Introduction To Sobolev Spaces

Introduction to Sobolev Spaces
Lecture Notes
MM692 2018-2
Joa Weber
UNICAMP
December 23, 2018

Contents
1 Introduction 1
1.1 Notation and conventions . . . . . . . . . . . . . . . . . . . . . . 2
2 Lp -spaces 5
2.1 Borel and Lebesgue measure space on Rn . . . . . . . . . . . . . 5
2.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 Convolution 13
3.1 Convolution of functions . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Convolution of equivalence classes . . . . . . . . . . . . . . . . . 15
3.3 Local Mollification . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.3.1 Locally integrable functions . . . . . . . . . . . . . . . . . 16
3.3.2 Continuous functions . . . . . . . . . . . . . . . . . . . . . 17
3.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4 Sobolev spaces 19
4.1 Weak derivatives of locally integrable functions . . . . . . . . . . 19
4.1.1 The mother of all Sobolev spaces L1loc . . . . . . . . . . . 19
4.1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.1.3 ACL characterization . . . . . . . . . . . . . . . . . . . . 21
4.1.4 Weak and partial derivatives . . . . . . . . . . . . . . . . 22
4.1.5 Approximation characterization . . . . . . . . . . . . . . . 23
4.1.6 Bounded weakly differentiable means Lipschitz . . . . . . 24
4.1.7 Leibniz or product rule . . . . . . . . . . . . . . . . . . . 24
4.1.8 Chain rule and change of coordinates . . . . . . . . . . . . 25
4.1.9 Equivalence classes of locally integrable functions . . . . . 27
4.2 Definition and basic properties . . . . . . . . . . . . . . . . . . . 27
4.2.1 The Sobolev spaces W k,p . . . . . . . . . . . . . . . . . . 27
4.2.2 Difference quotient characterization of W 1,p . . . . . . . . 29
4.2.3 The compact support Sobolev spaces W0k,p . . . . . . . . 30
k,p
4.2.4 The local Sobolev spaces Wloc . . . . . . . . . . . . . . . 30
4.2.5 How the spaces relate . . . . . . . . . . . . . . . . . . . . 31
4.2.6 Basic properties – products and coordinate change . . . . 31
i
ii CONTENTS
5 Approximation and extension 33

5.1 Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.1.1 Local approximation – any domain . . . . . . . . . . . . . 33
5.1.2 Global approximation on bounded domains . . . . . . . . 34
5.1.3 Approximation even up to ∂ on Lipschitz domains . . . . 36
5.2 Extensions and traces . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2.1 Extension . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2.2 Trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6 Sobolev inequalities 47
6.1 Sub-dimensional case kp < n . . . . . . . . . . . . . . . . . . . . 47
6.1.1 Gagliardo-Nirenberg-Sobolev inequality (p < n) . . . . . . 47
6.1.2 General Sobolev inequalities (kp < n) . . . . . . . . . . . 51
6.1.3 Compactness (Rellich-Kondrachov) . . . . . . . . . . . . . 53
6.2 Super-dimensional case kp > n . . . . . . . . . . . . . . . . . . . 53
6.2.1 Morrey’s inequality (p > n) – continuity . . . . . . . . . . 53
6.2.2 General Sobolev inequalities (kp > n) . . . . . . . . . . . 57
7 Applications 61
7.1 Poincaré inequalities . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.2 Lipschitz functions . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.3 Differentiability almost everywhere . . . . . . . . . . . . . . . . . 62
A Appendix 63
A.1 Allerlei . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
A.1.1 Inequalities via convexity and concavity . . . . . . . . . . 63
A.1.2 Continuity types and their relations . . . . . . . . . . . . 64
A.1.3 Distance function . . . . . . . . . . . . . . . . . . . . . . . 65
A.1.4 Modes of convergence . . . . . . . . . . . . . . . . . . . . 66
A.2 Banach space valued Sobolev spaces . . . . . . . . . . . . . . . . 66
Bibliography 67
Index 69
Chapter 1
Introduction
These are Lecture Notes 1 written for the last third of the course “MM692
Análise Real II” in 2018-2 at UNICAMP. Originally motivated by [MS04,
App. B.1] this text is essentially a compilation of the presentations in [AF03,
Bre11,Eva98,GT01]. Given a solid background in measure theory, we also highly
recommend the book [Zie89] whose geometric measure theory point of view of-
fers deep insights and strong results. An excellent source that analyses Sobolev
spaces as part of a wider context is [Ste70, Ch. V].
Differentiation and integration are reverse operations. The notion of weak
derivative combines both worlds. It is more general than the usual partial
derivative and at the same time it makes available the powerful tool box of inte-
gration theory. Weak derivatives and Sobolev spaces, the spaces of functions
that admit weak derivatives, are typically used in applications as an intermedi-
ate step towards solution. For instance, to solve a PDE it is often easier (bigger
space) to establish first existence of a weak solution (ask for weak derivatives
only) and then show in a second step, regularity, that the weak solution found
is actually differentiable in the usual sense.
We made an effort to spell out details in the case of Lipschitz domains D,
because often in the literature one only finds proofs for the C 1 case. The rel-
evant theorems are the Approximation Theorem 5.1.7 and the Extension The-
orem 5.2.1. The assumptions in these theorems determine the assumptions in
Section 6 on the Sobolev inequalities.
Acknowledgements. It is a pleasure to thank Brazilian tax payers for the ex-
cellent research and teaching opportunities at UNICAMP and for generous financial
support: “O presente trabalho foi realizado com apoio do CNPq, Conselho Nacional
de Desenvolvimento Cientı́fico e Tecnológico - Brasil, e da FAPESP, Fundação de Am-
paro à Pesquisa do Estado de São Paulo - Brasil.”
Many thanks to Andrey Antônio Alves Cabral Júnior and Matheus Frederico Stapen-
horst for interest in and many pleasant conversations throughout the lecture course
“MM692 Análise Real II” held in the second semester of 2018 at UNICAMP.
1 The present manuscript is not complete and won’t be in the near future.
1
2 CHAPTER 1. INTRODUCTION
1.1 Notation and conventions

partial derivative ∂i u = ∂xi or Dei – weak derivative uxi or uei
Ω open subset of Rn where n ≥ 1
∂A := Ā ∩ AC boundary of set A ⊂ Rn where Ā is the closure of A
Q b Ω pre-compact subset Q of Ω:
An open subset of Ω whose closure Q̄ is compact and contained in Ω
D b Ω Lipschitz domain (∂D is locally a Lipschitz graph; see Definition 5.1.5)
K compact set (i.e. bounded and closed)
u = [u] equivalence class; the boldface notation2 eases Dα [u] to Dα u, whereas the
bracket notation clarifies the definition Dα u := [Dα u] as opposed to D α u
C k spaces
A map taking values in the real line R is called a function.
C k (Ω) set of functions on Ω all of whose partial derivatives up to order k exist

and are continuous, equipped with (but incomplete under) the norm
k·kC k maximum of sup-norms of all partial derivatives up to order k
Cbk (Ω) := {f ∈ C k (Ω) : kf kC k < ∞} is the Banach space under k·kC k of k times
bounded continuously differentiable functions on Ω C k (K) = Cbk (K)
C k (Ω̄) consists 3 of those functions f : Ω → R in C k (Ω) whose partial derivatives

up to order k are uniformly continuous on bounded subsets of Ω; cf. [Eva98,
§A.3]. So each such derivative continuously extends to Ω̄
Note. In C k (Ω̄) the bar is notation only, it does not denote the closure.
Otherwise, there is ambiguity for unbounded sets, such as Ω = Rn = Rn .
Cbk (Ω̄) := {f ∈ C k (Ω̄) : kf kC k < ∞} is a closed (hence Banach) subspace of the

Banach space Cbk (Ω); cf. [AF03, §1.28] C k (Q̄) = Cbk (Q̄)
C·∞ (·) := ∩∞ k
k=0 C· (·)
2 concerning notation we distinguish functions u and their equivalence classes [u] = u, but
for operators on them we use the same symbol, e.g. D, since Du and Du indicate context
3 In the pre-compact case C k (Q̄) is the set of restrictions to Q of smooth functions de-
fined on a neighborhood of Q̄: uniformly continuous derivatives up to order k follows from

compactness of Ω̄.
1.1. NOTATION AND CONVENTIONS 3
Hölder spaces
Observe: Hölder ⇒ uniform continuity ⇒ continuous extension to boundary.
|f |C 0,µ (Ω) := supx6=y |f (x)−f

|x−y|µ
(y)|
Hölder coefficient of f , Hölder exponent µ ∈ (0, 1].
One calls f µ-Hölder continuous on Ω :⇔ |f |C 0,µ (Ω) < ∞. Hölder implies
uniform continuity, cf. Figure A.1, hence f extends continuously to Ω̄
kf kC k,µ (Ω) := max|β|≤k kDβ f kC 0 (Ω) + max|α|=k |Dα f |C 0,µ (Ω) Hölder norm4
C k,µ (Ω) := {f ∈ C k (Ω) : |Dβ f |C 0,µ < ∞ ∀|β| ≤ k} = C k,µ (Ω̄) vector space of func-
tions which are uniformly Hölder continuous on Ω and so are all derivatives
up to order k (hence they all extend continuously to the boundary)
k,µ
Cloc (Ω) := {f ∈ C k (Ω) : f ∈ C k,µ (Q) ∀Q b Ω} ⊃ C k,µ (Ω) local Hölder space
Cbk,µ (Ω) := {f ∈ C k,µ (Ω) : kf kC k,µ < ∞} Hölder Banach space with respect to
the norm k·kC k,µ ; idea of proof e.g. here or here Cbk,µ (Q) = C k,µ (Q)5
Integration will always be with respect to Lebesgue measure m on Rn , unless

mentioned otherwise. We follow the notation in [Eva98, App. A] for derivatives:
Derivatives
Suppose u : Ω → R is a function and x = (x1 , . . . , xn ) ∈ Ω is a point.
∂u u(x+hei )−u(x)
(a) ∂xi (x) := limh→0 h , provided this limit exists.
∂u ∂2u
(b) We abbreviate ∂xi by ∂xi u or ∂i u. Also ∂ij u := ∂xi ∂xj etc.
(c) Multi-index notation.
(i) A list α = (α1 , . . . , αn ) of integers αi ≥ 0 is a multi-index of order
|α| := α1 + · · · + αn .
α α!

Set α! := α1 ! . . . αn ! and for α ≥ β (all αi ≥ βi ) set β := β!(α−β)! .
(ii) For a multi-index α the associated partial derivative is denoted by
∂ |α| u(x)
Dα u(x) := = ∂xα11 . . . ∂xαnn u(x).
∂x1 . . . ∂xn αn
α1
(iii) The list of all partial derivatives of order k ∈ N0 is denoted by

k
Dk u(x) := (Dα u(x))|α|=k ∈ Rn .
4Why |α| = k suffices? If kf kC k is finite, so are all Hölder coefficients of Dβ f with |β| < k.
5’⊃’ All derivatives of f ∈ C k,µ (Q) up to order k are µ-Hölder, thus uniformly continuous,
hence they extend to the closure - which is compact - so they are bounded.
4 CHAPTER 1. INTRODUCTION
(iv) Special cases. If k = 1, one obtains a list D1 u denoted by
Du = (∂1 u, . . . , ∂n u) = gradient vector

q
2 2
and we set |Du| := |∂1 u| + · · · + |∂n u| . The list Du of first order
partial derivatives is also called the strong gradient.
If k = 2, one obtains a symmetric n × n matrix
 ∂2u 2 
∂x21
. . . ∂x∂1 ∂xu
n
D2 u = 
 .. 
 = Hessian matrix.
 . 
2 2
∂ u ∂ u
∂xn ∂x1 ... ∂x2
n
(iv) Weak derivatives. We usually denote the weak derivative of u

corresponding to α by uα . In the smooth case uα = Dα u coincides
with the ordinary derivative.
Order one. Weak derivatives in direction of the ith coordinate unit
vector α = ei := (0, . . . , 0, 1, 0, . . . , 0) are denoted by uei .
In abuse of notation we also denote by Du the list (ue1 , . . . , uen ) of
weak derivatives. But we call it the weak gradient for distinction.
Chapter 2
Lp-spaces
This chapter is an exposition of notions and results a.e. without proofs.
2.1 Borel and Lebesgue measure space on Rn

Topology
The standard topology on euclidean space Rn is the collection Un of subsets

of Rn that arises as follows. Consider the family of all open balls centered at the
points of Rn of all radii. Now add all finite intersections of these balls. Then
n
add all arbitrary unions of members of the enlarged family to obtain Un ⊂ 2R .
X
By 2 we denote the power set of a set X: the collection of all subsets of X.
The elements U of Un are called open sets and their complements U C := Rn \U
are called closed sets.
Lemma 2.1.1. Every open subset Ω of Rn is of the form Ω = ∪∞

i=1 Ki for a
nested sequence of compact subsets Ki ⊂ Ki+1 ⊂ Ω.
Proof. [Bre11, Cor. 4.23]: Ki := {x ∈ Rn | dist(x, ΩC ) ≥ 2/i and |x| ≤ i}.
Measure compatible with topology – Borel measure
There are many ways to enlarge a given collection C of subsets of Rn to obtain a

family of subsets that satisfies the axioms of a σ-algebra; see e.g. [Sal16, Ch. 1].
n
However, there is a smallest such family, denoted by AC ⊂ 2R and called the
n
smallest σ-algebra on R that contains the collection C. The smallest σ-algebra
on a topological space that contains all open sets is called the Borel σ-algebra.
In case of Rn with the standard topology Un we use the notation Bn := AUn for
the Borel σ-algebra on Rn .
A measure is a function µ : A → [0, ∞] whose domain is a σ-algebra such
that, firstly, at least one family member A ∈ A has finite measure µ(A) < ∞
5
6 CHAPTER 2. LP -SPACES
and, secondly, the function µ is σ-additive.1 The elements A of a σ-algebra

are called measurable sets, those of measure zero, i.e. µ(A) = 0, null sets.
On Bn there is a unique measure µ, called Borel measure on Rn , which
is translation invariant and assigns measure 1 to the unit cube [0, 1]n .
Completion of Borel measure – Lebesgue measure

It is reasonable to expect that the measure of a subset should not be larger
than the measure of the ambient set. If the ambient set has measure zero,
subsets should be of measure zero, too. However, the domain of a measure is
a σ-algebra, a family of sets. But a subset C of a member A, even if A is of
measure zero, is not necessarily a member itself, hence not in the domain of the
measure.
Such annoying kind of incompleteness happens in Borel measure space
(Rn , Bn , µ), namely, subsets of null sets are not necessarily measurable. How-
ever, there is a completion procedure that results in a larger collection An ⊃ Bn ,
the Lebesgue σ-algebra, together with a measure m : An → [0, ∞], the
Lebesgue measure. The desired completeness property holds true: subsets of
Lebesgue null sets are Lebesgue null sets. Moreover, the restriction of Lebesgue
measure to the Borel σ-algebra m|Bn = µ coincides with Borel measure. For
details see e.g. [Sal16, Ch. 2].
Functions – equality almost everywhere and measurability

One says that two functions f, g : Rn → R are equal almost everywhere, in
symbols f = g a.e., if the set {f 6= g} ⊂ Rn of points on which f and g differ
is a Lebesgue null set: an element of An of Lebesgue measure zero.
A function f : Rn → R is called Lebesgue measurable if Rn is equipped
n
with the Lebesgue σ-algebra An ⊂ 2R and R with the Borel σ-algebra B1 ⊂ 2R
and the pre-images f −1 (B) ∈ An of B1 -measurable sets, i.e. B ∈ B1 , are An -
measurable sets; see [Sal16, Def. 2.2]. The function is called Borel measurable
if the pre-images f −1 (B) ∈ Bn ⊂ An even lie in the Borel σ-algebra.
Recall that a map between topological spaces is continuous if pre-images
of open sets are open. Every continuous map f : Rn → R is Borel measurable;
see e.g. [Sal16, Thm. 1.20].
Measurable support
It is common to define the support of a function on a topological space to be
the complement of the largest open set on which f vanishes or, equivalently,
the closure of the set {f 6= 0}. The usual symbol is supp f . In contrast,
for measurable functions it is useful to replace ’vanishes’ by ’vanishes almost
everywhere’; cf. [Bre11, §4.4]. The symbol will be suppm f . This way functions
which are equal a.e. will have the same support in the new sense.
1 If (A )∞ ⊂ A is a countable collection of pairwise disjoint sets then µ assigns to their
i i=1 P∞
union the sum of all individual measures, in symbols µ(∪∞
i=1 Ai ) = i=1 µ(Ai ).
2.1. BOREL AND LEBESGUE MEASURE SPACE ON RN 7
Definition 2.1.2 (Support of measurable functions). Suppose f : Rn → R is

Lebesque measurable. Consider the family (Ωλ )λ∈Λ of all open subsets of Rn
such that f = 0 a.e. on each of them. By definition the (measurable) support
of f is the complement of the union Ω := ∪λ∈Λ Ωλ , in symbols
suppm f := ΩC .
Since Ω is open suppm f is closed.
Exercise 2.1.3. Check that supp χQ = R, while suppm χQ = ∅. The latter

seems more reasonable for a function that is zero almost everywhere.
Exercise 2.1.4 (Ordinary and measurable support).
(i) Show that f = 0 a.e. on Ω, equivalently, that N := {f 6= 0} is a null set.
(ii) If f = g almost everywhere, then suppm f = suppm g.
(iii) For continuous functions both supports suppm f = supp f coincide.
[Hints: (i) That f = 0 a.e. on Ωλ means Nλ := {f 6= 0} ∩ Ωλ is a null set. While

N = ∪λ∈Λ Nλ , the union is not necessarily countable. Recall that the standard
topology of Rn has a countable base: There is a countable family (On ) of open
sets such that any open set, say Ωλ , is the union of some On ’s.]
Integration of Lebesgue measurable functions

The construction of the Lebesgue integral starts out from characteristic func-
tions χA : Rn → R. By intuition, for a constant c ≥ 0, the value of the integral
of cχA should be the area below its graph, that is the measure of A (length if
n = 1) times the height c. But A Ronly has a measure if it is a member of the
σ-algebra. In this case one defines Rn cχA dm := c · m(A) and correspondingly
Pk
in case of finite sums s = 1 ci χAi ≥ 0 (called simple functions).
The fact that the characteristic function χA : Rn → R of a subset A ⊂ Rn
is Lebesgue measurable iff the set A is Lebesgue measurable, brings in measur-
ability of the functions to be integrated. A non-negative Lebesgue measurable
function f : Rn → [0, ∞] can be approximated from below by simple functions
and one defines the Lebesgue integral of such f ≥ 0, denoted by
Z
f dm ∈ [0, ∞],
Rn
as the supremum of the integrals of simple functions s with 0 ≤ s ≤ f . Decom-

pose a real valued Lebesgue measurable f = f + − f − : Rn →R R intoR its positive
and negative parts f ± (x) := max{±f (x), 0} ≥ 0 and define f := f + − f −
R
if at least one of the two terms is finite. Such fR is saidR to admit

R Lebesgue
integration. If both terms are finite, that is if |f | = f + + f − < ∞, the
measurable function f : Rn → R is called (Lebesgue) integrable. By L1 (Rn )

one denotes the set of all integrable functions on Rn . For A ∈ An define
Z Z
f dm := χA f dm.
A Rn
The frequent problem of interchanging integration and limit is settled by the
Theorem 2.1.5 (Lebesgue dominated convergence theorem). Let all functions

be defined on a measurable set E. Let fk be a sequence of measurable functions
that converges pointwise a.e. to a measurable function f and is dominated by
an integrable function g in the sense that
∀k ∈ N : |fk (x)| ≤ g(x) for a.e. x ∈ E.
Then the fk and also f are integrable and

Z Z
f dm = lim fk dm. (2.1.1)
E k→∞ E
Proof. E.g. [Sal16, Thm. 1.45 and text after Cor. 1.56] or [Fol99, Thm. 2.24].
Theorem 2.1.6 (Generalized Lebesgue dominated convergence theorem). Let

all functions be defined on a measurable set E. Let f, {fk } be measurable with
fk → f a.e. and g, {gk } be integrable with gk → g a.e. and |fk | ≤ gk pointwise2
for each k. Then the fk and also f are integrable
Z Z Z Z
lim gk = g, ⇒ lim fk = f. (2.1.2)
k→∞ E E k→∞ E E
Proof. Exercise.
Convention 2.1.7. Throughout we work with Lebesgue measure m : An →

[0, ∞]. Measurable
R means LebesgueR measurable, Runless mentioned otherwise.
Instead of Rn f dm we usually write Rn f or even Rf . If we wish toR emphasize
the variable f = f (x) of the function we shall write Rn f (x) dx or Rn f dx.
2.2 Definition
Vector space Lp of functions
Suppose f : Rn → R is a measurable function. For finite p ∈ [1, ∞) the Lp -
norm of f is the extended real defined by
Z 1/p
p
kf kp := |f | ∈ [0, ∞]. (2.2.3)
Rn
2 here pointwise, as opposed to a.e. pointwise, is crucial so to g − f ≥ 0 applies Fatou’s
k k
Lemma for non-negative measurable functions
2.2. DEFINITION 9
On the other hand, consider the (possibly empty) semi-infinite interval
If := {c ∈ [0, ∞) : the set {|f | > c} is of measure zero} ⊂ [0, ∞).
The elements c of If tell that |f | > c happens only along a null set. The infimum
of such c is called the L∞ -norm of f and denoted by
kf k∞ := inf If = ess sup |f |. (2.2.4)
By convention inf ∅ = ∞, hence kf k∞ ∈ [0, ∞]

Definition 2.2.1 (Lp -spaces). Suppose p ∈ [1, ∞]. A measurable function
f : Rn → R of finite Lp -norm, i.e. kf kp < ∞, is called p-integrable or an
Lp -function. A 1-integrable function is called integrable. The set Lp (Rn ) of
Lp -functions is a real vector space. For finite p closedness under addition holds
by Minkowski’s inequality. Failure of non-degeneracy causes that the function
k·kp is actually not a norm, at least not on Lp (Rn ).
Definition 2.2.2 (Local Lp -spaces). Suppose p ∈ [1, ∞]. Let Lploc (Rn ) be the
set of all measurable functions f : Rn → R such that χQ f is p-integrable for
every pre-compact Q b Rn or, equivalently, such that χK f is p-integrable for
every compact K ⊂ Rn .
Definition 2.2.3 (Hölder conjugates). Two reals p, q ∈ [1, ∞] with p1 + 1q = 1
are called Hölder conjugates of one another. It is also common to denote the
Hölder conjugate if p by p0 . Suppose p, q, r ∈ [1, ∞] satisfy p1 + 1q = 1r , then p, q
are called Hölder r-conjugates of one another.
Theorem 2.2.4 (Hölder). Suppose p, q ∈ [1, ∞] are Hölder conjugates and
f ∈ Lp (Rn ) and g ∈ Lq (Rn ). Then the product f g is integrable and
Z
|f g| ≤ kf kp kgkq . (2.2.5)
Rn
Proof. See e.g. [Sal16, Thm. 4.1].

