 and resp. its
minimum if p ą 2. We compute
FR1 “ α1 ` β 1 Rp
“ pp ´ 1qrp1 ` rqp´2 ´ p1 ´ rqp´2 s ´ pp ´ 1qrp1 ` rqp´2 ´ p1 ´ rqp´2 sr´p Rp
“ pp ´ 1qrp1 ` rqp´2 ´ p1 ´ rqp´2 sp1 ´ pR{rqp q
and the derivative vanishes only at r “ R inside p0, 1q and changes sign there.
Thus
αprq ` βprqRp ď p1 ` Rqp ` p1 ´ Rqp
if 0 ď R ď 1 and p ď 2 with the opposite inequality if p ě 2. Now let R ě 1.
Since α ` βRp is increasing on r0, 1s if p ď 2, we obtain
αprq ` βprqRp ď 2p´1 p1 ` Rp q ď pR ` 1qp ` pR ´ 1qp
where the second inequality is ensured by the relation between the deriva-
tives: 2p´1 pRp´1 ď pppR ` 1qp´1 ` pR ´ 1qp´1 q which follows from 2p´1 ď
αp R1 q “ p1 ` R1 qp´1 ` p1 ´ R1 qp´1 . The reverse inequality holds if p ą 2. This
implies (3.7) for real-valued f and g. We claim that (3.7) holds for complex
f and g. It suffices to consider f “ a ą 0 and g “ beiθ , b ą 0 and it follows
from the fact that
|f ` g|p ` |f ´ g|p “ pa2 ` b2 ` 2ab cos θqp{2 ` pa2 ` b2 ´ 2ab cos θqp{2
has its minimum at θ “ 0 (resp. its maximum if p ě 2). It suffices to consider
the function hpρq “ pa2 ` b2 ` 2abρqp{2 ` pa2 ` b2 ´ 2abρqp{2 on ρ P r0, 1s, and
notice that x Ñ xp{2 is concave if p ď 2 (convex if p ě 2) and h1 p0q “ 0.

38 [February 17, 2023]

3.3 Lebesgue spaces Lp pµq
In this section we will always assume that pX, A, µq is a measurable space,
and we will study the Lebesgue spaces Lp pµq.

3.3.1 Fischer-Riesz theorem

Theorem 3.20 (Fischer-Riesz). pLp pµq, }}Lp q is Banach space.
Proof. The fact that pLp pµq, }}Lp q is a normed space follows from Definition
3.6, Remark 3.7 and the triangle inequality (3.4).
It remains to show completeness. Let pfn q be representatives of a Cauchy
sequence. By taking subsequences if necessary we may assume
}fn ´ fm }Lp ď 2´ mintm,nu .
We define the monotone sequence of functions
n´1
ÿ
Fn pxq “ |f1 pxq| ` |fm`1 pxq ´ fm pxq|
m“1

and F “ limnÑ8 Fn pxq. F is measurable and by monotone convergence

theorem for 1 ď p ă 8,
´ˆ ¯1{p ´ˆ ¯1{p ÿ
p p
|F | dµ “ lim |Fn | dµ ď }f1 }Lp ` 2´m “ }f1 }Lp ` 1,
X nÑ8 X mě1

while for p “ 8 it holds also }F }L8 ď }f1 }L8 ` 1, and in particular F is

finite almost everywhere. Thus
n´1
ÿ
fn “ f1 ` pfm`1 ´ fm q
m“1

converges if F pxq ă 8. Let f be the limit if F pxq ă 8, and 0 otherwise. It

is measurable. Since max tf, fn u ď F we obtain by dominated convergence
ˆ
p
}f ´ fn }Lp “ |f ´ fn |p dµ Ñ 0,
X
ř8
for 1 ď p ă 8, and for p “ 8 it follows directly from } m“n pfm`1 ´
fm q}L8 Ñ 0 as n Ñ 8.
Remark 3.21. The Lebesgue space L2 pµq equipped with the inner produce
ˆ
xf, gyL2 pµq “ f ḡ dµ
X

is a Hilbert space.

39 [February 17, 2023]

3.3.2 Projection and duality
Recall the Hilbert projection Theorem 2.15 for the Hilbert space L2 pµq. We
have similar projection result in Lp pµq, 1 ă p ă 8.

Lemma 3.22 (Projection in Lp pµq). Let 1 ă p ă 8 and let K be a closed

convex set in Lp pµq. Let f P Lp pµq. Then there exists a unqiue g P K with

}f ´ g}Lp pµq “ distpf, Kq.

Moreover
ˆ
Re ph ´ gqpf¯ ´ ḡq|f ´ g|p´2 dµ ď 0, @h P K. (3.8)
X

Proof. Let thn u be a minimzing sequence. Since 21 phn ` hm q P K and

}hn ´ f ` hm ´ f }Lp ď }hn ´ f }Lp ` }hm ´ f }Lp , we see that

}hn ´ f ` hm ´ f }Lp Ñ 2 distpf, Kq.

Now let p ď 2, from the second Hanner’s inequality we obtain

` ˘p
}hn ´ f ` hm ´ f }Lp ` }hn ´ hm }Lp
ˇ ˇp
` ˇ}hm ´ f ` hm ´ f }Lp ´ }hn ´ hm }Lp ˇ
ď2p }hn ´ f }pLp ` }hm ´ f }pLp .
` ˘

Let A “ lim supn,mÑ8 }hn ´ hm }Lp . This limsup is obtained along two sub-
sequences n, m Ñ 8. Let D “ distpf, Kq. Then

p2D ` Aqp ` |2D ´ A|p ď 2p`1 Dp

which implies A “ 0 by the strict convexity of A Ñ p2D ` Aqp ` p2D ´ Aqp

for A P r0, 2Dq.
[02.12.2022]
[05.12.2022]

If p ą 2 we argue similarly with the first inequality:

}hn ´ f ` hm ´ f }pLp ` }hn ´ hm }pLp

ď p}hn ´ f }Lp ` }hm ´ f }Lp qp ` |}hn ´ f }Lp ´ }hm ´ f }Lp |p Ñ p2Dqp .

Thus phn q is a Cauchy sequence. Up to a subsequence, phn q converges to

some limit g P K in Lp pµq, and g P K is the point of minimal distance. The
same argument implies that the closest point is unique.

40 [February 17, 2023]

Now let h P K. Let
ˆ
N ptq “ |f ´ pg ` tph ´ gqq|p dµ.
X

Then N ptq attains its minimum at t “ 0 on the interval r0, 1s. We claim that
its derivative at t “ 0 is
ˆ
d
N |t“0 “ p Re |f pxq ´ gpxq|p´2 pf pxq ´ gpxqqpḡpxq ´ h̄pxqqdµ.
dt X

This implies the assertion N 1 p0q ě 0. To calculate the derivative we assume

that f, g P Lp pµq, then almost everywhere
d
|t“0 |f ` tg|p “ p|f |p´2 Re f ḡ
dt
and the pth power is convex such that the composition of G “ Gpz, z̄q “ |z|p
and the linear function Hptq “ f ` tg is convex in t P r´1, 1s:
1
|f |p ´ |f ´ g|p ď p|f ` tg|p ´ |f |p q ď |f ` g|p ´ |f |p .
t
The formula follows by Lebesgue’s dominated convergence theorem.
Remark 3.23. If K is a closed subspace of Lp pµq, 1 ă p ă 8, then the
projection relation (3.8) can be improved into
ˆ
hpf¯ ´ ḡq|f ´ g|p´2 dµ “ 0, @h P K. (3.9)
X

Theorem 3.24 (Duality of Lp pµq). Let pX, A, µq be a measure space, 1 ă

p, q ă 8, p1 ` 1q “ 1. Then
ˆ
q
j : L Q g Ñ pf Ñ f gdµq P pLp pµqq˚
X

is a linear isometric isomorphism.

Proof. The proof is the same as for Riesz representation theorem 2.25 in case
of Hilbert spaces. By Hölder’s inequality the map is well defined and

}jpgq}pLp q˚ ď }g}Lq .

Since
ˆ
}jpgq}pLp q˚ }g}q´1
Lq “ }jpgq}pLp q˚ }|g| q´2
ḡ}Lp ě jpgqp|g| q´2
ḡq “ |g|q dµ “ }g}qLq ,
X

41 [February 17, 2023]

we conclude as for Hilbert spaces that

}jpf q}pLp q˚ “ }f }Lq .

Surjectivity is proven exactly as for Hilbert spaces (see Theorem 2.25). Ex-
ercise.

Corollary 3.25 (Duality of L1 pµq). Suppose that pX, A, µq is σ finite mea-

sure space. Then
ˆ
8
L pµq Q g Ñ pf Ñ f gdµq P pL1 pµqq˚
X

is a linear isometric isomorphism.

Proof. Exercise.
The duality of the space L8 pµq is more involved. We are going to show Riesz
representation Theorem 3.41 below, which characterizes the linear functional
on Cc pXq where X is sigma compact metric space in terms of Radon measure.

3.3.3 Young’s inequality on Rd

Let pX, A, µq be a measure space and suppose that 1 ď p, q, r ď 8 and
1
p
` 1q ` 1r “ 1. If f P Lp pµq, g P Lq pµq and h P Lr pµq, then f gh is integrable
and ˇˆ ˇ
ˇ ˇ
ˇ f ghdµˇ ď }f }Lp pµq }g}Lq pµq }h}Lr pµq .
ˇ ˇ
X

This is a consequence of a multiple application of Hölder’s inequality:

ˇˆ ˇ
ˇ ˇ
ˇ f ghdµˇ ď }f }Lp }gh} p
ˇ ˇ L p´1
X

and if p ą 1,
ˆ p p
p p p p
|g| p´1 |h| p´1 dµ ď }|g| p´1 } qpp´1q }|h| p´1 } rpp´1q “ }g}Lp´1 p´1
q }h}Lr ,
X L p L p

since ˆ ˙
p 1 1 p 1
` “ p1 ´ q “ 1.
p´1 q r p´1 p
We denote Lp pRd q (or even Lp ) for Lp pmd q where md is the Lebesgue measure.

42 [February 17, 2023]

Lemma 3.26. Suppose that 1 ď p, q, r ď 8 satisfy
1 1 1
` ` “2
p q r

and that f P Lp pRd q, g P Lq pRd q and h P Lr pRd q. Then

Rd ˆ Rd Q px, yq Ñ f p´xqgpx ´ yqhpyq

is integrable and
ˆ
Ipf, g, hq :“ f p´xqgpx ´ yqhpyqdm2d px, yq
Rd ˆ Rd

satisfies
|Ipf, g, hq| ď }f }Lp }g}Lq }h}Lr
and I is invariant by changing the orders of f, g or of g, h:

Ipf, g, hq “ Ipg, f, hq “ Ipg, h, f q “ Iph, g, f q “ Iph, f, gq “ Ipf, h, gq.

Proof. Measurability is a consequence of the theorem of Fubini and an ap-

proximation argument (similar as in the proof of Minkowski’s inequality
3.2). It suffices to prove the estimate for nonnegative functions. We as-
sume 1 ă p, q, r ă 8 since the limit cases are simpler, and follow by obvious
p
modifications. We define p1 , q 1 and r1 by p1 ` p11 “ 1, i.e. p1 “ p´1 etc. Let
1 1
αpx, yq “|f p´xq|p{r |gpx ´ yq|q{r ,
1 1
βpx, yq “|f p´xq|p{q |hpyq|r{q ,
1 1
γpx, yq “|gpx ´ yq|q{p |hpyq|r{p .
1 1 1
Then p1
` q1
` r1
“ 1 and
ˆ
|I| ď αpx, yqβpx, yqγpx, yqdm2d
d d
R ˆR
ď}α}Lr1 }β}Lq1 }γ}Lp1
p q p r q r
1 1 1 1
“}f }Lr1p }g}Lr1q }f }Lq p }h}Lq r }g}Lp q }h}Lp r
“}f }Lp }g}Lq }h}Lr .
The second last equality is a consequence of the theorem of Fubini.
[05.12.2022]
[09.12.2022]

43 [February 17, 2023]

Theorem 3.27 (Young’s inequality). Suppose that 1 ď p, q, r1 ď 8 and
1 1 1
` “ 1 ` 1.
p q r
If f P Lp and g P Lq , then for almost all x
f px ´ yqgpyq
is integrable and
" ´
Rd
f px ´ yqgpyqdmd pyq if integrable
f ˚ gpxq :“
0 otherwise
1
defines a unique element in Lr pRd q and
}f ˚ g}Lr1 pRd q ď }f }Lp }g}Lq .
In particular, if q “ 1, then p “ r1 and
}f ˚ g}Lp ď }f }Lp }g}L1 .
´|x|2
Proof. We have e P Lr for all 1 ď r ď 8. Then
2
e´|x| f px ´ yqgpyq
´
is m2d integrable by Lemma 3.26. We apply Fubini to see that Rd f px ´
yqgpyqdmd pyq exists for almost all x. If r “ 8, i.e. p “ q “ r1 “ 1, then it
follows from Fubini Theorem.
For the case 1 ď r ă 8, by Theorem 3.24 and Corollary 3.25, the estimate
follows once we prove
ˇˆ ˇ
ˇ ˇ
dˇ
ˇ d f ˚ ghdm ˇ ď }f }Lp }g}Lq }h}Lr
ˇ
R

for
1 1
` 1 “1
r r
and all h P Lr . Since then
1 1 1
` ` “2
p q r
and, by Fubini and Lemma 3.26
ˇˆ ˇ ˆ
ˇ ˇ
dˇ d
ˇ
ˇ d f ˚ ghpxqdm ˇ ď d |f | ˚ |g||h|dm
R
ˆR
“ |f px ´ yq||gpyq||hpxq|dm2d px, yq
Rd ˆ Rd
ď}f }Lp }g}Lq }h}Lr .

44 [February 17, 2023]

3.3.4 Schur’s lemma
We have seen from Young’s inequality that any function g P L1 pRd q defines
a linear map from Lp pRd q to Lp pRd q, p P r1, 8s:

Lp pRd q Q f ÞÑ pf ˚ gq P Lp pRd q.

However a general measure space pX, A, µq is not necessarily a vector space

and the addition operator may make no sense. This motivates us to consider
more general integral kernel kpx, yq (instead of gpx ´ yq), and Schur’s lemma
gives a criterium for an integral kernel to define a linear map from Lp pνq to
Lp pµq for 1 ď p ď 8.

Theorem 3.28 (Schur’s lemma). Let pX, A, µq and pY, B, νq be σ finite

measure spaces and k : X ˆ Y Ñ C be µ ˆ ν measurable. Suppose that
C1 , C2 P r0, 8q and
ˆ ˆ
sup x |kpx, yq|dνpyq ď C1 , sup y |kpx, yq|dµpxq ď C2 .
Y X

If 1 ď p ď 8 and f P Lp pνq, then

ˆ
kpx, yqf pyqdνpyq
Y

exists for almost all x and

›ˆ › 1´ p1 1

› kpx, yqf pyqdνpyq› ď C1 C2p }f }Lp pνq .

› ›
Y Lp pµq

The map ˆ
p
L pνq Q f Ñ T f :“ kpx, yqf pyqdνpyq P Lp pµq
Y
is a continuous linear map which satisfies
1´ p1 1
}T }Lp pνqÑLp pνq ď C1 C2p .

Proof. As in the proof of Young’s inequality, it suffices to show the estimate.

If p “ 1 or p “ 8 this is an immediate consequence of the theorem of Fubini.
Repeating the argument of Young’s inequality we have to prove that
ˆ
1´ 1 1
|gpxq||kpx, yq||f pyq|dpµ ˆ νq ď C1 p C2p (3.10)
XˆY

45 [February 17, 2023]

where
1 1
` “ 1,
p q
if }f }Lp “ }g}Lq “ 1 and 1 ă p, q ă 8. It suffices to prove (3.10) for
nonnegative functions f , g and k, where f and g are bounded and 0 outside
a set of finite measure, and the general case follows by an approximation
argument.
For z P C with 0 ď Re z ď 1, we define
" pz´1
f f if f ‰ 0
fz “
0 otherwise,

and " p
g p´1 p1´zq´1 g if g ‰ 0
gz “
0 otherwise .
Then for σ P R,
}fiσ }L8 “ }g1`iσ }L8 “ 1,
and
}f1`iσ }L1 “ }f }pLp “ }giσ }L1 “ }g}qLq “ 1,
and hence ˆ
|giσ pxq||kpx, yq||fiσ pyq|dµ ˆ ν ď C1
XˆY
ˆ
|g1`iσ pxq||kpx, yq||f1`iσ pyq|dµ ˆ ν ď C2 .
XˆY

Moreover
f 1 “ f, g 1 “ g.
p p

Notice that fz and gz are bounded and zero outside a set of finite measure.
By dominated convergence
ˆ
z Ñ Hpzq “ gz pxqkpx, yqfz pyqdpµ ˆ νq
XˆY

is continuous in the strip C “ tz : 0 ď Re z ď 1u, differentiable and satisfies

the Cauchy-Riemann differential equations. The claim follows from the three
lines inequality, which ensures that the maximum is attained at the boundary:
Lemma 3.29 (Three lines inequality). Suppose that u P CpCq is bounded
and holomorphic in the interior. Then

sup C |u| “ sup BC |u|

46 [February 17, 2023]

We apply the lemma to

upzq “ C1z´1 C2´z Hpzq,

such that
ˆ
´p1´ p1 q ´ p1 1
C1 C2 gpxqkpx, yqf pyqdpµ ˆ νq “ up q ď 1.
XˆY p

Proof of Lemma 3.29. (Explained in the exercise class.) 1) Let U Ă C

be a bounded open connected set and u P CpU ; Cq be a holomorphic function
in the interior. We claim that then

sup xPU |upxq| “ sup xPBU |upxq|.

