Chapter 1

Preliminary general material

For the reader’s convenience, this chapter summarizes some concepts, definitions
and results which are mostly relevant to the undergraduate curriculum and are
thus assumed as basically known, or have specific roots in rather distant areas
and have rather auxiliary character with respect to the purpose of this book,
which is to push pure theory towards possible applications as fast as possible. As
such, all assertions in Chapter 1 are made without proofs and the scope has been
minimized to only material actually needed in the book. If a reader decides to skip
this Chapter, (s)he can easily come back through references in the Index or in the
text itself, if in need.

1.1 Functional analysis

The universal framework used in (even nonlinear) differential equations is based
on linear functional analysis, and also on convex analysis.

1.1.1 Normed spaces, Banach spaces, locally convex spaces

Considering a (real) linear space V ,1 a non-negative, degree-1 homogeneous, sub-
additive functional · V : V → R is called a norm if it vanishes only at 0; often,
we will write briefly · instead of · V if V is clear from the context.2 A linear
space equipped with a norm is called a normed linear space. If the last property
(i.e. u V = 0 ⇒ u = 0) is missing, we call such a functional a seminorm; i.e. a
1 This means V is endowed by a binary operation (v , v ) → v + v : V × V → V which
1 2 1 2
makes it a commutative group, i.e. v1 + v2 = v2 + v1 , v1 + (v2 + v3 ) = (v1 + v2 ) + v3 , ∃ 0 ∈ V :
v + 0 = v, and ∀v1 ∈ V ∃v2 : v1 + v2 = 0, and furthermore it is equipped with a multiplication
by scalars (a, x) → ax : R × V → V satisfying (a1 + a2 )v = a1 v + a2 v, a(v1 + v2 ) = av1 + av2 ,
(a1 a2 )v = a1 (a2 v), and 1v = v.
2 The mentioned properties mean respectively: v ≥ 0, av = |a| v, u+v ≤ u + v

for any u, v ∈ V and a ∈ R, and v = 0 ⇒ v = 0.

T. Roubíþek, Nonlinear Partial Differential Equations with Applications, 1

International Series of Numerical Mathematics 153,
DOI 10.1007/978-3-0348-0513-1_1, © Springer Basel 2013
2 Chapter 1. Preliminary general material

functional | · |ξ : V → R is a seminorm if it satisfies

∀u, v ∈ V ∀a ∈ R : 0 ≤ |u+v|ξ ≤ |u|ξ + |v|ξ and |au|ξ = |a| |u|ξ . (1.1)

Having V equipped with a collection {| · |ξ }ξ∈Ξ of seminorms | · |ξ with an arbitrary

index set Ξ, we call V a locally convex space. A sequence {uk }k∈N in V is then
called a Cauchy sequence if

∀ξ ∈ Ξ ∀ε > 0 ∃k0 ∈ N ∀k1 , k2 ≥ k0 : |uk1 − uk2 |ξ ≤ ε. (1.2)

Moreover, a sequence {uk }k∈N is called convergent to some u ∈ V if

∀ξ ∈ Ξ ∀ε > 0 ∃k0 ∈ N ∀k ≥ k0 : |uk − u|ξ ≤ ε. (1.3)

In this case, u is called its limit and we will write u = limk→∞ uk or uk → u (or,
depending on a particular collection of seminorms, uk  u). A subset A of a locally
convex space V is called closed if any limit of any convergent sequence contained
in A is itself in A.3 Moreover, A is called open if V \ A is closed. The closure of A,
denoted by cl(A), is the smallest closed set B ⊃ A, while int(A) := A \ cl(V \ A)
is called the interior of A. The concrete collection of seminorms will often be
specified by various adjectives as “strong” or “weak” or “weak*”. If |u|ξ = 0 for
all ξ ∈ Ξ implies u = 0, V is called a Hausdorff locally convex space. Then every
convergent sequence has a uniquely determined limit. The normed linear space
is a Hausdorff locally convex space with its only seminorm being then just the
norm. A subset A ⊂ V is called bounded if supx∈A x < +∞, and dense (in V ) if
cl(A) = V . If there is a countable dense subset of V , we say that V is separable.
If every Cauchy sequence in a normed linear space V converges, we say that this
space is complete and then V is called a Banach space.4 An example of a Banach
space is Rn
endowed by the norm, denoted usually by | · | instead of · , defined
by |s| = ( ni=1 s2i )1/2 ; such a Banach space is called an n-dimensional Euclidean
space.
If V is a Banach space such that, for any v ∈ V , V → R : u → u + v 2 −
u − v 2 is linear, then V is called a Hilbert space. In this case, we define the inner
product (also called scalar product) by

1 1
(u, v) := u + v 2 − u − v 2 . (1.4)
4 4

By the assumption, (·, ·):V ×V →R is a bilinear form which is obviously symmetric5

and satisfies (u, u) = u 2 . E.g., the Euclidean space Rn is a Hilbert space.
3 We will always work with Ξ at most countable and therefore, for simplicity, we define “closed-

ness” in terms of convergence of sequences, which would be in general situations called rather
“sequential closedness”.
4 This fundamental concept has been introduced in [33] in 1922.
5 This means both u → (u, v) and v → (u, v) are linear functionals on V and (u, v) = (v, u).
1.1. Functional analysis 3

Let us call a Banach space V strictly convex if the sphere6 in V does not
contain any line segment. The space V is uniformly convex if
 1
u = v =1  1 
∀ε > 0 ∃δ > 0 ∀u, v ∈ V : ⇒  u + v  ≤ 1 − δ. (1.5)
u − v ≥ ε 2 2
Any uniformly convex Banach space is strictly convex but not vice versa.

1.1.2 Functions and mappings on Banach spaces, dual spaces

Having in mind a specific mode of convergence, a function f : V → R ∪ {±∞} is
called lower (resp. upper) semicontinuous7 if

∀u ∈ V, uk → u : f (u) ≤ lim inf f (uk ) resp. f (u) ≥ lim sup f (uk ) . (1.6)
k→∞ k→∞

Having two normed linear spaces V1 and V2 and a mapping A : V1 → V2 , we say

that A is continuous if it maps convergent sequences in V1 onto convergent ones
in V2 , and is a linear operator if it satisfies A(a1 v1 + a2 v2 ) = a1 A(v1 ) + a2 A(v2 )
for any a1 , a2 ∈ R and v1 , v2 ∈ V1 . Often we will write briefly Av instead of A(v).
If V1 = V2 , a linear continuous operator A : V1 → V2 is called a projector if
A ◦ A = A. The set of all linear continuous operators V1 → V2 will be denoted
by L (V1 , V2 ), being itself a normed linear space when equipped with the addition
and multiplication by scalars defined respectively by (A1 + A2 )v = A1 v + A2 v and
(aA)v = a(Av), and with the norm
Av V2
A L (V1 ,V2 ) := sup Av V2 = sup . (1.7)
vV1 ≤1 v=0 v V1