The theorem also applies to functions f, g : A → R which are defined on a
measurable set A ∈ An and whose extensions ˜ g̃ : Rn → R to Rn by zero are
f,
measurable. In this just apply the theorem to the extensions.
Corollary 2.2.5 (Hölder-r). Suppose p, q ∈ [1, ∞] are Hölder r-conjugates and
f ∈ Lp (Rn ) and g ∈ Lq (Rn ). Then f g ∈ Lr (Rn ) and
kf gkr ≤ kf kp kgkq . (2.2.6)
Proof. Apply Hölder to |u|r |v|r ∈ L1 with Hölder conjugates p/r and q/r.
Exercise 2.2.6. Suppose p1 , . . . p` , r ∈ [1, ∞] satisfy p11 + · · · + p1` = 1r and
fi ∈ Lpi (Rn ) for i = 1, . . . , `. Then the product f1 f2 . . . f` lies in Lr (Rn ) and
kf1 f2 . . . f` kr ≤ kf1 kp1 . . . kf` kp` . (2.2.7)

Theorem 2.2.7 (Lebesgue dominated Lp convergence theorem). Let all func-

tions be defined on a measurable set E. Suppose fk is a sequence of measurable
functions that converges pointwise a.e. to a measurable function f and is dom-
inated by a p-integrable function g for some p ∈ [1, ∞) in the sense that
∀k ∈ N : |fk (x)| ≤ g(x) for a.e. x ∈ E.
Then the fk and also f are p-integrable and fk converges to f in Lp , in symbols
lim kfk − f kp = 0.
k→∞
p
Proof. Theorem 2.1.5 with Fk := |fk − f | , F := 0, and G := (2g)p .
Theorem 2.2.8 (Generalized Lebesgue dominated Lp convergence theorem).

Let p ∈ [1, ∞). Let all functions be defined on E. Let f, {fk } be measurable
with fk → f a.e. and g, {gk } be p-integrable with gk → g a.e. and |fk | ≤ gk
pointwise for each k. Then the fk and also f are p-integrable and
kgk − gkp → 0 ⇒ kfk − f kp → 0. (2.2.8)
Proof. Exercise.
Looks like a complification of Theorem 2.1.6 only? Well, the theorem can
be a real peacemaker, just wait for the action, e.g. in Proposition 4.1.21.
Banach space Lp of equivalence classes
The function Lp (Rn ) → [0, ∞), f 7→ kf kp , satisfies all norm axioms except non-
degeneracy. Indeed kf kp = 0 only tells that f = 0 almost everywhere. Thus it
is natural to quotient out by equality almost everywhere to obtain the vector
space of equivalence classes
Lp (Rn ) := Lp (Rn )/ ∼, Lploc (Rn ) := Lploc (Rn )/ ∼,
where by definition f ∼ g if f = g almost everywhere. We use either notation

[f ] or f for equivalence classes. Because the Lebesgue integral is insensitive to
sets of measure zero, the definition
kf kp := kf kp
makes sense and provides a norm on Lp (Rn ).
Exercise 2.2.9. Illustrate by examples that the property “u has a continuous

representative” is not the same as “u is continuous a.e.”.
2.3. BASIC PROPERTIES 11
2.3 Basic properties

Theorem 2.3.1. (i) Lp (Rn ) is a Banach space for p ∈ [1, ∞].
(ii) Lp (Rn ) is separable 3 for finite p ∈ [1, ∞).
(iii) Lp (Rn ) is reflexive 4 for finite p ∈ (1, ∞) larger 1.
(iv) C0 (Rn ) is a dense subset of Lp (Rn ) for finite p ∈ [1, ∞).5
Proof. See e.g. [Sal16, (i) Thm. 4.9, (ii) Thm. 4.13, (iv) Thm. 4.15] or [Bre11, (i)
Thm. 4.8, (ii) Thm. 4.13, (iii) Thm. 4.10].
The properties (i) completeness, (ii) separability, and (iii) reflexivity are
hereditary to Banach subspaces, that is they are inherited by closed subspaces
of a Banach space. Part (iv) will be improved from continuous to smooth as an
application of convolution in Theorem 3.2.3 (ii).
Lemma 2.3.2 (Continuity of shift operator in compact-open topology). Given
p ∈ [1, ∞), the shift operator is for ξ ∈ Rn defined by
τs : Lp (Rn ) → Lp (Rn ), f 7→ τs f := f (· + sξ), s ∈ [0, ∞).
It satisfies τt τs = τt+s and τ0 = 1l. Given f ∈ Lp (Rn ), the path
γ : [0, ∞) → Lp (Rn ), s 7→ τs f ,
is continuous.
Proof. It suffices to prove continuity at s = 0. Suppose f ∈ f ∈ Lp (Rn ) and let

ε > 0. It would be nice if f was uniformly continuous. So let’s approximate f
by φ ∈ C0 (Rn ) with kf − φkp < ε/3 using Theorem 2.3.1. Since φ is continuous
and of compact support it is uniformly continuous. Hence there is a constant
δ = δ(ε) > 0 such that |s| < δ implies kτs φ − φkp < ε/3. We get that
kτs f − f kp ≤ kτs (f − φ)kp + kτs φ − φkp + kφ − f kp < ε/3 + ε/3 + ε/3.
We used linearity of τs and kτs (f − φ)kp = kf − φkp since τs is an isometry.

Exercise 2.3.3 (Shift operator in operator norm topology). Is the shift operator
continuous with respect to the norm topology? I.e. is
τ : [0, ∞) → L(Lp (Rn )), s 7→ τs ,

kτs f kp
continuous? The operator norm is given by kτs k := supf 6=0 kf kp .
3 there must be a dense sequence

4 the canonical injection J : E E ∗∗ that assigns to x ∈ E the continuous linear functional
on E ∗ given by [E ∗ → R : f 7→ f (x)] ∈ E ∗∗ must be surjective, hence E ' E ∗∗ canonically
5 ’subset’ actually refers to the image of the natural injection C (Rn ) → Lp (Rn ), f 7→ [f ]
0
Chapter 3
Convolution
Roughly speaking, the convolution product associates to two adequately inte-

grable functions f and g on Rn an integrable function, denoted by f ∗ g, which
inherits nice properties from one of the factors, say smoothness of f , but still re-
sembles very much the other factor g, if f is chosen to have very small support. A
major application is to k·k1 -approximate an integrable g by the smooth function
f ∗ g by choosing f appropriately. One obtains the fundamental density result
that the smooth compactly supported functions form a dense subset among
the integrable ones. This remains valid for the class of functions with domain
Ω ⊂ Rn . We recommend the presentations in [Sal16, §7.5] and [Bre11, §4.4].
An integrable function f on Ω ⊂ Rn naturally corresponds to an integrable
function on Rn by extending f by zero outside Ω. We denote this natural zero
extension again by f . This does not work for locally integrable functions. A
way out is to restrict first to a closed subset of Ω and then extend the restriction
by zero. See Section 3.3 on mollification of functions with domain Ω.
3.1 Convolution of functions

In this section we deal with functions on Rn . If a function f is given on Ω ⊂ Rn
just replace it by the natural zero extension f : Rn → R which assigns constantly
the value zero outside Ω.
Definition 3.1.1. For Lebesgue measurable f, g : Rn → R define the bad set1

Z
n
E(f, g) := x ∈ R h(x) := |f (x − y)g(y)| dy = ∞ .

Rn
1 The function y 7→ f (x − y)g(y) = (f ◦ φ )(y)g(y) is Lebesgue measurable, independent

x
of x, since f, g are and φx (y) := x − y is a homeomorphism of Rn . Thus not to be Lebesgue
integrable, as required in [Sal16, (7.30)] to define E(f, g), just means ’infinite integral’.
13
14 CHAPTER 3. CONVOLUTION
The convolution of f and g is the function f ∗ g : Rn → R defined by

(R
Rn
f (x − y)g(y) dy , for x ∈ E(f, g)C ,
(f ∗ g) (x) :=
0 , for x ∈ E(f, g).
Of course, to have any chance that the function f ∗ g : Rn → R be measur-

able, let alone integrable, one needs that E(f, g) is a measurable set to start
with. Ideally E(f, g) should be of measure zero, so it becomes invisible when
integrating f ∗ g over Rn . Note that m(E(f, g)) = 0 means that the function
y 7→ f (x − y)g(g) is integrable for a.e. x ∈ Rn . If both f and g are compactly
supported, so is their convolution f ∗ g by Theorem 3.1.6 (i).
˜ g̃ : Rn → R it is true that
Theorem 3.1.2. For Lebesgue measurable f, g, f,
(i) if f = f˜ a.e. and g = g̃ a.e., then E(f, g) = E(f,

˜ g̃) and f ∗ g = f˜ ∗ g̃;
(ii) the bad set E(f, g) ∈ Bn is Borel and f ∗ g is Borel measurable;
(iii) convolution is commutative: indeed E(f, g) = E(g, f ) and f ∗ g = g ∗ f .
Proof. [Sal16, Thm. 7.32].
While the theorem asserts that the convolution of measurables is measur-

able, even Borel measurable, and so is the bad set, the question whether the
convolution of integrables is integrable is answered next. The answer is positive,
even for generalized (r 6= ∞) conjugate exponents.
Theorem 3.1.3 (Young’s inequality). Given Hölder r-conjugates p, q ∈ [1, ∞],

suppose f ∈ Lp (Rn ) and g ∈ Lq (Rn ). Then the bad set E(f, g) is of measure
zero, empty for r = ∞ i.e. p1 + 1q = 1, and the convolution satisfies the estimate
kf ∗ gkr ≤ kf kp kgkq . (3.1.1)
So the Borel measurable function f ∗ g lies in Lr (Rn ).
Proof. [Sal16,
R Thm. 7.33]. Emptiness of E for r = ∞ holds by Hölder (2.2.5)
as h(x) = Rn |f (x − y)g(y)| dm(y) ≤ kf (x − ·)kp kgkq = kf kp kgkq < ∞.
Remark 3.1.4 (Case r = ∞). Theorem 3.1.3 asserts that in case of Hölder
conjugate exponents, i.e. f ∈ Lp and g ∈ Lq where p, q ∈ [1, ∞] with p1 + 1q = 1,
one has an everywhere defined convolution f ∗ g ∈ L∞ (Rn ). So f ∗ g is bounded.
By Theorem 3.1.6 (ii) below f ∗ g is even uniformly continuous.
Remark 3.1.5 (Convolution and local integrability). Let p, q ∈ [1, ∞] be

Hölder r-conjugates. For locally p-integrable f ∈ Lploc (Rn ) and compactly sup-
ported g ∈ Lq (Rn ) the bad set E(f, g) is still of measure zero and the convolution
f ∗ g ∈ Lrloc (Rn ) is locally r-integrable; cf. paragraph after Def. 7.34 in [Sal16].
3.2. CONVOLUTION OF EQUIVALENCE CLASSES 15
Continuity, differentiability, support

Theorem 3.1.6. Let p, q ∈ [1, ∞] be conjugate. Then the following is true.
(i) If f ∈ Lp (Rn ) and g ∈ L1 (Rn ), then suppm (f ∗ g) ⊂ suppm f + suppm g.
(ii) If f ∈ Lp (Rn ) and g ∈ Lq (Rn ), then f ∗ g is uniformly continuous.
(iii) If f ∈ Lp (Rn ) has compact support and g ∈ Lqloc (Rn ), then f ∗g ∈ C 0 (Rn ).
(iv) If ϕ ∈ C0k (Rn ), k ≥ 1, and g ∈ L1loc (Rn ), then ϕ ∗ g ∈ C k (Rn ) and
∂ α (ϕ ∗ g) = (∂ α ϕ) ∗ g (3.1.2)
for every multi-index with |α| ≤ k.
Proof. [Bre11, (i) Prop. 4.18, (iv) Prop. 4.20] and [Sal16, (ii-iv) Thm. 7.35].
3.2 Convolution of equivalence classes

By Theorem 3.1.2 and A.1.1 convolution of functions descends to a map
Lp (Rn ) × Lq (Rn ) → Lr (Rn ), ([f ], [g]) 7→ [f ∗ g]. (3.2.3)

1 1
whenever p, q, r ∈ [1, ∞] are such that p + q = 1 + 1r .2
Remark 3.2.1. Convolution provides
- a commutative (and associative) product on L1 (Rn ); (p = q = r = 1)
- an action of L1 (Rn ) on Lp (Rn ). (q = 1, r = p)
Definition 3.2.2 (Mollifier). A mollifier on Rn is a smooth symmetric 3

func-
n
R
tion ρ : R → [0, ∞) supported in the open unit ball B1 and with Rn ρ = 1.
For each δ ∈ (0, 1] define the rescaling ρδ (x) := δ −n ρ(δ −1 x) forR x ∈ Rn . While
ρδ is even supported in Bδ ⊂ B1 , it is still of unit integral 1 = Rn ρδ = kρδ k1 .
Theorem 3.2.3. Suppose p ∈ [1, ∞) is finite. Then the following is true.
(i) Fix a mollifier ρ and f ∈ Lp (Rn ), then kρδ ∗ f − f kp → 0, as δ → 0.
(ii) C0∞ (Ω) is a dense subset of Lp (Ω) for Ω ⊂ Rn open, bounded or not.4
Proof. See e.g. [Bre11, (i) Thm. 4.22, (ii) Cor. 4.23] or [Sal16, (ii) Thm. 7.35].
Exercise 3.2.4. Check that the density result (ii) fails for p = ∞ on Ω = Rn .
pq pq q p
2 Note that r =
p+q−pq
≥ 1 iff p+q ≥ 12 iff p ≥ 2q−1 iff q ≥ 2p−1 .
q p
Also note that r < ∞ iff p + q > pq iff p < q−1 iff q < p−1 .
3 a function f is called symmetric if f (−x) = f (x) for every x
4 ’subset’ actually refers to the image of the natural injection C ∞ (Ω) → Lp (Ω), f 7→ [f ]
0
Remark 3.2.5. Concerning part (i): Note that ρδ ∗ f ∈ Lp (Rn ) by Young’s

inequality (A.1.1) with r = p and q = 1. As mentioned earlier, if f ∈ Lp (Ω)
use the natural zero extension, still denoted by f ∈ Lp (Rn ), to get convergence
ρδ ∗ f → f in the Lp -norm. If f ∈ Lploc (Ω), use the δ-extension f¯ in (3.3.4)
below to get k·kp -convergence along pre-compacts Q b Ω, see 3.3.1.
Concerning part (ii): The density result fails for Sobolev spaces on bounded
domains Q, whereas on Rn density does generalize to Sobolev spaces; see Re-
mark 4.2.8 and Theorem 5.1.1 (iii), respectively.
3.3 Local Mollification

Convolution with a compactly supported smooth function can be used to
smoothen out a locally integrable function on Rn , see (3.1.2). Let us detail
this in the general case of a locally integrable function that is only defined on
Ω ⊂ Rn . Note that, in contrast to integrability, local integrability is not neces-
sarily inherited by the natural zero extension – but by the zero δ-extension.
As Lploc ⊂ L1loc by Hölder, let us deal with the largest space L1loc right away.
3.3.1 Locally integrable functions

Suppose f ∈ L1loc (Ω) is a locally integrable function on a non-empty open subset
Ω ⊂ Rn . For δ ≥ 0 consider the open subset of Ω defined by
Ωδ := {x ∈ Ω | d(x, ∂Ω) > δ}, Ωi := Ω1/i , i ∈ N.
Observe that Ωi ⊂ Ωi+1 and ∪i∈N Ωi = Ω = Ω0 . By openness of Ω its subset
Ωδ is non-empty whenever δ > 0 is sufficiently small.
Note that the natural zero extension of f to Rn is not necessarily locally
integrable, as we do not have control how f behaves near ∂Ω. A way out is
to restrict f to a closed subset of Ω and extend this, still locally integrable,
restriction to Rn by zero. To do this in a mollification compatible way we define
the zero δ-extension of f : Ω → R for δ > 0 by
(
(δ) f , on Ωδ ,
f¯ := f := f¯ ∈ L1loc (Rn ) if f ∈ L1loc (Ω). (3.3.4)
0 , on Rn \ Ωδ ,
Cf. Fig. 3.1. If Ω = Rn , then f¯ = f . Define the mollification of f ∈ L1loc (Ω) by
f δ := ρδ ∗ f¯ ∈ C0∞ (Ω) ⊂ C0∞ (Rn ). (3.3.5)

Smoothness and compact support are guaranteed by Theorem 3.1.6. Since f¯ = 0
outside Ωδ we have f δ = 0 outside Ω. So E := E(f δ ) := E(ρδ , f¯), the bad set,
is contained in Ω. The value of the mollification is given by
(R
δ ¯
ρ (x − y)f (y) dy , x ∈ E C ∩ Ω2δ ,
Bδ (x)⊂Ωδ δ
f (x) := ρδ ∗ f (x) =
0 , x ∈ E C ∩ Ω2δ ,
at points x ∈ Ω2δ . The bad set E(f δ ) ⊂ Ω is of measure zero by Remark 3.1.5.
3.3. LOCAL MOLLIFICATION 17
(δ)
Figure 3.1: The zero δ-extension f¯ = f ∈ L1loc (Ω) of f ∈ L1loc (Ω)
Lemma 3.3.1 (Local L∞ convergence). Given p ∈ [1, ∞] and f ∈ Lploc (Ω),

then ρδ ∗ f¯ ∈ C0∞ (Ω) converges to f in k·kL∞ (Q) along pre-compacts Q b Ω.
Proof. Lemma 3.3.2 via approximation; see e.g. [GT01, Le. 7.2].
3.3.2 Continuous functions

Suppose u is a continuous function on Ω ⊂ Rn . Since C 0 (Ω) ⊂ L1loc (Ω) we define
the mollification uδ = ρδ ∗ ū ∈ C0∞ (Ω) by (3.3.5). It has empty bad set E(uδ ) =
∅, since the extension ū = u(δ) is locally bounded.5 So the mollification of a
continuous function, also called a C 0 mollification, is given by6
Z
δ
u (x) := (ρδ ∗ ū) (x) = ρδ (x − y)u(y) dy
B (x)
Z δ (3.3.6)
= ρ(z)u(x + δz) dz
B1 (0)
for any x ∈ Ω2δ = {x ∈ Ω | d(x, ∂Ω) > 2δ}.