We prove this by contradiction. Suppose that |u| attains its maximum M

at some interior point z0 and suppose that this is larger than sup BU |upzq|.
Then
f pzq “ Rerupzq{upz0 qs
satisfies 0 ď f ď 1 and f pz0 q “ 1. Moreover f is harmonic. Let

fε px ` iyq “ f px ` iyq ` εpx ´ Re z0 q2

where ε is so small that fε pzq ă 1 for z P BU . Then fε has a maximum in

an interior point z1 . At this point the Hessian is negative semidefinite by its
trace ∆fε pz1 q “ 2ε. This is a contradiction.
2) Let u be defined in C as in the lemma, and let
2
uε pzq “ eεz upzq.

Since uε pzq Ñ 0 as | Im z| Ñ 8

sup zPC |uε pzq| “ sup zPBC |uε | ď eε sup zPBC |upzq|.

Now we let ε tend to zero.

3.4 Radon measures on metric spaces

In this section we always assume that pX, dq is a metric space, and we restrict
ourselves to Borel measure space pX, BpXq, µq.
We recall that a subset K in a metric space pX, dq is compact (i.e. any open
covering of K contains a finite subcovering) is equivalent to (This will be
explained shortly in the exercise class.)

47 [February 17, 2023]

 • K is sequentially compact, i.e. every sequence in K contains a subse-
quence which converges to a limit in K;

• K is totally bounded and complete, i.e. for any ε ą 0 there exists a

finite covering YN
j“1 Bpxj , εq of K, and any Cauchy sequence in K has
a subsequence which converges to a limit in K.
Recall that X is locally compact if for every point x there is a neighborhood
whose closure is compact. pX, dq is called σ compact 4 if it is locally compact
and if it is a countable union of compact sets. If pX, dq is σ compact, then
there exists a sequence of compact set Kj so that Kj is contained in the
interior K˚
Ť8
j`1 of Kj`1 and X “ j“1 Kj (Exercise.) We denote by C0 pXq Ă
Cb pXq the continuous functions f with limit 0 at 8, i.e. for all ε ą 0 there
exists j so that f is at most of size ε outside Kj . We define Cc pXq as the
subspace of continuous functions with compact support.
We recall that the Borel sets BpXq are the smallest σ algebra containing all
open sets. A Borel measure is a measure on the Borel sets. A Borel measure
µ on a metric space pX, dq is called inner regular if for every Borel set A

µpAq “ sup tµpKq : K Ă A, K compactu,

while outer regular if for every Borel set A

µpAq “ inftµpU q : A Ă U, U openu.

It is called locally finite if for every x P X there exists an open neighborhood

U Q x so that µpU q ă 8.

3.4.1 Radon measure

Definition 3.30 (Radon measure). A Borel measure µ on a metric space
pX, dq is called a Radon measure, if it is locally finite and inner regular.
Lemma 3.31 (Compact sets are of finite Radon-measure). Let µ be a Radon
measure on a metric space pX, dq. Then the measure of a compact set is finite.
Furthermore for ε ą 0 and K compact, there exists an open set U Ą K of
finite measure with µpU q ă µpKq ` ε.
Proof. Let K be compact. For every x P K exists an open set Ux containing
x with µpUx q ă 8. Since K is compact and
ď
KĂ Ux ,
4
In some references σ compact metric space is not necessarily locally compact.

48 [February 17, 2023]

there exists a finite subcovering
N
ď
KĂ Uxj “: Ũ
j“1

such that
N
ÿ
µpKq ď µpŨ q ď µpUxj q ă 8.
j“1

We define Uj “ Ũ X tx : dpx, Kq ă 1j u, such that the finite-measure set

Uj Ą K. Since 1Uj Ñ 1K pointwisely, by the theorem of Lebesgue
ˆ ˆ
1Uj dµ “ µpUj q Ñ µpKq “ 1K dµ.
X X

[09.12.2022]
[12.12.2022]

Vise Versa: If pX, dq is a σ compact metric space, then a Borel measure is

also a Radon measure, if (and only if) any compact set is of finite measure.

Lemma 3.32 (Radon measure on σ compact metric space). Let pX, dq be a

σ compact metric space.
Then a Borel measure such that any compact set is of finite measure is Radon
measure. And a Radon measure is σ finite and outer regular.

Proof. We claim first that for any Borel measure µ on a metric space pX, dq,
for any Borel set B with µpBq ă 8 and any ε ą 0, there exists a closed set
C Ă B with µpBzCq ă ε. (Exercise.) If in addition any compact set if
of finite µ-measure, then µ is locally
Ť finite since X is locally compact. Let
Kj be compact subsets with X “ Kj and Kj contained in the interior of
Kj`1 . Then for any closed set C with finite-measure

µpC X Kj q Ñ µpCq

and C X Kj is compact. If µpBq ă 8, then its measure is the supremum of

the measures of a sequence of compact sets Ck Y Kj . If µpBq “ 8, then the
same holds since the Borel set Kj X B is of finite measure. Thus µ is inner
regular and hence Radon measure.

49 [February 17, 2023]

Let now µ be Radon. Then it is σ finite since X is σ compact. For any
Borel set B, Kj zB is Borel with µpK̊j zBq ă 8. Then there exists a closed
set Cj Ă K̊j zB with µppK̊j zCj qzBq ă ε2´j . Let
8
ď
U“ pK̊j zCj q.
j“1

It is open and
8
ď 8
ď
B“ pK̊j X Bq Ă K̊j zCj “ U.
j“1 j“1

Moreover ď
µpU zBq “ µp pK̊j zCj qzBq ă ε.
Thus a Radon measure is outer regular.
Remark 3.33. The Lebesgue measure restricted to Borel sets is Radon mea-
sure. The counting measure on R is not a Radon measure.
Continuous functions on compact metric spaces are p´integrable, p P r1, 8s
with respect to Radon measures, which are finite measures. In particular,
continuous functions with compact supports are p´integrable with respect to
the Radon measure µ on a metric space. However, without the assumption
of compact supports, continuous functions are not necessarily in Lp pµq: For
example, the constant function 1 is a continuous function defined on Rd with
support Rd , but is not in Lp pRd q for any p P r1, 8q.

3.4.2 Density of Lip.-functions with comp. supp. in Lp pµq

Continuous functions with compact supports belong to Lp pµq, p P r1, 8s with
respect to Radon measure µ on a σ-compact metric space. They are indeed
dense in Lp pµq if 1 ď p ă 8.
Theorem 3.34 (Density of Lipschitz functions with compact supports in
Lp pµq). Let pX, dq be a σ compact metric space, µ a Radon measure on X
and 1 ď p ă 8. Then the set of Lipschitz-continuous functions with compact
supports are dense in Lp pµq.
Proof. Step 1. Continuous functions with compact supports are
dense. We claim that the simple function set is dense in Lp pµq, 1 ď p ă 8,
(Exercise.) and it suffices to approximate a characteristic function of a
measurable set A of finite measure by a continuous function in Lp pµq. Let
ε ą 0. By the inner and outer regularity of a Radon measure on a σ compact
metric space, there exists a compact set K and an open set U so that
KĂAĂU µpU q ă µpKq ` ε.

50 [February 17, 2023]

Then the distance between a compact set K and a disjoint closed set XzU
is positive dpK, XzU q “: d0 ą 0. For L sufficiently large such that d0 L ě 1,
we define
$
& 1 if x P K,
fL pxq “ max t1 ´ Ldpx, Kq, 0u “ 1 ´ Ldpx, Kq P r0, 1s if x P KL ,
0 if x P XzKL ,
%

where KL :“ tx P X|dpx, Kq ď L1 u Ă U is compact. Then fL P Cc pXq with

Supp pf q “ KL and it approximates the characterisitic function χA in the
sense of Lp pµq:
1 1
}fL ´ χA }Lp ď pµpU zKqq p ă ε p .
Step 2. Lipschitz continuous functions with compact supports are
dense. We prove that for every ε ą 0 and f P Cc pXq with compact support,
there exists g Lipschitz continuous with
supp g Ă supp f
and
sup |g ´ f | ă ε.
Then g approximates f also in Lp pµq-sense. It suffices to do this for f ě 0.
Since supp f is compact it is uniformly continuous: There exists δ ą 0 so
that |f pxq ´ f pyq| ă ε if dpx, yq ă δ. With
" *
|f pxq ´ f pyq|
L̃ “ sup x,y : dpx, yq ě δ ,
dpx, yq
which is finite since it is the supremum of a continuous function on a compact
set, we obtain the inequality
|f pxq ´ f pyq| ď ε ` L̃dpx, yq, @x, y P X.
We define
gpxq “ mintf pyq ` 2L̃dpx, yqu.
y

One easily checks that

• The mimimum is attained in Bδ pxq, since the values evaluated outside
Bδ pxq is bigger than f pxq ´ L̃dpx, yq ` 2L̃dpx, yq ě f pxq ` L̃δ;
• g ě 0 and takes value between f ´ ε and f as
max t0, f pxq ´ εu ď gpxq ď f pxq.
In particular, Supp pgq Ă Supp pf q and sup |g ´ f | ă ε;

51 [February 17, 2023]

 • g has Lipschitz constant 2L̃, since for any x1 , x2 P X, if gpx1 q “ f py1 q`
2L̃dpx1 , y1 q for some y1 P Bδ px1 q, then

gpx2 q´gpx1 q ď f px2 q ´ pf py1 q ` 2L̃dpx1 , y1 qq

ď f py1 q ` L̃dpx2 , y1 q ´ pf py1 q ` 2L̃dpx1 , y1 qq ď 2L̃dpx1 , x2 q,

if dpx2 , y1 q ě δ, and similarly for dpx2 , y1 q ă δ (by bounding gpx2 q by

f py1 q ` 2L̃dpx2 , y1 q).

[12.12.2022]
[16.12.2022]

3.4.3 Arzela-Ascoli theorem and separability of Lp pµq

Theorem 3.35 (Arzela-Ascoli). Let pX, dq be a compact metric space. Then
a closed set A Ă Cb pXq is compact if and only if

1. A is bounded.

2. A is equicontinuous, i.e. for ε ą 0 there exists δ ą 0 so that

|f pxq ´ f pyq| ă ε if f P A and dpx, yq ă δ.

Proof. Let A be compact. Since

Cb pXq Q f Ñ }f }Cb pXq

is continuous and hence attains its maximum in A, we deduce that A is

bounded. Let ε ą 0. For every continuous function f P Cb pXq defined on
the compact metric space X, there exist δf ą 0 and an open neighborhood
Uf Ă Cb pXq so that

|gpxq ´ gpyq| ă ε if g P Uf and dpx, yq ă δf .

Then A Ăf PA Uf , and since A is compact there is a finite subcovering, A ĂN

j“1
Ufj . We define δ “ min δfj .
Now assume that A is closed, bounded and equicontinuous. Let pfj q Ă A
be a sequence. Let ε ą 0 and δ ą 0 as in the second condition. Then
there exist a finite number N of points xk so that Bδ pxk q cover X since
X is compact. There exists a subsequence so that fjl pxk q converges for all

52 [February 17, 2023]

xk . In particular, after relabeling, there are infinitely many tfjl ulPN so that
|fjl pxk q ´ fjm pxk q| ă ε. Then for any x P X there exists Bδ pxk q Q x such that

|fjl pxq´fj1 pxq| ď |fjl pxq´fjl pxk q|`|fjl pxk q´fj1 pxk q|`|fj1 pxq´fj1 pxk q| ă 3ε.

Hence pfjl q Ă B3ε pfj1 q and A is compact.

Corollary 3.36. If pX, dq is compact, then Cb pXq is separable.

Proof. By the proof of Theorem 3.34, the Lipschitz continuous functions are
dense in Cb pXq. The countable union of separable sets is separable and its
closure is separable. Hence it suffices to prove that

tf P Cb pXq : }f pxq}Cb pXq ď n, |f pxq ´ f pyq| ď ndpx, yqu

is separable. This set is compact by Theorem 3.35 and hence separable.

Corollary 3.37. Let pX, dq be σ compact metric space and µ a Radon mea-
sure. If 1 ď p ă 8 then Lp pµq is separable.

Proof. Since continuous functions with compact supports are dense, it suf-
fices to show that Cc pKn q is separable in Lp pµq. It follows from Corollary
3.36.

3.4.4 Compact sets in Lp pRd q

The space Rd is σ compact and the Lebesgue measure is σ finite Radon mea-
sure. We observe first the continuity of the translation operator on Lp pRd q,
p P r1, 8q.

Lemma 3.38. Suppose that 1 ď p ă 8, f P Lp pRd q, ε ą 0. Then there exist

δ ą 0 and R ą 0 so that for all |h| ă δ

}f p. ` hq ´ f p.q}Lp ă ε, }χRd zBR p0q f }Lp ă ε.

Proof. The second claim is a consequence of monotone convergence. For the

first we approximate f by a Lipschitz continuous function g with compact
support, }g ´ f }Lp ă ε{4 and estimate

}f p. ` hq ´ f }Lp ď}f p. ` hq ´ gp. ` hq}Lp ` }f ´ g}Lp ` }gp¨ ` hq ´ g}Lp

` ˘1{p
ďε{4 ` ε{4 ` |h|}g}Lip 2md psupp gq
ďε

by choosing |h| ď δ for some small δ.

53 [February 17, 2023]

 [16.12.2022]
[19.12.2022]
We hence can approximate f P Lp pRd q, p P r1, 8q by a sequence of smooth
functions which are convolutions of f itself and the Dirac-sequence (which is
a sequence of smooth functions with compact supports, and take values in
r0, 1s with the integral 1 on the whole space).

Corollary 3.39. Let η be a radial´ smooth nonnegative function supported

in the unit ball with unit integral: Rd ηdmd “ 1, and let ηr pxq “ r´d ηpx{rq,
r ą 0. Let f P Lp pRd q, p P r1, 8q and fr :“ ηr ˚ f as in Young’s Theorem
3.27. Then fr P C k pRd q, @k P N, @r ą 0, and for any ε ą 0, there exists δ
so that
}fr ´ f }Lp ă ε, @r ď δ.
In particular fr Ñ f in Lp as r Ñ 0.

Proof. Exercise (and give an example of such η). It suffices to control

}fr ´ f }Lp by sup |h|ďr }f p. ` hq ´ f }Lp .
We characterize compact subsets of Lp spaces.

Theorem 3.40 (Kolmogorov). Let 1 ď p ă 8. A closed subset C Ă Lp pRn q

is compact iff

1. C is bounded.

2. For every ε ą 0 there exists δ so that for all |h| ă δ and all f P C

}f p. ` hq ´ f }Lp pRd q ă ε

3. For every ε ą 0 there exists R so that for all f P C

}χRd zBR p0q f }Lp ă ε.

Proof. Let C be compact. Since f Ñ }f }Lp is continuous it attains its

maximum and hence C is bounded. We argue by contradiction to show the
second and third parts. Suppose there exists ε ą 0, hj Ñ 0 and fj P C so
that
}fj p¨ ` hj q ´ fj }Lp ě ε.
Since C is compact, there exists a subsequence pfjl q Ă C which converges to
the limit f in C. By Lemma 3.38 there exists δ ą 0 so that

}f p¨ ` hq ´ f }Lp ă ε{2

54 [February 17, 2023]

for |h| ă δ. This contradicts the previous inequality. Similarly we deduce
the third part.
Vice versa: Suppose that C Ă Lp pRd q is closed, bounded, and satisfies the
three claims. Let ε, δ, R be as in the claims. Since a closed set in a Banach
space is complete, it suffices to show the existence of finite 10ε-covering for
C (i.e. there are finite covering balls of radius 10ε for C). By the third claim,
the set C1 :“ tf χBR`2δ p0q | f P Cu lies in ε-neighborhood of C, and vice versa.
Hence it suffices to show the existence of finite 4ε-covering in Lp pRd q for C1 .
We claim that the set C δ :“ tf δ :“ ηδ ˚ f1 | f1 P C1 u is in 3ε-neighborhood
of C1 . This claim follows from the following inequality (similarly as in the
proof of Corollary 3.39)

}ηδ ˚ pf χBR`2δ p0q q ´ f χBR`2δ p0q }Lp

›ˆ ›
› ` ˘ ›
“›› f χBR`2δ p0q px ´ δzq ´ f 1BR`2δ p0q pxq ηpzqdz ››
B1 p0q Lpx
›ˆ ›
› ` ˘ ›
ď ›› f px ´ δzq ´ f pxq ηpzqdz ››
B1 p0q Lp pB p0qq
›ˆ x R`δ
›
› ` ˘ ›
` ›› f χBR`2δ p0q px ´ δzq ´ f χBR`2δ p0q pxq ηpzqdz ››
d
R Lpx pRd zBR`δ p0qq

ď sup |h|ăδ }f p¨ ` hq ´ f p¨q}Lp ` 2}f χRd zBR p0q }Lp .