A continuous (possibly even nonlinear) mapping A : V1 → V2 is called a homeo-

morphism if the inverse A−1 : V2 → V1 does exist and is continuous. Moreover,
A : V1 → V2 is called a homeomorphical embedding if A : V1 → A(V1 ) is a homeo-
morphism and A(V1 ) is dense in V2 .
As R itself is a linear topological space8 , we can consider the linear space
L (V, R), being also denoted by V ∗ and called the dual space to V . The original
space V is then called predual to V ∗ . For an operator (=now a functional) f ∈ V ∗ ,
we will write f, v instead of f v. The bilinear form ·, ·V ∗ ×V : V ∗ × V → R is
called a canonical duality pairing. Instead of ·, ·V ∗ ×V , we will occasionally write
briefly ·, ·. Always, V ∗ is a Banach space if endowed by the norm (1.7), denoted
often briefly · ∗ instead of · V ∗ , i.e. f ∗ = supv≤1 f, v. Obviously,

u 
f, u = u f, ≤ u sup f, v = f ∗ u . (1.8)
u v≤1

6A “sphere” {v ∈ V ; v = } (of the radius  > 0) is a surface of a “ball” {v ∈ V ; v ≤ }.
7 The property (1.6) is often called sequential lower semicontinuity.
8 The conventional norm on R is the absolute value | · |.
4 Chapter 1. Preliminary general material

If V is a Hilbert space, then (u → (f, u)) ∈ V ∗ for any f ∈ V , and the mapping
f → (u → (f, u)) identifies V with V ∗ . Then (1.8) turns into a so-called Cauchy-
Bunyakovskiı̆ inequality (f, u) ≤ f u .
Having two Banach spaces V1 , V2 , we have
 ∗ 
V1 ∩ V2 ∼ = V1∗ + V2∗ := f = f1 +f2 ; f1 ∈ V1∗ , f2 ∈ V2∗ (1.9)

if the duality pairing is defined by f, v = f1 , vV1∗ ×V1 + f2 , vV2∗ ×V2 with f =
f1 + f2 , and if the norm on V1 ∩ V2 is taken as v V1 ∩V2 := max( v V1 , v V2 )
while f V1∗ +V2∗ := inf f =f1 +f2 ( f1 V1∗ + f2 V2∗ ).
Theorem 1.1 (Banach-Steinhaus principle [36]). Let {Aα }α∈S be a family in
L (V1 , V2 ), V1 a Banach space, V2 a normed linear space. Then boundedness of
{Aα (v)}α∈S ⊂ V2 for any v ∈ V1 implies boundedness of {Aα }α∈S ⊂ L (V1 , V2 ).
One can consider a normed linear space V equipped with the collection of
seminorms {v → |f, v|}f ∈V ∗ , which makes it a locally convex space. Sequences
converging in this locally convex space will be called weakly convergent. Any se-
quence {uk }k∈N converging in the original norm, i.e. limk→∞ uk − u V = 0, will
be called norm-convergent or also strongly convergent and is automatically weakly
convergent.9
If the convergence in (1.6) refers to the weak one, the function f : V → R
will be called weakly lower (resp. upper) semicontinuous.
Theorem 1.2. 10 If V is uniformly convex, uk  u, and uk → u , then uk → u.
Likewise, also the Banach space V ∗ can be endowed by the collection of
seminorms {f → |f, v|}v∈V , which makes it a locally convex space. Sequences
converging in this locally convex space will be called weakly* convergent. Any
sequence {fk }k∈N converging in the original norm, i.e. limk→∞ fk − f ∗ =
0, is automatically weakly* convergent. The duality pairing is continuous if
V ∗ × V is equipped with weak*×norm or norm×weak topology11 and separately
(weak*,weak)-continuous.12
Proposition 1.3. If V ∗ is separable, then so is also the space V .
The Banach-Steinhaus Theorem 1.1 has immediately the following
Corollary 1.4. Every weakly* convergent sequence in V ∗ must be bounded if V
is a Banach space. In particular, every weakly convergent sequence in a reflexive
Banach V must be bounded.
9 The converse implication holds only if the normed linear space V is finite-dimensional, i.e. iso-

metrically isomorphic to an Euclidean space.

10 See Fan and Glicksberg [140] for thorough investigation and various modifications.

k→∞ fk , uk = f, u if either fk → f weakly* and uk → u strongly or

11 This means lim

fk → f strongly and uk → u weakly.

k→∞ liml→∞ fk , ul = f, u and liml→∞ limk→∞ fk , ul = f, u if
12 This means both lim

fk → f weakly* and uk → u weakly.

1.1. Functional analysis 5

Having two locally convex spaces V1 and V2 and an operator A ∈ L (V1 , V2 ),

we define the so-called adjoint operator A∗ ∈ L (V2∗ , V1∗ ) by the identity A∗ f, v =
f, Av to be valid for any v ∈ V1 and f ∈ V2∗ . If V1 and V2 are normed linear
spaces, then A → A∗ realizes an isometrical (i.e. norm-preserving) isomorphism
between L (V1 , V2 ) and L (V2∗ , V1∗ ).
Besides, forgetting that V ∗ has been created by its predual V , one can think
about the weak convergence on V ∗ , induced by the collection of seminorms {f →
|φ, f |}φ∈(V ∗ )∗ . The space V ∗∗ := (V ∗ )∗ is called a bidual to V and the space
V itself is embedded into it by the canonical embedding i : V → V ∗∗ defined
by i(v), f  = f, v. We will often identify13 V with its image i(V ) in V ∗∗ so
that always V ⊂ V ∗∗ , and therefore any weakly convergent sequence in V ∗ is also
weakly* convergent (to the same limit). The converse implication holds if and
only if V ≡ i(V ) = V ∗∗ , at which case the Banach space V is called reflexive. The
Milman-Pettis theorem [285, 335] says that every uniformly convex Banach space
must be reflexive. The Asplund theorem [23] says that any reflexive Banach space
can equivalently be renormed14 so that it is, together with its dual, strictly convex.
By the Clarkson theorem [97], also separable Banach spaces can be renormed to
be strictly convex.15
A mapping A : V1 → V2 is called bounded if it maps bounded sets in V1
into bounded sets in V2 . The normed structure of V1 and V2 allows us to define
a mapping A : V1 → V2 to be Lipschitz continuous if, for some ∈ R and all
u, v ∈ V , it holds that A(u) − A(v) V2 ≤ u − v V1 ; in this case, is called a
Lipschitz constant. Obviously, any Lipschitz continuous mapping is bounded and
uniformly continuous in the sense that A(u) − A(v) V2 ≤ ζ( u − v V1 ) with ζ
increasing, ζ(0) = 0. If ≤ 1 (resp. < 1), A is called non-expansive (resp. a
contraction).
The linear structure of V1 and V2 allows us to investigate smoothness of A.
We say that A : V1 → V2 has the directional derivative at u ∈ V in the direction
h ∈ V , denoted by DA(u, h), if there is the limit

A(u + εh) − A(u)

lim =: DA(u, h). (1.10)
ε0 ε

If the mapping h →DA(u, h) is linear and continuous, then we say that A has a
Gâteaux [173] differential at u ∈ V , denoted by A (u) ∈ L (V1 , V2 ). If the Gâteaux
differential exists in any point, A is called Gâteaux differentiable and A : V →
L (V1 , V2 ). In the special case V2 = R, a Gâteaux-differentiable functional Φ :
V1 → R has the differential Φ : V1 → V1∗ .
13 This identification is indeed very natural because the mapping i : V → V ∗∗ realizes a

(weak,weak*)- as well as (norm,norm)-homeomorphical embedding.