Lemma 3.3.2 (Local uniform convergence). For continuous functions u on Ω
the mollification uδ := ρδ ∗ ū converges to u uniformly on compact sets K ⊂ Ω.
Proof. Given K ⊂ Ω, pick δ > 0 sufficiently small such that Ω2δ ⊃ K. Then the
value of uδ := ρδ ∗ ū ∈ C0∞ (Ω) defined by (3.3.5) is given along K by (3.3.6). So
Z

sup u − uδ = sup

ρ(z) (u(x) − u(x + δz)) dz

K x∈K {|z|≤1}
≤ sup sup |u(x) − u(x + δz)|
x∈K |z|≤1 | {z }
≤cx |δz|
≤ cδ.
5
R R
Indeed Rn |ρδ (x − y)ū(y)|dy = B (x) |ρδ (x − y)ū(y)|dy ≤ kūkC 0 (Bδ (x)) < ∞.
δ
6 Change of variable z(y) := − x−y
δ
with dz = δ −n dy using the symmetry ρ(−z) = ρ(z).
We used continuity of u and compactness of the closure of the δ-neighborhood

Uδ (K) of K to replace the continuity constant cx of u at x by a constant c
uniform in K. It enters that K ⊂ Ω2δ implies Uδ (K) ⊂ Ωδ ⊂ Ω = dom u.
3.4 Applications
Lemma 3.4.1. Suppose v ∈ L1loc (Ω) has the property that
Z
vφ = 0 (3.4.7)
Ω
for all φ ∈ C0∞ (Ω). Then v vanishes almost everywhere, in symbols v = 0 a.e.
The lemma guarantees uniqueness of weak derivatives almost everywhere;
cf. Lemma 4.1.6.
The lemma is easy to prove for continuous v in which case the conclusion is that
v even vanishes pointwise. The lemma is already less easy to prove for non-
negative v ∈ L1loc (Ω). Pick φ ∈ C0∞ (Ω, [0, 1]) which is 1 over a given compact
K ⊂ Ω. That Lebesgue integral zero implies that a non-negative integrand
is zero a.e. indeed requires a bit of work; see e.g. [Sal16, Le. 1.49 (iii) ⇒(i)].
Consequently over K we get v = φv = 0 a.e., but K was arbitrary.
Proof. Pick a compact subset K ⊂ Ω of the open subset Ω of Rn . Fix χ ∈

C0∞ (Ω) with χ ≡ 1 along K. Extending χ and v to Rn by zero the product
vχ := χv ∈ L1 (Rn ) is an integrable function on Rn . Pick a mollifier ρ on Rn
with rescalings ρδ . Now, on the one hand, Theorem 3.2.3 (i) asserts that
[ρδ ∗ vχ ] → [vχ ] in L1 (Rn ), as δ → 0.
On the other hand, at any x ∈ Rn the function

Z Z
(ρδ ∗ vχ ) (x) = ρδ (x − y)χ(y)v(y) dy = ρδ (x − y)χ(y) v(y) dy = 0
Rn Ω| {z }
=:φx (y)
vanishes by hypothesis (3.4.7) since φx ∈ C0∞ (Ω): Indeed supp φx ⊂ supp χ ⊂ Ω.

By uniqueness of limits [χv] = [0] in L1 (Rn ). Thus along any compact subset
K ⊂ Ω we get that v = χv = 0 a.e. on K, in symbols m({v 6= 0} ∩ K) = 0.
It remains to conclude that v = 0 a.e. on Ω. To see this use the nested
sequence of compact sets Ki ⊂ Ki+1 ⊂ Ω = ∪∞ i=1 Ki provided by Lemma 2.1.1.
Set A := {v 6= 0} and consider the Lebesgue null sets Ai := A ∩ Ki to obtain that
Ai ⊂ Ai+1 and A = ∪∞ i=1 Ai . Thus m({v 6= 0}) = limi→∞ m(Ai ) = 0 by [Sal16,
Thm. 1.28 (iv)]. For a slightly different proof see [Bre11, Cor. 4.24].
Chapter 4
Sobolev spaces
In typical applications, say in the analysis of PDEs, whereas the property of

being continuously differentiable is of fundamental importance, it is often diffi-
cult to establish this property. Therefore its is desirable to introduce a weaker
version of differentiability, thereby enlarging the spaces C k , hence making it
easier to establish membership in the larger space.
The idea comes from Lebesgue integration theory where the value at an indi-
vidual point, in fact, along sets of measure zero (null sets), is not seen by the
integral. But insensibility to null sets ruins non-degeneracy of the natural norm
candidate. The way out is to identify functions that differ only along a null set.
In Section 4.1 we introduce the new weak concept of differentiability on the
level of functions. A key tool is local mollification from Section 3.3.
In Section 4.2 we quotient out by equality up to null sets and look at the
resulting Banach spaces, called Sobolev spaces, which come in the three flavours
k,p
Wloc , W k,p , and W0k,p .
4.1 Weak derivatives of locally integrable fcts

4.1.1 The mother of all Sobolev spaces L1loc
Key observation: Given two smooth compactly supported functions on Ω ⊂ Rn ,
the formula for partial integration still makes sense if one of the functions is just
a locally integrable function. Use the formula as criterion in a definition!
Definition 4.1.1 (Weak derivatives of locally integrable functions). Let α =

(α1 , . . . , αn ) be a multi-index of non-negative integers αi ∈ N0 . A locally inte-
grable function u : Ω → R, in symbols u ∈ L1loc (Ω), is said to admit a weak
derivative corresponding to α if there is some uα ∈ L1loc (Ω) such that
Z Z
∀φ ∈ C0∞ (Ω) : u(Dα φ) dm = (−1)|α| uα φ dm. (4.1.1)
Ω Ω
19
20 CHAPTER 4. SOBOLEV SPACES
The locally integrable function uα is called a weak derivative of u corre-

sponding to α. By linearity of the integral (u+v)α = uα +vα and (cu)α = cuα .
We call u weakly differentiable if it admits weak derivatives up to order 1.
Definition 4.1.2 (Vector spaces of functions admitting weak derivatives). For
k ∈ N0 and p ∈ (1, ∞] consider the vector spaces
k,1
Wloc (Ω) := {u ∈ L1loc (Ω) | all weak derivatives uα exist up to order k}
k,p k,1
Wloc (Ω) := {u ∈ Wloc (Ω) | Q b Ω, |α| ≤ k ⇒ kuα kLp (Q) < ∞}.
k,p k,p
We say that a sequence u` ∈ Wloc (Ω) converges in Wloc if there is an element
k,p
u ∈ Wloc (Ω) such that for any pre-compact Q b Ω
k(u` )α − uα kLp (Q) → 0, as ` → ∞,
k,p
for any weak derivative of order |α| ≤ k. In symbols u` → u in Wloc .
Definition 4.1.3 (Pre-Banach spaces). For k ∈ N0 and p ∈ [1, ∞] define
W k,p (Ω) := {u ∈ L1loc (Ω) | all uα exist up to order k and kuα kp < ∞}.
Remark 4.1.4 (Ordinary derivatives are weak derivatives). If u is of class C k ,

then any partial derivative of order ` ≤ k is a weak derivative by the theorem
of partial integration and compact support of the φ’s.
Remark 4.1.5 (Weak derivatives are not unique on the level of functions).
Suppose uα is a weak derivative of u corresponding to the multi-index α. Then
so is any ũα that differs from uα on a set of measure zero (same RHS in (4.1.1))
and besides these there are no other weak derivatives of u by Lemma 4.1.6.
Furthermore, any ũ that differs from u on a set of measure zero has the same
weak derivatives as u (same LHS in (4.1.1)).
Lemma 4.1.6 (Uniqueness almost everywhere). If u ∈ L1loc (Ω) has weak deriva-
tives uα and ũα , then uα = ũα a.e.
Proof. Lemma 3.4.1 with v = uα − ũα .
4.1.2 Examples
Exercise 4.1.7. The Cantor function c : [0, 1] → [0, 1], cf. [Sal16, Exc. 6.24],
does not admit a weak derivative, although the derivative of c exists almost
everywhere (namely along the complement of the Cantor set C) and is zero.
Exercise 4.1.8. Let I = (−1, +1). For any p ∈ [1, ∞] show the following:
(i) The absolute value function u(x) = |x| is weakly differentiable. A weak
derivative u1 is given by the sign function

−1 , x < 0,

sign (x) := 0 , x = 0,

+1 , x > 0.

4.1. WEAK DERIVATIVES 21
(ii) The sign function does not have a weak derivative.
[Hint: In case you get stuck consult [Eva98, §5.2.1].]
Note that for weak differentiability the corner of |·| did not matter, but the
jump discontinuity of sign was a problem. Check that sign is not absolutely
continuous. What is the problem with the Cantor function c?
Exercise 4.1.9. Let Q ⊂ R2 be the open unit ball and consider the function
u(x) := |x|−γ for x 6= 0 and u(0) := 0. For which values of γ > 0 and p ∈ [1, ∞]
• is u ∈ Lp (Q)?
• does u admit a weak derivative?
[Hint: In case you get stuck consult [Eva98, §5.2.2].]
4.1.3 ACL characterization

The notion of weak derivative is more general than partial derivative. This
leads to the question if some property of classical differentiability survives. An
excellent treatment of Sobolev spaces from a more geometric measure theory
point of view emphasizing the role of absolute continuity is presented in Ziemer’s
book [Zie89].
Theorem 4.1.10 (ACL characterization). Let p ∈ [1, ∞] and u ∈ Lp (Ω). Then

a function u lies in W 1,p (Ω) iff it admits a version u∗ , i.e. u = u∗ a.e., that has
the ACL property, i.e. u∗ is absolutely continuous on almost all line segments
in Ω parallel to the coordinate axis and whose partial derivatives ∂1 u∗ , . . . , ∂n u∗
exist pointwise a.e. (extend by zero) and are p-integrable.
Proof. [Zie89, Thm. 2.1.4]
In the terminology of Section A.1.2 there are the inclusions1
LC ⊂ AC ⊂ ACL ∩{∂1 u∗ , . . . , ∂n u∗ p-integrable} ⊂ Diff-a.e.

| {z }
W 1,p
For u ∈ Lp (Ω) one can reformulate the ACL theorem as follows: u ∈ W 1,p (Ω)
iff u has a version u∗ that lies in W 1,p (Λ) for almost all line segments Λ in Ω
parallel to the coordinate axes and the strong gradient satisfies |Du| ∈ Lp (Ω).
To obtain an equivalent statement replace almost all line segments Λ by
almost all k-dimensional planes Λk in Ω parallel to the coordinate k-planes.
For a characterization of W 1,p (Ω) through approximation see Section 4.1.5.
For a characterization of the Sobolev spaces W 1,p (Rn ), p ∈ (1, ∞), in terms of
difference quotients see Section 4.2.2.
1 strictly speaking, the identification is u∗ 7→ [u∗ ] ∈ W 1,p = W 1,p / ∼
4.1.4 Weak and partial derivatives

Lemma 4.1.11. Suppose a continuous function u : Ω → R is weakly differen-
tiable and ue1 , . . . , uen are even continuous (thus unique). Show that u ∈ C 1 (Ω)
and the weak derivatives coincide with the partial ones, in symbols
(ue1 , . . . , ue1 ) = (∂x1 u, . . . , ∂xn u).
Proof. As differentiation is a local problem, fix any pre-compact subset Q b Ω

and pick a constant δ > 0 such that Q b Ω2δ := {x ∈ Ω | dist(x, Ω) > 2δ}.
Recall that uδ := ρδ ∗ ū ∈ C0∞ (Ω) is given along Ω2δ by formula (3.3.6). Along
the closure of Ω2δ the formula does not see the choice of extension, so in that
domain it is safe to use the notation ρδ ∗ u ∈ C ∞ (Ω2δ ). The partial derivative
∂i uδ is according to (3.1.2) and at x ∈ Ω2δ given by
Z
∂i uδ (x) = (∂i ρδ ∗ u) (x) = (∂i ρδ ) (x − y) ·u(y) dy
Bδ (x)⊂Ωδ | {z }
=−∂i (ρδ (x−y))
Z
= (−1)2 ρδ (x − y) · uei (y) dy
Bδ (x)⊂Ωδ
= (ρδ ∗ uei ) (x).
The second step is by definition (4.1.1) of a weak derivative. By assumption u

and its weak derivative uei are continuous, so by Lemma 3.3.2 we have that
uδ = ρδ ∗ u → u, ∂i uδ = ρδ ∗ ue1 → uei , as δ → 0, (4.1.2)
uniformly along the compact set K := Q.

It remains to show that the partial derivatives of u exist on Q and are equal
to the (continuous) weak derivatives. But this means that u ∈ C 1 (Q). Because
Q b Ω2δ and δ > 0 were both arbitrary, this proves that u ∈ C 1 (Ω).
To this end pick z ∈ Q, then for any |h| > 0 so small that z + hei ∈ Q we get
u(z + hei ) − u(z) = lim uδ (z + hei ) − uδ (z)

δ→0
Z h
= lim ∂i uδ (z + sei ) ds
δ→0 0
Z h
= uei (z + sei ) ds.
0
Step one is by pointwise convergence uδ → u on Q. Step two holds by the

fundamental theorem of calculus. Step three holds by uniform convergence
∂i uδ → uei on Q.2 Divide by h and take the limit h → 0 to get that the partial
derivative of u at z ∈ Q exists and equals the weak one ∂i u(z) = uei (z).
2 By uniform convergence along Q the sequence ∂ uδ , modulo finitely many members, is
i
dominated by the integrable function uei + 1, so by the Lebesgue dominated convergence
Theorem 2.1.5 limit and integral commute.
4.1.5 Approximation characterization

Theorem 4.1.12. Suppose p ∈ [1, ∞] and u, v ∈ Lploc (Ω). Equivalent are
v = uα ⇔ ∃(u` ) ⊂ C0∞ (Ω) : u` → u, (u` )α → v in Lploc (Ω).
Proof. Proposition 4.1.13 and (4.1.2); cf. [GT01, Thm. 7.4].
k,p
A function u ∈ Wloc (Ω), with all weak derivatives up to order k, can be Lp
approximated along any pre-compact by a sequence of smooth functions u` .
Proposition 4.1.13 (Local approximation by smooth functions). Let k ∈ N

k,p
and p ∈ [1, ∞] and suppose u ∈ Wloc (Ω).3 Then there is a sequence of smooth
k,p
functions u` ∈ C0∞ (Ω) such that u` → u in Wloc , as ` → ∞.
Proof. Recall from (3.3.5) the definition of the mollification uδ := ρδ ∗ ū ∈

C0∞ (Ω) of a function u on Ω ⊂ Rn using the δ-extension ū = u(δ) : Rn → R, as
defined by (3.3.4). Along the closure of Ω2δ := {x ∈ Ω | dist(x, Ω) > 2δ} the
formula does not see the choice of extension, so in that domain it is safe to use
the notation ρδ ∗ u ∈ C ∞ (Ω2δ ).
Pick Q b Ω and let δ > 0 be small such that Ω2δ c Q. Now consider any
multi-index of order |α| ≤ k and let x ∈ Q. By Theorem 3.1.6 (iv) the derivative
of the convolution can then be thrown on the smooth factor, hence
Dα (ρδ ∗ u)(x) = (Dα ρδ ∗ u)(x)

Z
= Dxα ρδ (x − y) · u(y) dy
Ω
Z
|α|
= (−1) Dyα ρδ (x − y) ·u(y) dy
Ω | {z } (4.1.3)
=:φx ∈C0∞
Z
|α|+|α|
= (−1) ρδ (x − y) · uα (y) dy
Ω
u ∈ W k,p (Ω), x ∈ Q, Q b Ω .

= (ρδ ∗ uα )(x)
Here Step four is by definition (4.1.1) of the weak derivative of u. Note

that (4.1.3) proves that one can throw the derivative of a mollification onto
the factor of class W k,p in which case it turns into a weak derivative. Also
Dα uδ = ρδ ∗ uα → uα in L∞ (Q), so in Lp (Q), as δ → 0, (4.1.4)
by Lemma 3.3.1. The choice of extension is invisible along Q. For ` ∈ N set

u` := u1/` . This proves Proposition 4.1.13.
3 Without ’loc’ the assertion fails for p = ∞, as u ≡ 1 : R → R shows.

4.1.6 Bounded weakly differentiable means Lipschitz

Definition 4.1.14. A function u : Ω → R is called Lipschitz continuous if
|u(x) − u(y)| ≤ C |x − y|
for some constant C and all x, y ∈ Ω. The smallest such constant, say L, is called
the Lipschitz constant of u on Ω. It is called locally Lipschitz continuous
if it is Lipschitz continuous on every pre-compact Q b Ω or, equivalently, on
every compact K ⊂ Ω.
Exercise 4.1.15. Confirm “equivalently” above. [Hint: Figure A.1.]
0,1 1,∞ 0,1
Proposition 4.1.16 (Cloc = Wloc ). A function u : Ω → R belongs to Cloc iff
u is weakly differentiable with locally bounded weak derivatives.
Proof. See e.g. [Eva98, §5.8.2]. Cf. also [MS04, Exc. B.1.8].
4.1.7 Leibniz or product rule

For interesting different approaches and further information concerning the
present subsection and the previous one we refer to [Zie89, §2.1 §2.2].
Proposition 4.1.17 (Leibniz rule for weak derivatives). Given p ∈ [1, ∞], let
1,p 1,∞ 1,p
u ∈ Wloc (Ω) and v ∈ Wloc (Ω). Then uv lies in Wloc (Ω) with weak derivatives
(uv)α = uα v + uvα , |α| ≤ 1. (4.1.5)
Corollary 4.1.18. Given p ∈ [1, ∞], let u ∈ W 1,p (Ω) and v ∈ W 1,∞ (Ω). Then
the product uv lies in W 1,p (Ω) and it satisfies Leibniz (4.1.5) and the estimate
kuvk1,p ≤ kuk1,p kvk1,∞ . (4.1.6)

1,q
Proof. Since W 1,q ⊂ Wloc for q ∈ [1, ∞] Leibniz (4.1.5) holds. Now use the
estimate k(uv)α kp ≤ kuα kp kvk∞ + kukp kvα k∞ .
Proof of Proposition 4.1.17. The following is true for i = 1, . . . , n. By Proposi-
tion 4.1.13 there are sequences uk , vk ∈ C0∞ (Ω) such that, as k → ∞, it holds
∂uk
uk → u, = (uk )ei → uei in Lploc (Ω)
∂xi
∂vk p
vk → v, = (vk )ei → vei in L∞
loc (Ω), thus in Lloc (Ω).
∂xi
Note that Lploc convergence for finite p implies pointwise convergence almost
everywhere for a subsequence (same notation); cf. Figure A.3. By partial
integration and Leibniz for smooth functions we get for each k that
Z Z
∂φ ∂uk ∂vk
uk vk =− vk + uk φ ∀φ ∈ C0∞ (Ω).
Ω ∂xi Ω ∂x i ∂xi
Apply φ-wise the dominated convergence Theorem 2.1.5 to obtain in the limit4
Z Z
∂φ
uv =− (uei v + uvei ) φ ∀φ ∈ C0∞ (Ω).
Ω ∂xi Ω
This proves the formula (4.1.5) for the weak derivative of the product. The
product uv is of class Lploc (Ω) since v ∈ L∞
loc (Ω). The right hand side of (uv)ei
is of the same type ’Lp L∞ + Lp L∞ ’, hence of class Lploc (Ω).
Exercise 4.1.19. Deal with terms 2 and 3 in the previous proof; cf. footnote.
Remark 4.1.20 (Alternative hypotheses). Suppose p ∈ [1, ∞]. Then the space
W 1,p (Ω) ∩ L∞ (Ω) is preserved under taking the product uv of two elements u
and v. The product rule (4.1.5) still holds true; see [Bre11, Prop. 9.4].
4.1.8 Chain rule and change of coordinates

Proposition 4.1.21 (Chain rule – composition). Let p ∈ [1, ∞] and suppose
1,p
u ∈ Wloc (Ω). 5 Then for every function F ∈ C 1 (R) with bounded derivative F 0
1,p
the post-composition F ◦ u lies in Wloc (Ω) and its weak derivatives are given by
(F ◦ u)ei = F 0 (u) · uei , i = 1, . . . , n.