Hence it suffices to show the existence of ε-covering in Lp pRd q for C δ . By

Corollary 3.39, the approximated functions f δ are Lipschitz continuous with
compact supports in BR`3δ p0q, and all the functions in C δ have uniform
supremum norm, Lipschitz constant and support. By Theorem 3.35 the set
C δ is compact in Cb pBR`3δ p0qq and hence in Lp pRd q, and we can cover it by
a finite ε-covering in Lp pRd q.

3.4.5 Riesz representation theorem for Cc pXq

Let X “ r0, 1s Ă R be a finite interval, and Cb pXq “ Cc pXq “ Cpr0, 1sq
be the set of all continuous functions on r0, 1s, which is a separable Banach
space. The ordinary Riemann integral
ˆ 1
Lpf q “ f pxqdx
0

• defines a linear map (also called linear functional) from Cb pXq to K;

• satisfies Lpf q ě 0 if f ě 0;

55 [February 17, 2023]

 • is bounded |Lpf q| ď }f }Cb pXq .

We observe that for any open interval pa, bq Ă r0, 1s, and any continuous
function f with compact support inside pa, bq and with value f P r0, 1s, it
holds

Lpf q ă b ´ a,

and since we can choose f such that Lpf q approximates b ´ a,

b ´ a “ sup tLpf q : f P Cc pXq, Supp pf q Ă pa, bq, 0 ď f ď 1u.

This reveals a connection between the (Lebesgue) measure of pa, bq and the
values of the nonnegative linear map L on continuous functions with support
in pa, bq and values in r0, 1s.
[19.12.2022]
[23.12.2022]
More generally, let pX, dq be metric space, and let L : Cc pXq ÞÑ K be a
nonnegative linear map, then we define the variation measure of L by

µ˚ pU q “ sup tLpf q : f P Cc pXq, supp f Ă U, 0 ď f ď 1u

for open sets U and for general sets

µ˚ pAq “ inftµ˚ pU q : A Ă U, U openu.

If we assume further the boundedness for L in the sense that for any compact
set K, there exists a constant CK ą 0 such that

|Lpf q| ď CK }f }Cb pXq (3.11)

holds for all f P Cc pXq with Supp f Ă K, we have the following representa-
tion of this nonnegative linear map on Cc pXq by the Lebesgue integral with
respect to the Radon measure µ induced by µ˚ .

Theorem 3.41 (Riesz representation theorem for nonnegative linear func-

tionals on Cc pXq). Let pX, dq be a σ compact metric space and let the nonneg-
ative linear map L : Cc pXq Ñ K satisfy (3.11). Then there exists a unique
Radon measure µ on pX, dq so that for all f P Cc pXq,
ˆ
Lpf q “ f dµ.
X

56 [February 17, 2023]

Remark 3.42. One can relax the assumption that X is a σ compact metric
space to that X is a locally compact Hausdorff space (with Radon measure
defined more generally), and the condition (3.11) can be removed: The proof
will be more delicate. In probability some interesting measures (e.g. the
so-called Wiener measure) occur on topological spaces that are not locally
compact, and Riesz representation theorem is not applicable.
The reverse holds obviously: For
´ any Radon measure µ on a σ-compact metric
space X, the integral Lpf q :“ X f dµ defines a nonnegative linear map from
Cc pXq to K satisfying (3.11). That is, there exists a bijective map

J : tRadon measure µ on Xu Q µ ÞÑ L P tL : Cc pXq Ñ K | L is nonnegative linear map,

L is bounded on Cc pKq, @K Ă Xcompactu

and for all open sets U ,

µpU q “ sup tLpf q | f P Cc pXq, supppf q Ă U, 0 ď f ď 1u.

Notice that in general it is not natural to assume that L is a continuous/bounded

linear map from Cb pXq to´ K: For example, for the Lebesgue measure on
Rd , the Lebesgue integral Rd f dmd is not bounded on the unit closed ball
t}f }Cb pRd q ď 1u.
The Lebesgue measure on Rd can be constructed from the Lebesgue integral
as in the theorem.
Proof. Step 1: µ˚ is an outer metric measure, and its restriction
µ on Caratheodory measurable sets and then on Borel sets is a
Radon measure. It is easy to see that µ˚ pAq ď µ˚ pBq
Ť if A Ă B. To show
˚
that µ is an outer measure, let Uj be open sets, U “ j Uj and we have to
show that ÿ
µ˚ pU q ď µ˚ pUj q,
j

which is equivalent to show for all 0 ď f ď 1 with Supp f Ă U ,

ÿ
Lpf q ď µ˚ pUj q.
j

Let K “ Supp f which is compact. Thus K is covered by finitely many

Ť N
j“1 Uj for some N ă 8. We claim that there exist gj , 0 ď gj ď 1, Supp gj Ă
Uj and N
ř
gj “ 1 on K. (Exercise.) We define fj “ gj f . Then 0 ď fj ď 1,
řN j“1
f “ j“1 fj and
N
ÿ N
ÿ
Lpf q “ Lpfj q ď µ˚ pUj q.
j“1 j“1

57 [February 17, 2023]

 Ť
For general set A “ j Aj , for any ε ą 0, there exist pUj q such that µ˚ pUj q ă
Ť
µ˚ pAj q ` ε2´j , and hence A Ă Uj and
ď ÿ ÿ
µ˚ pAq ď µ˚ p Uj q ď µ˚ pUj q ď µ˚ pAj q ` ε.
j j j
ř
Thus µ˚ pAq ď j µ˚ pAj q.
To show that µ˚ is an outer metric measure, let A, B be two disjoint sets with
positive distance. If A, B are both open sets, then µ˚ pAYBq ě µ˚ pAq`µ˚ pBq,
since for all f P Cc pA; r0, 1sq and g P Cc pB; r0, 1sq the sum f ` g P Cc pA Y
B; r0, 1sq. For general disjoint sets A, B, there exist disjoint open sets V and
W containing A resp. B. Then for all open set U Ą A Y B, the two open
sets U X V and U X W are disjoint so that
µ˚ pU q ě µ˚ pU X V q ` µ˚ pU X W q ě µ˚ pAq ` µ˚ pBq.
Thus µ˚ pA Y Bq “ µ˚ pAq ` µ˚ pBq. Similarly it follows µ˚ pHq “ 0.
Let µ be the measure constructed from the outer metric measure µ˚ by the
Caratheodory Theorem 3.4. Then
µpBq “ µ˚ pBq
for Borel sets. By (3.11), µ is bounded on compact sets: For any K Ă X
compact, there exist an open set V and a compact set K0 such that K0 Ą V Ą
K (since X is σ compact metric space), and hence for any f P Cc pX; r0, 1sq
with f “ 1 on V and supppf q Ă K0 ,
µpKq ď µpV q ď Lpf q ď CK0 }f }Cb pXq “ CK0 ă 8.
Thus by Lemma 3.32 its restriction to Borel sets is a Radon measure.
Step 2: Representation of the renormalized map Λpf q by the Lebesgue
integral. 5 Let f P Cc pXq be nonnegative. We define
Λpf q “ sup t|Lpgq| : g P Cc pXq, |g| ď f u ě 0.
Clearly 0 ď f1 ď f2 implies Λpf1 q ď Λpf2 q and for c ą 0, Λpcf q “ cΛpf q. We
claim that
Λpf1 ` f2 q “ Λpf1 q ` Λpf2 q
for f1 , f2 P Cc pXq nonnegative. Indeed, if |g1 | ď f1 and |g2 | ď f2 then
|g1 ` g2 | ď f1 ` f2 , and, if in addition Lpg1 q, Lpg2 q P r0, 8q,
|Lpg1 q ` Lpg2 q| “ |Lpg1 q| ` |Lpg2 q| ď Λpf1 ` f2 q.
5
We can simply take Λ to be L itself, when it applies on nonnegative functions. For
general case, we notice by linearity that Lpf q “ LpRe f q ` iLpIm f q and for real-valued
function f , Lpf q “ Lpf ` q ´ Lpf ´ q.

58 [February 17, 2023]

Since for any g P Cc pXq, there exists α P C with |α| “ 1 so that |g| “ |αg|
and |Lpgq| “ αLpgq “ Lpαgq, the above gives

|Lpg1 q| ` |Lpg2 q| “ Lpα1 g1 q ` Lpα2 g2 q ď Λpf1 ` f2 q,

and hence Λpf1 q ` Λpf2 q ď Λpf1 ` f2 q. Now let |g| ď f1 ` f2 . We define

" f1 g
if f1 ` f2 ą 0
g1 “ f1 `f2
0 otherwise

and similarly g2 . Then |gi | ď fi and hence

|Lpgq| “ |Lpg1 q ` Lpg2 q| ď Λpf1 q ` Λpf2 q

which gives
Λpf q “ Λpf1 q ` Λpf2 q.
We claim that ˆ
Λpf q “ f dµ.
X
It suffices to consider f P Cc pXq such that 0 ď f ď 1. We approximate f by
step function from below so that
N ´1
1 ÿ 1
0ďf´ χUj ď (3.12)
N j“1 N

with
j
Uj “ tx : f pxq ą
u.
N
By continuity Uj`1 Ă Uj . We approximate the characteristic function χUj
by continuous and compactly supported functions ηj so that supp ηj Ă Uj´1 ,
ηj “ 1 on Uj , N1 N
ř ´1
j“1 ηj ď f and µpsupp ηj zUj q ă 1{N . Hence f can be
approximated by N N
1
ř ´1
j“1 ηj :

N ´1 N ´1 N ´1
1 ÿ ` 1 ÿ ˘ 1 ÿ` ˘ 1
0ďf´ ηj “ f ´ χj ´ ηj ´ χUj ď .
N j“1 N j“1 N j“1 N

Since
´1 ˆ ˆ
ˇ ˇ ˇ ˇ
N N ´1 N ´1 ˇ ˇ
ˇ 1 ÿ ˇ ˇ 1 ÿ ˇ 1 ÿ ˇˇ ˇ
ˇΛpf q ´ ηj dµˇ ď ˇΛpf q ´ Λpηj qˇ ` Λpη η dµ
ˇ ˇ ˇ ˇ
j q ´ j
ˇ
ˇ N j“1 X ˇ ˇ N j“1 ˇ N j“1 ˇ X
ˇ
ˇ ˆ ˇ
1 ˇ ˇ
ď Csupppf q ` sup j ˇΛpηj q ´
ˇ ηj dµˇˇ ,
N X

59 [February 17, 2023]

it suffice to verify ˇ ˆ ˇ
ˇ ˇ
ˇΛpηj q ´ η j dµ ˇ ď 1{N,
ˇ ˇ
X
which follows from µpsupp ηj zUj q ă 1{N and
ˆ
µpUj q ď Λpηj q ď µpsupp ηj q, µpUj q ď ηj dµ ď µpsupp ηj q.
X

Step 3: Final check. Now

ˆ
|Lpf q| ď Λp|f |q “ |f |dµ, @f P Cc pXq.
X

Since Cc pXq is dense in L1 pµq, we can extend L to an element in pL1 pµqq˚ :

For any g P L1 pµq, there exists a sequence pfn q Ă Cc pXq so that fn Ñ g in
L1 pµq, and we define
Lpgq “ lim Lpfn q.
nÑ8
´
This is welldefined since |Lpfn ´ fm q| ď X |fn ´ fm |dµ. It also holds
}L}pL1 pµqq˚ ď 1 since
ˆ ˆ
|Lpgq| “ lim |Lpfn q| ď lim |fn |dµ “ |g|dµ.
nÑ8 nÑ8 X X
8
´
By Corollary 3.25, there exists σ P L pµq such that Lpf q “ X
f σdµ for
f P L1 pµq and
}σ}L8 pµq “ }L}pL1 pµqq˚ ď 1.
We claim that σ “ 1 almost everywhere. By definition
"ˆ *
µpU q “ sup f σdµ “ Lpf q : f P Cc pXq, 0 ď f ď 1, supp f Ă U .
X

We choose a sequence of functions with

ˆ
fj σdµ “ Lpfj q Ñ µpU q.
X
´ ´
Since X fj σdµ ď U |σ|dµ ď µpU q, we deduce σ “ 1 almost everywhere.
To show the uniqueness of the Radon measure, it suffices to show that the
measure of any compact set is unique: Let µ1 , µ2 be two corresponding Radon
measures. For any compact set K and open set U Ą K, there exists a function
f P Cc pU ; r0, 1sq with f “ 1 on K, so that
ˆ ˆ
µ1 pKq ď f dµ1 “ Lpf q “ f dµ2 ď µ2 pU q,
X X

and hence µ1 pKq ď µ2 pKq, which implies µ1 “ µ2 .

60 [February 17, 2023]

3.4.6 Radon-Nikodym Theorem
In this section we restrict ourselves to the Lebesgue measure space.6 The
space Rd is σ compact and the Lebesgue measure is σ finite Radon measure.
Definition 3.43. Let µ and ν be Radon measures on Rd . For x P Rd we
define
#
νpBr pxqq
lim suprÑ0 µpB if µpBr pxqq ą 0 for all r ą 0
Dµ νpxq “ r pxqq
8 if for some r ą 0, µpBr pxqq “ 0
#
νpBr pxqq
lim inf rÑ0 µpBr pxqq
if µpBr pxqq ą 0 for all r ą 0
Dµ νpxq “
8 if for some r ą 0, µpBr pxqq “ 0.
We say that ν is differentiable with respect to µ and x if Dµ νpxq “ Dµ νpxq,
for which we write Dµ νpxq and we call it the density of ν with respect to µ.
We say the measure ν is absolutely continuous with respect to µ, denoted by
ν ăă µ, if µpAq “ 0 implies νpAq “ 0 for any Borel set A. We say that the
measure ν is singular with respect to µ, denoted by ν K µ, if there exists a
Borel set B such that µpXzBq “ νpBq “ 0.
Remark 3.44. Let f P Cc pRd q and µ Radon measure. Then
ˆ
x Ñ f py ´ xqdµpyq

is continuous and hence Borel measurable.

Since the characteristic function of open and closed balls can be obtained as
pointwise limit of continuous functions with compact support, the map

x Ñ νpBr pxqq, x Ñ µpBr pxqq

are measurable. Thus

#
νpBr pxqq
µpBr pxqq
if µpBr pxqq ą 0
xÑ
8 if µpBr pxqq “ 0

is Borel measurable. The map

#
νpBr pxqq
µpBr pxqq
if µpBr pxqq ą 0
rÑ
8 if µpBr pxqq “ 0
6
This section is for interested readers, and the content will not be taken into account
in the exam.

61 [February 17, 2023]

is continuous from the left and right if µpBr pxqq ą 0 by inner and outer
regularity. Thus also Dµ νpxq and Dµ νpxq are Borel measurable since we can
write them as inf ’s and sup’s over rational radii. Moreover by inner and
outer regularity we obtain the same Dµ νpxq and Dµ νpxq if we use closed
balls.

Theorem 3.45 (Radon-Nikodym). Let ν and µ be Radon measures on Rd .

Then

• Dµ νpxq exists and is finite µ almost everywhere.

• If ν ăă µ, then for all Borel set A,

ˆ
νpAq “ Dµ νdµ.
A

Proof. It suffices to consider the case µpRd q ă 8 and νpRd q ă 8. We claim

without proof that for two Radon measures µ, ν on Rd , for all Borel sets B
and all t ą 0,

νpB X tx : Dµ νpxq ă tuq ď tµpB X tx : Dµ νpxq ă tuq

and
νpB X tx : Dµ νpxq ą tuq ě tµpB X tx : Dµ νpxq ą tuq.
Similarly for Dµ ν. It follows from a covering theorem in Rd and regularity
of Radon measures.
We now show the existence of Dµ νpxq. For s ă t we define

Rps, tq “ tx : Dµ νpxq ă s ă t ă Dµ νpxqu.

Then
tµpRps, tqq ď νpRps, tqq ď sµpRps, tqq,
which implies µpRps, tqq “ 0. Since
ď
tx : Dµ νpxq ă Dµ νpxqu “ Rps, tq,
săt,s,tPQ

we see that Dµ νpxq “ Dµ νpxq for µ almost all x. To show that Dµ ν is finite
µ-a.e., let B “ tx : Dµ νpxq “ 8u. Then µpBq “ 0 follows from

νpBq ě tµpBq, @t ą 0.