14 Two norms  ·  and  ·  on V are called equivalent to each other if ∃ε > 0 ∀v ∈ V :
1 2
εv1 ≤ v2 ≤ ε−1 v1 .
15 These renormalization results can be still improved for so-called locally uniformly convex

spaces, i.e. δ in (1.5) depends on u; cf. Troyanski [412].

6 Chapter 1. Preliminary general material

1.1.3 Convex sets

A set A in a linear space is called convex if λu + (1−λ)v ∈ A whenever u, v ∈
A and λ ∈ [0, 1], and it is called a cone (with the vertex at the origin 0) if
λv ∈ A whenever v ∈ A and λ ≥ 0. A function f : V → R is called convex if
f (λu + (1−λ)v) ≤ λf (u) + (1−λ)f (v) for any u, v ∈ V , λ ∈ [0, 1]. If u = v and
λ ∈ (0, 1) imply a strict inequality, then f is called strictly convex. A function
f : V → R is convex (resp. lower semicontinuous) if and only if its epigraph
epi(f ) := {(x, a) ∈ V × R; a ≥ f (x)} is a convex (resp. closed) subset of V × R.
Theorem 1.5 (Hahn [196] and Banach [34]). Let K be an open convex nonempty
subset of a locally convex space V and L be a linear manifold that does not intersect
K. Then there is a closed hyperplane L̄ such that L ⊂ L̄ and K ∩ L̄ = ∅. In other
words, there is f ∈ V ∗ such that f, u > f, v whenever u ∈ K and v ∈ L.
Proposition 1.6. A closed set A is convex if and only if 1
2u + 12 v ∈ A whenever
u, v ∈ A. Closed convex sets are weakly closed.
This is related to the fact that a lower semicontinuous functional f is convex
if and only if
1 1 u + u
1 2
f (u1 ) + f (u2 ) ≥ f , (1.11)
2 2 2
cf. Exercise 4.18 below.
Having a convex subset K of a locally convex space V and u ∈ K, we define
the tangent cone TK (u) ⊂ V by

TK (u) := cl a(K−u) . (1.12)
a>0

Obviously, TK (u) is a closed convex cone and v ∈ TK (u) means precisely that
u + ak vk ∈ K for suitable sequences {ak }k∈N ⊂ R and {vk }k∈N ⊂ V such that
limk→∞ vk = v. Besides, we define the normal cone NK (v) as

NK (u) := f ∈ V ∗ ; ∀v ∈ TK (u) : f, v ≤ 0 . (1.13)

Again, the normal cone is always a closed convex cone in V ∗ .

TK (u) TK (u)

NK (u)
NK (u)

Figure 1. Two examples of the tangent and the normal cones to a convex set K⊂R2
≡ (R2 )∗ ; for illustration, the cones are shifted to the respective points u’s.
1.1. Functional analysis 7

1.1.4 Compactness
An important notion on which a lot of powerful tools are based is compactness. Let
us agree, for a certain simplicity16 , to say that a set V in a locally convex space
V is compact if every sequence in A contains a convergent subsequence whose
limit belongs to A. Having in mind a Banach space V (possibly having a predual)
with a structure of the norm (resp. weak or possibly weak*) locally convex space,
we will specify compactness by the adjective “norm” (resp. “weak” or possibly
“weak*”); if no adjective is mentioned, then the “norm” one will implicitly be
understood. A weakening of the notion of compactness is precompactness: we say
that a set A is precompact in the sense that every sequence in such a set contains
a Cauchy subsequence. Another modification is the following: we say that A is
relatively compact if its closure is compact. The reader can easily guess what
e.g. “relatively weakly* compact” or “relatively norm compact” mean. If every
Cauchy sequence converges (in particular in a Banach space), the prefix/adjective
“pre-” and “relatively” coincide with each other. Thus, in a Banach space, relative
norm compactness and norm precompactness are the same.
A mapping A : V1 → V2 , V1 , V2 Banach spaces, is called totally continu-
ous if it is (weak,norm)-continuous, i.e. it maps weakly convergent sequences to
strongly convergent ones. It is called compact if it maps bounded sets in V1 into
precompact17 sets in V2 . If V1 is reflexive, then any totally continuous mapping is
compact18 but not conversely19.
Theorem 1.7 (Selection principle, Banach [35, [Link], Thm.3]20 ). In a
Banach space with a separable predual, any bounded sequence contains a weakly*
convergent subsequence.
Often, Theorem 1.7 is also called, not completely exactly however, Alaoglu-
Bourbaki’s theorem21 .
16 In fact, our “compactness” is in general topology called “sequential compactness” while

compactness means that every net (=a generalized sequence) contains a convergent finer net
(=a suitable generalization of the notion of “subsequence”), or, equivalently, that every covering
of A by open sets contains a finite sub-covering. These two concepts coincide with each other
in a lot of important cases, primarily in normed spaces (considering norm compactness). More
generally, it happens if the structure of a locally convex spaces can be equally induced by a
countable collection of seminorms {| · |ξ }ξ∈N , e.g. it concerns weak (resp. weak*) compactness
if V has a separable dual (resp. predual). A far less trivial fact, known as the Eberlein-Šmuljan
theorem, is that in a Banach space the relatively weakly compact (in the general-topology sense)
sets coincide with the relatively weakly sequentially compact ones.
17 As V is assumed a Banach space, precompact sets are just those which are relatively com-
2
pact.
18 For B ⊂ V bounded, any sequence in A(B), say {A(v )}
1 k k∈N with vk ∈ B contains a
subsequence convergent in V2 , e.g. {A(vkl )}l∈N with {vkl }l∈N weakly convergent in V1 ; here
both reflexivity of V1 guaranteeing existence of {vkl }l∈N and compactness of A guaranteeing
convergence of {A(vkl )}l∈N has been used.
19 See Remark 2.39 and Exercise 2.64 below.
20 For the proof cf. Exercise 2.51.
21 In fact, Alaoglu-Bourbaki’s theorem says that the polar set to a neighbourhood of the origin

in a locally convex space is weakly* compact, see Alaoglu [13] and Bourbaki [63].
8 Chapter 1. Preliminary general material

Theorem 1.8 (Bolzano and Weierstrass22 ). Every lower (resp. upper) semi-
continuous function X → R on a compact set attains its minimum (resp. maxi-
mum) on this set.