1,1
For necessary and sufficient conditions for the chain rule in Wloc (Rn , Rd )
see [LM07].
Proof. The following is true for every i = 1, . . . , n. By Proposition 4.1.13 there

is a sequence uk ∈ C0∞ (Ω) such that, as k → ∞, there is convergence
∂uk
uk → u, = (uk )ei → uei in Lploc (Ω), thus in L1loc (Ω).
∂xi
Note that Lploc convergence for finite p yields a.e. pointwise convergence of some
subsequence (for which we use the same notation); cf. Figure A.3.
Instead of doing things again ’by hand’ as in the previous proof, let us check
out the comfort of Theorem 4.1.12. Following [GT01, Le. 7.5] let us show that
fk := F ◦ uk → F ◦ u =: f in Lploc (Ω),
(fk )ei := F 0 (uk ) (uk )ei → F 0 (u) uei =: v in Lploc (Ω),
and that f, v ∈ Lploc (Ω). Theorem 4.1.12 tells that v = fα and we are done.
4
R R
Fix φ. Pick Q with supp φ b Q b Ω. Replace everywhere Ω by Q . There are
three terms. Term 1: The sequence fk := uk vk ∂i φ lies in L (Q), since vk , φ ∈ L∞ (Q), and fk
p
converges to f := uv∂i φ ∈ Lp (Q) in Lp , so a.e. Indeed along Q we get adding zero (φi := ∂i φ)
kuk vk φi − uvφi kp ≤ kuk − ukp kvk k∞ kφi k∞ + kukp kvk − vk∞ kφi k∞ → 0.
Now g := |f | + 1 ∈ L1 (Q) dominates |fk (x)| ≤ |f (x)| + 1 a.e. (since fk → f in Lp , so a.e.).

5 The notation W 1 (Ω) in [GT01, Le. 7.5] corresponds to our notation W 1,1 (Ω).
loc
a) To this end pick Q b Ω. Norms are over Q from now on. We obtain
|F (u(x))| ≤ |F (u(x)) − F (0)| + |F (0)| ≤ c0 |u(x)| + |F (0)|
where c0 := kF 0 kL∞ (Ω) . So f = F ◦ u ∈ Lp (Q) since u is. Also v ∈ Lp (Q):
kvkp = kF 0 (u) · uei kp ≤ c0 kuei kp < ∞.
b) Similarly as above we get that
kfk − f kp = kF (uk ) − F (u)kp ≤ c0 kuk − ukp → 0.
c) For finite p ∈ [1, ∞) we show that kF 0 (uk ) (uk )ei − F 0 (u) uei kp → 0.
From Lp convergence we get, after taking subsequences, that both uk → u and
(uk )ei → uei converge a.e. along Q, similarly for absolute values. Thus
gk := c0 |(uk )ei | → c0 |uei | =: g a.e. along Q
Together with continuity of F 0 we still get a.e. pointwise convergence
hk := F 0 (uk ) (uk )ei → F 0 (u) uei =: h a.e. along Q.
Note that |hk | ≤ gk pointwise along Q. Since (uk )ei → uei in Lp (Q) the hy-
pothesis in (2.2.8) is satisfied, hence the conclusion, namely that khk − hk1 =
kF 0 (uk ) (uk )ei − F 0 (u) uei kp → 0. This illustrates the usefulness of the general-
ized Lebesgue dominated convergence Theorem 2.2.8, doesn’t it?
d) Case p = ∞. Pick Q b Ω. Note that u ∈ W 1,∞ (Q) ⊂ W 1,q (Q) for any
q ∈ [1, ∞) by Hölder. So by b) the weak derivatives of F ◦ u exist and are given
by F 0 (u) · uei . We verify that F ◦ u and F 0 (u) · uei both lie in L∞ (Q). Let all
norms be over Q. Then kF 0 (u) · uei k∞ ≤ c0 kuei k∞ < ∞. And F ◦ u ∈ L∞ (Q),
because the sequence F ◦ uk ∈ C 1 (Q) ⊂ L∞ (Q) converges to F ◦ u. Indeed
kF ◦ u − F ◦ uk k∞ ≤ c0 kuk − uk∞ → 0.
But L∞ (Q) is a Banach space, so it includes the limit F ◦ u.
Remark 4.1.22 (Lipschitz chain rule). At first glance it looks like the proof
might go through for Lipschitz F with bounded derivative (which exists a.e.
by Rademacher’s Theorem 7.3.2). In a) replace c0 by the Lipschitz constant.
However, in b) continuity of F 0 enters crucially to get a.e. convergence hk → h,
doesn’t it?
Slightly modifying the assumptions an a.e. chain rule for Lipschitz functions
F is given in [Zie89, Thm. 2.1.11].
Proposition 4.1.23 (Change of coordinates). Let U, V ⊂ Rn be open and let
ψ : V → U be a bijection which is Lipschitz continuous and so is its inverse.
Let p ∈ [1, ∞] and u = u(x) ∈ W 1,p (U ). Write x = ψ(y). Then v = v(y) :=
u ◦ ψ ∈ W 1,p (V ) and there is the identity for the weak derivatives 6 vyi and uxj
n
X
(u ◦ ψ)yi = (uxj ◦ ψ) · (ψj )yi a.e. on V . (4.1.7)
j=1
6 for obvious reasons we use the notation vyi and uxj , and not our usual one vei and uej
4.2. DEFINITION AND BASIC PROPERTIES 27
Proof. Details of the proof are given in [Zie89, Thm. 2.2.2]. For finite p ∈ [1, ∞)
the proof is based on the transformation law
Z Z
f ◦ ψ · |det dψ| dy = f dx (4.1.8)
Y ψ(Y )
valid for measurable functions f : U → R and measurable subsets Y ⊂ V . In

fact, bi-Lipschitz maps preserve (Lebesgue) measurability of sets.
4.1.9 Equivalence classes of locally integrable functions

Definition 4.1.24 (Weak derivative of u = [u] ∈ L1loc = L1loc /∼). Suppose a
representative u of an equivalence class [u] ∈ L1loc (Ω) admits a weak derivative
uα corresponding to a multi-index α. In this case, the equivalence class
[u]α := [uα ] ∈ L1loc (Ω)
is called the weak derivative of [u] corresponding to α and |α| = α1 +

· · · + αn is called the order of the weak derivative.
Notation. Writing [u] for equivalence classes is often, but not always, very
convenient, e.g. when it comes to take derivatives. We shall alternatively use
boldface u := [u] to denote equivalence classes. The above then reads
uα := uα ∈ L1loc (Ω).
By Remark 4.1.5 the definition does not depend on the choice of the rep-
resentative u of [u]. By Lemma 4.1.6 any choice of a weak derivative uα of u
provides the same equivalence class [uα ], i.e. the same element of L1loc (Ω).
To summarize, if some, hence any, representative of an element [u] ∈ L1loc (Ω)
admits a weak derivative, then there is a unique element of L1loc (Ω), denoted
by [u]α , whose representatives are the weak derivatives of some, hence any,
representative of [u].
4.2 Definition and basic properties

Throughout Ω ⊂ Rn is open and Q b Rn pre-compact where n ≥ 1. Unless
mentioned otherwise, we suppose p ∈ [1, ∞] and k ∈ N0 . We write u = [u].
4.2.1 The Sobolev spaces W k,p

By definition the Sobolev space W k,p (Ω) consists of all locally integrable classes
[u] that admit weak derivatives [u]α ∈ L1loc (Ω) up to order k and each of them
is p-integrable on Ω. In symbols,7
W k,p (Ω) := u ∈ L1loc (Ω) : ∀|α| ≤ k ∃ uα ∈ Lp (Ω) ⊂ Lp (Ω)

7 writing ∃uα ∈ Lp (Ω) means that the weak derivative exists and it is p-integrable
where p ∈ [1, ∞] and k ∈ N0 .8 For finite p ∈ [1, ∞) define the value of the
W k,p -norm of a a class u ∈ W k,p (Ω) by the value
 1/p  1/p
Z X X
p p
kukk,p :=  |uα (x)| dx = kuα kp  ∈ [0, ∞)
Ω |α|≤k |α|≤k
where u is any representative of the class and uα : Ω → R is the weak derivative

corresponding to the multi-index α = (α1 , . . . , αn ). For infinite p = ∞ the
W k,∞ -norm of a class u ∈ W k,∞ (Ω) is defined as the maximum
kukk,∞ := max kuα k∞

|α|≤k
where u is any representative of the class.

A representative u of a class u ∈ W k,p (Ω) is called a W k,p function on Ω.
To indicate that a sequence of W k,p functions u` converges to u we just write
u` → u in W k,p (Ω) :⇔ ku` − ukW k,p (Ω) → 0, as ` → ∞.
Proposition 4.2.1 (Weak derivatives). Let u, v ∈ W k,p (Ω) and |α| ≤ k, then
(i) Dα u ∈ W k−|α|,p (Ω) and for all multi-indices α, β with |α| + |β| ≤ k
Dβ (Dα u) = Dα (Dβ u) = Dα (Dβ u);
(ii) for λ, µ ∈ R the linear combination λu + µv also lies in W k,p (Ω) and
Dα (λu + µv) = λDα u + µDα v , whenever |α| ≤ k;
(iii) for open sets U ⊂ Ω one has the inclusion W k,p (U ) ⊂ W k,p (Ω);
(iv) multiplication with a smooth compactly supported function φ ∈ C0∞ (Ω)
preserves W k,p (Ω) and the weak derivatives of the product are given by
X α
Dα (φu) = Dβ φ · Dα−β u.
β
β≤α
Proof. [Eva98, §5.2.3 Thm. 1].

Exercise 4.2.2. For k ∈ N0 and p ∈ [1, ∞], there is the equivalence
u ∈ W k+1,p (Ω) ⇔ u ∈ W 1,p (Ω), ue1 , . . . , uen ∈ W k,p (Ω).
[Hints: ’⇒’ Proposition 4.2.1 (i).

’⇐’ Write down (4.1.1) for the weak derivative (uei )β of uei , then use that uei
is the weak derivative of u utilizing once more (4.1.1).]
8 To see the inclusion ’⊂’ note that for α = (0, . . . , 0) one has u
0 α0 = u in (4.1.1). Hence
u = uα0 ∈ Lp (Ω) in order for u to be element of W k,p (Ω).
Exercise 4.2.3. Let Q ⊂ R2 be the open unit ball. Find an example of p ≥ 1

and a W 1,p function u : Q → R which is unbounded with respect to the L∞ -
norm (2.2.4) on each open subset U of Q. Conclude that your example is not
Lipschitz continuous.
Convention 4.2.4 (Natural inclusion). Notation such as
C ∞ (Q) ⊂ W k,p (Q)
makes sense if (and indicates that) we identify C ∞ (Q) with its image in W k,p (Q)
under the map u 7→ [u]. Of course, whenever a class [u] ∈ W k,p contains a
continuous representative, notation u∗ , we tacitly choose u∗ to represent [u].
Theorem 4.2.5. For every integer k ≥ 0 the following is true.
(i) W k,p (Ω) is a Banach space for p ∈ [1, ∞].
(ii) W k,p (Ω) is separable for finite p ∈ [1, ∞).
(ii) W k,p (Ω) is reflexive for finite p ∈ (1, ∞) larger 1.
(iv) C ∞ (D̄) is a dense subset of W k,p (D) for finite p ∈ [1, ∞) and Lipschitz
domain D b Rn ; cf. Remark 4.2.8 and Theorem 5.1.3.
Proof. (i-iii) The idea is to isometrically embedd W k,p (Ω) as a closed subspace
of Lp (Ω × · · · × Ω) by mapping u to the list that contains all weak derivatives
Dα u up to order k. See e.g. [AF03, §3.5]. Completeness and reflexivity are
inherited by closed subspaces of a Banach space. Furthermore, every subspace9
A of a separable metric space X is separable.10 Recall Theorem 2.3.1 (i-iii). All
three properties are transferred back to W k,p (Ω) via the isometry. For a direct
proof of (i) see e.g. [Eva98, §5.2.3 Thm. 2].
4.2.2 Difference quotient characterization of W 1,p

Observe that by continuity of u ∈ Lp (Rn ) under the shift map, see Exercise 2.3.3,
we have for each ξ ∈ Rn that ku(· + hξ) − ukp → 0, as h → 0.
Theorem 4.2.6. Let p ∈ (1, ∞). Then u ∈ W 1,p (Rn ) iff u ∈ Lp (Rn ) and the
−1
function h 7→ |h| ku(· + hξ) − ukp is bounded for any ξ ∈ Rn .
Proof. See e.g. [Zie89, Thm. 2.1.6] or [GT01, §7.11] or [Ste70, Ch. V §3.5].
9 A subspace of a metric space is a subset equipped with the induced metric, namely,
the restriction of the ambient metric.

10 Consider the set of open balls of rational radii centered at the elements of the countable
dense subset of X. Consider the collection of intersections of these balls with A. In each such
intersection select one element. The set S of selected elements is countable and dense in A.
4.2.3 The compact support Sobolev spaces W0k,p

By definition the compact support Sobolev space W0k,p (Ω) is the closure with
respect to the norm k·kk,p of the vector subspace ι(C0∞ (Ω)) ⊂ W k,p (Ω) that
consists of the compactly supported smooth functions on Ω. Here ι(u) := [u] is
the natural injection. In symbols,
k,p k·kk,p
W0k,p (Ω) := C0∞ (Ω) := ι(C0∞ (Ω)) .
In other words, the space W0k,p (Ω) is the completion of the linear subspace
ι(C0∞ (Ω)) of the Banach space (W k,p (Ω), k·kk,p ) where ι(u) := [u].
Remark 4.2.7. Being a closed Banach subspace W0k,p (Ω) inherits properties
(i) completeness, (ii) separability, and (iii) reflexivity in Theorem 4.2.5; see
e.g. [AF03, Thm. 1.22].
Remark 4.2.8 (One difference between Lp and Sobolev spaces). Part (ii) of
Theorem 3.2.3 in case of non-empty pre-compact sets Q b Rn is in sharp contrast
to what happens for Sobolev spaces with k ≥ 1. Namely, it says that
p
Lp0 (Q) := C0∞ (Q) = Lp (Q)
for finite p ∈ [1, ∞), whereas for Lipschitz D the two Sobolev spaces
k,p k,p
W0k,p (D) := C0∞ (D) ( C ∞ (D̄) = W k,p (D)
are different; cf. Theorem 5.1.3. Indeed the characteristic function χD repre-
sents an element of W k,p (D) ⊂ Lp (D), but not of W0k,p (D).
Exercise 4.2.9. For non-empty intervals I b R and finite p show that χI ∈

W 1,p (I) \ W01,p (I). Illustrate graphically why the characteristic function can be
approximated through elements of C0∞ (I) in the Lp , but not in the W k,p norm.
k,p
4.2.4 The local Sobolev spaces Wloc
k,p
By definition the local Sobolev space Wloc (Ω) consists of all locally integrable
classes [u] whose restriction to any pre-compact Q b Ω lies in W k,p (Q), i.e.
k,p
(Ω) := [u] ∈ L1loc (Ω) | ∀Q b Ω : [u|Q ] ∈ W k,p (Q) ⊂ Lploc (Ω).

Wloc
k,p
To see the inclusion ’⊂’ note that u ∈ [u] ∈ Wloc (Ω) restricted to any pre-
compact Q b Ω is an element u|Q ∈ Lp (Q). In other words, any representative
u is p-integrable over all pre-compact subsets of Ω.
4.2.5 How the spaces relate

To summarize, for k ∈ N and p ∈ [1, ∞] there are the, in general strict, inclusions
Lp (Ω) ⊂ Lploc (Ω) ⊂ L1loc (Ω)
⊂
⊂
k,p 1,p
W k,p (Ω)
⊂ ⊂ Wloc (Ω) ⊂ Wloc (Ω)
W0k,p (Ω)
In each line the second inclusion holds by Hölder’s inequality (2.2.5).
4.2.6 Basic properties – products and coordinate change

The elements of L∞ (Ω) are represented by almost everywhere bounded func-
tions. Let us call the elements of W k,∞ (Ω) k-bounded Sobolev functions.
Exercise 4.2.10 (Products with k-bounded Sobolev functions). Let k ∈ N0

and p ∈ [1, ∞]. Then the following is true. If u ∈ W k,p (Ω) and v ∈ W k,∞ (Ω),
then the product class uv := [uv] lies again in W k,p (Ω) and
kuvkk,p ≤ c kukk,p kvkk,∞
for some constant c that depends only on k and n.

[Hint: Induction based on Exercise 4.2.2 and Corollary 4.1.18.
Suppose k ∈ N. A C k−1,1 -diffeomorphism is a C k−1 -diffeomorphism ψ
such that the partial derivatives of ψ and ψ −1 up to order (k − 1) are Lipschitz
continuous. See Definition 5.1.5 for Lipschitz domains.
Exercise 4.2.11 (Change of coordinates). Let k ∈ N and p ∈ [1, ∞). Let
U, V ⊂ Rn be open and let ψ : V → U be a C k−1,1 -diffeomorphism. Then
the following is true. Pull-back ψ ∗ : W k,p (U ) → W k,p (V ), [u] 7→ [u ◦ ψ], is an
isomorphism of Banach spaces. Indeed any [u] ∈ W k,p (U ) satisfies the estimate
ku ◦ ψkW k,p (V ) ≤ c kukW k,p (U )
where c > 0 is independent of [u]. For weak derivatives of u ◦ ψ see (4.1.7).