62 [February 17, 2023]

Now let ν ăă µ, then
νptDµ νpxq “ 8uq “ νptDµ νpxq “ 0uq “ µptDµ νpxq “ 8uq “ µptDµ νpxq “ 0uq “ 0.
Let A be a Borel set. For t ą 1 we define
An “ A X ttn ď Dµ ν ă tn`1 u.
Then
8
ÿ 8
ÿ ˆ 8 ˆ
m`1
νpAq “ νpAm q ď t µpAm q ď t µptDµ ν ą suqds “ t Dµ νdµ
m“´8 m“´8 0 A

and
8
ÿ 8
ÿ ˆ 8 ˆ
m ´1 ´1
νpAq “ νpAm q ě t µpAm q ě t µptDµ ν ą suqds “ t Dµ νdµ.
m“´8 m“´8 0 A

We let now t Ñ 1.
Corollary 3.46 (Lebesgue points). Let µ be a Radon measure on Rd and
f˜ P L1loc pµq. Then
ˆ
f pxq :“ lim µpBr pxqq´1
f˜dµ
rÑ0 Br pxq

exists almost everywhere and we define f pxq “ 0 if it does not exist. Then f
is in the equivalence class of f˜.
Proof. It suffices to consider nonnegative f˜ and µpRd q ă 8. We define
ˆ
νpAq “ f˜ dµ.
A
This is a Radon measure which is absolutely continuous with respect to µ.
Thus ˆ ˆ
νpAq “ Dµ νdµ “ f˜dµ
A A
and Dµ ν lies in the equivalence class. Now the first claim follows from The-
orem 3.45.

Remark 3.47. Similarly if f˜ P Lploc pµq, p P r1, 8q, then f P Lploc and almost
everywhere
ˆ
lim µpBr pxqq ´1
|f pyq ´ f pxq|p dµpyq “ 0.
rÑ0 Br pxq

Then there exists a cannonical representative of the equivalence class.

[23.12.2022]
[09.01.2023]

63 [February 17, 2023]

4 Distributions and Sobolev spaces
In this chapter we restrict ourselves to Euclidean space Rd , and we are going
to introduce distributions and Sobolev spaces defined on open sets in Rd .
Recall the Riesz representation theorem describing the (continuous) dual
space X ˚ consisting of (continuous) linear functionals x˚ P X ˚ 7 from some
topological K-vector space X to K:
• Theorem 2.25 for Hilbert spaces H: Any linear functional on H can
be uniquely identified with an element in H itself, in the sense that
J : H Q x ÞÑ py ÞÑ xy, xyq P H ˚ is an antilinear isomorphism (isometric
bijective map).

• Theorem 3.24 (and Corollary 3.25) for the Lebesgue spaces Lp pRd q,
1
1 ď p ă 8: Any function in Lp pRd q can be uniquely identified with
1
a linear
´ functional in pLp pRd qq˚ , in the sense that j : Lp pRd q Q g ÞÑ
pf ÞÑ Rd f gdmd q P pLp pRd qq˚ is a linear isomorphism.
Here the Lebesgue spaces read

Lp pRd q “ tf : Rd Ñ K is p ´ integrableu{ „, 1 ď p ď 8,

equipped with the norm }f }Lp , where „ denotes the equivalence rela-
tion between two measurable functions which differ only on a measure-
2 d
zero set. In particular, the Lebesgue space
´ L pR q dis a Hilbert space,
equipped with the inner product xf, gy “ Rd f ḡdm .

• Theorem 3.41 for the continuous compactly supported function space

d ˚
d 8
` ˘
Cc pR q: Any nonnegative continuous linear map L P Cc pR q is
represented
´ uniquely by a Radon measure µ on Rd , in the sense that
Lpf q “ Rd f dµ holds for all f P Cc pRd q.
Here the continuous bounded function space

Cb pRd q “ tf : Rd Ñ K continuous and boundedu,

is equipped with the norm }f }Cb “ sup xPRd |f pxq|, and its subspace

Cc pRd q “ tf : Rd Ñ K continuous and with compact supportu,

7
In the following we also denote by X 1 the dual space of X.
8
If fn Ñ f in Cc pRd q means that there exists a compact set K Ă Rd such that
supppfn q Ă K for all n and }fn ´ f }Cb pRd q Ñ 0, then a continuous map L : Cc pRd q ÞÑ K
means that Lpfn q Ñ Lpf q holds for all fn Ñ f in Cc pRd q. Thus a continuous linear map
on Cc pRd q is bounded on Cb pKq for all compact set K Ă Rd (with the proof the same as
Lemma 4.9 below).

64 [February 17, 2023]

 is a vector subspace but not a closed subset of Cb pRd q. Cc pRd q is a dense
1
subset in Lp pRd q, 1 ď p ă 8, and any nonnegative Lp pRd q-function
g can then be viewed as a Radon measure gdmd on Rd , which defines
(uniquely) a (continuous) linear functional on Cc pRd q.
This motivates to generalize the definition of “functions” as continuous linear
maps on the test function space Cc8 pRd q. The vector space Cc8 pRd q is not a
Banach space, and its topology is complicated, which we do not discuss in
detail here.
We start with some celebrated theorems for continuous linear maps between
two Banach spaces in Section 4.1, and we will see that the ideas can apply
in Section 4.2.

4.1 Baire category theorem, Banach-Steinhaus theo-

rem and open mapping theorem
Theorem 4.1 (Baire category theorem). A countable intersection of dense
open subsets of a complete metric space is dense.
Proof. Let pX, dq be a complete metric space and Aj open dense sets. Let
x P X and ε ą 0. Take x1 P A1 so that

dpx, x1 q ă δ0 :“ ε,

and take δ1 ą 0 with 2δ1 ă δ0 such that

B2δ1 px1 q Ă A1 X Bδ0 pxq.

Recursively given pxk , δk qn´1

k“1 , there are xn P An , δn ą 0 with 2δn ă δn´1 such
that

dpxn´1 , xn q ă δn´1 , B2δn pxn q Ă An X Bδn´1 pxn´1 q.

We arrive at a sequence pδn q:

1
δn ă δn´1 ă ¨ ¨ ¨ ă 2´n δ0 “ 2´n ε,
2
and a sequence pxn q such that for all m ą n,
m´n´1
ÿ m´n´1
ÿ
dpxn , xm q ď dpxn`j , xn`j`1 q ă δn`j ă 2´n`1 ε,
j“0 j“0
m
č
Bδm pxm q Ă B2δm pxm q Ă Am X Bδm´1 pxm´1 q Ă ¨ ¨ ¨ Ă Aj X Bδn´1 pxn´1 q.
j“n

65 [February 17, 2023]

Thus pxn q is a Cauchy sequence, which converges to a limit y, such that
y P Bδn pxn q, @n, and
č
dpx, yq ď dpx, x1 q ` dpx1 , yq ă 2ε, y P Aj .
j
Ş
Thus j Aj is dense in pX, dq.

Recall the definition of Banach spaces (i.e. complete normed spaces) in

Section 1.1, and the set LpX, Y q consisting of all continuous linear maps
between two normed spaces X, Y in Section 1.3.

Theorem 4.2 (Banach-Steinhaus, Principle of uniform boundedness). Let

X and Y be Banach spaces, F Ă LpX, Y q. If for each x P X

sup t}T x}Y : T P Fu ă 8,

then
sup t}T }XÑY : T P Fu ă 8.
In particular, for a sequence of continuous linear maps tTn u Ă LpX, Y q such
that

Tn x converges in Y, @x P X,

we have

sup n t}Tn }XÑY u ă 8.

Proof. Let
Cn “ tx P X : sup T PF }T x}Y ď nu.
This set is closed since both T andŤ}} are continuous, and Cn is an intersection
of closed sets. By assumption Cn “ X. We claim that some Cn has
nonempty open interior. If not then the sets Un “ XzCn are open and dense,
andŤhence with nonempty intersection by Lemma 4.1. This is a contradiction
to Cn “ X. Let Br px0 q Ă Cn0 be an open ball. If }x} ă r then

}T x}Y ď }T px ´ x0 q}Y ` }T px0 q}Y ď n0 ` sup T PF }T px0 q}Y “: R.

Then
}T }XÑY ď R{r
for all T P F.

66 [February 17, 2023]

Remark 4.3. The assumption that Y is Banach space can be relaxed to
normed space.
Obviously for normed spaces X, Y , the uniform boundedness of the operator
norms
sup T PF t}T }XÑY u ď c0 ă 8
implies the (uniform) boundedness of the application on some fixed element
x P X:
sup T PF t}T x}Y u ď c0 }x}X ă 8.
The two boundedness are equivalent if X is Banach spaces. Given an example
to show the failure of the equivalence if X is not complete. (Exercise)

[09.01.2023]
[13.01.2023]

The Baire category theorem has further interesting consequences. The fol-
lowing open mapping theorem shows the equivalence between surjectivity
and open property of a continuous linear map between two Banach spaces.

Theorem 4.4 (Open mapping theorem). Let X and Y be Banach spaces

and T P LpX, Y q. T is surjective if and only if it is open, i.e. if the image
of open sets is open.

Proof. Let T be open. Ť Then T pB1X p0qq

Ť is open. In particular it contains a
Y Y X
ball Br p0q. Then Y “ n Bn p0q “ n T pBn{r p0qq and T is surjective.
Now suppose that T is surjective. It suffices to show that T pB1X p0qq contains
some ball BεY p0q, since for any open ball U Ă X and any x P U , there exists
BδX pxq Ă U such that Bεδ Y
pT xq Ă T U :
Y Y
Bεδ pT xq “ T x ` Bεδ p0q “ T x ` δBεY p0q Ă T x ` δT pB1X p0qq “ T pBδX pxqq Ă T U.

Let

Yn :“ T pBnX p0qq “ tT x|}x}X ă nu “ ntT x | }x}X ă 1u “ nY1 .

Ť
It is closed, Yn Ă Yn`1 and Y “ Yn . As in the proof of Theorem 4.2 we
conclude that one (and hence all) of the Yn contains an open ball. Hence
there exists BrY py0 q Ă Y1 , and thus BrY p0q Ă Y2 , since

y P BrY p0q ùñ y ` y0 P BrY py0 q Ă Y1 “ T pB1X p0qq

ùñ y P T pB1X p0qq ´ T pB1X p0qq Ă 2T pB1X p0qq “ Y2 .

67 [February 17, 2023]

In particular, BrY p0q Ă T pB4X p0qq, since for any y P Y2 “ T B2X p0q,

Dx1 P B2X p0q s.t. y ´ T x1 P B2Y´1 r p0q Ă T B1X p0q,

and inductively

Dxn P B2X2´n p0q s.t. y ´ T x1 ´ ¨ ¨ ¨ ´ T xn P B2Y´n r p0q Ă T B2X1´n p0q,

ř
which implies y “ T x with x :“ n xn P B4X p0q.
The following corollary implies that the inverse map of an invertible contin-
uous linear map is also continuous linear map.

Corollary 4.5. Let X, Y be two Banach spaces. Suppose that T P LpX, Y q

is bijective. Then T ´1 P LpY, Xq.

Proof. Linearity of the inverse map follows immediately from the linearity
and injectivity of T . By Theorem 4.4 T is open. So T pB1X p0qq contains a
ball BrY p0q and hence
1
@y P BrY p0q, Dx P B1X p0q s.t. y “ T x, that is, }x}X “ }T ´1 y}X ď }y}Y .
r
T ´1 is a linear and bounded map.

Corollary 4.6. The set of nowhere differentiable functions in Cb p0, 1q is

dense.

Proof. Exercise.

4.2 Distributions
In this section, we are going to introduce the definition of the distribution
space first, and then to generalize the definitions of product, differentiation,
support and convolution from test function space to distribution space by
the idea of duality. Finally, we are going to approximate a distribution by
test functions.

4.2.1 Definitions
Definition 4.7 (Test function space and distribution space). Let U Ă Rd be
open.
We denote by DpU q “ Cc8 pU q the vector space of all test functions on U ,
which are infinitely differentiable functions with compact support in U .

68 [February 17, 2023]

We call φj Ñ φ in Cc8 pU q “ DpU q if there is a compact set K Ă U so that
supp φj Ă K for all j, and for all multiindices α

B α φj Ñ B α φ in Cb pU q.

A distribution T on U is a continuous linear map from Cc8 pU q Ñ K, where

by continuity we mean

T φj Ñ T φ, if φj Ñ φ in Cc8 pU q.

We denote the space of distributions by D1 pU q, (i.e. the dual space of DpU q),
which is a K-vector space.
We call Tn Ñ T in D1 pU q if

Tn φ Ñ T φ, @φ P DpU q.

Remark 4.8. The test function space DpU q is not a Banach space. Nev-
ertheless, one can introduce a countable family of seminorms on DpU q such
that it becomes a Fréchet space. For example, for U “ R, we define a se-
quence of seminorm (i.e. a function a ÞÑ ras from a K-vector space to R
satisfying ras ě 0, rc ¨ as “ |c|ras and ra ` bs ď ras ` rbs)

rφsk,n “ }φpkq }Cb pr´n,nsq , k, n P N Yt0u,

and introduce a metric

ÿ rφ ´ ψsk,n
d1 pφ, ψq “ 2´pk`nq ,
k,n
1 ` rφ ´ ψsk,n

such that pDpRq, d1 q becomes a metric space (but not complete). The con-
vergence φj Ñ φ in DpRq (defined above) is not induced by this metric, and
hence T P D1 pU q is not continuous on pDpU q, d1 q.
Fréchet spaces can be viewed as generalization of Banach spaces, but with
more complicated topological structure. Nevertheless, Banach-Steinhaus the-
orem and the open mapping theorem still hold on Fréchet spaces.

When restricted on a compact set, the definition of a distribution becomes

more clear. For any fixed compact set K Ă U , we consider the subspace of
DpU q:
XK “ tφ P DpU q : supp φ Ă Ku.
We define a metric on XK

dK pφ, ψq “ sup kě0 2´k mint1, }φ ´ ψ}Cbk pU q u.

69 [February 17, 2023]

With this metric XK is a complete metric space:

dK pφj , φq Ñ 0 iff φj Ñ φ in Cbk pU q, @k ě 0.

Hence if T P D1 pU q and dK pφj , φq Ñ 0, then T φj Ñ T φ, and T determines

a continuous linear map on XK .
[13.01.2023]
[16.01.2023]
Furthermore we have the following description.
Lemma 4.9. Let T P D1 pU q. For every compact set K Ă U there exists k
and C ą 0 so that
|T φ| ď C}φ}Cbk pU q , @φ P XK .

Proof. Since T P D1 pU q defines a continuous linear map from pXK , dq Ñ K,

|T φ| ă 8, @φ P XK .

Now we argue as for the uniform boundedness principle of Banach-Steinhaus:

There exists m so that the set

Cm :“ tφ P XK : |T φ| ď mu

contains an open ball Br pφ0 q. Hence the set C2m contains an open ball Br p0q,
since

@φ P Br p0q, |T φ| ď |T pφ ´ φ0 q| ` |T φ0 | ď m ` m “ 2m.

Let k0 be such that 2´k0 ă r. If φ P DpU q such that }φ}C k0 pU q ď 2´k0 and
b
supp φ Ă K, then

dpφ, 0q ď max tsup kďk0 2´k }φ}C k0 pU q , sup kąk0 2´k u ă r,

b

and hence
|T φ| ď 2m, @φ P XK with }φ}C k0 pU q ď 2´k0 .
b

1`k0
The result follows with C “ m2 .
Lemma 4.10. Suppose that Tn Ñ T in D1 pU q and that K Ă U is compact.
Then there exists k and C so that

sup n |Tn pφq| ď C}φ}Cbk pU q , @φ P XK ,

and
sup t|Tn pf q ´ T pf q| : f P XK , }f }Cbk ď 1u Ñ 0.

70 [February 17, 2023]

Proof. Exercise.
D1 pU q is a rather big space, which includes almost all the interesting “func-
tions”. For example, CpU q, C k pU q, Lp pU q, p P r1, 8s are all subspace of
L1loc pU q, which is subspace of D1 pU q:
Example 4.11 (L1loc pU q Ă D1 pU q). Let

L1loc pU q “ tf : U Ñ K measurable | f P L1 pKq for all K Ă U compactu,

which is a K-vector space. We call fj Ñ f in L1loc pU q if

fj χK Ñ f χK in L1 pKq, @ compact set K Ă U.

(Exercise:) Any f P L1loc pU q determines a unique distribution Tf P D1 pU q

given by
ˆ
Tf pφq “ f φ dmd , @φ P DpU q.
U

In the following we will identify an L1loc pU q function f with the distribution

Tf P D1 pU q.
Similarly, any Radon measure µ on U defines a unique distribution T µ P
D1 pU q:
ˆ
µ
T pφq “ φdµ, @φ P DpU q.
U

4.2.2 Products, derivatives and supports

One can easily define the product between two smooth functions: If φ P Dpuq
and ϕ P C 8 pU q then ϕφ P DpU q, and
ϕPC 8 pU q
φj Ñ φ in DpU q ùñ ϕφj Ñ ϕφ in DpU q.

Hence for any T P D1 pU q and ϕ P C 8 pU q, T pϕ ¨q : DpU q Ñ K determines a

continuous linear map, which we define as the product between T and ϕ.
Similarly, if φ P DpU q, then Bxk φ P DpU q, and φj Ñ φ in DpU q implies
Bxk φj Ñ Bxk φ in DpU q. Hence for any T P D1 pU q, the map T p´Bxk ¨q :
DpU q Ñ K determines a distribution in D1 pU q, which we define as Bxk T .
For φ P DpU q, its zero points are those where it vanishes, and we define
its support as supppφq “ tx P U | φpxq ‰ 0u. For T P D1 pU q, we say that T
vanishes near x P U if there exists r ą 0 so that T pφq “ 0 for all φ P DpU q
with support in Br pxq.