1.1.5 Fixed-point theorems

A point u ∈ V is called a fixed point of a mapping M : V → V if M (u) = u.
Theorem 1.9 (Schauder fixed-point theorem [379]). A continuous compact
mapping on a closed, bounded, convex set in a Banach space has a fixed point.
Alternatively, a continuous mapping on a compact, convex, set in a Banach space
has a fixed point.
As a special case for V = Rn , one gets a historically older achievement which
has been tremendously generalized by the previous Theorem 1.9:
Theorem 1.10 (Brouwer fixed-point theorem [72]). A continuous mapping
on a compact convex set in Rn has a fixed point.
Further generalization of Theorem 1.9 is very useful in applications. A useful
concept of a set-valued mapping M : V ⇒ V just means that M : V → 2V , the
set of all subsets of V . Such M is called upper semicontinuous if it has a closed
graph, i.e. uk → u, fk → f , and fk ∈ M (uk ) implies f ∈ M (u).
Theorem 1.11 (Kakutani fixed-point theorem [224]23 ). An upper semicon-
tinuous set-valued mapping M : V ⇒ V with nonempty closed convex values which
maps a nonempty compact convex set K in a locally convex space into itself has a
fixed point, i.e. ∃u ∈ K: u ∈ M (u).
A completely different principle is based on metric properties and exploits
completeness instead of compactness:
Theorem 1.12 (Banach fixed-point theorem [33]24 ). A contractive mapping
on a Banach space has a fixed point which is even unique.

1.2 Function spaces

We consider the Euclidean space Rn , n ≥ 1, endowed with standard Euclidean
topology and for Ω a subset of Rn we will define various spaces of functions Ω →
22 In
fact, this is rather a tremendous generalization of the original assertion by Bolzano [60]
who showed, rather intuitively (because completion of irrational numbers by transcendental ones
which locally compactifies R has been discovered only much later) that any real continuous
function of a bounded closed interval is bounded. An essence, called the Bolzano-Weierstrass
principle, is that every sequence in a compact set has a cluster point, i.e. a point whose each
neighbourhood contains infinitely many members of this sequence.
23 The original version [224] formulated in Rn has been generalized for locally convex spaces

by Fan [139] and Glicksberg [181].

24 In fact, this theorem works (and was formulated in [33]) in a complete metric space, too.
1.2. Function spaces 9

Rm . If endowed by a pointwise addition and multiplication25 the linear-space struc-

ture of Rm is inherited by these spaces. Besides, we will endow them by norms,
which makes them normed linear (or, mostly even Banach) spaces. Having two
such spaces U ⊂ V , we say that the mapping I : U → V : u → u is a continuous
embedding (or, that U is embedded continuously to V ) if the linear operator I is
continuous (hence bounded). This means that u V ≤ N u U ; for N one can take
the norm I L (U,V ) . If I is compact, we speak about a compact embedding and use
the notation U  V . If U is a dense subset in V , we will speak about a dense em-
bedding; this property obviously depends on the norm of V but not of U . It follows
by a general functional-analysis argument that the adjoint mapping I ∗ : V ∗ → U ∗
is continuous and injective provided U ⊂ V continuously and densely26 ; then we
can identify V ∗ as a subset of U ∗ .

1.2.1 Continuous and smooth functions

The notation C(·), C 0 (·), and C 0,1 (·) will indicate sets of continuous, bounded
continuous, and Lipschitz continuous functions, respectively.27 E.g. C 0 (Ω; Rm )
denote the set of bounded continuous functions Ω → Rm . We denote by Ω̄ the
closure of Ω in the Euclidean space Rn . By the Bolzano-Weierstrass Theorem 1.8,
C 0 (Ω; Rm ) = C(Ω̄; Rm ) if Ω is bounded. Equally, we can understand C(Ω̄; Rm ) as
composed from functions Ω → Rm having a continuous extension on the closure Ω̄.
For m = 1, we will write briefly C 0 (Ω̄) (resp. C 0,1 (Ω̄)). If equipped by a pointwise
addition and multiplication and by the norm
 
|u(x) − u(ξ)|
u C 0 (Ω̄;Rm ) := max |u(x)|, u C 0,1 (Ω̄;Rm ) := sup |u(x)| +
x∈Ω̄ x,ξ∈Ω̄ |x − ξ|
x=ξ

with |·| denoting the Euclidean norm in the corresponding spaces, both C 0 (Ω̄; Rm )
and C 0,1 (Ω̄; Rm ) become Banach spaces.
Furthermore, for k ≥ 1, we define spaces of smooth functions, having deriva-
tives up to k-th order continuous up to the boundary, i.e.

C k (Ω̄; Rm ) := u ∈ C 0 (Ω̄; Rm ); ∀(i1 , . . . , in ) ∈ (N ∪ {0})n ,

n
∂ i1 +···+in u 
iα ≤ k : ∈ C 0
(Ω̄; R m
) . (1.14)
α=1
∂ i1 x1 · · · ∂ in xn

If endowed by the linear structure of C 0 (Ω̄; Rm ) and by the norm u C k (Ω̄;Rm ) :=

    i1 +···+in 
u 0 + i1 +···+in ≤k  ∂ i∂1 x1 ···∂ in xn uC 0 (Ω̄;Rm ) , they become Banach spaces.
C (Ω̄;Rm )
25 This means, for u, v : Ω → Rm and λ ∈ R, we define [u + v](x) := u(x) + v(x) and

[λu](x) := λu(x) for all x ∈ Ω̄.

26 Indeed, I ∗ is injective (because two different linear continuous functionals on V must have

also different traces on any dense subset, in particular on U ).

27 For Ω = Rn , the adjectives “continuous” and “Lipschitz continuous” can be understood as

in Section 1.1 because both Rn and Rm are normed spaces, while for a general Ω it is just a
restriction of these definitions on Ω̄.
10 Chapter 1. Preliminary general material

The continuous embedding C k (Ω̄; Rm ) ⊂ C l (Ω̄; Rm ) holds for any k ≥ l ≥ 0,

i.e. u C l(Ω̄;Rm ) ≤ N u C k (Ω̄;Rm ) . In this particular case, N = 1. This embedding is
even dense and, if k > l ≥ 0, also compact. As functions with bounded derivatives
are Lipschitz continuous, we have also C 1 (Ω̄; Rm ) ⊂ C 0,1 (Ω̄; Rm ).
By Riesz’s theorem, the dual space C 0 (Ω̄)∗ is isometrically isomorphic with
the space M (Ω̄) of regular Borel measures on Ω̄; a measure is a σ-additive set
function28 and a regular Borel measure μ ∈ M (Ω̄) is a measure on the so-
called Borel σ-algebra29 which is regular30 and has a finite variation31 |μ| over
∗
Ω̄, i.e. |μ|(
 Ω̄) < +∞. The mentioned isomorphism f → μ : C (Ω̄) → M (Ω̄) is by
0

f, v = Ω̄ v μ(dx).32 Then f C 0(Ω̄)∗ = |μ|(Ω̄). For x ∈ Ω̄, a measure δx ∈ M (Ω̄)

defined by δx , v = v(x) is called Dirac’s measure supported at x.
Considering an increasing sequence of compact subsets Ki ⊂ Ω such that
Ω = i∈N Ki , we put
 