Note that the change of coordinates transformation shows that one can define
Sobolev spaces on manifolds.
Chapter 5
Approximation and
extension
5.1 Approximation
Major application of density of smooth functions in a given Sobolev space:
• Avoid to work with weak derivatives in proofs
– first approximate a Sobolev function u by smooth functions u`

– so you can use usual derivatives on the smooth functions u`
– often a limit argument leads back to the original Sobolev function u
The quality of possible approximations depends on geometrical properties of

the function domain, such as boundedness or having a smooth boundary. The
present chapter follows very closely [Eva98, §5.3], including the proof of Theo-
rem 5.1.7 which, however, we spell out for the Lipschitz case, not only C 1 .
The key tool in this chapter is mollification. Whereas Section 5.1.1 is about
local approximation, so we need to work with δ-extensions as defined by (3.3.4),
in Sections 5.1.2 and 5.1.3 we get away with natural zero extensions, because of
the much stronger hypothesis on the functions to be of class W k,p on the whole
domain, not just on compact subsets. (Thus no dangerous non-integrability can
be lurking near the boundary. ;-)
5.1.1 Local approximation – any domain

The following generalization of Lemma 3.3.1 to Sobolev spaces is about “inte-
rior/local approximation by smooth functions with domain any open set”.
Theorem 5.1.1. Let u ∈ u ∈ W k,p (Ω) with k ∈ N0 and p ∈ [1, ∞]. Let
uδ := ρδ ∗ u in Ω2δ := {x ∈ Ω | dist(x, Ω) > 2δ}
33
34 CHAPTER 5. APPROXIMATION AND EXTENSION
be the convolution of u with a mollifier {ρδ }δ>0 . Then uδ ∈ C ∞ (Ω2δ ) and

k,p
uδ → u in Wloc (Ω), as δ → 0.
Recall from (3.3.5) the definition of the mollification uδ := ρδ ∗ ū ∈ C0∞ (Ω)

of a function u on Ω ⊂ Rn using the δ-extension ū = u(δ) : Rn → R, as defined
by (3.3.4). Along the closure of Ω2δ := {x ∈ Ω | dist(x, Ω) > 2δ} the formula
does not see the choice of extension, so in that domain it is safe to use the
notation ρδ ∗ u ∈ C ∞ (Ω2δ ).
Proof. Proposition 4.1.13.
Remark 5.1.2 (Case Ω = Rn ). Let k ∈ N0 and p ∈ [1, ∞). Recall from The-
orem 3.2.3 (i) that (4.1.4) is valid with Q replaced by Rn . Thus any W k,p (Rn )
function u can be W k,p approximated by a family uδ ∈ C ∞ (Rn ). Why does this
not prove that C ∞ (Rn ) is dense in the Banach space W k,p (Rn )? How about
density of C0∞ (Rn ) in W k,p (Rn )? In symbols, how about
k,p
W k,p (Rn ) = C0∞ (Rn ) =: W0k,p (Rn )?
5.1.2 Global approximation on bounded domains

The next theorem is about “global approximation by smooth functions on a
bounded open domain.” For k = 0 it is covered by Theorem 3.2.3 which asserts
that already the subset C0∞ of C ∞ is dense in Lp . For k ≥ 1 and pre-compact
Q b Rn this is not any more so, but C ∞ is still dense.
Theorem 5.1.3. Let Q b Rn be pre-compact. Let u ∈ u ∈ W k,p (Q) for k ∈ N
and finite p ∈ [1, ∞). Then there is a sequence of smooth functions u` ∈ C ∞ (Q)
with p-integrable derivatives of all orders |α| ≤ k, i.e. u` ∈ W k,p (Q), such that
C ∞ (Q) 3 u` → u in W k,p (Q).
Equivalently, over pre-compact sets Sobolev spaces are closures of the form
k,p
W k,p (Q) = C ∞ (Q) .
Remark 5.1.4 (Wrong for p = ∞). A counterexample for k = 1 and Ω =

(−1, 1) is u = |·|; cf. [AF03, Ex. 3.18].
Because the functions u : Q → R are of class W k,p along their whole domain
k,p
Q, as opposed to Wloc , we can work in the mollification of u with the natural
zero extension to Rn , still denoted by u.
Proof. Pick a function u ∈ u ∈ W k,p (Q) and let ε > 0. There are three
steps. Firstly, we construct a locally finite open cover (Vi )∞ i=0 of Q and fix a
subordinated partition of unity (χi ). Secondly, we enlarge each set Vi to some
Wi , still in Q but containing an εi -neighborhood of Vi for some εi (ε) > 0, in order
to have space for the convolution of χi u along Vi by a mollifier ρεi ∈ C0∞ (Bεi ).
5.1. APPROXIMATION 35
Figure 5.1: Locally finite covers Vi := Qi+3 \ Q̄i+1 and Wi := Qi+4 \ Q̄i of Q
This provides the smooth functions ui := ρεi ∗ (χi u) ∈ C0∞P (Wi ). Thirdly, we
globalize by considering the sum of smooth functions v = ui which is itself
smooth since the sum will be locally finite. By construction along any given
V b Q one gets kv − ukW k,p (V ) ≤ ε. Now take the sup over all such V .
I. For i ∈ N set Qi := {x ∈ Q | dist(x, ∂Q) > 1i } b Qi+1 to get a nested open
cover of Q. (If Q1 = ∅, replace 1i by i+c 1
for some large constant c so Q1 6= ∅.)
As we illustrated in Figure 5.1 define Vi := Qi+3 \cl Qi+1 for i ≥ 1 and V0 := Q3 .
Let (χi )∞
i=0 be a partition of unity P subordinate to the open cover of Q
∞
by the Vi ’s, that is χi ∈ C0∞ (Vi ) and i=0 χi ≡ 1. Enlarge the Vi ’s by setting
Wi := Qi+4 \ cl Qi for i ≥ 1 and W0 := Q4 . Obviously supp (χi u) ⊂ Vi ⊂ Wi
and [χi u] ∈ W k,p (Q); cf. Proposition 4.2.1 (iv).
II. Let ρ be a mollifier. Fix i ∈ N0 . By Theorem 3.2.3 (i) there is convergence
uδ := ρδ ∗ (χi u) → χi u in Lp (Q), as δ → 0,
and analogously by (iv) for Dα uδ = ρδ ∗ (χi u)α , see (4.1.3), and the weak
derivative (χi u)α whenever |α| ≤ k. Suppose δ = δ(i) > 0 is so small that
a) Uδ (Vi ) ⊂ Wi , so convolution is well defined and uδ ∈ C0∞ (Wi );
ε
b) kρδ ∗ (χi u) − χi ukW k,p (Q) ≤ 2i+1 .
Set εi := δ(i) and ui := uεi P∈ C0∞ (Wi ).

∞
III. The function v := i=0 ui on Q is well defined and smooth, because
about any point of Q there is an open neighborhood which meets at most five
of the Wi ’s, hence the support of at most five ui ’s; cf. Figure
P 5.1. Note that
any of the Qj ’s from I. is pre-compact in Q. Writing u = χi u we get that
∞
X
kv − ukW k,p (Qj ) ≤ kρεi ∗ (χi u) − χi ukW k,p (Qj ) ≤ ε
i=0
where the finalPstep uses the inclusion Qj ⊂ Q and b).

With f := |α|≤k |Dα (v − u)|p the function
Z
A 7→ mf (A) := f dm
A
is a measure on the Lebesgue σ-algebra on Q. Indeed the axioms of σ-additivity

and non-triviality hold by σ-additivity of the integral and as mf (Qj ) ≤ ε < ∞.
Since the sequence of sets Qj ⊂ Qj+1 is ascending with union Q, Theorem 1.28
part (iv) in [Sal16] guarantees the second identity in the following
p p
kv − ukW k,p (Q) = mf (Q) = lim mf (Qj ) = lim kv − ukW k,p (Qj ) ≤ εp .
j j
This completes the proof of Theorem 5.1.3.
5.1.3 Approximation even up to ∂ on Lipschitz domains

Imposing mild smoothness assumptions on the boundary of a pre-compact
D b Rn , namely, Lipschitz continuity (thus differentiable almost everywhere;
cf. Figure A.1), approximation of a W k,p (D) function is possible by smooth
functions u` whose domain is the compact set D, i.e. their domain even in-
cludes the boundary. This is in sharp contrast to Theorem 5.1.3 which provides
approximations u` defined on the open set D only, so finiteness of ku` kC k (D) is
not guaranteed at all by Theorem 5.1.3.
Definition 5.1.5 (Lipschitz domain D). A pre-compact D b Rn is called a
Lipschitz domain 1 if for each point x0 ∈ ∂D there is a radius r > 0 and a
Lipschitz continuous map γ : Rn−1 ⊃ ω → R with Lipschitz constant, say M ,
such that,2 firstly, the part of the boundary ∂D inside the open ball Br (x0 ) is
the graph of γ and, secondly, the part of D inside the ball is of the simple form
D ∩ Br (x0 ) = {x ∈ Br (x0 ) | xn > γ(x1 , . . . , xn−1 )},
as illustrated in Figure 5.2. The Lipschitz constant M of γ is called the Lip-
schitz bound of the local parametrization (γ, ω) of the Lipschitz domain D.
Definition 5.1.6 (Smooth domain D). A pre-compact D b Rn is called a C k -
domain if all local graph maps γ can be chosen of class C k . In case k = ∞ we
call D a smooth domain.
The next theorem is about “global approximation by functions smooth up to
the boundary of Lipschitz domains”. It is wrong for p = ∞; cf. Remark 5.1.4.
Theorem 5.1.7. Let D b Rn be a Lipschitz domain. Suppose u ∈ u ∈ W k,p (D)
with k ∈ N and finite p ∈ [1, ∞). Then there is a sequence u` ∈ C ∞ (D) with 3
C ∞ (D) 3 u` → u in W k,p (D).
Equivalently, over Lipschitz domains Sobolev spaces are closures of the form
k,p
W k,p (D) = C ∞ (D) .
1 These domains are called special Lipschitz domains in [Ste70, Ch. VI §3.2]. If one re-
laxes Lipschitz to α-Hölder with α ∈ (0, 1), then the Extension Theorem 5.2.1 fails as shown
in [Ste70, Ch. VI §3.2].
2 after possibly renaming and reorienting some coordinate axes, here ω ⊂ Rn−1 is open
3 Observe that any given derivative D α u extends continuously to the compact set D, so
`
it is bounded on D, hence ku` kC k (D) < ∞.
5.1. APPROXIMATION 37
Figure 5.2: Lipschitz domain D ⊂ Rn and upward cone C0
Proof. The proof takes three steps. Firstly, we generate room for mollification
by shifting u : D → R locally near ∂D ’upward’, meaning inside D. Here the
Lipschitz condition enters which tames the geometry of ∂D in that ∂D is forced
to stay in the horizontal part of one and the same double cone C̃0 translated
along ∂D, notation C̃x = x + C̃0 for x ∈ ∂D. The vertical upward cone Cx will
then be available for mollification.
Step 1: Upward shift to generate room for mollification within D
Given x0 ∈ ∂D, pick a Lipschitz continuous map γ : Rn−1 ⊃ ω → R whose
graph is the part of ∂D inside the open ball Br (x0 ). By compactness of ∂D the
map γ : ω → R is uniformly Lipschitz with Lipschitz constant, say c > 0. The
closed horizontal double cone C̃0 and the open upward cone C0
C̃0 := {(x0 , xn ) | ± |xn | ≤ c |x0 |} C0 := {(x0 , xn ) | ± |xn | ≤ c |x0 |}
are illustrated by Figure 5.2. The significance of the horizontal double cone C̃0
is that wherever you put it along ∂D – consider the translate C̃y := y + C̃0 at
y ∈ ∂D – that translated horizontal double cone C̃y contains the graph of γ and
therefore ∂D (locally near y). Hence the open upward cone Cy lies in D, at
least its part that lies within some small radius from y, say r(y). See Figure 5.3
in which y is denoted by x̃.
Let V := D ∩ Br/2 (x0 ) be the open region strictly above the graph but still
inside the ball of radius r/2; see Figure 5.3. For any x ∈ V define the by ε > 0
upward shifted point
xε := x + ελen , x ∈ V , ε > 0.
As illustrated by Figure 5.3, let us put the unit ball B1 sufficiently far out,
centered at x+λen where λ = λ(x) >> 1 is sufficiently large such that eventually
the ball fits into the upward cone. But then the whole family of balls Bε (x +
ελen ) for ε ∈ (0, 1) stays in the upward cone. Moreover, for all ε > 0 sufficiently
small the family is located near x, hence in the open neighborhood V of x.
Choose λ larger, if necessary, so it works for all x ∈ γ(ω) ⊂ ∂D.
Figure 5.3: Fitting a ball Bε (x + ελen ) into Cx̃ ∩ Br (x0 ): λ large, ε small
Define the translate uε (x) := u(xε ), for x ∈ V . This is the function u moved
by distance ελ in the en direction. Pick a mollifier ρ = {ρε }. The convolution4
v ε = ρε ∗ uε ∈ C ∞ (V )
is well defined on V since by moving up we have created space for mollification.

To see that v ε ∈ C ∞ (V ), as opposed to C ∞ (V ), let us check that v ε is defined
not only on the pre-compact V , but even on a small neighborhood. Let us
illustrate this by considering the case x̃ ∈ ∂V ∩ ∂D, cf. Figure 5.3. In this case
Z Z
v ε (x̃) = ρε (x̃ − y) · uε (y) dy = ρε (x̃ − y) · u( y + ελen ) dy.
Bε (x̃) Bε (x̃) | {z }
∈Bε (x̃+ελen )
But Bε (x̃ + ελen ) is contained in the upward cone Cx̃ , so v ε (x̃) is well defined.
Argument and conclusion remain valid for points x∗ slightly below x̃, i.e. slightly
outside V , because Bε (x∗ + ελen ) still remains in Cx̃ , thus in the domain of u.
Step 2: Convergence to u along V
To see that v ε → u in W k,p (V ) pick a multi-index of order |α| ≤ k, then
kDα v ε − Dα ukLp (V ) ≤ kDα v ε − Dα uε kLp (V ) + kDα uε − Dα ukLp (V )

≤ kρε ∗ Dα uε − Dα uε kLp (V ) + k(uα )ε − uα kLp (V )
where the second inequality holds by (4.1.3) and since weak derivative com-
mutes with the linear operation of translation, in symbols (uε )α = (uα )ε .5 Now
4
Use the natural zero extension of u, so uε R∈ Lp (Rn ) and ρε ∗ uε makes sense by (3.1.2).
To see that V uε (x)Dα φ(x) dx = (−1)|α| V (uα )ε (x)φ(x) dx introduce the new variable
5
R
y := x + ελen , take weak derivative of u(y) according to (4.1.1), then go back to x := y − ελen .
5.2. EXTENSIONS AND TRACES 39
consider the two terms in line two. As ε → 0, term one goes to zero by The-
orem 3.2.3 (i) and term two by Lemma 2.3.2. The lemma tells that the path
γ : ε 7→ τε uα − uα is continuous with respect to the Lp norm. Clearly γ(0) = 0.
Step 3: Globalization via partition of unity
Pick δ > 0. By compactness of ∂D there are finitely many points x0i ∈ ∂D,
radii ri > 0, corresponding set Vi = D ∩ Bri /2 (x0i ), and functions v i ∈ C ∞ (V i ),
where i = 1, . . . , N , such that the balls Bri /2 (x0i ) cover ∂D and (by Step 2)
i
v − u k,p ≤ δ. (5.1.1)
W (V )
i
Pick V0 b D such that (Vi )N i=0 is an open cover of D. Utilize Theorem 5.1.1 to
get a function v 0 ∈ C ∞ (V 0 ) that also satisfies (5.1.1).
Pick a smooth partition of unity (χi )N i=0 subordinate to the open cover
PN
(Vi )N
i=0 . The finite sum of smooth functions v := i=0 i vi is smooth and
χ
defined on some neighborhood of D. Thus v ∈ C ∞ (D). Suppose |α| ≤ k and
use the Leibniz rule Proposition 4.2.1 (iv) to get that
N
X
α α
α
D (χi v i ) − Dα (χi u) p

kD v − D ukLp (D) ≤ L (V i)
i=0
N X
X α β α−β i β α−β

= D χ i · D v − D χi · D u
β
i=0 β≤α

Lp (Vi )
N
X i
≤C v − u k,p
W (Vi )
i=0
≤ C(N + 1)δ
where the constant C = C(k, p) > 0 also incorporates the C k norms of the
cutoff functions χi . This completes the proof of Step 3 and Theorem 5.1.7.
5.2 Extensions and traces

5.2.1 Extension
Whereas in the realm of Lp spaces extending an Lp function on a domain Ω ⊂ Rn
to all Rn within Lp is trivial, just extend naturally by zero. This does not work
for Sobolev spaces, already not for those of first order W 1,p . Roughly speaking,
jump singularities obstruct existence of weak derivatives, while corners are still
digestible; cf. Exercise 4.1.7. The following extension theorem works for any
p ∈ [1, ∞], finite or not. The same smoothness assumption on the boundary
as below will be required in both Sobolev inequalities, the sub- and the super-
dimensional one, Theorems 6.1.3 and 6.2.3, respectively, because both proofs
build on Theorem 5.2.1. Therefore we spell out details in the Lipschitz case,
as opposed to the common C 1 assumption; cf. [Eva98, §5.4]. The C 1 condition
allows to deal with the open unit ball – but not even the open unit cube in Rn .
The extension theorem for Lipschitz domains is due to Calderon [Cal61,

1 < p < ∞] and Stein [Ste70, Ch. VI §3.4, p = 1, ∞].
Theorem 5.2.1 (Extension). Suppose p ∈ [1, ∞] and let D b Rn be a Lipschitz
domain. Then the following is true. For any pre-compact Q b Rn that contains
the closure of D, in symbols D b Q b Rn , there is a bounded linear operator
E : W 1,p (D) → W01,p (Q) ,→ W 1,p (Rn ), u 7→ Eu = ū, (5.2.2)
6
such that ū|D = u, equivalently ū|D = u a.e. along D for representatives, and
kūkW 1,p (Q) ≤ c kukW 1,p (D)
whenever ū ∈ ū and u ∈ W 1,p (D); the constant c > 0 only depends on n, D, Q.

In the case p = ∞ the constant is actually independent of n.
The function Eu is called an extension of u ∈ u ∈ W 1,p (D) to Q. For
Lipschitz domains of class C k−1,1 the extension operators are of the form
E : W k,p (D) → W0k,p (Q) and the constant c also depends on k; see [GT01,
Thm. 7.25] and [MS04, Prop. B.1.9].
Idea of proof. For details see e.g. [Eva98, §5.4]. The idea is to deal first with
the model case of a function, say v, defined on an open upper half ball V ; see
Figure 5.4. One extends v of class C 1 up to the boundary to the closure of the
whole unit ball B by explicit formulas (called reflections and depending on the
desired degree of regularity). Secondly, “flatten out” the boundary through a
coordinate change given by the natural bi-Lipschitz map7 that comes from
the Lipschitz graph map γ of ∂D; see Figure 5.5. Here D being a Lipschitz do-
main enters. Thirdly, construct local extensions along the boundary. Fourthly,
globalize via partition of unity and cut off between D and Q.
Proof. The proof is in three steps. For finite p we follow largely [Eva98, §5.4].
Step 1 (Extension operator in the half ball model). Suppose B ⊂ Rn
is an open ball with center y 0 in the plane {yn = 0}. Let B+ be the open upper
half-ball of B. Assume p ∈ [1, ∞]. Then there is a linear map
E0 : W 1,p (B+ ) → W 1,p (B), v 7→ E0 v = v̄, (5.2.3)
that satisfies the identity v̄|B+ = v, i.e. acts by extension. The estimates hold
kv̄kW 1,p (B) ≤ cn kvkW 1,p (B+ ) , kv̄kW 1,∞ (B) = kvkW 1,∞ (B+ ) , (5.2.4)
for every v ∈ v and finite p. The constant is given by cn = 16 · 22n−1 .

Proof of Step 1. Let Σ = B ∩ {yn = 0} be the (n − 1) ball that divides B into
B+ and B− ; see Figure 5.4.
6 The notation ū := Eu is useful. It helps to avoid misleading notation such as writing
Eu ∈ Eu for an element of the extension equivalence class. There is no naturally defined

operator of the form u 7→ Eu on the level of Sobolev functions, only on subcategories. In
contrast, the theorem provides an operator u 7→ Eu on the level of equivalence classes.
7 Lipschitz homeomorphism whose inverse is Lipschitz
Figure 5.4: Open upper half B+ of open ball B
Finite p ∈ [0, ∞). Consider the inclusions C ∞ (B + ) ⊂ C 1 (B + ) ⊂ W 1,p (B+ ).

By Theorem 5.1.7, here finite p enters, the first set is dense in W 1,p , hence so is
the second one, abbreviated C 1 . For v ∈ C 1 we define the extension E0 v := v̄
pointwise and show that it lies in C 1 (B), thus in W 1,p (B). We then prove the
estimate (5.2.4) for any v in the dense set C 1 . By density E0 extends uniquely
to a bounded linear operator with domain W 1,p (B) and we are done.
To this end we define the extension of v ∈ C 1 (B + ) to B by8
(
v(y) , y ∈ B+,
v̄(y) := yn
−3v(y∗ , −yn ) + 4v(y∗ , − 2 ) , y = (y∗ , yn ) ∈ B − .
Here and throughout y∗ = (y1 , . . . , yn−1 ) and y = (y∗ , yn ) ∈ Rn . To see that
v̄ ∈ C 1 (B) write v + := v̄|B + = v and v − := v̄|B − . So v − = v + along Σ. Now
∂n v − (y) = 3∂n v(y∗ , −yn ) − 2∂n v(y∗ , − y2n ).