71 [February 17, 2023]

Definition 4.12 (Products, derivatives and supports). We define the product
ϕT P D1 pU q of a distribution T P D1 pU q and a smooth function ϕ P C 8 pU q
as
pϕT qpφq “ T pϕφq, @φ P DpU q.
We define the derivative Bxj T P D1 pU q of a distribution T P D1 pU q as

pBxj T qpφq “ T p´Bxj φq, @φ P DpU q.

We define the support of T P D1 pU q as the complement of the points near

which T vanishes. We denote by E 1 pU q the set of distributions with compact
supports.
We can easily check
• Associative law: ϕpψT q “ ψpϕT q “ pϕψqT , @T P D1 pU q, ϕ, ψ P C 8 pU q;

• Distributive law: ϕpT ` Sq “ ϕT ` ϕS, @T, S P D1 pU q, @ϕ P C 8 pU q;

• Schwarz theorem: Bxj Bxk T “ Bxk Bxj T , @T P D1 pU q;

• Leibniz’ product rule: Bxj pϕT q “ pBxj ϕqT ` ϕpBxj T q, @ϕ P C 8 pU q,

T P D1 pU q:

Bxj pϕT qpφq “ ´T pϕBxj φq “ T ppBxj ϕqφq ´ T pBxj pϕφqq

“ rpBxj ϕqT spφq ` pϕBxj T qpφq.

• Coincidence with the classical definitions: If T “ Tϕ P D1 pU q, with

ϕ P DpU q Ă L1loc pU q, then

ψTϕ “ Tϕψ for ψ P C 8 pU q,

TBxj ϕ “ Bxj Tϕ .

We first have the following approximation result for the constant function 1
on compact sets.
Lemma 4.13 (Local approximation of 1). Let U Ă Rd be an open set, and
Kj be a monotone sequence
Ť of compact sets so that Kj is contained in the
interior of Kj`1 and U “ Kj . Then there exists a sequence of test functions
pϕj q Ă DpU q so that

supp ϕj Ă Kj`1 , ϕj “ 1 on Kj´1 , and B α ϕj Ñ B α 1 on compact sets.

In particular, for any φ P DpU q, ϕj φ Ñ φ in DpU q, and hence for any

T P D1 pU q, ϕj T Ñ T in D1 pU q, where supppϕj T q Ă supppϕj q.

72 [February 17, 2023]

Proof. For any j, let rj ą 0 so that

mintdistpKj´1 , Rd zKj q, distpKj , Rd zKj`1 qu ě rj ą 0.

´
Let η P Cc8 pB1 p0qq be radial with Rd φdx “ 1, and ηr pxq “ r´d ηpx{rq, as in
Corollary 3.39. Then
ϕj :“ ηrj ˚ χKj P DpU q,
is the searched sequence: For any compact set K Ă U , there exists j0 such
that K Ă Kj0 ´1 and ϕj “ 1 on K for all j ě j0 .

Unlike the product of two continuous functions, the product of two distribu-
tions are not always well-defined. But one can always define the derivatives
of a distribution.

Example 4.14 (Bx H “ δ0 ). Let"U “ R.

0 if x ă 0
The Heaviside function Hpxq “ belongs to L8 pU q Ă L1loc pU q Ă
1 if x ě 0
D1 pU q. We calculate its derivative: For φ P DpU q,
ˆ 8
1
pBx Hqpφq “ pBx TH qpφq “ ´TH pφ q “ ´ Hpxqφ1 pxqdx “ φp0q “ δ0 pφq,
0

and hence
Bx H “ δ0 ,
where δ0 P D1 pRq is the Dirac measure, which belongs to E 1 pRq with compact
support supppδ0 q “ t0u.
One can further calculate Bx δ0 P D1 pRq as

pBx δ0 qpφq “ ´δ0 pφ1 q “ ´φ1 p0q, @φ P DpRq,

and iteratively pBxn δ0 qpφq “ p´1qn φpnq p0q.

[16.01.2023]
[20.01.2023]

Lemma 4.15 (Disjoint supports). Let φ P DpU q and T P D1 pU q with disjoint

supports. Then
T φ “ 0.

73 [February 17, 2023]

Proof. Let K be the support of φ. Given x P K there exists r so that T ψ “ 0
for every ψ P DpU q supported in Br pxq. Since K is compact there is a finite
covering of such balls Brj pxj q with 1 ď j ď N . We choose a partition
of unity ηj P C 8 pU q supported in Brj pxj q so that (recalling the proof of
Theorem 3.41)
ÿN
ηj pxq “ 1, @x P K,
j“1

Then
N
ÿ
Tφ “ T pηj φq “ 0.
j“1

4.2.3 Convolution in Rd and density of DpU q in D1 pU q

Definition 4.16 (Convolution between DpRd q and D1 pRd q). Let T P D1 pRd q
and ϕ P DpRd q. We define their convolution by

pϕ ˚ T qpxq “ T pϕpx ´ ¨qq, @x P Rd .

The righthand side denotes T acting on ϕpx ´ yq as a function of y.

Lemma 4.17. With the notation above, we know ϕ ˚ T P C 8 pRd q and

Bxj pϕ ˚ T q “ pBxj ϕq ˚ T “ ϕ ˚ pBxj T q.

If ψ P L1 pRd q then
ϕ ˚ Tψ pxq “ ϕ ˚ ψpxq.
Moreover, if supp ϕ “ K1 and supp T “ K2 then

supppϕ ˚ T q Ă K1 ` K2 “ tx ` y : x P K1 and y P K2 u.

In particular, if ϕ P DpRd q and T P E 1 pRd q, then ϕ ˚ T P DpRd q.

Proof. Exercise.

Remark 4.18. We can equivalently define the convolution of ϕ P DpRd q and

T P D1 pRd q as the distribution:

pϕ ˚ T qpφq “ T pϕ̃ ˚ φq, @φ P DpRd q,

74 [February 17, 2023]

where ϕ̃pxq “ ϕp´xq, since by the linearity of T ,
ˆ ´ˆ ¯
d d
pϕ ˚ T qpφq “ T pϕpx ´ ¨qqφpxqdm “ T ϕ̃p¨ ´ xqφpxqdm “ T pϕ̃ ˚ φq.
Rd Rd

In the same way we can easily define the convolution ϕ ˚ T P D1 pRd q when
ϕ P C k pRd q has compact support. By an abuse of notation we write the
convolution evaluated at x whenever it is defined, even if it is not defined on
all of Rd .
Definition 4.19 (Convolution between D1 pRd q and E 1 pRd q). Let T P D1 pRd q
and let S P E 1 pRd q which consists of distributions with compact support. We
define their convolution by
pS ˚ T qpϕq “ T pS̃ ˚ ϕq
for ϕ P DpRd q. Here S̃pψq “ Spψ̃q, @ψ P DpRd q.
Example 4.20 (Convolution with δ0 ). Recall the Dirac sequence pηr qrą0 in
Corollary 3.39, such that for any f P Lp , 1 ď p ă 8, we have ηr ˚ f Ñ f in
Lp .
For any φ P DpU q, ηr ˚ φ Ñ φ in DpU q. Similarly for any f P Cck pU q, there
exist compact set K Ă U and ε ą 0 such that supppηr ˚ f q is contained in K
for all r ď ε, and ηr ˚ f Ñ f in Cbk pU q.
Show that δ0 P E 1 pRd q and δ0 ˚ T “ T for all T P D1 pRd q. In particular
ηr Ñ δ0 , and ηr ˚ T Ñ T in the sense of distribution. (Exercise).
Recall Lemma 4.13 that any distribution T can be approximated by a se-
quence of distributions with compact supports ϕj T . Lemma 4.17 and the
above example tells us that any distribution with compact support S P E 1 pU q
can be approximated by pηr ˚ Sqr Ă DpU q in the distribution sense. Hence
we have proved
Theorem 4.21. Let U Ă Rd be open. Then DpU q Ă D1 pU q is dense.
By approximation argument we can also prove
Corollary 4.22. Suppose that U Ă Rd is an open connected set and Bxj T “
0, j “ 1, ¨ ¨ ¨ , d. Then there eixsts a constant c so that T “ Tc .
Proof. Exercise.

4.3 Schwartz functions and tempered distributions

We briefly cover the definitions of Schwartz functions and tempered distribu-
tions, which are the proper functional framework for the Fourier transform
on the whole space Rd .

75 [February 17, 2023]

4.3.1 Definitions
Definition 4.23. The Schwartz space SpRd q consists of all Schwartz func-
tions, which are smooth functions f P C 8 pRd q satisfying for any k P N

}f }k,S :“ sup xPRd , |α|ďk p1 ` |x|k q|B α f pxq| ă 8,

which is equivalent to say

SpRd q “ tf P C 8 pRd q | sup xPRd |xα B β f pxq| ă 8, @ multiindices α, βu.

We define d : S ˆ S Ñ r0, 8q by

dpf, gq “ sup k 2´k mint1, sup |α|`|β|“k }xα B β pf ´ gq}sup u,

such that pS, dq is a complete metric space.

A tempered distribution is a continuous linear map from pSpRd q, dq to K. We
denote the space of tempered distributions by S 1 pRd q. We say that Tj Ñ T
in S 1 pRd q if Tj pϕq Ñ T pϕq for all ϕ P SpRd q.

It is straightforward to check that

• For f P S, }f }k,S may depend on k P N, and for any N there exists CN

such that }f }CbN ď CN and |f pxq| ď CN p1 ` |x|q´N ;

• DpRd q Ă SpRd q;

• SpRd q Ă C k pRd q for all k P N and SpRd q Ă Lp pRd q for all p P r1, 8s;

• If f P SpRd q, then xα f, B α f P SpRd q for any multiindex α;

• If g P C 8 and for any multiindex α there exist c|α| and κ|α| so that

|B α g| ď c|α| p1 ` |x|qκ|α| ,

then f P SpRd q implies f g P SpRd q;

• If f, g P SpRd q, then f ˚ g P SpRd q;

1 2
• The Gaussian function e´ 2 |x| P SpRd q;

• E 1 pRd q Ă S 1 pRd q, since for any S P E 1 pRd q, Spf q “ Spηf q, @f P C 8 pRd q

and η P DpRd q with η “ 1 on supppSq.

Similar as Lemma 4.9 and Lemma 4.10, we have (Exercise)

76 [February 17, 2023]

Lemma 4.24. Let T P S 1 pRd q. Then there exist k and c so that

|T ϕ| ď c}ϕ}k,S , @ϕ P SpRd q.

If Tn Ñ T in S 1 pRd q then there exist C and l so that

sup n |Tn ϕ| ď C}ϕ}l,S ,

and
sup t|Tn pϕq ´ T ϕ| | ϕ P SpRd q, }ϕ}l,S ď 1u Ñ 0.

[20.01.2023]
[23.01.2023]

For T P S 1 pRd q, we define the multiplication by a smooth function with

controlled derivatives gT and the derivative Bxj T as we did it for distributions
in Definition 4.12.
Similarly, since compactly supported distributions S P E 1 pRd q can act on
Schwartz functions f P SpRd q, we can define the convolution pS ˚ f qpxq “
Spf px ´ ¨qq P SpRd q. We then can define, as in Definition 4.16 and Definition
4.19, the convolution of a tempered distribution T P S 1 pRd q with Schwartz
functions f P SpRd q: pf ˚ T qpxq “ T pf px ´ ¨qq P C 8 pRd q, and with compactly
supported distributions S P E 1 pRd q: pS ˚ T qpϕq “ T pS̃ ˚ ϕq, @ϕ P SpRd q.
By approximation argument, one has (Exercise)

Theorem 4.25. For 1 ď p ď 8, there are the embeddings

DpRd q Ă SpRd q Ă Lp pRd q Ă S 1 pRd q Ă D1 pRd q.

The embeddings are dense if p ă 8.

4.3.2 Fourier transformation

Recall the definition of the Fourier transform Fpf q of a Lebesgue integrable
function f P L1 pRd ; Cq:
ˆ
ˆ 1
f pξq :“ Fpf qpξq “ d e´ix¨ξ f pxq dx , @ξ P Rd .
p2πq R
2 d

By the density of the embedding DpRd q Ă L1 pRd q, one can easily show
Riemann-Lebesgue theorem:

F : L1 pRd q Ñ C0 pRd q Ă Cb pRd q.

77 [February 17, 2023]

The Fourier transform F : L1 pRd q Ñ Cb pRd q is a linear continuous map and
1 1
}Fpf q}Cb pRd q ď d }f }L1 pRd q , i.e. }F}L1 pRd qÞÑCb pRd q ď d .
p2πq 2 p2πq 2

Since SpRd q Ă L1 pRd q, we can define the Fourier transform on SpRd q:

Theorem 4.26. The Fourier transform maps continuously from SpRd q to

SpRd q: For any integer k P N, there exist a constant C and an integer N P N
such that

}Fpf q}k,S ď C}f }N,S , @f P SpRd q.

Furthermore, for any f P SpRd q,

1
Fpp Bx qα f q “ ξ α Fpf q, Fpxα f q “ piBξ qα Fpf q, @multiindex α. (4.1)
i
Proof. Exercise.
Roughly speaking, the Schwartz functions have two properties: they have
infinite bounded derivatives and they decay fast at infinity. In view of (4.1),
the Fourier transform works well in the framework of Schwartz space, and
hence of tempered distributions by duality: We define the Fourier transform
of a tempered distribution T P S 1 pRd q as a tempered distribution FpT q P
S 1 pRd q:
FpT qpf q “ T pFpf qq, @f P SpRd q.
We have

Corollary 4.27. The Fourier transform F : S 1 pRd q Ñ S 1 pRd q is a continu-

ous linear map, and (4.1) holds on S 1 pRd q.

Example 4.28. We calculate the Fourier transform δˆ0 of the delta function
δ0 P E 1 pRd q Ă S 1 pRd q:
ˆ
ˆ ˆ ´ d2
δ̂0 pf q “ δ0 pf q “ f p0q “ p2πq f pξq dξ “ T ´ d2 pf q, @f P SpRd q,
p2πq
Rd

that is, the Fourier transform of the delta function is a constant function:
d
Fδ0 “ p2πq´ 2 .

By (4.1), we have
d
α δ “ p2πq´ 2 piξqα .
By 0

78 [February 17, 2023]

Since (see Example 4.11) any function f P Lp pRd q, p P r1, 8s can be identified
with a tempered distribution Tf P S 1 pRd q which is defined as
ˆ
Tf pϕq “ f ϕ dx , @ϕ P SpRd q,
Rd

we can define the Fourier transform of f as the following tempered distribu-

tion
Fpf qpϕq “ Tf pFpϕqq, @ϕ P SpRd q.
We have seen F : L1 pRd q Ñ L8 pRd q. We have further (Exercise)
Theorem 4.29. The Fourier transform defines a unitary operator from
L2 pRd q to L2 pRd q, i.e. F is continuous linear map satisfying F ˚ F “ FF ˚ “
Id . In particular, the Fourier transform of f P L2 pRd q is a function in
L2 pRd q which reads as the following limit (almost everywhere)
ˆ
1
Fpf qpξq “ lim d e´ix¨ξ f pxq dx . (4.2)
RÑ8 p2πq 2 |x|ăR

By Riesz-Thorin Interpolation theorem (see Exercise sheet),

1
F : Lp pRd q Ñ Lp pRd q is continuous and linear, @p P r1, 2s,
´ d2 p p2 ´1q
such that }Fpf q}Lp1 pRd q ď p2πq }f }Lp pRd q .

In general the Fourier transform of a Lp pRd q, p ą 2-function does not neces-

sarily belong to Lebesgue spaces: We keep in mind that the Fourier transform
d
of the constant function f pxq “ p2πq´ 2 is the Dirac-function δ0 P E 1 pRd q. We
remark that one can define the inverse Fourier transform as F ´1 pT q “ FpTrq
on S 1 pRd q, such that F ´1 F “ FF ´1 “ Id on S 1 pRd q.
[23.01.2023]
[27.01.2023]

4.4 Sobolev spaces W k,p

4.4.1 Definitions
Definition 4.30. Let U Ă Rd be open, k P N.
• For 1 ď p ď 8, the Sobolev space W k,p pU q Ă Lp pU q is the set of all
Lp pU q functions, so that for all multiindices α of length at most k there
exists fα P Lp pU q so that

B α Tf “ Tfα .

79 [February 17, 2023]

 By an abuse of notation we identify B α Tf with fα , and hence g “ Bxj f
in Lp pU q for f P Lp pU q means
ˆ ˆ
gϕdm “ ´ f Bxj ϕdmd , @ϕ P DpU q.
d
(4.3)
U U

For p “ 2, we denote H k pU q “ W k,2 pU q.

• We define for p P r1, 8q

¨ ˛ p1
ÿ
}f }W k,p pU q “ ˝ }B α f }pLp pU q ‚ ,
|α|ďk

and
}f }W k,8 pU q “ sup |α|ďk }B α f }L8 pU q .
Then pW k,p pU q, }}W k,p pU q q is normed space.
For p “ 2, we define
ÿ
xf, gyH k pU q “ xB α f, B α gyL2 pU q .
|α|ďk

Then pH k pU q, x, yH k pU q q is pre-Hilbert space.