D(Ω) := CKk
i
(Ω), (1.15)
i∈N k∈N

where k
CK i
denotes the space of all functions Ω → R which are continuous
(Ω)
together with all their derivatives
 up to the order k and which have the support
∞ k
contained in Ki . Each CK :=  i  is endowed by the collection of semi-
C (Ω)
 i k∈N K
norms | · |k,Ki k∈N with |u|k,Ki := u|Ki C k (K ) , which makes it a locally convex
 ∞
i
space. Then D(Ω) = i∈N CK i
is equipped with the finest topology that makes
∞
all the embeddings CK i
→ D(Ω) continuous,33 which makes it a locally convex
space. The elements of the dual space D(Ω)∗ are called distributions on Ω.

1.2.2 Lebesgue integrable functions

The n-dimensional outer Lebesgue measure measn (·) on the Euclidean space Rn ,
n ≥ 1, is defined as
∞ 
n ∞
 
measn (A) := inf bki −aki : A ⊂ [ak1 , bk1 ]× · · · ×[akn , bkn ], aki ≤bki (1.16)
k=1 i=1 k=1

and then we call a set A ⊂ R Lebesgue measurable if measn (A) = measn (A∩S)+
n

measn (A \ S) for any subset S ⊂ Rn .34 The collection Σ of Lebesgue measurable

28 This
 
means μ( i∈N Ai ) = i∈N μ(Ai ) for any mutually disjoint Ai from an algebra in
question. 
29 A collection Σ of subsets of Ω̄ is called a σ-algebra if A ∈Σ ⇒
i i∈N Ai ∈Σ, ∅∈Σ, and
A∈Σ ⇒ Ω\A∈Σ. Borel’s σ-algebra is the smallest σ-algebra containing all open subsets of Ω̄.
30 A set function μ is regular if ∀A∈Σ ∀ε>0 ∃A , A ∈Σ: cl(A )⊂A⊂int(A ) and |μ|(A \A )≤ε.
1 2
31 For μ additive, we define the variation |μ| of μ by |μ|(A) = sup
1 2
I 2 1
(A1 ,...,AI )∈M (A) i=1 |μ(Ai )|,
where M (A) denotes the set of all finite collections (A1 , . . . , AI ) of mutually disjoint Ai ∈ Σ
such that Ai ⊂ A for any i = 1, . . . , I.
32 The integral via μ is defined by limit of simple functions similarly as in Section 1.2.2.
33 The convergence f → f in D(Ω) means that ∃K ⊂ Ω compact ∃k ∈ N ∀k ≥ k : supp(f ) ⊂
k 0 0 k
K and fk → f in CK l (Ω) for any l ∈ N; cf. e.g. [425, Sect.I.1].
34 For example, all closed sets are Lebesgue measurable, hence every open set too, as well as

their countable union or intersection, etc.

1.2. Function spaces 11

subsets of Ω forms a so-called σ-algebra35 which, together with the function

measn : Σ → R ∪ {+∞}, have (and are characterized by) the following prop-
erties:
1. A open implies A ∈ Σ,
n
2. A = [a1 , b1 ] × · · · × [an , bn ] with ai ≤ bi implies measn (A) = i=1 (bi − ai ),
  
3. measn is countably additive, i.e. measn k∈N Ak = k∈N measn (Ak ) for any
countable collection {Ak }k∈N of pairwise disjoint sets Ai ∈ Σ,
4. A ⊂ B ∈ Σ and measn (B) = 0 implies A ∈ Σ and measn (A) = 0.
The function measn : Σ → R ∪ {+∞} is called the Lebesgue measure. Having a
set Ω ∈ Σ, we say that a property holds almost everywhere on Ω (in abbreviation
a.e. on Ω) if this property holds everywhere on Ω with the possible exception of a
set of Lebesgue-measure zero; referring to those x where this property holds, we
will also say that it holds at almost all x ∈ Ω (in abbreviation a.a. x ∈ Ω).
A function u : Rn → Rm is called (Lebesgue) measurable if u−1 (A) := {x ∈
Rn ; u(x) ∈ A} is Lebesgue measurable for any A ∈ Rm open.
We call u : Rn → Rm simple if it takes only a finite number of values
vi ∈ Rm and  u−1 (vi ) = {x; u(x) = vi } ∈ Σ; then we define the integral Rn u(x) dx
naturally as finite measn (Ai )vi . Furthermore, a measurable u is called integrable
if there is a sequence of simple functions {uk }k∈N such that limk→∞ uk (x) = u(x)
for a.a. x ∈ Ω and limk→∞ Rn uk (x) dx does exist in R. Then, this limit will
be denoted by Rn u(x) dx and we call it the (Lebesgue) integral of u. It is then
independent of the particular choice of the sequence {uk }k∈N .
We will now consider Ω ⊂ Rn measurable with measn (Ω) < +∞. The notions
of measurability and the integral of functions Ω → Rm can be understood as before
provided all these functions are extended on Rn \ Ω by 0. By Lp (Ω; Rm ) we will
denote the set of all measurable functions36 u : Ω → Rm such that u Lp(Ω;Rm ) <
+∞, where37
⎧ 
⎨ p
Ω |u(x)| dx for 1 ≤ p < +∞ ,
p
u Lp(Ω;Rm ) := (1.17)
⎩ ess sup |u(x)| for p = +∞
x∈Ω

and | · | is the Euclidean norm on Rm . The set Lp (Ω; Rm ), endowed by a pointwise

addition and scalar multiplication, is a linear space. Besides, · Lp(Ω;Rm ) is a norm
35 We call Σ an algebra if ∅ ∈ Σ, A ∈ Σ ⇒ Ω \ A ∈ Σ, and A , A ∈ Σ ⇒ A ∪ A ∈ Σ.
 1 2 1 2
If also Ai ∈ Σ ⇒ i∈N Ai ∈ Σ, then Σ will be called a σ-algebra. An example is the so-called
Borel σ-algebra: the smallest σ-algebra containing all open subsets of Ω. In fact, Σ is the so-
called Lebesgue extension of the Borel σ-algebra, created by adding all subsets of sets having
the measure zero.
36 As usual, we will not distinguish between functions that are equal to each other a.e., so that,

strictly speaking, Lp (Ω; Rm ) contains classes of equivalence of such functions.

37 The “essential supremum” is defined as

ess sup f (x) = inf sup f (x).

N⊂Ω x∈Ω\N
x∈Ω measn (N)=0
12 Chapter 1. Preliminary general material

on Lp (Ω; Rm ) which makes it a Banach space, called a Lebesgue space.