Along {yn = 0} this becomes (3 − 2)∂n v(y∗ , 0) = ∂n v + (y∗ , 0). So ∂n v − = ∂n v +
along {yn = 0}. As v − = v + along {yn = 0}, we also have ∂i v − = ∂i v + along
{yn = 0} for i = 1, . . . , n − 1. Hence Dα v − = Dα v + along {yn = 0} for |α| ≤ 1.
This proves that v̄ ∈ C 1 (B). Straightforward calculation9 proves (5.2.4).
Infinite p = ∞. Consider the linear map (W 1,∞ ' C 0,1 by Theorem 7.2.1)
E0 : C 0,1 (B+ ) → C 0,1 (B), v 7→ E0 v := v̄,
where the extension v̄ of v from B + to B is given by simple horizontal reflection
(
v(y) , y ∈ B+,
v̄(y) :=
v(y∗ , −yn ) , y = (y∗ , yn ) ∈ B − ⊂ Rn−1 × R.
Check that v̄ is indeed Lipschitz, even with the same Lipschitz constant as v.10
Check that kv̄kW 1,∞ (B) = kvkW 1,∞ (B+ ) for every v ∈ W 1,∞ (B+ ).
8 Try to pointwise define v̄ for v ∈ v ∈ W 1,p (B+ ). Which values to assign along Σ?
9 With |Dv̄|2 = |∂1 v̄|2 +· · ·+|∂n v̄|2 get kv̄kLp (B) ≤ (1+8p )1/p kvkLp (B+ ) ≤ 16 kvkLp (B+ )
and kDv̄kLp (B) ≤ (κ(1 + 8p ))1/p 2n−1 kDvkLp (B+ ) ≤ 16 · 22n−2 kDvkLp (B+ ) . By Exercise
A.1.2 κ1/p ≤ 2n−1 . Thus cn = 16 · 22n−1 (depends on n as we used `2 norm to define |Dv|).
10 Hints: This is clear if x, y are in the same half ball. If one is in B , the other in B ,
+ −
consider the unique point z of the line segment between x and y that lies in Σ. In |v̄(x) − v̄(y)|
add 0 = −v̄(z) + v̄(z) and note that |x − z| + |z − y| = |x − y|.
Step 2 (Boundary flattening coordinates).

To “flatten out” the boundary of D locally near x0 ∈ ∂D recall that by Defini-
tion 5.1.6 the part of ∂D in a small open ball Br (x0 ) is the graph of a Lipschitz
map γ : Rn−1 ⊃ ω → R with Lipschitz constant M . The domain D lies above
the graph of γ; see Figure 5.5.
Figure 5.5: Coordinates that locally flatten out ∂D
Consider the open neighborhoods X := ω×R of x0 = (x0∗ , x0n ) and Y := ω×R

of y 0 = (x0∗ , 0). Define a bi-Lipschitz map Φ : X → Y by the formula
(
yi := xi = Φi (x) , i = 1, . . . , n − 1,
y = Φ(x) n
(5.2.5)
yn := xn − γ(x∗ ) =: Φ (x) .
Note that |Φ(x) − Φ(z)| = |x − z| and Φ(x0 ) = y 0 . The inverse Ψ : Y → X is

(
xi := yi = Ψi (y) , i = 1, . . . , n − 1,
x = Ψ(y) n
xn := yn + γ(y∗ ) =: Ψ (y) .
Note that |Ψ(y) − Ψ(z)| = |y − z| and Φ = Ψ−1 . Observe that the map x 7→
Φ(x) “flattens out” ∂D near x0 in the sense that the boundary image is the
open subset ω of Rn−1 . By Rademacher’s Theorem 7.3.2 the graph map γ is
differentiable for a.e. x∗ ∈ ω. Hence the linearizations of Φ and Ψ exist pointwise
a.e. and, furthermore, they are triangular matrizes with diagonal elements 1.
Thus det dΦ = 1 = det dΨ pointwise a.e.
Step 3 (Local extension (x0 , ū := v̄ ◦ Ψ−1 ) of u|U to A about x0 ∈ ∂D).
Choose u ∈ u ∈ W 1,p (D). Pick a ball B centered at the point y 0 = Φ(x0 ) and
contained in the open neighborhood Φ(Br (x0 )) of y 0 , as illustrated by Figure 5.5.
Let V := B+ be the upper open half ball of B. Now consider the restriction of
u : D → R to the open set U := Ψ(V ), still denoted by u : U → R. Observe
that u ∈ W 1,p (U ); just restrict the weak derivatives.
Next pull back u : U → R to the y coordinates to obtain the function v := u◦Ψ :
V → R which lies in W 1,p (V ) by the change of coordinates Proposition 4.1.23.
The identity kvkW 1,p (V ) = kukW 1,p (U ) holds since dΨ is of unit determinant;
use (4.1.7) in case p = ∞ and the transformation law (4.1.8) in case of finite p.
Next employ the extension operator constructed in Step 1 to pick an exten-

sion v̄ ∈ v̄ := E0 v of v = u ◦ Ψ from the upper half ball V = B+ to the whole
ball B. The extension of u from U = Ψ(V ) to A := Ψ(B) is defined by
ū := v̄ ◦ Ψ−1 = u ◦ Ψ ◦ Ψ−1 ∈ W 1,p (A), ū := [ū] ∈ W 1,p (A).
By the argument five lines above kūkW 1,p (A) = kv̄kW 1,p (B) . The estimate (5.2.4)
for the extension operator E0 in the half ball model of Step 1 shows that
kūkW 1,p (A) = kv̄kW 1,p (B) ≤ cn kvkW 1,p (B+ ) = cn kukW 1,p (U ) (5.2.6)
for all u ∈ u ∈ W 1,p (D) and finite p. For p = ∞ equality holds and cn = 1.
Step 4 (Globalising via finite partition of unity).
Let p ∈ [1, ∞] and u ∈ u ∈ W 1,p (D). By Step 3 and compactness of ∂D
there are finitely many boundary points x1 , . . . , xN ∈ ∂D and local extensions
(xi , ūi = v̄ i ◦Ψ−1 : Ai → R)Ni=1 covering ∂D, that is ∂D ⊂ A1 ∪ . . . ∪ AN . Recall
from Step 3 that Ui = Ψ(Vi ) is the part of Ai = Ψ(Bi ) in D. Pick U0 b D in
order to get an open cover (Ui )N i=0 of D. Set A0 := U0 in order to get a pre-
compact set A := ∪N i=0 Ai containing the closure of D, in symbols D b A b R .
n
Pick a smooth partition of unity χ = (χi )N i=0 subordinate to the open cover
(Ai )Ni=0 of A. Extend u ∈ u ∈ W
1,p
(D) to A by
N
X
ū := χi ūi : A → R, ū := [ū] ∈ W 1,p (A), (5.2.7)
i=0
where ū0 := u : A0 = U0 → R and where as in Step 3
ūi := v̄ i ◦ Ψ−1 = ui ◦ Ψ ◦ Ψ−1 ∈ W 1,p (Ai ), Ai = Ψ(Bi ).
Indeed ū ∈ W 1,p (A) by Proposition 4.2.1 (iv). There are the estimates
N
X
kūkW 1,p (A) ≤ kχi ūi kW 1,p (Ai )
i=0
N
X
≤ kχi kW 1,∞ (Ai ) kūi kW 1,p (Ai )
i=0
N
X
i
≤ cn max kχi k1,∞ u 1,p
W (Ui )
i=0,...,N
i=0

≤ cn (N + 1) max kχi k1,∞ kukW 1,p (D)
i=0,...,N
=: C kukW 1,p (D)
where C depends only on n and D for finite p and only on D in case p = ∞.

Inequality two uses the product rule (4.1.6). Inequality three holds by (5.2.6)
for ūi ; recall that cn = 16 · 22n−1 for finite p and cn = 1 for p = ∞. The final
inequality four uses that ui : Ui → R is just the restriction of u : D → R and
this happens N + 1 times.
Given u ∈ W 1,p (D) and Q such that D b Q b Rn , then D b (Q ∩ A) b Rn .
Pick a cutoff function χ ∈ C0∞ (Q∩A, [0, 1]) with χ ≡ 1 on D. Then χū = [χū] ∈
W 1,p (Q) by Proposition 4.2.1 (iv) where ū = [ū] ∈ W 1,p (Q) is the restriction
to Q of (5.2.7). That χū even lies in a subspace, namely
1,p
Eu := χū ∈ W01,p (Q) := C0∞ (Q) ,
follows by approximating ū ∈ ū ∈ W 1,p (Q) by smooth functions v` ∈ C ∞ (Q)

according to Theorem 5.1.3. So the classes of χv` ∈ C0∞ (Q) approximate χū.
As for the estimate, since supp χ ⊂ (Q ∩ A) and using the previous estimate
kχūkW 1,p (Q) = kχūkW 1,p (Q∩A)

≤ kχūkW 1,p (A)
≤ kχkW 1,∞ (A) kūkW 1,p (A)
≤ C kχkW 1,∞ (A) kukW 1,p (D) .
This concludes the proof of Theorem 5.2.1.
5.2.2 Trace
While the usual notion of differentiability is hereditary – the restriction of a C 1
function to a codimension one submanifold is C 1 again – for weak derivatives
the situation is more subtle: The values of a function on a set of measure zero
are irrelevant, but a codimension one submanifold is of measure zero itself.
The following trace theorem works for finite p ∈ [1, ∞).
Theorem 5.2.2 (Trace). Let p ∈ [1, ∞) be finite and D b Rn be Lipschitz.

Then there is a bounded linear operator T : W 1,p (D) → Lp (∂D), u 7→ T u, such
that for any u ∈ u ∈ W 1,p (D) one has 11
kT ukLp (∂D) ≤ c kukW 1,p (D)
where the constant c > 0 only depends on p and D. In case u has a uni-
formly continuous representative u∗ , in symbols u∗ ∈ C 0 (D), the trace T u has
a (uniformly) continuous representative, denoted by (T u)∗ , which is given by
restricting u∗ to the boundary, in symbols (T u)∗ = u∗ |∂D , likewise of u ∈ u.
Any representative function T u ∈ T u ∈ Lp (∂D) is called a trace of u ∈

1,p
W (D) on ∂D.
11 We use the abusive notation T u ∈ T u = T [u], abusive since T u is not an operator applied
to u, but T u just denotes an element of Lp (∂D). We should really write f ∈ T u ∈ Lp (∂D).

Proof. Details are given in [Eva98, §5.5] for C 1 domains; just use the Lipschitz
coordinate change (5.2.5) and the Lipschitz approximation Theorem 5.1.7. The
skeleton of the proof is the same as the one of the Extension Theorem 5.2.1.
One starts with the model case where ∂D is locally flat.
Theorem 5.2.3 (Trace zero functions in W 1,p ). Let p ∈ [1, ∞) and let D b Rn
be Lipschitz. For [u] ∈ W 1,p (D) there is the equivalence
[u] ∈ W01,p (D) ⇔ T u = 0 : ∂D → R.
Proof. ’⇒’ Theorem 5.2.2. ’⇐’ Harder, see [Eva98, §5.5] for details.
Chapter 6
Sobolev inequalities
6.1 Sub-dimensional case kp < n

6.1.1 Gagliardo-Nirenberg-Sobolev inequality (p < n)
Throughout this section k = 1, so sub-dimensional case means that
1≤p<n (thus n ≥ 2).
A powerful tool, as we shall see, would be an estimate of the form
kukLq (Rn ) ≤ c kDukLp (Rn )
for all u ∈ C0∞ (Rn ). Compact support is necessary as u ≡ 1 shows. For which
q and c can such estimate be expected? Suppose 0 6≡ u ∈ C0∞ (Rn ), pick a
constant λ > 0 and insert the rescaled function uλ (x) := u(λx) in the estimate
above. Details are given in [Eva98, §5.6.1]. The upshot is that validity of the
estimate implies the identity 1q = p1 − n1 . Now for this particular q the estimate
indeed holds true – it is the Gagliardo-Nirenberg-Sobolev inequality.
Definition 6.1.1. If p ∈ [1, n), the Sobolev conjugate of p is
np
p∗ := ∈ (p, ∞)
n−p
or equivalently
1 1 1
∗
= − , p ∈ [1, n), p∗ ∈ (p, ∞). (6.1.1)
p p n
Theorem 6.1.2 (Gagliardo-Nirenberg-Sobolev inequality). If p ∈ [1, n), then
kukLp∗ (Rn ) ≤ γp kDukLp (Rn ) (6.1.2)

p(n−1)
for every compactly supported u ∈ C01 (Rn ) and with γp = γp (n) := n−p .
47
48 CHAPTER 6. SOBOLEV INEQUALITIES
np
Figure 6.1: Sobolev conjugate p∗ := n−p ∈ (p, ∞) of p ∈ [1, n)
Note that γ1 = 1 and γp>1 > 1 explodes for p → n. Compact support is

necessary as u ≡ 1 shows. But the constant γp doesn’t see the size of supp u.
Proof. There are two steps I (p = 1) implying II (1 < p < n); cf. [Eva98, §5.6.1].
I. Case p = 1. By the fundamental theorem of calculus and compact support
Z xi

|u(x)| = ∂xi u(x1 , . . . , yi , . . . , xn dyi )
−∞
Z ∞
≤ |Du(x1 , . . . , yi , . . . , xn dyi )| dyi .
−∞
1
Multiply the inequalities for i = 1, . . . , n and take the power n−1 to get
n Z 1
n Y ∞ n−1
|u(x)| n−1
≤ |Du(x1 , . . . , yi , . . . , xn dyi )| dyi
i=1 −∞
Integrate the variable x1 over R, then extract the term corresponding to i = 1

(it does not depend on x1 ) to obtain that
Z ∞
n
|u(x)| n−1 dx1
x1 =−∞
n Z 1
Z ∞ Y ∞ n−1
≤ |Du(x1 , . . . , yi , . . . , xn )| dyi dx1
x1 =−∞ i=1 yi =−∞
1
Z ∞ n−1
= |Du(y1 , x2 . . . , xn )| dy1
y1 =−∞
fi (x1 )
z }| {
n Z ∞ 1
Z ∞ Y n−1
|Du(x1 , . . . , yi , . . . , xn )| dyi dx1
x1 =−∞ i=2 yi =−∞
| {z }
1 1
f2 (x1 )...fn (x1 ), n−1 +···+ n−1 =1
n
Z ∞ n−1
≤ |Du(y1 , x2 . . . , xn )| dy1
y1 =−∞
n Z n
Y ∞ Z ∞ n−1
|Du(x1 , . . . , yi , . . . , xn )| dyi dx1
i=2 x1 =−∞ yi =−∞
| {z }
|fi (x1 )|n−1
6.1. SUB-DIMENSIONAL CASE KP < N 49
where the last step is by the generalized Hölder inequality (2.2.7).

Now integrate this estimate with respect to x2 over R, integrate the resulting
estimate with respect to x3 and so on until finally you have integrated the xn
variable over R and have gotten to the desired estimate (6.1.2) for p = 1, namely
n Z 1
Z
n Y ∞ Z ∞ n−1
|u| n−1 dx ≤ ... |Du| dx1 . . . dyi . . . dxn
Rn i=1 −∞ −∞
Z n
n−1 (6.1.3)
= |Du| dx .
Rn
For a few more details of the iteration we refer to [Eva98, §5.6.1 Thm. 1].
γ
II. Case p ∈ (1, n). Apply (6.1.3) to v := |u| to determine γ, namely
Z n−1
n
Z Z
γn
γ−1
|u| n−1
≤ |D |u|γ | = γ |u| |Du|
Rn Rn Rn | {z } |{z}
p−1 1
p + p =1
Z p−1 Z p1
p p
(γ−1) p−1 p
≤γ |u| |Du|
Rn Rn
where we used Hölder (2.2.5). Set γ := p(n−1)

n−p to equalize the two exponents.
p pn ∗ γ p−1 γ−1
Note that (γ − 1) p−1 = γ n−1 = n−p = p , hence n−1
n
n = p∗ and p = p∗ . So
the above estimate becomes kukγp∗ ≤ γkukγ−1
p∗ kDukp which is (6.1.2).
The proof of the following theorem is based on the Extension Theorem 5.2.1
and therefore requires a Lipschitz domain D.
Theorem 6.1.3 (Sub-dimensional W 1,p (D) estimates). Suppose D b Rn is
∗
Lipschitz. If p ∈ [1, n) and u ∈ W 1,p (D), then u ∈ Lp (D) with the estimate
kukLp∗ (D) ≤ c kukW 1,p (D)
where the constant c depends only on p, n, and D.

The idea of proof is to extend u ∈ W 1,p (D) to a compact support Sobolev
class ū ∈ W01,p (Q) where D b Q b Rn . Approximate a representative ū ∈ ū
in W 1,p by a Cauchy sequence um ∈ C0∞ (Q). Use the Gagliardo-Nirenberg-
Sobolev inequality to see that the um , extended to Rn by zero, also form a
∗ ∗
Cauchy sequence in Lp with the same limit ū. Hence ū ∈ Lp and the inequality
kum kLp∗ (Rn ) ≤ γp kDum kLp (Rn ) also holds for the limit element ū.
1,p
Proof. Extend u ∈ W 1,p (D) to ū := Eu ∈ W01,p (Q) := C0∞ (Q) for any
pre-compact Q with D b Q b Rn using the Extension Theorem 5.2.1. Pick a
representative ū ∈ ū and approximate it by a Cauchy sequence um ∈ C0∞ (Q)
that converges to ū in the W 1,p -norm, in symbols
um → ū, Dum → Dū, both in Lp (Q).

Extend the um to Rn by zero, then the Gagliardo-Nirenberg-Sobolev inequality

kum −u` kp∗ ≤ γp kDum −Du` kp shows that the um also form a Cauchy sequence
∗ ∗
in Lp (Q) with limit, say v ∈ Lp (Q). By compactness of Q and since p∗ > p
∗
Hölder shows that Lp convergence implies Lp convergence; cf. Figure 6.2.
Hence v = ū by uniqueness of the limit. Taking limits on both sides of the
Gagliardo-Nirenberg-Sobolev inequality kum kLp∗ (Q) ≤ γp kDum kLp (Q) and, in
the first step, monotonicity of the integral shows that
kūkLp∗ (D) ≤ kūkLp∗ (Q) ≤ γp kDūkLp (Q) ≤ γp kūkW 1,p (Q) ≤ γp C kukW 1,p (D)
where C = C(p, D, Q) is the constant in the Extension Theorem 5.2.1.
The difference between Theorem 6.1.3 and the next estimate is that only the
gradient of u appears on the right hand side if we work with compact support
Sobolev spaces. Lipschitz boundary is not needed, but pre-compactness Q b Rn
since the Gagliardo-Nirenberg-Sobolev inequality requires compact support.
Theorem 6.1.4 (Poincaré inequality – sub-dimensional W01,p (Q) estimate). Let

Q b Rn . If p ∈ [1, n) and u ∈ W01,p (Q), then u ∈ Lq (Q) with the estimate
kukLq (Q) ≤ c kDukLp (Q)
whenever q ∈ [1, p∗ ] and where the constant c depends only on p, q, n, and Q.
The Poincaré inequality tells that on compact support Sobolev spaces the
norm kukW 1,p (Q) is equivalent to kDukLp (Q) .
1,p
Proof. Given u ∈ W01,p (Q) := C0∞ (Q) , pick a Cauchy sequence (um ) ⊂
C0∞ (Q) that converges in the W 1,p -norm to a representative u of u. Extend
the um to Rn by zero. Then the Gagliardo-Nirenberg-Sobolev inequality (6.1.2)
kum kLp∗ (Q) = kum kLp∗ (Rn ) ≤ γp kDum kLp (Rn ) = γp kDum kLp (Q)
∗ ∗
tells that (um ) is a Cauchy sequence in Lp (Q). Thus um converges in Lp to
∗
some v ∈ Lp (Q), but also in Lp to u ∈ Lp (Q). Hence um converges in L1loc to
v and also to u. Thus u = v a.e. and therefore
kukLp∗ (Q) = kvkLp∗ (Q) ≤ γp kDukLp (Q) .
1
But q-Hölder (2.2.6) for q = ( 1q − 1
p∗ ) + 1
p∗ tells that u ∈ Lq (Q) and
1
− p1∗
kukLq (Q) ≤ |Q| q kukLp∗ (Q) .
1
− n−p
np p(n−1)
This proves Theorem 6.1.4 with c = |Q| q n−p .
6.1. SUB-DIMENSIONAL CASE KP < N 51
6.1.2 General Sobolev inequalities (kp < n)

The following hypotheses require dimension n ≥ 2.
Theorem 6.1.5 (General Sobolev inequalities k < np ). Let D b Rn be Lipschitz

and k ∈ {1, . . . , n − 1} and p ∈ [1, nk ). Then any u ∈ W k,p (D) lies in Lq (Q) for
1 1 k n
= − , q =p· (> p),
q p n n − kp
and there is the estimate
kukLq (D) ≤ c kukW k,p (D)
where the constant c depends only on k, p, n, and D.
Before entering the proof let us see what the theorem actually means. Since
p < q and the domain D is pre-compact Hölder’s inequality (2.2.5) tells that
Lp ⊃ Lq whereas the Sobolev inequality tells W k,p ⊂ Lq . In other words, the
sub-dimensional Sobolev inequalities invert the Hölder inclusions - at the cost
of asking existence of some derivatives.
Figure 6.2: Sub-dimensional Sobolev inclusions run opposite to Hölder
Proof. If k = 1 we are done by Theorem 6.1.3 and q = p∗ . Thus suppose k ≥ 2.