• For 1 ď p ă 8, we define W0k,p pU q as the closure of Cc8 pU q with respect

to the norm } ¨ }W k,p pU q .
1
• For 1 ă p ă 8, we define W ´k,p pU q “ pW0k,p pU qq˚ . The norm is
defined by

}T }W ´k,p pU q “ sup f PW k,p1 pU q, }f } 1 ď1

|T pf q|.
0 W k,p pU q

Remark 4.31. • If V Ă U open, then the restriction map

χV : W k,p pU q Q f ÞÑ f |V P W k,p pV q

is a continuous linear map with the norm less than 1: We have B α pf |V q “

pB α f q|V and }f |V }W k,p pV q ď }f }W k,p pU q for all f P W k,p pU q.
The (trivial) extension by 0
"
f inside V
W0k,p pV q Q f ÞÑ f˜ “ P W0k,p pU q
0 outside V

80 [February 17, 2023]

 defines a canonical (nonsurjective) isometry from W0k,p pV q to W0k,p pU q:
}f }W k,p pV q “ }f˜}W k,p pU q .
Hence the restriction map χV : W k,p pU q Ñ W k,p pV q has indeed norm
1.
• We denote Lp “ W 0,p . Obviously the embedding W k1 ,p pU q Ă W k2 ,p pU q
holds if k2 ď k1 and p P p1, 8q.
• One can check (Exercise)
– If g P Cbk pU q and f P W k,p pU q, then gf P W k,p pU q and the Leibniz’
rule holds;
– For f P W k,p pU q and ϕ : V Ñ U a Cbk diffeomorphism, the chain
rule holds and
}f ˝ ϕ}W k,p pV q ď C}f }W k,p pU q . (4.4)
Theorem 4.32. Let U Ă Rd be open, k P N and 1 ď p ď 8.
The Sobolev space pW k,p pU q, }}W k,p pΩq q is a Banach space. In particular for
p “ 2, pH k pΩq, x, yH k pΩq q is a Hilbert space.
}}W k,p pRd q
If U “ Rd and 1 ď p ă 8, then W0k,p pRd q “ W k,p pRd q “ Cc8 pRd q .
Proof. Let Σk “ tα : |α| ď ku. There is an obvious isometry (nonsurjective)
W k,p pU q Q f Ñ pB α f q|α|ďk P Lp pU ˆ Σk q,
where
}f }W k,p pU q “ }pB α f q}Lp pU ˆΣk q .
Let fj be a Cauchy sequence in W k,p pU q with limit f P Lp pU q. Then B α fj Ñ
fα in Lp pU q and hence in D1 pU q. It is easy to check that fα “ B α f in D1 pU q:
For any ϕ P DpU q,
ˆ ˆ
α d
Tfα pϕq “ lim p´1q |α|
fj B ϕ dm “ p´1q |α|
f B α ϕ dmd “ TBα f pϕq.
jÑ8 U U

Thus f P W pU q and W pU q is complete and we can identify W k,p pU q

k,p k,p

with a closed subspace of Lp pU ˆ Σk q.

[27.01.2023]
[30.01.2023]

Now let 1 ď p ă 8. Density of DpRd q Ă W k,p pRd q follows by the same argu-
ment (first smooth cutoff and then convolution) as in the proof of Theorem
4.21.

81 [February 17, 2023]

Remark 4.33. • If U Ă Rd is bounded, we can also follow the argument
as in the proof of Theorem 4.21 to show that any function f P W k,p pU q,
k P N, p P r1, 8q can be approximated by a sequence of test functions
k,p
ηrj ˚ pϕj f q in Wloc pU q, i.e. in W k,p pKq, for all compact sets K Ă
U . The approximation holds only locally, otherwise we have shown
W k,p pU q “ W0k,p pU q, which is in general not true for bounded open set
U (keeping in mind of the constant function 1 P W 1,p pU q but not in
W01,p pU q, p ą d, where W 1,p -norm is finer than Cb -norm).

• In order to show the global approximation by smooth functions, we need

some regularity assumption on the boundary of U . For example, if U is
an open bounded Lipschitz domain, then f P W k,p pU q, k P N, p P r1, 8q
can be approximated by a sequence of smooth functions fn P C 8 pU q
(globally) in W k,p pU q, see Corollary 4.36 below.

• We may interpret W0k,p pU q as the set of f P W k,p pU q such that “B α U “

0 on BU , @|α| ď k´1.” Let U Ă Rd be bounded and open and k P N. We
assume that f P Cbk pU q and its derivatives of order up to k ´ 1 extend
to continous functions in Ū which vanish at BU . Then f P W0k,p pU q for
1 ď p ă 8.

• The map J : Lp pU ˆ Σk q Ñ W ´k,p pU q defined below has norm 1:

ÿ ˆ 1
Jppfα qqpuq “ fα B α u dmd , @u P W0k,p pU q,
|α|ďk U

1
where u P W0k,p pU q can be identified with the element pB α uq P Lp pU ˆ
1

Σk q. If U is a Lipschitz domain, then for any f P W ´k,p pU q, p P p1, 8q,

there exists a unique F P W0k,p pU q, such that
ÿ ÿ ˆ 1
f“ |α| 2α
p´1q B F, i.e. f pgq “ B α F B α gdmd , @g P W0k,p pU q,
|α|ďk |α|ďk U

and hence DpU q is dense in W ´k,p pU q.

4.4.2 Extension and traces

In order to extend a Sobolev function defined in a bounded domain U to the
whole space Rd or to take the trace of it on the boundary, we need some
regularity assumption of the boundary.
Definition 4.34 (Boundary regularity of a domain). A domain in Rd is a
non-empty open connected set.

82 [February 17, 2023]

A bounded domain U Ă Rd is said to have a Lipschitz boundary if for each
point x0 P BU there exists r ą 0 and a one-to-one function ϕ : Br px0 q Ñ
D Ă Rd such that

• ϕpBr px0 q X U q Ă Rd´ ;

• ϕpBr px0 q X BU q Ă B Rd´ ;

• ϕ, ϕ´1 are Lipschitz continuous.

We then call U a bounded Lipschitz-domain. Similarly a bounded domain

has C k -boundary if the functions ϕ, ϕ´1 P C k .

Theorem 4.35 (Whitney extenstion theorem). Let 1 ď p ă 8 and U Ă

Rd be a bounded Lipschitz domain. Then there exists a linear (continuous)
extension map
W k,p pU q Ñ W k,p pRd q.

Sketchy proof. We sketch the proof under the stronger assumption that BU
is a C k manifold. By use of compactness, we take the partition of unity pηm q
on U such that for any function f defined on U :
N
ÿ
f“ pηm f q, Supp pη0 q ĂĂ U, Supp pηm q Ă Brm pxm q, m “ 1, ¨ ¨ ¨ , N.
m“0

By the one-to-one C k -map ϕm : Brm pxm q Ñ Dm Ă Rd and the Leibniz’ rule

as well as the chain rule (4.4), it suffices to extend Sobolev functions gm :“
pηm f q ˝ pϕ´1
m q P W
k,p
pRd´ q defined on the lower half space Rd´ continuously to
Sobolev functions g̃m P W k,p pRd q defined the whole space Rd , since then
N
ÿ
f˜ “ pη0 f q ` g̃m ˝ ϕm P W k,p pRd q,
m“1

}f˜}W k,p pRd q ď C}f }W k,p pU q .

Case g P C k . Let V :“ Rd´ , and g P C k pV q. We make the Ansatz

$
& gpxq if xd ď 0
k`1
g̃pxq “ ř
% aj gpx1 , . . . xd´1 , ´jxd q if xd ą 0.
j“1

If g P C k pV q we want to choose the aj so that B α g̃ is continuous for |α| ď k.

It suffices to do this for Bxj d for 0 ď j ď k when txd “ 0u, which leads to the

83 [February 17, 2023]

Vandermonde matrix
¨ ˛¨ ˛ ¨
1 1 1 ... 1 a1
˛
1
˚1 2 3 ... k`1 ‹ ‹ ˚ a2 ‹ ˚
˚ ‹
˚ ´1 ‹
˚ ‹
˚12 22 2 2‹ ˚
3 . . . pk ` 1q ‹ ˚ a3 ‹ “ ˚ 1 ‹ .
‹
˚
˚ .. .. .. . . .. ‹ ˚ .. ‹ ˚ ‹
˝. . . . . ‚˝ . ‚ ˝ . . . ‚
1k 2k 3k . . . pk ` 1qk ak`1 p´1qk

The Vandermonde matrix is invertible and we can solve this system. The
coefficients paj q hence exist and depend only on k. Then
}g̃}W k,p pRd q ď C}g}W k,p pV q .

Case g P W k,p pRd´ q. We define the extension g̃ in the same way. Then we
have to prove that for |α| ď k the distributional derivative is given by the
distribution defined by the distributional derivatives on both Rd` and Rd´ . By
an application of the theorem of Fubini and using distributional derivatives
in d ´ 1 variables this is true for all α with αd “ 0, and it suffices to consider
d “ 1. Let 1 ď κ ď k and φ P DpRq. Then, with
$ dκ
& dxκ gpxq if x ă 0
k`1
g̃κ pxq :“ ř ` dκ
˘
% aj p´jqκ dx κ g p´jxq if x ą 0
j“1

we have to check (Exercise)

ˆ ˆ
κ dκ
p´1q g̃ κ φdx “ g̃κ φpxqdx.
R dx R

[30.01.2023]
[03.02.2023]
We have the following density result for the embedding C pU q Ă W k,p pU q
8

from Theorem 4.32.

Corollary 4.36 (Density of smooth function space in Sobolev spaces). Sup-
pose that U is a open, bounded with Lipschitz boundary, k P N and 1 ď p ă 8.
Then the restrictions of Cc8 pRd q functions is dense in W k,p pU q.
We can take the trace f |BU of a continuous function f P CpU q on the bound-
ary in the usual sense. Recall the Gaussian Integral Theorem: Any function
u P C 1 pU q defined in a C 1 -domain U satisfies
ˆ ˆ
uνj dσ “ Bxj udx, @1 ď j ď d,
BU U

84 [February 17, 2023]

where ν denotes the outer normal vector of BU . Consequently for f, F P
C 1 pU q,
ˆ ˆ ˆ
pf F q|BU νj dσ “ Bxj pf F qdx “ pBxj f F ` f Bxj F qdx, @1 ď j ď d.
BU U U

Hence if F |BU “ 0, then

ˆ ˆ
pBxj f qF dx “ ´ pf Bxj F qdx, @1 ď j ď d,
U U

which coincides with (4.3). If F j P C 1 pŪ q, j “ 1, ¨ ¨ ¨ , d, then

ˆ d
ÿ ˆ ´ ÿ d ÿd ¯
j j j j
pf |BU qp F ν qdσ “ f B xj F ` Bxj f F dx.
BU j“1 U j“1 j“1

This motivates us to define the trace of f P W 1,p pU q as follows.

Theorem 4.37 (Traces). Let U be a bounded C 1 domain and let f P W 1,p pU q,
1 ď p ă 8. Then there is a unique trace g P Lp pBU q so that
ˆ ÿ d ˆ ´ ÿ d d
ÿ ¯
j j d´1
p F ν qgdH “ f B xj F j ` Bxj f F j dmd , (4.5)
BU j“1 U j“1 j“1

where F j P C 1 pU q and ν denotes the outer normal vector of BU . We denote

f |BU “ g, and it satisfies
p´1
1
}g}Lp pBU q ď c}f }Lpp pU q p}f }Lp pU q ` }Df }Lp q p , }Df }Lp :“ } |pBxj f q| }Lp . (4.6)

Sketchy proof. Case f P C 1 pU q. If f P C 1 pU q then g “ f |BU satisfies (4.5).

By a partition of unity and the one-to-one map pϕm q, we may assume that
U is below B Rd´ and f has compact support. If p ă 8, then (4.6) follows:
ˆ
p
}g}Lp pBU q “ |g|p dσ
Rd´1ˆ

“p Re |f |p´2 f¯Bxd f dmd

Rd
p´1
ďp}f }Lp }Bxd f }Lp .
Case f P W 1,p pU q. For f P W 1,p pU q, we approximate f by smooth functions
fn P C 1 pU q in W 1,p pU q. Since pfn q is a Cauchy sequence in W 1,p pU q, pfn |BU q is
a Cauchy sequence in Lp pBU q and hence we obtain a limit function g P Lp pBU q
at the boundary,´which satisfies (4.5) and (4.6). The uniqueness follows from
that fact that if BU hg “ 0 for all h P C 1 pU q, then g “ 0 in Lp pBU q.

85 [February 17, 2023]

Remark 4.38. If U is a bounded C 1 -domain and 1 ď p ă 8, then W01,p pU q
can be characterized as
W01,p pU q “ tf P W 1,p pU q | f |BU “ 0u.

4.4.3 Poincaré inequality and Sobolev inequality

We have seen in the trace theorem that, with some extra regularity assump-
tion on the Lebesgue functions, we can take the trace of a Sobolev func-
tion f P W 1,p pU q at the boundary (it does not make sense to take trace of
f P Lp pU q). We are interested in the relation between a Sobolev function
itself and its derivatives, in the framework of Lebesgue
´x spaces.
Recall the Newton-Leibniz’ formular: f pxq “ 0 f 1 ptqdt, for x P r0, 1s, if
f P C 1 pr0, 1sq and f p0q “ 0. This implies }f }Cb pr0,1sq ď }f 1 }L8 . We have also
similar estimates for f P W01,p pU q.
Theorem 4.39 (Poincaré inequality). Let U Ă Rd be a bounded domain.
Then
}f }Lp pU q ď diam pU q}Df }Lp pU q , @f P W01,p pU q, p P r1, 8q,
where diam pU q “ sup t|x ´ y| | x, y P U u.
Proof. Let a :“ diam pU q ă 8. Without loss of generality we assume U Ă
tx “ px1 , xd q|0 ă xd ă au. By the density of Cc8 pU q in W01,p pU q, it suffices to
show the inequality for f P Cc8 pU q. We extend f trivially to Rd , then
ˆ ˆ ˆ a
p
|f pxq| dx “ |f px1 , xd q|p dx d dx 1
d´1
U
ˆR ˆ0 a ˇˆ xd ˇp
1
f , yq dy ˇ dx d dx 1
ˇ ˇ
“ ˇ B xd px
d´1
ˆR ˆ0 a ˆ 0xd
ď |Bxd f px1 , yq|p dy pxd qp´1 dx d dx 1 , by Hölder’s inequality,
Rˆd´1
0 0

ď ap |Bxd f |p dx .
U

Remark 4.40. We have assumed “zero boundary condition” in Theorem

4.39. Similarly, we can assume “zero average condition” and show the Poincaré
inequality in a ball (Exercise): If 1 ď p ď 8 and f P W 1,p pBR p0qq then
d
}f ´ fB }Lp pBR p0qq ď 2 p R}Df }Lp pBR p0qq ,
´
where fB :“ pmd pBR p0qqq´1 BR p0q f dmd denotes the average of f in BR p0q.

86 [February 17, 2023]

The following Sobolev inequality implies the Sobolev embedding W 1,p pRd q Ă
dp
Lq pRd q, for 1 ď p ă d and Sobolev conjugate q “ d´p P pp, 8q.

Theorem 4.41 (Sobolev inequality). Suppose that 1 ď p ă d and q P pp, 8q

such that
1 1 1
` “ .
q d p
Then there exists c ą 0 such that

}f }Lq pRd q ď c}Df }Lp pRd q , @f P W 1,p pRd q.