 If a measure μ ∈ M (Ω̄) possesses a density dμ ∈ L1 (Ω), which means μ(A) =
A dμ (x) dx for any measurable A ⊂ Ω, then μ has a certain special property,
namely it is absolutely continuous with respect to the Lebesgue measure38 and
also the converse assertion is true: every absolutely continuous measure possesses
a density belonging to L1 (Ω). This fact is known as the Radon-Nikodým theorem.
An important question is how to characterize concretely the dual spaces. The
“natural”duality pairing comes from the inner product in L2 -spaces, which means
m
u, v := Ω u(x)·v(x) dx, where u·v := i=1 ui vi is the inner product in Rm . If
1 < p < +∞, then Lp (Ω; Rm ) is reflexive. From the algebraic Young inequality
1 p 1 
ab ≤ a +  bp (1.18)
p p
one gets39 the Hölder [208] (or, for p=2, the Hölder-Bunyakovskiı̆ [85]) inequality:40
  
 
u(x)·v(x) dx ≤ p |u(x)| p
dx p 
|v(x)|p dx, (1.19)
Ω Ω
Ω

where p is the so-called conjugate exponent defined as
⎧
⎨ p/(p−1) for 1 < p < +∞,
p := 1 for p = +∞, (1.20)
⎩
+∞ for p = 1.

In fact, the modification of (1.19) for p = 1 or p = +∞ looks trivially as Ω |u ·
v|dx ≤ u L∞ (Ω;Rn ) u L1(Ω;Rn ) . From (1.19), it can be shown that the dual space

is isometrically isomorphic with Lp (Ω; Rm ) if 1 ≤ p < +∞. On the other hand,
the dual space to L∞ (Ω; Rm ) is substantially larger than L1 (Ω; Rm ).41 Applying
the algebraic Young inequality (1.18) to |u(x)·v(x)| ≤ |u(x)| |v(x)| and integrating
it over Ω, we obtain another important inequality, the integral Young inequality
  
 
u(x) · v(x) dx ≤ 1 1
|u(x)|p dx + 

|v(x)|p dx. (1.21)
Ω p Ω p Ω
Instead of (1.18), we will often apply the modified Young inequality
√ p√
p
p−1
ε p
∀ε > 0 : ab ≤ εa + Cε b ,
p
where Cε := . (1.22)
p−1
Moreover, for 1 < p < +∞, the space Lp (Ω; Rm ) is uniformly convex.42
38 This means that ∀ε > 0 ∃δ > 0 ∀A ⊂ Ω measurable: measn (A) ≤ δ =⇒ |μ(A)| ≤ ε.
 −p  −p  p
39 Hint: u−1
Lp (Ω)
v−1
p Ω |u · v|dx ≤ p uLp (Ω) Ω |u| dx + p v p
1 p 1
Ω |u| dx = 1.
L (Ω) L (Ω)
40 Originally,Hölder [208] states this in a less symmetrical form for sums in place of integrals.
41 The elements of L∞ (Ω; Rm )∗ are indeed very abstract objects and can be identified with
finitely-additive measures vanishing on zero-measure sets, see Yosida and Hewitt [426].
42 This result is due to Clarkson [97], see also e.g. Adams [3, Corollary 2.29] or Kufner at

al. [245, Remark 2.17.8].

1.2. Function spaces 13

If measn (Ω) < +∞ and 1 ≤ q ≤ p ≤ +∞, the embeddings C 0 (Ω̄; Rm ) ⊂

L (Ω; Rm ) ⊂ Lq (Ω; Rm ) are continuous43 . Moreover, for p < +∞ these embed-
p

dings are dense and then, as C 0 (Ω̄; Rm ) is separable, Lp (Ω; Rm ) is separable, too.
On the other hand, L∞ (Ω) is not separable.44
Hölder’s inequality also allows for an interpolation between Lp1 (Ω) and
L (Ω): for p1 , p2 , p ∈ [1, +∞], λ ∈ [0, 1], it holds that45
p2

1 λ 1−λ
= + ⇒ v Lp(Ω) ≤ v λLp1 (Ω) v 1−λ
Lp2 (Ω) . (1.23)
p p1 p2
Moreover, if p is as in (1.23), p1 ≤ p2 , and similarly q −1 = λq1−1 + (1−λ)q2−1 ,
q1 ≤ q2 , and A is a bounded linear operator Lp1 (Ω) → Lq1 (Ω) whose restriction
on Lp2 (Ω) belongs to L (Lp2 (Ω), Lq2 (Ω)), then
   λ  1−λ
A ≤ C AL (Lp1 (Ω),Lq1 (Ω)) AL (Lp2 (Ω),Lq2 (Ω)) (1.24)
L (Lp (Ω),Lq (Ω))

for some constant C = C(p1 , p2 , q1 , q2 , λ).

We say that a sequence uk : Ω → Rm converges in measure to u if
 
∀ε > 0 : lim measn {x ∈ Ω; |uk (x)−u(x)| ≥ ε} = 0. (1.25)
k→∞

Naturally, the convergence a.e. means that uk (x) → u(x) for a.a. x ∈ Ω. Let us em-
phasize that convergence in measure does not imply convergence a.e.46 . Anyhow:
Proposition 1.13 (Various modes of convergences).
(i) Any sequence converging a.e. converges also in measure.
(ii) Any sequence converging in measure admits a subsequence converging a.e.
(iii) Any sequence converging in L1 (Ω) converges in measure.
Theorem 1.14 (Lebesgue [254]). Let {uk }k∈N ⊂ L1 (Ω) be a sequence converging
a.e. to some
 u and |uk (x)|
 ≤ v(x) for some v ∈ L (Ω). Then u lives in L (Ω) and
1 1

limk→∞ A uk (x) dx = A u(x) dx for any A ⊂ Ω measurable.

Theorem 1.15 (Fatou [143]). Let {uk }k∈N  ⊂ L1 (Ω) be a sequence of non-
47
negative functions such that lim inf k→∞ Ω uk (x) dx < +∞. Then the function
x → lim inf k→∞ uk (x) is integrable and
  
lim inf uk (x) dx ≥ lim inf uk (x) dx. (1.26)
k→∞ Ω Ω k→∞
43 Cf. Exercise 2.68 below.
44 To see it, consider Ω = (0, 1) and a collection {χ(0,a) }a∈(0,1) of characteristic functions of an
interval (0, a), i.e. χ(0,a) (x) = 1 for x ∈ (0, a) and χ(0,a) (x) = 0 for x ∈ [a, 1). This collection is
an uncountable subset of the unit sphere in L∞ (0, 1) and χ(0,a) − χ(0,b) L∞ (0,1) = 1 for a = b,
hence L∞ (0, 1) cannot be separable.
45 Cf. Exercise 2.59 below. See e.g. [268, p.26].
46 An example for a sequence converging in the measure on [0, 1] to 0 but not a.e. is u
(l,m)
a characteristic function of the interval of the type [(m − 1)2−l , m2−l ] for 1 ≤ m ≤ 2l , ar-
ranged lexicographically as (l, m) = (1, 1), (1, 2), (2, 1), (2, 2), (2, 3), (2, 4), (3, 1), . . .. Selection of
a subsequence converging a.e. to 0 can be done, e.g., by taking m = 1 only.
47 Obviously, existence of a common integrable minorant can weaken this assumption.
14 Chapter 1. Preliminary general material