The proof is a k ≥ 2 step inclusion process as illustrated by Figure 6.2. Given
u ∈ W k,p (D) with kp < n, the idea is to give away one order of derivative in
exchange of enlarging p to its Sobolev exponent p∗ defined by (6.1.1). More
precisely, reduce 1/p by 1/n to get 1/p∗ . After k steps all derivatives are gone,
corresponding to subtracting k/n from 1/p. Due to the assumption kp < n the
difference p1 − nk > 0 is still positive and its inverse will be q.
Let us abbreviate W `,p := W `,p (D) and k·k`,p := k·kW `,p (D) . Proposi-
tion 4.2.1 (i) tells that if |α| = `, then Dα u ∈ W k−`,p .
Step 1. Whenever |α| ≤ k − 1 we have Dα u ∈ W 1,p and since p < n/k ≤ n is
∗
sub-dimensional Theorem 6.1.3 tells that Dα u ∈ Lp with the estimate

kDα ukLp∗ ≤ c1 kDα ukW 1,p ≤ c1 kDα ukp + kDDα ukp
∗
1 1
for some constant c1 = c1 (p, n, D). So u ∈ W k−1,p where p∗ = p − n1 . So
n/k
p∗ = p · n
n−p > p, p∗ = n · p
n−p <n· n−n/k =n· 1
k−1 ≤ n.
If k = 2 we are done by Theorem 6.1.3 and q = p∗∗ . Suppose k ≥ 3.

∗
Step 2. By Step 1 whenever |β| ≤ k − 2 we have Dβ u ∈ W 1,p and since p∗ < n
∗∗
is sub-dimensional Theorem 6.1.3 tells that Dβ u ∈ Lp with the estimate
β
D u ∗∗ ≤ c2 Dβ u ∗ ≤ c2 Dβ u ∗ + DDβ u ∗

p 1,p p p
∗∗
for some constant c2 = c2 (p∗ (p, n), n, D) = c2 (p, n, D). Hence u ∈ W k−2,p
where p1∗∗ = p1∗ − n1 = p1 − n2 . Thus
p∗ n/(k−1)
p∗∗ = p∗ · n
n−p∗ > p∗ , p∗∗ = n · n−p∗ <n· n−n/(k−1) =n· 1
k−2 ≤ n.
If k = 3 we are done by Theorem 6.1.3 and q = p∗∗∗ . Suppose k ≥ 4.

..
.
Let us abbreviate p2 := p∗∗ and p3 := p∗∗∗ and so on.
..
.
Step k − 1. By Step k − 2 whenever |γ| ≤ 1 we have Dγ u ∈ W 1,pk−2 and as
pk−2 < n is sub-dimensional Theorem 6.1.3 tells that Dγ u ∈ Lpk−1 and

kDγ ukpk−1 ≤ ck−1 kDγ uk1,pk−2 ≤ c2 kDγ ukpk−2 + kDDγ ukpk−2
1 1 k−1
for some ck−1 = ck−1 (p, n, D). So u ∈ W 1,pk−1 where pk−1 = p − n and
n pk−2 n/2 1
pk−1 = pk−2 · n−pk−2 > pk−2 , pk−1 = n n−p k−2
< n n−n/2 = n 2−1 = n.
Step k. By Step k − 1 we have u ∈ W 1,pk−1 and pk−1 < n is sub-dimensional.

Thus Theorem 6.1.3 tells that u ∈ Lq , where 1q := p1k = p1 − nk > 0 since kp < n,
and with the estimate

kukq ≤ ck kuk1,pk−1 ≤ ck kukpk−1 + kDukpk−1
for some constant ck = ck (p, n, D). Note that pk = pk−1 · n−pnk−1 > pk−1 .
To conclude the proof set q := pk and consider the Lq estimate. Estimate
its RHS by the estimate in Step k − 1 whose RHS is estimated by Step k − 2
and so on until we get to Step 1 whose RHS involves terms DDα u of order k
and in the Lp norm. This concludes the proof of Theorem 6.1.5.
6.2. SUPER-DIMENSIONAL CASE KP > N 53
6.1.3 Compactness (Rellich-Kondrachov)
6.2 Super-dimensional case kp > n

6.2.1 Morrey’s inequality (p > n) – continuity
Throughout this subsection k = 1, so super-dimensional finite case means
1 ≤ n < p < ∞.
For infinite p = ∞ see Section 7.2 on the relation to Lipschitz continuity.

Theorem 6.2.1 (Morrey’s inequality). Given a finite p > n, every u ∈ C ∞ (Rn )
satisfies the estimates
|u(x) − u(y)|
[u]C 0,µ (Rn ) := sup µ ≤ c kDukLp (Rn )
x6=y |x − y|
n
where µ := 1 − p and

kukC 0 (Rn ) := sup |u| ≤ c kukLp (Rn ) + kDukLp (Rn )
n+1
1− p1
p−1
where in both estimates c := σ2n 1/p p−n . Here σn = nβn is the area of the
n
unit sphere in R and βn denotes the volume of the unit ball.
Finiteness of the right hand sides is guaranteed for compactly supported u’s.
To name the theorem we followed [Eva98, §5.6.2]. Since the relation between
weak differentiability and continuity is one, if not the, fundamental pillar of the
theory of Sobolev spaces we include the proof following [MS04, Le. B.1.16].
Corollary 6.2.2. Theorem 6.2.1 asserts that for finite p > n it holds that
kuk 0,1− n ≤ 3c kukW 1,p (Rn ) < ∞ (6.2.4)

C p (Rn )
for every u ∈ C0∞ (Rn ).

Proof of Theorem 6.2.1. The proof takes three steps.
Step 1. Suppose B b Rn is a non-empty and convex set. Then every smooth
function u on B of mean value zero, in symbols
Z
(u)B := u = 0,
B
satisfies at every point x of B the estimate

1 1− p1
dn σn 1− p

p−1 n
|u(x)| ≤ d1− p kDukLp (B) (6.2.5)
n |B| p−n
where d := diam B denotes the diameter of B and |B| := vol B the volume.
To see this pick x, y ∈ B. For t ∈ [0, 1] the path γ(t) := x + t(y − x) takes values
in the convex domain B of u. As γ̇ = y − x and γ(1) = y and γ(0) = x, we get
Z Z 1 Z Z
− Du|x+t(y−x) [y − x] dt dy = − u(y) dy + u(x) dy = |B| · u(x).
B 0
| B {z } B
| {z }
d
= dt u◦γ(t)
=(u)B =0
Without changing notation extend the gradient Du from B to Rn by zero. So

Z Z 1
|B| · |u(x)| ≤ |Du(x + t(y − x))| · |y − x| dt dy
B 0
Z Z 1
≤ |Du(x + t(y − x))| · |y − x| dt dy
Bd (x) 0 | {z }
z(y)
Z Z 1
= |Du(x + tz)| · |z| dt dz.
{|z|≤d}
|0 {z }
=:f (z)
Step two uses that x ∈ B ⊂ Bd (x) := {|y| < d} since d = diam B. Observe that
dy = dy1 ∧ · · · ∧ dyn . In step three we introduced the new variable z(y) := y − x.
Next let dS(r) denote the area form on the radius r sphere {|z| = r} and
introduce polar coordinates 1 r = |z| and η = |z|−1 z with rη = z to get that
=f (rη)
z }| { !
Z d Z Z 1
n−1
= r |Du(x + |{z}
tr η)| r dt dS(1) dr
0 {|η|=1} 0
ρ(t)
Z d Z r Z !
= rn−1 ρn−1 |Du(x + ρη )| · ρ1−n dS(1) dρ dr
0 0 {|η|=1} |{z}
y
!
Z d Z
= rn−1 |Du(x + y)| · |y|1−n dy dr.
0 {|y|≤r}
To get line two we chose as a new variable the dilation ρ(t) := rt, so r dt = dρ,
we interchanged the order of integration, and we multiplied by 1. To obtain line
three we interpreted (ρ, η) as the polar coordinates of y = ρη ∈ {|y| ≤ r}.
Next use the inclusion {|y| ≤ r} ⊂ Bd (0) and monotonicity of the integral
1 For (r, η) : {0 < |z| ≤ d} → (0, d] × Sn−1 , z 7→ (|z|, |z|−1 z), Fubini implies the formula
! !
Z Z d Z Z d Z
f (z) dz = f (z) dS(r) dr = rn−1 f (rη) dS(1) dr.
Bd (0) 0 {|z|=r} 0 {|η|=1}
Concerning polar coordinates we recommend [Fol99, §2.7].

to obtain a new dy integral which is independent of r. Thus

!Z
Z d
n−1 1−n
≤ r dr |Du(x + y )| · |y| dy
0 Bd (0) | {z }
z(y)
dn
Z
1−n
= |Du(z)| · |z − x| dz.
n Bd (x) | {z }
1/p + 1/q =1
Now apply Hölder (2.2.5) to get that
! q1
dn
Z
≤ kDukLp (Bd (x)) |z − x|(1−n)q dz
n Bd (x)
| {z }
=:y
!1− p1
dn
Z
(1−n)q
= kDukLp (B) |y| dy
n {|y|≤d}
To obtain the last line we used the inclusion B ⊂ Bd (x) and the fact that we
have defined Du = 0 on the complement of B. Straightforward calculation2
then proves (6.2.5) and thus Step 1.
Step 2. Let p ∈ (n, ∞). Given u ∈ C ∞ (Rn ) and x, y ∈ Rn , set x0 := 12 (x + y)

and B := Br (x0 ) where 2r := |x − y| = diam B =: d. Then
2n+1 p−1 1− p1 1− n

|u(x) − u(y)| ≤ 1 p−n |x − y| p
kDukLp (B) .
σn p
Step 2 proves the first estimate in Theorem 6.2.1. To see Step 2 use the triangle
inequality to get |u(x) − u(y)| ≤ |u(x) − (u)B | + |(u)B − u(y)|. Now apply Step
n n
1 with dn|B|
σn
= (2r) nβn n
nr n βn = 2 to each of the two terms in the sum.
Step 3. Let p ∈ (n, ∞). Given u ∈ C ∞ (Rn ) and x ∈ Rn , set B := B1 (x). Then
2n 1
p−1 1− p 1
|u(x)| ≤ 1 p−n kDukLp (B) + βn 1− p kukLp (B) .
σn p
2 introduce polar coordinates, the radial coordinate being r := |y|, to obtain that
Z Z d σn 1−n p−n
+1 p−1 p−1
|y|(1−n)q dy = σn r| n−1{z
rq−nq} dr = 1−n
d p−1 = σn p−n d
{|y|≤d} 0 p−1
+1
(n−1)(1−q)= 1−n
r p−1
Step 3 proves the second estimate in Theorem 6.2.1.3 Apply Step 1 to get
|u(x)| ≤ |u(x) − (u)B | + |(u)B |

2n p−1 1− p1 1− np 1
≤ 1 p−n 2| {z } kDukLp (B) + βn 1− p kukLp (B) .
σn p ∈(1,2)
where we used that |B| := volB = βn andR d := diam B = 2. Moreover, we

applied Hölder to get the estimate |(u)B | ≤ B |1 · u| ≤ |B|1−1/p kukLp (B) .
This completes the proof of Theorem 6.2.1.
Theorem 6.2.3 (Super-dimensional W 1,p (D) estimates). Suppose D b Rn is

Lipschitz. For super-dimensional p ∈ (n, ∞) any equivalence class u ∈ W 1,p (D)
admits a µ-Hölder continuous representative u∗ ∈ C 0,µ (D) where µ = 1 − np and
ku∗ kC 0,µ (D) ≤ c ku∗ kW 1,p (D)
where the constant c depends only on p, n, and D.
For infinite p = ∞ see Section 7.2 on Lipschitz continuity.
Proof. The idea of proof is precisely the same as the one for Theorem 6.1.3,
have a look there, just replace Gagliardo-Nirenberg-Sobolev by Morrey.
1,p
Extend u ∈ W 1,p (D) to ū := Eu ∈ W01,p (Q) := C0∞ (Q) for any pre-
compact Q with D b Q b Rn using the Extension Theorem 5.2.1. Pick a rep-
resentative ū ∈ ū and approximate it by a Cauchy sequence um ∈ C0∞ (Q) that
converges to ū in the W 1,p , hence the Lp , norm. The Morrey inequality (6.2.4)
for the um extended to Rn by zero, namely kum − u` kC 0,µ ≤ 3ckum − u` kW 1,p
where µ = 1 − n/p, tells that the um also form a Cauchy sequence in the Hölder
Banach space C 0,µ (Q). Thus it admits a limit, say u∗ ∈ C 0,µ (Q).
Note that k·kLp (Q) ≤ |Q|1/p k·kC 0 (Q) . Hence C 0,µ (Q) convergence implies Lp (Q)
convergence. Thus u∗ = ū a.e. by uniqueness of limits in Lp (Q). But along D
we have that u∗ = ū = u a.e. and therefore [u∗ ] = u ∈ W 1,p (D).
Taking limits on both sides of the Morrey inequality kum kC 0,µ ≤ 3ckum kW 1,p
and, in the first step, monotonicity of the supremum shows that
ku∗ kC 0,µ (D)) ≤ ku∗ kC 0,µ (Q) ≤ 3c kūkW 1,p (Q) ≤ γp C kukW 1,p (D)
where C = C(p, D, Q) is the constant in the Extension Theorem 5.2.1.

3 The coefficient of kDukp is obviously smaller than c. For the one of kukp we get that
2n+1 p−1 1−1/p n+1
2n+1 2n+1 βn
> 2 1/p = 1/p 1−1/p

c :=
σn1/p p−n σn 1/p ≥
n 1/p >
βn 1/p = βn
n·βn βn
.
The final inequality is equivalent to βn < 2n+1 /n which is true: For even n = 2k Wiki tells
β2k = π k /k! which is indeed smaller than 4k /k = 22k+1 /2k. For odd n = 2k + 1 one has that
2 2 2 2 4 2(2k+1)+1
β2k+1 = 1 3 5
... 2k+1
πk < 2k+1
4k = 2k+1
.
6.2.2 General Sobolev inequalities (kp > n)

Given a non-integer real r ∈ R \ Z, then ` < r < ` + 1 for some ` ∈ Z. Let us
call ` and ` + 1 the floor and ceiling integer of r, respectively, in symbols
brc := `, dre := ` + 1.
For an integer real r ∈ Z set brc := r =: dre.

Theorem 6.2.4 (General Sobolev inequalities k > np ). Let D b Rn be Lipschitz
and k ∈ N and p ∈ [1, ∞). If kp > n then any u ∈ W k,p (D) has a k − d np e times
n
γ-Hölder continuously differentiable representative u∗ ∈ C k−d p e,γ (D) where
(
d np e − np , np ∈
/ N,
γ :=
any µ ∈ (0, 1) , otherwise (in this case k − d np e ≥ 1).
Moreover, the Hölder representative u∗ satisfies the estimate
ku∗ k k−d n e,γ

p
≤ c ku∗ kW k,p (D) (6.2.6)
C (D)
where the constant c depends only on k, p, n, γ, and D.

For infinite p = ∞ see Section 7.2 on Lipschitz continuity.
Proof. Case n/p ∈ / N. The idea is to employ the general sub-dimensional
Sobolev inequalities to get from the largest sub-dimensional case `p < n to the
smallest super-dimensional one 1 · r > n where 1r = p1 − n` and then contract the
super-dimensional Sobolev W 1,r estimate that we just proved.
More precisely, let b np c := ` < np < ` + 1 ≤ k be the first integer below np . So
we are in the sub-dimensional case `p < n and from our hypothesis u ∈ W k,p (D)
we conclude that u ∈ W k−`,r (D) by the same iteration process as in the proof
of the general sub-dimensional Sobolev inequalities, Theorem 6.1.5. Note that
n/p < ` + 1 iff `p > n − p. Thus
p p
r =n· >n· =n
n − `p n − (n − p)
is super-dimensional and Dα u ∈ W 1,r (D) whenever |α| ∈ {0, . . . , k − ` − 1}.

Such Dα u admits a representative (Dα u)∗ ∈ C 0,µ (D) by Theorem 6.2.3 and
k(Dα u)∗ kC 0,µ (D) ≤ c k(Dα u)∗ kW 1,r (D)
where the constant c depends only on p, n, D and µ = 1− nr = 1− np +` = d np e− np .

Hence all weak derivatives up to order k−`−1 = k−d np e are µ-Hölder continuous
n n n
by Lemma 4.1.11. This shows that u∗ ∈ C k−d p e,d p e− p (D) as we had to prove.
Take the maximum over all these estimates to obtain (6.2.6).
Case n/p ∈ N. Suppose k > np , i.e. k − 2 ≥ np − 1 ≥ 0, and u ∈ W k,p (D). To
apply the general sub-dimensional Sobolev inequalities, Theorem 6.1.5, we need
n
Figure 6.3: Case p ∈ N - choice of ` ∈ N0
an integer ` < np . For best result let us choose the largest one ` := np − 1 as
illustrated by Figure 6.3 Observe that ` ∈ {0, . . . , k − 2} and Dα u ∈ W `,p (D)
whenever |α| ≤ k − ` (≥ 2).
Case ` ≥ 1. a) In this case Dα u ∈ W 1,p (D). So by sub-dimensionality (`p < n)
Theorem 6.1.5 tells that Dα u ∈ Lq (D) with 1q = p1 − n` (= n1 ) and it provides
the constant c = c(`(p, n), p, n, D) and the second of the two inequalities
kDα ukLr (D)
1 1
−n
≤ kDα ukLn (D) ≤ c kDα ukW `+1,p (D) , |α| ≤ k − ` (≥ 2).
|D| r
Here the first inequality is by Hölder (2.2.6) and holds for any r ∈ [1, n). This
shows that Dβ u ∈ W 1,r (D) whenever |β| ≤ k − ` − 1 (≥ 1) and r ∈ [1, n).
b) Hence the consequence of Gagliardo-Nirenberg-Sobolev, the sub-dimensional
∗
W 1,r estimate Theorem 6.1.3, asserts that Dβ u ∈ Lr (D) with the estimate
|β| ≤ k − ` − 1 (≥ 1), r∗ ∈ (n, ∞),

β
≤ C Dβ u W 1,r (D) ,

D u r∗
L (D)
where C = C(r, n, D).4

∗
But this means that Dβ u ∈ W 1,r (D) for every super-dimensional r∗ ∈ (n, ∞).
Thus D u admits a representative (Dβ u)∗ ∈ C 0,µ (D) by Theorem 6.2.3 and
β
β ∗
≤ c0 (Dβ u)∗ W 1,r (D) .

(D u) 0,µ (6.2.7)
C (D)
Here the constant c0 depends only on r∗ , n, D and

1
µ : (n, ∞) → (0, 1), r∗ 7→ 1 − ,
r∗
is a bijection. Hence all weak derivatives of u∗ = (D0 u)∗ up to order k − ` − 1 =
n
k − d np e are γ-Hölder continuous by Lemma 4.1.11, in symbols u∗ ∈ C k−d p e,γ ,
and this is true for any γ ∈ (0, 1).
Take the maximum over |β| ≤ k − `−1 of the estimates (6.2.7) to obtain that
ku∗ k k−d n e,γ

p
≤ C 0 ku∗ kW k−`,r (D) ≤ C 0 ku∗ kW k,r (D) (6.2.8)
C (D)
for all γ ∈ (0, 1) and r ∈ [1, n) and where C 0 = C 0 (r∗ , n, D, k). But p < n
` ≤ n,
so we can choose r := p to finish the proof of (6.2.6) in the case ` ≥ 1.
4 The map r∗ : ( n
2
, n) → (n, ∞), r 7→ nr
n−r
, is a bijection.
Case ` = 0 (p = n). In this case we have u ∈ W k,n (D) with k ≥ 2. Then

Dβ u ∈ W 1,n (D) whenever |β| ≤ k − 1. By Hölder (2.2.6) and finite measure
of D (compact closure) we get that Dβ u ∈ W 1,r (D) whenever |β| ≤ k − 1 and
any sub-dimensional r ∈ [1, n). Now continue as in part b) of case ` ≥ 1 above
just setting ` = 0 throughout. Only in the very last sentence arises a difference,
because in the case at hand p = n is not sub-dimensional, so we can’t set r := p
to finish the proof. However, again by Hölder (2.2.6) we can estimate the RHS
of (6.2.8) by the W k,p (D) norm since r < n = p.
This concludes the proof of Theorem 6.2.4.
Chapter 7
Applications
7.1 Poincaré inequalities

As an application of the compactness theorem in Section 6.1.3 one obtains
Theorem 7.1.1 (Poincaré’s inequality). Suppose D b Rn is Lipschitz and
connected. Let p ∈ [1, ∞]. Then there is a constant c = c(n, p, D) such that
ku − (u)D kLp (D) ≤ c kDukLp (D)
for every u ∈ W 1,p (D).