Proof. By the density embedding DpRd q Ă W 1,p pRd q, it suffices to show the
inequality for f P DpRd q. We prove the estimate first for p “ 1 and q “ d´1
d
.
More precisely we prove the estimate
d
ź
}f }d d ď 2´d }Bj f }L1 pRd q (4.7)
L d´1 pRd q
j“1

by induction on the dimension. The case d “ 1 follows straightforward from

Newton-Leibniz’ formular:
ˆ x ˆ 8
1
f pxq “ f ptqdt “ ´ f 1 ptqdt, @f P DpRq.
´8 x

Suppose we have proven the estimate for d ď k ´ 1. Then by Fubini and

1 k´2
Hölder’s inequality (since k´1 ` k´1 “ 1)
k
ˆ ˆ
1
k´1
}f } k “ |f | k´1 |f |dmk´1 dm1 px1 q
k
L k´1 pR q k´1
ˆR R 1
ď }f px1 , . . . q}Lk´11 pRk´1 q }f px1 , . . . q} k´1 dm1 px1 q
L k´2 pRk´1 q
R
1
ˆ
k´1
ďsup x1 }f px1 , .q}L1 pRk´1 q }f px1 , .q} k´1 k´2 k´1
dm1 px1 q
R L pR q

ˆ ˜ k
1
¸ k´1
1 ź
ď2´ k´1 }Bx1 f }L1 pRk q 2´pk´1q }Bxj f px1 , .q}L1 pRk´1 q dm1 px1 q
R j“2

We take the inequality to the power k ´ 1, and apply Hölder’s inequality in

the form ˜ˆ
k ˆ
¸k´1
k
ź 1 ź
1
|gj | k´1 dm ď |gj |dm1
j“2 j“2

87 [February 17, 2023]

to arrive at
k
ź
}f }k k ď 2´k }Bj f }L1 pRk q .
L k´1 pRk q
j“1
pd´1qp
Now let 1 ă p ă d. We apply the above inequality (4.7) to |f | d´p . Then
pd´1qp pd´1qp
}f }Lqd´p
pRd q
“}|f | d´p } d
L d´1 pRd q
ˆ
pd´1qp pd ´ 1qp dpp´1q
ď}D|f | d´p }L1 pRd q ď |f | d´p |Df |dmd
d´p
dpp´1q
pd ´ 1qp
ď }f }Lqd´p
pRd q
}Df }Lp pRd q
d´p
where we first argue for smooth functions, and where we used Hölder’s in-
equality in the last step.
Remark 4.42. For p ą d, we have Morrey’s inequality for f˜ P W 1,p pU q, U Ă
Rd open: Every point is a Lebesgue point and the canonnical representative
f is continuous, such that the following is true for all x, y P U with |x ´ y| ă
distpx, Rd zU q
d
|f pxq ´ f pyq| ď c|x ´ y|1´ p }Df }Lp pB|x´y| pxqq .
d
This implies indeed the Hölder continuity f P C 1´ p pU q for p P pd, 8q, which
is almost everywhere differentiable, with the derivative the same as its weak
derivative almost everywhere. If p “ 8, then Lipschitz continuous functions
f P W 1,8 pU q are almost everywhere differentiable.
[03.02.2023]
[06.02.2023]

5 Hahn-Banach Theorem, Reflexivity and Weak

Topology
Recall the normed space of continuous linear maps LpX, Y q between two
normed spaces X, Y , where the norm in LpX, Y q is given by
}T }XÑY “ sup }x}X ď1 }T x}Y .
If Y “ K, then we denote X ˚ “ LpX, Kq (as the dual space of X), which is
a Banach space, equipped with the norm
}x˚ }X ˚ “ sup }x}X ď1 |x˚ pxq|.

88 [February 17, 2023]

 1
We have seen that plp pNqq˚ is isomorphic to lp pNq, p1 ` p11 “ 1 for p P r1, 8q
(see Theorem 3.24 and Corollary 3.25). In the exercise sheet pc0 pNqq˚ can
be characterized with l1 pNq. We are going to show that pl8 q˚ ‰ l1 pNq, as
a consequence of the celebrated Hahn-Banach Theorem. It is one of the
fundamental theorems in functional analysis.

5.1 Hahn-Banach Theorems

In this section we will introduce Hahn-Banach theorems, first in the purely
algebraic setting (for both real and complex cases), and then in the framework
of normed spaces (which is referred as Hahn-Banach theorem by default). We
will then give some interesting consequences of Hahn-Banach Theorem.

5.1.1 Hahn-Banach theorems

Definition 5.1. Let X be a K vector space. A map p : X Ñ R is called
sublinear if

• ppλxq “ λppxq for x P X and λ ě 0,

• ppx ` yq ď ppxq ` ppyq for x, y P X.

Example 5.2. 1. The norm of a normed space is sublinear.

2. If K “ R, any element in the dual space of a normed space X ˚ “

tcontinuous linear functional T : X Ñ Ru is sublinear.

3. Let C Ă X be convex such that for every x P X there exists λ ą 0 so

that λx P C. The Minkowski functional
1
pC pxq “ inftλ ą 0 : x P Cu P r0, 8q
λ
is sublinear. If X is a normed space, then its norm is the Minkowski
functional of the unit ball.

Theorem 5.3 (Hahn-Banach, real case). Let X be a real vector space, Y Ă X

a linear subspace, p : X Ñ R sublinear and l : Y Ñ R linear such that

lpyq ď ppyq for all y P Y.

Then there exists L : X Ñ R linear so that

1. Lpyq “ lpyq for all y P Y

89 [February 17, 2023]

 2. Lpxq ď ppxq for all x P X.

Proof. There are two very different steps.

Step 1. Suppose that Y ‰ X. Then there exists x0 P XzY . Let Y1 be the
space spanned by Y and x0 . Every element of Y1 can uniquely be written as

y ` sx0 , y P Y, s P R .

In Step 1 we aim to find a linear map l1 : Y1 Ñ R such that

• l1 pyq “ lpyq for y P Y

• l1 py ` sx0 q ď ppy ` sx0 q for s P R and y P Y .

It suffices to find t “ l1 px0 q so that

lpyq ` st ď ppy ` sx0 q, @y P Y, s P R .

Notice the following two equivalence relations

st ď spppy{s ` x0 q ´ lpy{sqq, @s ą 0, y P Y
ðñ t ď inf tppy ` x0 q ´ lpyqu,
yPY

and

´ st ě ´splp´y{sq ´ pp´y{s ´ x0 qq, @s ă 0, y P Y

ðñ t ě sup y tlpyq ´ ppy ´ x0 qu.

Hence we can find t if and only if

lpyq ´ ppy ´ x0 q ď ppỹ ` x0 q ´ lpỹq for all y, ỹ P Y

which follows from the inequality on Y and sublinearity of p:

lpyq ` lpỹq “ lpy ` ỹq ď ppy ` ỹq ď ppy ` x0 q ` ppỹ ´ x0 q.

Step 2. We need the axiom of choice in the form of Zorn’s lemma: Let Z be
a partially ordered set which contains an upper bound for every chain. Then
there is a maximal element. A partially ordered set Z is a set equipped with
the relation ď, which satisfies a ď a for all a P Z, a ď b&b ď a ñ a “ b
and a ď b&b ď c ñ a ď c. Here a chain is a totally ordered subset, that is
a subset A where either a ď b or b ď a holds for any two elements a, b P A.
An element b is an upper bound for the chain A, if a ď b for all a P A. An
element a P Z is maximal if b P Z, b ě a implies b “ a.

90 [February 17, 2023]

We define
Z “ tpW, lW q : Y Ă W Ă X, lW |Y “ l, lW pwq ď ppwq for w P W u,
with the ordering
pW, lW q ď pV, lV q if W Ă V and lV |W “ lW .
This is a partial order. If Z̃ is a chain then
ď
V “ W
pW,lW qPZ̃

with the obvious lV being an upper bound for the chain. Now let pV, lV q
be a maximal element. If V “ X we are done. Otherwise we obtain a
contradiction by the first step.
Theorem 5.4 (Hahn-Banach, complex version). Let X be a complex vector
space, Y a subvector space, p : X Ñ R sublinear and l : Y Ñ C linear so
that
Re lpyq ď ppyq for y P Y.
Then there exists L : X Ñ C linear so that
1. L|Y “ l
2. Re Lpxq ď ppxq.
Proof. We first notice that any linear map T : X Ñ C is uniquely determined
by its real part R “ Re T :
T pxq “ Rpxq ´ iRpixq.
Let X be viewed as R-vector space, and Y R-subvector space. We apply
Theorem 5.3 to the real part Re l : Y Ñ R and derive A : X Ñ R such that
Ay “ Re plpyqq for all y P Y and Ax ď ppxq for all x P X.
The linear map Lpxq “ Apxq ´ iApixq is the searched extension map.
The above Hahn-Banach theorems are purely algebraic. We formulate the
consequences for normed vector spaces, making use of the fact that norms
are sublinear.
Theorem 5.5 (Hahn-Banach). Let X be a normed K vector space, Y a
subspace endowed with the norm of X, and y ˚ P Y ˚ “ LpY, Kq a continuous
linear map.
Then there exists x˚ P X ˚ “ LpX, Kq so that

91 [February 17, 2023]

 1. x˚ |Y “ y ˚ , i.e. x˚ pyq “ y ˚ pyq for all y P Y ,

2. }x˚ }X ˚ “ }y ˚ }Y ˚ .

Proof. We define
ppxq “ }y ˚ }Y ˚ }x}X ,
such that
Re y ˚ pyq ď ppyq, @y P Y.
We apply Theorem 5.4 to obtain x˚ : X Ñ K a linear map so that x˚ |Y “ y ˚
and
Re x˚ pxq ď ppxq.
This implies that for all x P X, there exists α P C with |α| “ 1 such that

|x˚ pxq| “ Re αx˚ pxq “ Re x˚ pαxq ď ppαxq “ ppxq “ }y ˚ }Y ˚ }x}X ,

and hence }x˚ }X ˚ ď }y ˚ }Y ˚ such that x˚ P X ˚ . Since x˚ |Y “ y ˚ ,

}x˚ }X ˚ “ }y ˚ }Y ˚ .

Example 5.6. pl8 pNqq˚ ‰ l1 pNq.

The space of converging sequences c is a closed subspace of l8 . Let l : c Ñ K
be defined by
lppxj qq “ lim xj .
jÑ8

Then l P c˚ and
|lppxj qq| ď }pxj q}l8
for every converging sequence. By Theorem 5.5 it has an extension L P pl8 q˚ .
Clearly Lpej q “ lpej q “ 0. Hence L cannot be represented by a sequence
y “ pyj q P l1 since otherwise
8
ÿ
Lppxj qq “ yj xj ùñ Lpen q “ yn “ 0 for all n.
j“1

[06.02.2023]
[10.02.2023]

92 [February 17, 2023]

5.1.2 Consequences of Hahn-Banach theorem
We state some consequences of Hahn-Banach Theorem.
Lemma 5.7. If X is a normed space, Y Ă X is a closed subspace with
Y ‰ X and x0 P XzY , then there exists x˚ P X ˚ , with x˚ |Y “ 0, }x˚ }X ˚ “ 1
and x˚ px0 q “ distpx0 , Y q ą 0.
Proof. Let Ỹ be the span of Y and x0 . Since Y is closed, a :“ distpx0 , Y q ą 0.
We define
y ˚ py ` tx0 q “ ta, @py ` tx0 q P Ỹ.
We claim that y ˚ P Ỹ ˚ with }y ˚ }Ỹ ˚ “ 1. Obviously y ˚ is linear. For t “ 0,
|y ˚ py ` tx0 q| “ 0 ď }y ` tx0 }X . For t ‰ 0,
}y ` tx0 }X }y ` tx0 }X
|y ˚ py ` tx0 q| “ |t|a “ |t|a “d y ď }y ` tx0 }X .
}y ` tx0 }X } t ` x0 } X

Now let pyn q Ă Y be a sequence such that a “ limn }x0 ´ yn }X . Then

ˇ ´ x ´ y ¯ˇ a
0 n
}y ˚ }Ỹ ˚ ě ˇy ˚ Ñ 1,
ˇ ˇ
ˇ“
}x0 ´ yn }X }x0 ´ yn }X
which implies }y ˚ }Ỹ ˚ “ 1.
Hence by Theorem 5.5 we can extend it to x˚ P X ˚ , with x˚ |Y “ y ˚ |Y “ 0,
x˚ px0 q “ y ˚ px0 q “ a and }x˚ }X ˚ “ }y ˚ }Ỹ ˚ “ 1.
We have the following interesting descriptions of some specific subsets of X
by use of elements in X ˚ .
Corollary 5.8. Let X be a normed space. Then
(a) For any x P X, there exists x˚ P X ˚ with }x˚ }X ˚ “ 1 and x˚ pxq “ }x}X .
Hence for x1 , x2 P X with x1 ‰ x2 , there exists x˚ P X ˚ with }x˚ }X ˚ “ 1
and x˚ px1 q ‰ x˚ px2 q.

(b) The norm can be characterized as

}x}X “ sup tRe x˚ pxq : x˚ P X ˚ , }x˚ }X ˚ “ 1u.

Hence there is a cannonical isometry

J : X Ñ X ˚˚ “ pX ˚ q˚ , Jpxqpx˚ q “ x˚ pxq with }Jpxq}X ˚˚ “ }x}X .

(c) For A Ă X,

A is bounded ðñ x˚ pAq is bounded for all x˚ P X ˚ .

93 [February 17, 2023]

(d) For Y Ă X a subvector space,

Y is not dense in X ðñ Dx˚ P X ˚ zt0u with x˚ pyq “ 0, @y P Y.

(e) If X ˚ is separable, then X is separable.

Proof. (a) Let Y “ t0u. If x ‰ 0, then by Lemma 5.7 there exists x˚ P X ˚

such that }x˚ } “ 1 and x˚ pxq “ distpx, Y q “ }x}. If x “ 0 we choose x0 ‰ 0
and find x˚ for x0 . Consequently, if x1 ‰ x2 , we find x˚ for x0 “ x1 ´ x2 .
(b) The characterization for the norm on X follows from (a) and the fact
|Rex˚ pxq| ď }x˚ }X ˚ }x}X .
Correspondingly we compute

}Jpxq}X ˚˚ “ sup }x˚ }X ˚ “1 Re Jpxqpx˚ q “ sup }x˚ }X ˚ “1 x˚ pxq “ }x}X .

(c) “ùñ” If A Ă X is bounded and x˚ P X ˚ , then obviously x˚ pAq is bounded

in K.
“ðù” Now if x˚ P X ˚ and x˚ pAq is bounded, then

sup xPA |Jpxqpx˚ q| “ sup xPA |x˚ pxq| ă 8.

As X ˚ is a Banach space, Banach-Steinhaus Theorem 4.2 implies

sup xPA }Jpxq}X ˚˚ ă 8.

Hence A is bounded.
(d) “ùñ” It follows from Lemma 5.7.
“ðù” If Y is dense, then x˚ |Y “ 0 implies x˚ |X “ 0 by continuity.
(e) Since X ˚ is separable, its subset is also separable, and in particular the
unit sphere in X ˚ is separable: Let px˚j q be a dense sequence of unit vectors
in the unit sphere. We can take a sequence pxj q Ă X with }xj }X “ 1 such
that x˚j pxj q ě 12 . We claim that the span Y of pxj q is dense in X. Indeed
otherwise by (d) there exists x˚ P X ˚ such that x˚ |Y “ 0 (by continuity) and
}x˚ }X ˚ “ 1 (by linearity). Hence

x˚ pxj q “ x˚j pxj q ` px˚ ´ x˚j qpxj q,

which can be chosen bigger than 14 in view of x˚j pxj q ě 12 and the density of
px˚j q in the unit sphere. This is contradiction to x˚ |Y “ 0. Hence Y is dense
in X and tqxj | j P N, q P Q or Q ` iQu is a countable dense set in X.

94 [February 17, 2023]

5.2 Reflexivity, weak and weak-˚ topology
5.2.1 Reflexive spaces
Recall the linear isometry

J : X Ñ X ˚˚ , x ÞÑ x˚ pxq, @x˚ P X ˚ . (5.1)

}} ˚˚ }} ˚˚
As JpXq X is a closed linear subspace of X ˚˚ , JpXq X is a Banach space.
The map J is not necessarily surjective, as seen from the example that X “
l1 pNq, X ˚ “ l8 pNq, X ˚˚ ‰ l1 pNq. This motivates us to define
Definition 5.9. A normed space X is called reflexive if the evaluation map
J : X Ñ X ˚˚ is surjective.
Hence a reflexive normed space is automatically Banach space. lp pNq is
reflexive space for 1 ă p ă 8, while not reflexive if p “ 1 or 8.
Lemma 5.10. The following hold:
(a) Let X be reflexive Banach space, and Y Ă X be a closed subspace. Then
Y is reflexive.

(b) A Banach space X is reflexive if and only if X ˚ is reflexive.

Proof. (a) Let Y Ă X be a closed subspace and Y ‰ X. For any x˚ P X ˚
the restriction on Y belongs to Y ˚ with

}x˚ |Y }Y ˚ ď }x˚ }X ˚ .

Let y ˚˚ P Y ˚˚ , we define x˚˚ P X ˚˚ by

x˚˚ px˚ q “ y ˚˚ px˚ |Y q, @x˚ P X ˚ . (5.2)

Since X is reflexive, there exists x P X such that

x˚˚ px˚ q “ x˚ pxq, @x˚ P X ˚ . (5.3)

We claim that x P Y . If x R Y , then by Lemma 5.7 there exists x̃˚ P X ˚

such that x̃˚ |Y “ 0 and x̃˚ pxq “ 1. This is impossible due to (5.2) and (5.3).
Thus JY is surjective: For any y ˚ P Y ˚ , by Hahn-Banach Theorem 5.5 there
exists x˚1 P X ˚ such that x˚1 |Y “ y ˚ , and hence

y ˚˚ py ˚ q “ x˚˚ px˚1 q “ Jpxqpx˚1 q “ x˚1 pxq “ y ˚ pxq “ Jpxqpy ˚ q, @y ˚ P Y ˚ .

Thus Y is reflexive.