In particular, Theorem 1.14 says that the set {uk ; k ∈ N} is relatively weakly
compact48 in L1 (Ω). This property is related to both equi-absolute continuity and
uniform integrability:
Theorem 1.16 (Dunford and Pettis [128]). Let M ⊂ L1 (Ω; Rm ) be bounded.
Then the following statements are equivalent to each other:
(i) M is relatively weakly compact in L1 (Ω; Rm ),
(ii) the set M is uniformly integrable, which means:

∀ε > 0 ∃K ∈ R+ : sup |u(x)| dx ≤ ε , (1.27)
u∈M {x∈Ω;|u(x)|≥K}

(iii) the set M is equi-absolutely-continuous, which means:


∀ε > 0 ∃δ > 0 : sup sup |u(x)| dx ≤ ε . (1.28)
u∈M |A|≤δ A

Useful generalizations of the Lebesgue dominated-convergence Theorem 1.14

and the Fatou Theorem 1.15 are:
Theorem 1.17 (Vitali [421]49 ). Let {uk }k∈N ⊂ L1 (Ω) be a sequence converging
a.e. to some u. Then u ∈ Lp (Ω) and uk → u in Lp (Ω) if and only if {|uk |p }k∈N is
uniformly integrable.
Theorem 1.18 (Fatou, generalized50 ). The conclusion of Theorem 1.15 holds
if uk ≥ 0 is replaced by uk ≥ vk with {vk }k∈N being uniformly integrable.
Theorem 1.19 (Fubini [160]). Considering two Lebesgue measurable sets Ω1 ⊂
Rn1 and Ω2 ⊂ Rn2 , the following identity holds provided g ∈ L1 (Ω1 × Ω2 ) (in
particular, each of the following double-integrals does exist and is finite):
      
g(x1 , x2 ) d(x1 , x2 ) = g(x1 , x2 ) dx2 dx1 = g(x1 , x2 ) dx1 dx2 .
Ω1 Ω2 Ω2 Ω1
Ω1 ×Ω2

1.2.3 Sobolev spaces

Modern theory of differential equations is based on spaces of functions whose
derivatives exist in a generalized sense and enjoy a suitable integrability. Those
spaces, developed since the works by Beppo Levi [258], Leray [256], Sobolev [389],
48 Note that the linear hull of all characteristic functions χ with A ⊂ Ω measurable is dense
A
in L∞ (Ω) ∼= L1 (Ω)∗ , so that the sequence {uk }, being bounded in L1 (Ω), converges weakly in
L1 (Ω) and, as such, it is relatively sequentially weakly compact, hence by the Eberlein-Šmuljan
theorem relatively weakly compact, too.
49 More precisely, in [421], the integration of summable series is investigated rather than mere

sequences.
50 See Ash [21, Thm.7.5.2], or also Klei and Miyara [233] or Saadoune and Valadier [377].
1.2. Function spaces 15

and Tonelli [406], have thus an absolute importance in this context and their
theory is presently very broad. We present here only a rather minimal extent; for
more detailed exposition we refer to Adams [3], Adams, Fournier [4], Giusti [180],
Kufner, Fučı́k, John [245], Maz’ya [277], or Ziemer [430].
Having a function u ∈ Lp (Ω), we define its distributional derivative
∂ u/∂xk11 · · · ∂xknn with k1 + · · · +kn = k and ki ≥ 0 for any i = 1, . . . , n as a
k

distribution such that

 
∂ku ∂kg
∀g ∈ D(Ω) : , g = (−1)k
u k1 dx. (1.29)
∂x1 · · · ∂xn
k1 kn
Ω ∂x1 · · · ∂xknn
 ∂ ∂

The n-tuple of the first-order distributional derivatives ∂x 1
u, . . . , ∂x n
u is de-
noted by ∇u and called a gradient of u. We will now consider Ω ⊂ Rn open with
measn (Ω) < +∞; such a set will be called a domain. For p < +∞, we define a
Sobolev space [390]

W 1,p (Ω) := u ∈ Lp (Ω); ∇u ∈ Lp (Ω; Rn ) , equipped with the norm (1.30a)
⎧
⎪
⎪ u Lp(Ω)
p p
+ ∇u pLp(Ω;Rn )
⎪
⎪ 
⎨   p  p
u W 1,p (Ω) := = p Ω u(x) + ∇u(x) dx if p < +∞, (1.30b)
⎪
⎪  
⎪
⎩ max u L∞ (Ω) , ∇u L∞(Ω;Rn )
⎪ 
= ess supx∈Ω max |u(x)|, |∇u(x)| if p = +∞.

As, by Rademacher’s theorem, Lipschitz functions are a.e. differentiable, it holds

that W 1,∞ (Ω) = C 0,1 (Ω).
Analogously, for k > 1 integer, we define
 k
W k,p (Ω) := u ∈ Lp (Ω; Rm ); ∇k u ∈ Lp (Ω; Rn ) , (1.31)

where ∇k u denotes the set of all k-th order partial derivatives of u understood
in the distributional sense. The standard norm on W k,p (Ω) is u W k,p(Ω) =
 1/p
u pLp(Ω) + ∇k u p p nk
, which makes it a Banach space. Likewise for
L (Ω;R )
Lebesgue spaces, for 1 ≤ p < +∞ the Sobolev spaces W k,p (Ω; Rm ) are separa-
ble and, if 1 < p < +∞, they are uniformly convex,51 hence by Milman-Pettis’
theorem also reflexive.
The spaces of Rm -valued functions, W k,p (Ω; Rm ), are defined analogously.52
To give a good sense to traces on the boundary Γ := ∂Ω := Ω̄ \ Ω, we must
qualify Ω suitably. We say that Ω is a Lipschitz domain if there is a finite number
of overlapping parts Γi of the boundary of Γ and corresponding coordinate systems
(i.e. transformation unitary matrices Ai and open sets Gi ∈ Rn−1 ) such that each
51 See Adams [3, Theorem 3.5].
52 This means W k,p (Ω; Rm ) := {(u1 , . . . , um ); ui ∈ W k,p (Ω)}.
16 Chapter 1. Preliminary general material

Γi can be expressed as a graph of a Lipschitz function gi ∈ C 0,1 (Rn−1 ) in the

sense that
 
Γi = Ai ξ; ξ ∈ Rn , (ξ1 , . . . , ξn−1 ) ∈ Gi , ξn = gi (ξ1 , . . . , ξn−1 ) (1.32)

and Ω lies on one side of Γ in the sense that A i ξ; ξ ∈ Rn , (ξ1 , . . . , ξn−1 ) ∈
Gi , gi (ξ1 , . . . , ξn−1 )−ε < ξn < gi (ξ1 , . . . , ξn−1 ) ⊂ Ω and simultaneously

Ai ξ; ξ ∈ Rn , (ξ1 , . . . , ξn−1 ) ∈ Gi , gi (ξ1 , . . . , ξn−1 ) < ξn < gi (ξ1 , . . . , ξn−1 )+ε ⊂
Rn \ Ω̄ for some ε> 0. For an example of a Lipschitz domain Ω ⊂ R2 , see Figure 2a
where three indicated coordinate systems are sufficient to cover Γ, and Figure 2b
where this condition fails in a variety of ways.