Proof. [Eva98, §5.8.1]
7.2 Lipschitz functions

Theorem 7.2.1 (Identification of W 1,∞ (D) with C 0,1 (D)). Let D b Rn be
Lipschitz. Then every class u ∈ W 1,∞ (D) admits a (unique) Lipschitz contin-
uous representative, notation u∗ . Vice versa, any Lipschitz continuous map is
weakly differentiable with a.e. bounded weak derivatives. In symbols, the map
C 0,1 (D) W 1,∞ (D), u 7→ [u],
is a bijection with inverse u 7→ u∗ .

Proof. See e.g. [Eva98, §5.8.2].
For finite p there is no such bijection. For super-dimensional finite p ∈ (n, ∞)
there is the injection provided by the Sobolev inequality Theorem 6.2.3, namely
W 1,p (D) C 0,1−n/p (D), [u] 7→ u∗ .
Exercise 7.2.2. Find examples that show non-surjectivity of this injection.
61
62 CHAPTER 7. APPLICATIONS
7.3 Differentiability almost everywhere

Theorem 7.3.1 (Super-dimensional differentiability a.e.). Given Ω ⊂ Rn , let
1,p
p ∈ (n, ∞] be super-dimensional. Then the following is true. Let u ∈ Wloc (Ω).
1 ∗
Then its (1−n/p)-Hölder continuous representative u ∈ u, also any other rep-
resentative u, is differentiable almost everywhere and strong and weak gradient
coincide almost everywhere, in symbols (∂1 u, . . . , ∂n u) = (ue1 , . . . , uen ) a.e.
Proof. See e.g. [Eva98, §5.8.3]

Theorem 7.3.2 (Rademacher’s Theorem). Suppose u is locally Lipschitz con-
tinuous in Ω ⊂ Rn . Then u is differentiable almost everywhere in Ω.
Proof. Theorem 4.1.16 and Theorem 7.3.1.
1 see Theorem 6.2.3 for p ∈ (n, ∞) and Theorem 7.2.1 for p = ∞

Appendix A
Appendix
A.1 Allerlei
A.1.1 Inequalities via convexity and concavity
Lemma A.1.1 (Young’s inequality). Let a, b ≥ 0 and p, q ∈ (1, ∞) with p1 + 1q =
1, then
ap bq
ab ≤ + . (A.1.1)
p q
Proof. If one of a, b is zero, the inequality holds trivially. Assume a, b > 0.
Observe that
p
bq

1 p 1 q a
log(ab) = log a + log b = log a + log b ≤ log +
p q p q
where the last step holds by concavity of the logarithm. Since the logarithm is
strictly increasing the result follows.
Add (a + b)2 = a2 + 2ab + b2 and 0 ≤ (a − b)2 = a2 − 2ab + b2 , or use
Young (A.1.1), to get that
(a + b)2 ≤ 2a2 + 2b2 a, b, ∈ R. (A.1.2)
More generally, for a, b ∈ R and p ∈ [1, ∞) it is true that
p p p p
|a + b| ≤ (|a| + |b|) ≤ 2p−1 (|a| + |b| ) . (A.1.3)
This follows from convexity of t on (0, ∞). Hence c +d ≤ (c+d) for c, d ≥ 0.1
p p p p
It is useful to keep in mind the following consequences, namely

1
p p p
kuk1,p := kukp + kDukp ≤ kukp + kDukp
1 1
(A.1.4)
p p p
kukp + kDukp ≤ 2p−1 kukp + kDukp = |21− p kuk
{z } 1,p .
≤2
p
1 1 c+d 1
Indeed by convexity 2p
(c + d)p = 2
≤ 2
(cp + dp ); cf. [Rud87, Thm. 3.5 Pf.].
63
64 APPENDIX A. APPENDIX
2
Exercise A.1.2. For x = (x1 , . . . , xn ) ∈ Rn let |x| := x21 + · · · + x2n . Then
|x| ≤ |x1 | + · · · + |xn | .
More generally, for p ∈ [1, ∞) there is the estimate

p p p p
|x| ≤ (|x1 | + · · · + |xn |) ≤ κ (|x1 | + · · · + |xn | ) (A.1.5)
where κ = 2(p−1)(n−1) and κ1/p ≤ 2n−1 . For p = 2 we get that

n−1
|x1 | + · · · + |xn | ≤ 2 2 |x| ≤ 2n−1 |x| . (A.1.6)
A.1.2 Continuity types and their relations

For a function f : I → R on a compact interval, say I = [0, a] with a > 0 in view
of the examples, there are various types of continuity (Figures A.1 and A.2)
(C) continuous
(UC) uniformly continuous
(α-HC) α-Hölder continuous where α ∈ (0, 1)
(LC) Lipschitz continuous
(Diff ) differentiable
(AC) absolutely continuous
(BV) bounded variation
(Diff-a.e.) differentiable almost everywhere
Figure A.1: Relations among continuity types (compact domain)

A.1. ALLERLEI 65
Figure A.2: Relations among continuity types
A.1.3 Distance function

Saying that Y is a metric space means that Y is a set equipped with a metric2
d : Y × Y → [0, ∞) and the metric topology Td whose basis are the open balls
about the points of Y .
Exercise A.1.3. Let Y be a metric space and equipp Y × Y with the product
topology. Then the metric is continuous as a function d : Y × Y → [0, ∞).
The distance of a point y ∈ Y to a subset A ⊂ Y is the infimum dA (y) of
the distances from y to the points a of A, in symbols
dA (y) := d(y, A) := inf d(y, a).

a∈A
The map dA : Y → [0, ∞] is called the distance function of A. We use the

convention that the infimum over the empty set is infinite, in symbols
inf := ∞.
∅
Lemma A.1.4. Suppose Y is a metric space. Then for any subset A ⊂ Y the
distance function dA : Y → [0, ∞] is continuous.
Proof. Following [Dug66, Ch. IX Thm. 4.3] pick elements x, y ∈ Y . Then
dA (x) := inf d(x, a) ≤ d(x, y) + inf d(y, a) = d(x, y) + dA (y)

a∈A a∈A
which shows that dA (x) − dA (y) ≤ d(x, y). Interchange x and y to see that
|dA (x) − dA (y)| ≤ d(x, y).
Given x ∈ Y and ε > 0, then for every y ∈ Y with d(x, y) < δ := ε we get that
|dA (x) − dA (y)| ≤ d(x, y) < ε. This shows that dA is continuous.
2 axioms for a metric: non-degeneracy, symmetry, triangle inequality
66 APPENDIX A. APPENDIX
Figure A.3: Modes of convergence
A.1.4 Modes of convergence

See [Fol99, §2.4] for details.
A.2 Banach space valued Sobolev spaces

Bibliography
[AF03] Robert A. Adams and John J. F. Fournier. Sobolev spaces, volume 140
of Pure and Applied Mathematics (Amsterdam). Elsevier/Academic
Press, Amsterdam, second edition, 2003.
[Bre11] Haı̈m Brezis. Functional analysis, Sobolev spaces and partial differen-
tial equations. Universitext. Springer, New York, 2011.
[Cal61] A.-P. Calderón. Lebesgue spaces of differentiable functions and distri-
butions. In Proc. Sympos. Pure Math., Vol. IV, pages 33–49. American
Mathematical Society, Providence, R.I., 1961.
[Dug66] James Dugundji. Topology. Allyn and Bacon, Inc., Boston, Mass.,
1966.
[Eva98] Lawrence C. Evans. Partial differential equations, volume 19 of Grad-
uate Studies in Mathematics. American Mathematical Society, Provi-
dence, RI, 1998.
[Fol99] Gerald B. Folland. Real analysis. Pure and Applied Mathematics
(New York). John Wiley & Sons, Inc., New York, second edition, 1999.
Modern techniques and their applications, A Wiley-Interscience Pub-
lication.
[GT01] David Gilbarg and Neil S. Trudinger. Elliptic partial differential equa-
tions of second order. Classics in Mathematics. Springer-Verlag, Berlin,
2001. Reprint of the 1998 edition.
[LM07] Giovanni Leoni and Massimiliano Morini. Necessary and sufficient con-
1,1
ditions for the chain rule in Wloc (RN ; Rd ) and BVloc (RN ; Rd ). J. Eur.
Math. Soc. (JEMS), 9(2):219–252, 2007.
[MS04] Dusa McDuff and Dietmar Salamon. J-holomorphic curves and sym-
plectic topology, volume 52 of American Mathematical Society Collo-
quium Publications. American Mathematical Society, Providence, RI,
2004.
[Rud87] Walter Rudin. Real and complex analysis. McGraw-Hill Book Co., New
York, third edition, 1987.
67
68 BIBLIOGRAPHY
[Sal16] Dietmar A. Salamon. Measure and integration. EMS Textbooks in

Mathematics. European Mathematical Society (EMS), Zürich, 2016.
[Ste70] Elias M. Stein. Singular integrals and differentiability properties of
functions. Princeton Mathematical Series, No. 30. Princeton University
Press, Princeton, N.J., 1970.
[Zie89] William P. Ziemer. Weakly differentiable functions. Sobolev spaces and

functions of bounded variation., volume 120 of Graduate Texts in Math-
ematics. Springer-Verlag, New York, 1989.
Index
Ωi := Ω1/i , 16
R
(u)B := u mean value of u : B → R,
B
53 Un standard topology on Rn , 5
k,p
2X power set of X, 5 Wloc (Ω), 30
C ∞ (Q) ⊂ W k,p (Q, u 7→ [u], 29 k,1
Wloc (Ω) ⊂ L1loc (Ω), 20
C k -domain, 36 Wlock,∞
(Ω) ⊂ Wloc k,1
(Ω), 20
C k (Ω), 2 Wlock,p
-convergence, 20
C k (Ω̄), 2 |f |C 0,µ (Ω) µ-Hölder coefficient of f , 3
Cbk (Ω) – bounded C k , 2 α ≥ β iff αi ≥ βi ∀i, 3
Cbk (Ω̄), 2 Ā closure of set A, 2
C k,µ (Ω) uniformly Hölder continuous (δ)
f¯ := f zero δ-extension of f ∈
with derivatives up to order
L1loc (Ω), 16
k, 3
βn volume of unit ball in Rn , 53
C k−1,1 -diffeomorphism, 31
u := [u] equivalence class, 27
C0 (Rn ) continuous compact support, C
U
R complement of set U , 5
11
Rn
f (x) dx Lebesgue integral, 8
Du = (ux1 , . . . , uxn ) gradient, 4
dre ceiling integer, 57
D2 u Hessian, 4
brc floor integer, 57
E(f, g) bad set of convolution, 14
µ = m |Bn Borel measure, 6
L∞ -norm, 9
k[f ]kp := kf kp norm on Lp , 10
Lp -function, 9
k·kp Lp -norm of f , 8
Lp -norm, 9 kf kC k,µ (Ω) Hölder norm, 3
Q b Ω pre-compact subset, 2 σ-additive, 6
W k,∞ -norm, 28 σn = nβn area of unit sphere in Rn , 53
W k,p function, 28 supp f usual support, 6
W k,p -norm, 28 suppm f support of measurable func-
W k,p (Ω), 27 tion, 7
W0k,p (Ω), 30 R := R ∪ {±∞} extended reals, 8
[f ] ∗ [g] convolution of classes, 15 dA (y) := inf a∈A d(y, a) distance to A,
n
An ⊂ 2R Lebesgue σ-algebra, 6 65
Bn = AUn ⊂ An Borel σ-algebra, 5 ei := (0, . . . , 0, 1, 0, . . . , 0) ∈ Rn is the
N := {1, 2, . . . } natural numbers, 19 ith unit vector, 4
N0 := N ∪{0} = {0, 1, 2, . . . }, 19 f ∗ g convolution of functions, 14
kukk,∞ Sobolev norm, 28 f = g a.e., 6
kukk,p Sobolev norm, 28 k-bounded Sobolev functions, 31
Ω open subset of Rn , 2 m : An → [0, ∞] Lebesgue measure, 6
Ωδ := {x ∈ Ω | d(x, ∂Ω) > δ}, 16 p-integrable, 9
69
70 INDEX
np
p∗ := n−p Sobolev conjugate of p, 47 exponent, 3
∗
u continuous representative of Hölder r-conjugates p, q, 9
Sobolev class u = [u], 29 Hölder conjugates p, q, 9
uδ := ρδ ∗ ū mollification, 17 hereditary, 11
u` → u in W k,p , 28 Hessian, 4
uei weak derivative in direction ei , 4
inequality
admit Lebesgue integration, 7 Gagliardo-Nirenberg-Sobolev, 47
almost everywhere, 6 Hölder, 9
approximation Morrey, 53
k,p
local – in Wloc (Ω), 23 Sobolev sub-dimensional, 51
Sobolev super-dimensional, 57
bi-Lipschitz map, 40 Young ab, 63
Borel measurable, 6 Young f ∗ g, 14
Borel measure, 6 integrable
boundary of set, 2 Lp –, 9
function, 8, 9
ceiling integer, 57 integral
change of coordinates, 31 Lebesgue –, 7
characteristic function, 7
closed sets, 5 Lebesgue integral, 7
continuous map, 6 Lebesgue measurable, 6
continuous representative, 29 Lebesgue measure, 6
k,p
convergence in Wloc , 20 Lebesgue null set, 6
convolution, 14 LHS left hand side, 3
Lipschitz
diffeomorphism
bi-, 40
C k−1,1 - –, 31
Lipschitz constant, 24
distance function of a subset, 65
Lipschitz continuous, 24
extension, 40 locally –, 24
natural zero –, 13 Lipschitz domain, 36
zero δ- –, 16 local Lipschitz bound of –, 36
locally Lipschitz continuous, 24
floor integer, 57
function, 2 mean value, 53
W k,p –, 28 measurable function
symmetric –, 15 Borel –, 6
Lebesgue –, 6
gradient measurable means Lebesgue measur-
strong –, 4 able, 8
weak –, 4 measurable sets, 6
gradient vector, 4 measure, 5
metric space, 65
Hölder mollification, 16
coefficient, 3 mollification of a continuous function,
continuous, 3 17
INDEX 71
mollifier, 15 Lebesgue dominated convergence

multi-index, 3 generalized, 8
trace, 44
natural zero extension, 13 transformation law, 27
null sets, 6, 19
weak derivative, 4
open sets, 5 Dα u := uα of u ∈ L1loc (Ω), 27
operator norm, 11 uα of u ∈ L1loc (Ω), 19
weak gradient, 4
partial derivative, 3
weakly differentiable, 20
partition of unity
subordinate to open cover, 35 zero δ-extension, 16
polar coordinates, 54
power set, 5
pre-compact subset, 2
representative
continuous –, 29
RHS right hand side, 3
shift operator, 11
sign function, 20
smooth domain, 36
Sobolev conjugate of p, 47
Sobolev inequalities
sub-dimensional –, 51
super-dimensional –, 57
Sobolev space
W k,p (Ω), 27
W0k,p (Ω), 30
k,p
Wloc (Ω), 30
Sobolev spaces, 1
standard topology, 5
strong gradient, 4
sub-dimensional case, 47
super-dimensional case, 53
support
measurable –, 7
symmetric function, 15
theorem
Lebesgue dominated Lp conver-
gence, 10
Lebesgue dominated Lp conver-
gence generalized, 10
Lebesgue dominated convergence,
8

Introduction To Sobolev Spaces

Uploaded by

Copyright:

Available Formats

Introduction To Sobolev Spaces

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Introduction To Sobolev Spaces

Uploaded by

Copyright:

Available Formats

Introduction to Sobolev Spaces

December 23, 2018

5 Approximation and extension 33

1.1 Notation and conventions

Ω open subset of Rn where n ≥ 1

∂A := Ā ∩ AC boundary of set A ⊂ Rn where Ā is the closure of A

D b Ω Lipschitz domain (∂D is locally a Lipschitz graph; see Definition 5.1.5)

K compact set (i.e. bounded and closed)

A map taking values in the real line R is called a function.

C k (Ω) set of functions on Ω all of whose partial derivatives up to order k exist

k·kC k maximum of sup-norms of all partial derivatives up to order k

C k (Ω̄) consists 3 of those functions f : Ω → R in C k (Ω) whose partial derivatives

Cbk (Ω̄) := {f ∈ C k (Ω̄) : kf kC k < ∞} is a closed (hence Banach) subspace of the

fined on a neighborhood of Q̄: uniformly continuous derivatives up to order k follows from

|f |C 0,µ (Ω) := supx6=y |f (x)−f

Integration will always be with respect to Lebesgue measure m on Rn , unless

(c) Multi-index notation.

(i) A list α = (α1 , . . . , αn ) of integers αi ≥ 0 is a multi-index of order

(ii) For a multi-index α the associated partial derivative is denoted by

(iii) The list of all partial derivatives of order k ∈ N0 is denoted by

(iv) Special cases. If k = 1, one obtains a list D1 u denoted by

Du = (∂1 u, . . . , ∂n u) = gradient vector

(iv) Weak derivatives. We usually denote the weak derivative of u

This chapter is an exposition of notions and results a.e. without proofs.

2.1 Borel and Lebesgue measure space on Rn

The standard topology on euclidean space Rn is the collection Un of subsets

Lemma 2.1.1. Every open subset Ω of Rn is of the form Ω = ∪∞

Proof. [Bre11, Cor. 4.23]: Ki := {x ∈ Rn | dist(x, ΩC ) ≥ 2/i and |x| ≤ i}.

Measure compatible with topology – Borel measure

There are many ways to enlarge a given collection C of subsets of Rn to obtain a

and, secondly, the function µ is σ-additive.1 The elements A of a σ-algebra

Completion of Borel measure – Lebesgue measure

Functions – equality almost everywhere and measurability

Definition 2.1.2 (Support of measurable functions). Suppose f : Rn → R is

Since Ω is open suppm f is closed.

Exercise 2.1.3. Check that supp χQ = R, while suppm χQ = ∅. The latter

Exercise 2.1.4 (Ordinary and measurable support).

(i) Show that f = 0 a.e. on Ω, equivalently, that N := {f 6= 0} is a null set.

(ii) If f = g almost everywhere, then suppm f = suppm g.

(iii) For continuous functions both supports suppm f = supp f coincide.

[Hints: (i) That f = 0 a.e. on Ωλ means Nλ := {f 6= 0} ∩ Ωλ is a null set. While

Integration of Lebesgue measurable functions

as the supremum of the integrals of simple functions s with 0 ≤ s ≤ f . Decom-

if at least one of the two terms is finite. Such fR is saidR to admit

measurable function f : Rn → R is called (Lebesgue) integrable. By L1 (Rn )

The frequent problem of interchanging integration and limit is settled by the

Theorem 2.1.5 (Lebesgue dominated convergence theorem). Let all functions

∀k ∈ N : |fk (x)| ≤ g(x) for a.e. x ∈ E.

Then the fk and also f are integrable and

Theorem 2.1.6 (Generalized Lebesgue dominated convergence theorem). Let

Convention 2.1.7. Throughout we work with Lebesgue measure m : An →

On the other hand, consider the (possibly empty) semi-infinite interval

If := {c ∈ [0, ∞) : the set {|f | > c} is of measure zero} ⊂ [0, ∞).

kf k∞ := inf If = ess sup |f |. (2.2.4)

By convention inf ∅ = ∞, hence kf k∞ ∈ [0, ∞]

Proof. See e.g. [Sal16, Thm. 4.1].

kf gkr ≤ kf kp kgkq . (2.2.6)

kf1 f2 . . . f` kr ≤ kf1 kp1 . . . kf` kp` . (2.2.7)