95 [February 17, 2023]

 [10.02.2023]
[13.02.2023]

(b) “ñ” We observe that for x˚˚˚ P X ˚˚˚ , we can define x˚ P X ˚ by

x˚ pxq “ x˚˚˚ pJpxqq, @x P X.
If J : X Ñ X ˚˚ is surjective, then for any y ˚˚ P X ˚˚ there exists a unique
y P X such that y ˚˚ “ Jpyq, and hence
x˚˚˚ py ˚˚ q “ x˚˚˚ pJpyqq “ x˚ pyq “ y ˚˚ px˚ q, @y ˚˚ P X ˚˚ .
Thus X ˚ is reflexive.
“ð” Since X is a Banach space, JpXq is a closed subspace of X ˚˚ . If X is
not reflexive, then by Lemma 5.7 there exists x˚˚˚ P X ˚˚˚ zt0u such that
x˚˚˚ pJpyqq “ 0, @y P X.
Since X ˚ is reflexive, there exists x˚ P X ˚ such that
x˚˚˚ pJpyqq “ Jpyqpx˚ q “ x˚ pyq, @y P X.
Thus x˚ “ 0, which is contradiction to x˚˚˚ ‰ 0. Hence X is reflexive.
Remark 5.11. If Y Ă X are two normed spaces such that }y}X ď C}y}Y ,
then X ˚ Ă Y ˚ and for any x˚ P X ˚ ,
}x˚ }Y ˚ “ sup }y}Y ď1 |x˚ pyq| ď sup yPY, }y}X ďC |x˚ pyq| ď C}x˚ }X ˚ .
Correspondingly Y ˚˚ Ă X ˚˚ with
}y ˚˚ }X ˚˚ ď C}y ˚˚ }Y ˚˚ .
If }}X “ }}Y , then }x˚ }Y ˚ ď }x˚ }X ˚ but not necessarily equal.
Corollary 5.12. Let U Ă Rd be open, k P N, p P p1, 8q. Then W k,p pU q and
W0k,p pU q are reflexive.
Proof. We first claim that reflexivity of a Banach space is preserved by an
isomorphic map. Let j : X Ñ Y between two normed spaces is isomorphic,
then j ˚ : Y ˚ Q y ˚ ÞÑ y ˚ ˝ j P X ˚ is also isomorphic, and hence j ˚˚ : X ˚˚ Q
x˚˚ ÞÑ x˚˚ ˝ j ˚ P X ˚˚ is isomorphic. Thus if X is a reflexive Banach space,
then JX : X Ñ X ˚˚ is isomorphic, which implies JY “ j ˚˚ ˝ JX ˝ j ´1 : Y Ñ
Y ˚˚ is surjective.
W k,p pU q is isomorphic to a closed subspace of the reflexive Banach space
Lp pU ˆ Σk q:
W k,p pU q Q f ÞÑ pB α f q|α|ďk P Lp pU ˆ Σk q,
and is hence reflexive. W0k,p pU q is a closed subspace of W k,p pU q, and is hence
reflexive.

96 [February 17, 2023]

5.2.2 Separation theorems
Let X be a normed space. Recall the sublinear Minkowski functional
1
pC pxq “ inftλ ą 0 : x P Cu P r0, 8s
λ
where C Ă X is convex and we define the infimum of an empty set to be 8.
Lemma 5.13. Let X be a normed space and C Ă X convex.
(a) If 0 P C̊, then there exists δ ą 0 such that Bδ p0q Ă C, and hence
pC pxq ď δ ´1 }x}X for all x P X.
(b) If C is open with 0 R C, then there exists x˚ P X ˚ such that Re x˚ pxq ă 0
for all x P C.
Proof. (a) follows immediately from the definition. (b)Exercise.
Theorem 5.14 (Separation theorem). Let X be normed space. Let V, W be
disjoint convex sets.
(a) If V, W are open, there exists x˚ P X ˚ such that

Re x˚ pwq ă Re x˚ pvq, @v P V, w P W.

We say that x˚ separates W and V .

(b) If V is closed and W is compact. Then there exists x˚ P X ˚ such that

sup wPW x˚ pwq ă inf x˚ pvq.

vPV

We say that x˚ separates V and W strictly.

Proof. (a) Let U “ W ´ V “ tw ´ v | v P V, w P W u, then U is convex
and open, and 0 R U . By Lemma 5.13, there exists x˚ P X ˚ such that
Re x˚ |U ă 0, that is, x˚ separates W and V .
(b) Since V is closed, for any w P W there exists ε ą 0 such that Bε pwq
is disjoint with V . Since W is compact there are finite open covering balls
pBεj {2 pwpjq qqN
j“1 .
If N “ 1, by (a) there exists x˚ P X ˚ such that

Re x˚ pw1 q ă Re x˚ pvq, @w1 P Bε pwq, @v P V.

Since there exists y P B1 p0q such that

1 ε ε
Re x˚ pyq ą }x˚ }X ˚ ñ Re x˚ p yq ą }x˚ }X ˚ “: δ ą 0,
2 2 4
97 [February 17, 2023]
we have for all w2 P Bε{2 pwq,
ε ε
Re x˚ pw2 q “ Re x˚ pw2 ` yq ´ Re x˚ p yq ă Re x˚ pvq ´ δ, @v P V,
2 2
and hence x˚ separates W and V strictly.
If N ą 1, it follows by replacing Bε pwq by YN pjq
j“1 Bεj pw q, and taking δ “
εj
minN ˚
j“1 4 }x }X ˚ .

5.2.3 Weak topology

Definition 5.15. Let X be a normed space. We call a sequence pxn q Ă X
converges weakly against x P X, denoted by xn á x, if

x˚ pxn q Ñ x˚ pxq, @x˚ P X ˚ .

Lemma 5.16. Let X be a normed space.

(a) Convergence in norm implies weak convergence.

(b) Weakly convergent sequences are bounded.

(c) Let C be a closed convex set in X, and pxn q Ă C be a sequence converging

weakly to x P X. Then x P C.

(d) If X is Banach space, and the unit ball is uniformly convex, that is, for
any ε ą 0,
1
δpεq :“ inft1 ´ }x ` y}X | }x}X , }y}X ď 1, }x ´ y}X ą εu ą 0,
2
then weak convergence xn á x and convergence in norm }xn } Ñ }x}X
imply strong convergence }xn ´ x}X Ñ 0.

Proof. (a) It follows from the continuity of the map x˚ P X ˚ .

(b) Recall x˚ pxq “ Jpxqpx˚ q for all x P X, x˚ P X ˚ . If xn á x in X, then for
any fixed x˚ P X ˚ the continuous linear maps Jpxn q : X ˚ Ñ K between two
Banach spaces converges:

Jpxn qpx˚ q Ñ Jpxqpx˚ q.

By Banach-Steinhaus Theorem 4.2, we have

sup n }xn }X “ sup n }Jpxn q}X ˚˚ ă 8.

98 [February 17, 2023]

(c) It suffices to consider K “ R. We assume by contraction that x R C, then
by Theorem 5.14 there exists x˚ P X ˚ such that

x˚ pxq ă inf x˚ pyq.

yPC

This is a contradiction to the weak convergence of xn to x.

(d) If }x} “ 0, then }xn } Ñ }x} implies xn Ñ x.
If }x} ą 0, then without loss of generality we assume }x}X “ 1, and xn á x
with }xn }X Ñ 1. Replacing xn by }xxnn}X we may assume }xn }X “ 1.
[13.02.2023]
[17.02.2023]

By applying Lemma 5.13-(b) with C “ B1 pxq, there exists x˚ P X ˚ with

Re x˚ px ´ yq ą 0 ñ Re x˚ pxq ą Re x˚ pyq, @y P B1 p0q.

Hence Re x˚ pxq ą 0, and without loss of generality one may assume 1 “

Re x˚ pxq, which implies }x˚ }X ˚ “ 1.
We assume by contradiction that pxn q is not a Cauchy sequence, and hence
there exists ε ą 0 such that for any n P N there exists m ą n with }xn ´
xm }X ą ε. By uniform convexity of the unit ball, 12 }xn ` xm }X ď 1 ´ δpεq,
and hence
ˇ 1 ˇ
ˇ ˚
x x m ˇ ď 1 ´ δpεq.
ˇ
ˇ p px n ` qq
2
This is a contradiction to Re x˚ p 21 pxn ` xm qq Ñ Re x˚ pxq “ 1.
Similarly, the strong convergence for pxn q which converges weakly to x follows
from the uniform convexity.
Remark 5.17. By Hanner’s inequality, the unit ball of the Lebesgue space
Lp pU q or the Sobolev space W k,p pU q with p P p1, 8q is uniformly convex, and
hence (d) is applicable. Exercise.
There exist weak converging sequence in a Banach space whose norms do not
converge. Nevertheless, by Mazur’s Theorem below, there exist its convex
linear combinations which converges strongly.
Theorem 5.18 (Mazur). Let X be a Banach space, and pxn q converges
pnq
weakly to x P X. Then there exist real numbers pλk q Ă r0, 1s and pN pnqq Ă
N such that
Nÿ
pnq Nÿ
pnq
pnq pnq
λk “ 1 and λk xk Ñ x.
k“n k“n

99 [February 17, 2023]

Proof. For any n P N, let Cn be the convex hull of txk ukěn :
8
!ÿ ˇ 8
ÿ )
Cn “ µ k xk ˇ 0 ď µ k , µk “ 1, only finitely many µk are nonzero .
ˇ
k“n k“n

Then Cn is closed and convex in X. Since xk P Cn for all k ě n, x P Cn

by Lemma 5.16-(c). Thus for any n P N, there exists yn P Cn such that
}x ´ yn } ď n1 .

5.2.4 Weak-* topology

Definition 5.19. Let X be a normed space. A sequence px˚n q Ă X ˚ converges
weakly* to x˚ P X if

x˚n pxq Ñ x˚ pxq, @x P X.

By Banach-Steinhaus Theorem 4.2, any weak-* convergent sequence is bounded

in X ˚ . Conversely, any bounded sequence in X ˚ has a weak-* convergent
subsequence:
Theorem 5.20 (Banach-Alaoglu, a simplified version). Let X be a separable
normed space. Then every bounded sequence x˚n Ă X ˚ contains a weak*
convergent subsequence.
Proof. Let pxk q be dense in X, and px˚n q be a bounded sequence in X ˚ . Since
x˚n px1 q is bounded, there exists a converging subsequence x˚p1q px1 q. Itera-
nj
pkq pk´1q
` ˚ ˘
tively we obtain subsequences n Ă n such that x pkq pxk q j converges
nj
for each k P N. Let
˚
ym “ x˚npmq , @m P N,
m
˚ ˚
then pym pxk qqm converges for all k P N. Since pxk q is dense and pym q Ă X ˚ (as
˚ ˚
a subsequence of pxn q) is bounded, by continuity pym q weakly-* converges.
Remark 5.21. Without the separability assumption, and one can show that
˚
the closed unit ball in the dual space of a normed space B1X p0q Ă X ˚ is
compact in the weak-* topology. The version of Banach-Alaoglu Theorem
5.20 can fail without separability assumption. Let X “ ℓ8 , and φn pxq “ xn P
X ˚ with }φn }X ˚ “ 1. For any subsequence pφnk qk , with y :“ p´1q1 en1 `
p´1q2 en2 ` ¨ ¨ ¨ P ℓ8 ,

φnk pyq “ p´1qk diverges.

Hence there is no weak* convergent subsequence.

100 [February 17, 2023]

If X is reflexive, then J : X Ñ X ˚˚ is isometry, such that
x˚ pxq “ Jpxqpx˚ q, @x P X, @x˚ P X ˚ ,
and hence the weak and weak-* convergence are the same, both in X and
X ˚ . Thus we have immediately the following
Corollary 5.22. If X is reflexive and separable Banach space, then any
bounded sequence has a weakly converging subsequence.
Hence any bounded sequence in Lp , 1 ă p ă 8 has weakly converging
subsequence.
Remark 5.23. The separability assumption can be dropped: Let pxn q be
a bounded sequence in a reflexive Banach space, then the closure Y of the
span of pxn q is reflexive and separable, which yields the existence of weakly
converging subsequence by Corollary 5.22.
Example 5.24. 1. Let pX, µq be measurable space, and p P p1, 8q. Then
Lp pµq is reflexive Banach space.
Lp pµq
• The strong convergence of a sequence fn ÝÑ f : }fn ´ f }Lp pµq Ñ 0
is equivalent to
Lp pµq
fn á f & }fn }Lp pµq Ñ }f }Lp pµq ;
• If pX, µq “ pU, md q, then the weak (i.e. weak-*) convergence of a
Lp pU q ´ ´ 1
sequence fn á f : X fn gdµ Ñ X f gdµ, @g P Lp pµq, is equiva-
lent to
ˆ ˆ
sup n }fn }Lp pU q ă 8 & fn φdµ Ñ f φdµ, @φ P DpU q,
U X
D1 pU q
that is, sup n }fn }Lp pU q ă 8 & Tfn Ñ Tf .
2. Let U Ă Rd be open, k P N and p P p1, 8q. Then W k,p pU q is isometric
to a closed subspace of Lp pU ˆ Σk q and hence is reflexive, separable
W k,p
Banach space. The weak convergence fn á f is equivalent to
D1 pU q
sup n }fn }W k,p pU q ă 8 & Tfn Ñ Tf .

5.3 Introduction to spectral theory

We introduce here some basic definitions and facts (without proof) in spectral
theory, which will be generalized and discussed in more detail in the lecture
“Spectral Theory”. 9
9
This section is not taken into account in the exam.

101 [February 17, 2023]

5.3.1 Compact operators
Definition 5.25 (Kernel and Range). Let X, Y be two normed spaces and
T P LpX, Y q. We denote the kernel of T by

N pT q “ tx P X | T x “ 0u

and the range of T by

RpT q “ tT x | x P Xu.

We have the following facts concerning the kernel and range of a continuous
linear operator and its adjoint operator.

Lemma 5.26. Let X, Y be normed spaces.

• For T P LpX, Y q, let T ˚ : Y ˚ Ñ X ˚ be a linear map given by

T ˚ py ˚ qpxq “ y ˚ pT xq, @x P X.

Then }T ˚ }Y ˚ ÑX ˚ “ }T }XÑY , and

N pT ˚ q “ RpT qK :“ ty ˚ P Y ˚ | y ˚ pyq “ 0, @y P RpT qu,

N pT q “ RpT ˚ qK :“ tx P X | x˚ pxq “ 0, @x˚ P RpT ˚ qu

We call T ˚ : Y ˚ Ñ X ˚ the adjoint of T .

• Let P P LpX, Xq be a projection (i.e. P 2 “ P ). Then I ´ P P LpX, Y q

is also a projection, such that

RpP q “ N pI ´ P q, N pP q “ RpI ´ P q,

and X “ RpP q ‘ N pP q (i.e. N pP q, RpP q are closed subspace such that

N pP q ` RpP q “ X and N pP q X RpP q “ t0u).

Definition 5.27 (Compact operator). Let X, Y be Banach spaces and T P

LpX, Y q. We call T compact if for every bounded sequence pxj q Ă X,
pT pxj qq Ă Y has a convergent subsequence.

Remark 5.28. • If X, Y are finite dimensional, then T P LpX, Y q is

compact. More generally, if RpT q is finite dimensional then T is com-
pact.

• The composition of a compact operator and a continuous operator is

still compact.

102 [February 17, 2023]

 • T P LpX, Y q is compact ðñ T ˚ P LpY ˚ , X ˚ q is compact.

• The embedding W01,p pU q Ă Lp pU q with U Ă Rd a bounded open set is

compact.

Theorem 5.29 (Riesz-Schauder). Let X be Banach space and K P LpXq be

compact. Let T “ I ´ K, then dim N pT q “ dim N pT ˚ q ă 8 and RpT q is
closed. Furthermore, T is invertible iff T is injective.

5.3.2 Spectrum of compact operators

We generalize the eigenvalues of a matrix (as linear map from Cn to Cn ) to
the spectrum of a continuous linear operator.

Definition 5.30. Let X be Banach space and T P LpX, Xq “: LpXq. The

resolvent set of T is given by

ρpT q “ tλ P C | λI ´ T : X Ñ X is bijectiveu,

and its spectrum by

σpT q “ C zρpT q.

We further define the point spectrum of T by

σp pT q “ tλ P C | Dx P Xzt0u with λx “ T xu Ă σpT q,

where we call λ P σp pT q an eigenvalue of T and the corresponding x an

eigenvector or eigenfunction of T . For λ P ρpT q the operator

Rpλ, T q :“ pλI ´ T q´1 : X Ñ X

and the set tRpλ, T q | λ P ρpT qu are called resolvent.

Notice here that unlike the case of a n ˆ n-matrix M ,

pλI ´ M q is bijective ðñ λ P C is not an eigenvalue of M,

the operator T defined on a (infinite-dimensional) Banach space may have

λ P σpT qzσp pT q, which is in the spectrum of T (i.e. pλI ´ T q is not bijective)
but is not an eigenvalue. For example, the right and left shift operators
R, L : ℓp Ñ ℓp

Rx “ p0, x1 , x2 , ¨ ¨ ¨ q and Lx “ px2 , x3 , ¨ ¨ ¨ q

103 [February 17, 2023]

are continuous linear operators with the norm }R}ℓp Ñℓp “ }L}ℓp Ñℓp “ 1.
Obviously LR “ I and RLx “ p0, x2 , x3 , ¨ ¨ ¨ q, and hence R is injective (not
surjective) and has left inverse (but no right inverse), while L is surjective
(not injective) and has right inverse (but no left inverse). The right shift
operator R has no eigenvalue, since for any λ P C,

Rx “ λx ùñ x “ 0,

while 0 P σpRq, since R is not bijective.

Theorem 5.31 (Spectrum of compact operators). Let X be Banach space

and K P LpXq be compact. Then σpKq is countable set with 0 as the only
possible accumulation point.

104 [February 17, 2023]