Ω
Ω

Fig. 2. a) Example of a Lipschitz do- b) Examples of a 2D and a 3D domains

main whose boundary can which are not Lipschitz because of 3 and
be covered by three charts. 4 spots, respectively (marked by circles).

If the mappings gi belong to C k,α (Rn−1 ), we say that the domain Ω is of C k,α -
class. Lipschitz domains are then naturally called of the C 0,1 -class. We will always
assume Ω to be Lipschitz and bounded.
Theorem 1.20 (Sobolev embedding [389]). The continuous embedding
∗
W 1,p (Ω) ⊂ Lp (Ω) (1.33)

holds provided the exponent p∗ is defined as

⎧ np
⎪
⎨ for p < n ,
n−p
p∗ := an arbitrarily large real for p = n , (1.34)
⎪
⎩
+∞ for p > n .

Theorem 1.21 (Rellich, Kondrachov [348, 236]53 ). The compact embedding

∗
−
W 1,p (Ω)  Lp (Ω) ,  ∈ (0, p∗ −1], (1.35)

holds for p∗ from (1.34).

53 The pioneering Rellich’s work deals with p = 2 only.
1.2. Function spaces 17

For higher-order Sobolev spaces, one gets, e.g., W 2,p (Ω) ⊂ W 1,np/(n−p) (Ω)
by applying Theorem 1.20 on first derivatives, and applying Theorem 1.20 once
again for np/(n − p) instead of p one comes to W 1,np/(n−p) (Ω) ⊂ Lnp/(n−2p) (Ω)
provided 2p < n. It identifies p∗∗ := (p∗ )∗ as np/(n−2p) if 2p < n. Proceeding
further by induction, one comes to:
Corollary 1.22 (Higher-order Sobolev embedding54 ).
(i) If kp < n, the continuous embedding W k,p (Ω) ⊂ Lnp/(n−kp) (Ω) and the com-
pact embedding W k,p (Ω)  Lnp/(n−kp)− (Ω) hold for any  ∈ (0, n−kp
np
− 1].
(ii) For kp = n, it holds that W k,p (Ω)  Lq (Ω) for any q < +∞.
(iii) For kp > n, it holds that W k,p (Ω)  C(Ω̄).
Theorem 1.23 (Trace operator55 ). There is exactly one linear continuous op-
erator T : W 1,p (Ω) → L1 (Γ) such that, for any u ∈ C 1 (Ω̄), it holds that T u = u|Γ
(=the restriction of u on Γ). Moreover, T remains continuous (resp. is compact)
as the mappings
#
u → u|Γ : W 1,p (Ω) → Lp (Γ), resp., (1.36a)
#
−
u → u|Γ : W 1,p (Ω) → Lp (Γ) ,  ∈ (0, p# −1], (1.36b)

provided the exponent p# is defined as

⎧ np − p
⎪
⎨ for p < n ,
# n−p
p := (1.37)
⎪
⎩ an arbitrarily large real for p = n ,
+∞ for p > n .
We will call the operator T from Theorem 1.23 the trace operator, and
write simply u|Γ instead of T u even if u ∈ W 1,p (Ω) \ C 1 (Ω̄). Then we de-
fine W01,p (Ω) := {v ∈ W 1,p (Ω); v|Γ = 0}. For k > 1, we define similarly
i
W0k,p (Ω) := {v ∈ W k,p (Ω); ∇i v ∈ W0k−i,p (Ω; Rn ), i = 0, . . . , k − 1}.
Another useful result allows for a certain interpolation between the Sobolev
space W k,p (Ω) and the Lebesgue space Lq (Ω):
Theorem 1.24 (Gagliardo-Nirenberg inequality [161, 312, 313]). Let β =
β1 + · · · + βn , β1 , . . . , βn ∈ N ∪ {0}, k ∈ N, r, q, and p satisfy
1 β 1 k 1 β
= +λ − + (1 − λ) , ≤ λ ≤ 1, 0 ≤ β ≤ k − 1, (1.38)
r n p n q k
54 In fact, the point (iii) can be improved. E.g. for k = 1 and Ω smooth, one has W 1,p (Ω) ⊂

C 0,α (Ω̄) with α = 1 − n/p, which is referred to a Morrey inequality. In general, W k,p (Ω) ⊂
C k−[n/p]−1,α (Ω̄) with α < 1 if n/p ∈ N or α = [n/p] + 1 − n/p if n/p ∈ N.
55 In fact, u → u| 1,p (Ω) → W 1−1/p,p (Γ), where the Sobolev-Slobodeckiı̆ space
Γ : W
W 1−1/p,p (Γ) is defined as in (1.42) but on an (n−1)-dimensional manifold Γ instead of
n-dimensional domain Ω. Then, similarly as in Theorem 1.20, we have the embedding
# #
W 1−1/p,p (Γ) ⊂ Lp (Γ), resp. W 1−1/p,p (Γ)  Lp − (Γ). For the rather less standard nota-
#
tion of the exponent p , see e.g. Ciarlet [93, Thm 6.1-7].
18 Chapter 1. Preliminary general material

then it holds
 
 ∂β v   λ  1−λ
  ≤ CGN v W k,p (Ω) v Lq (Ω) , (1.39)
 β1 βn 
∂x1 · · · ∂x1 Lr (Ω)

provided k − β − n/p is not a negative integer (otherwise it holds for λ = |β|/k).

For example, if n = 2, then
  
v L4 (Ω) ≤ CGN v L2 (Ω) v L2 (Ω) + ∇v L2 (Ω;R2 ) . (1.40)

There is another definition

 of Sobolev
 spaces, not so explicit as that one we
used. Namely, Wk,p (Ω) := cl C ∞ (Ω̄) where “cl” stands for the closure in, say,
Lp (Ω) with respect to the W k,p -norm introduced above; for k = 1 see (1.30b).
 ∞  ∞
Similarly, Wk,p
0 (Ω) := cl C0 (Ω) where C0 (Ω) stands for the set of infinitely
smooth functions with a compact support in Ω. For the following important as-
sertion an important (here permanently accepted) assumption is that Ω has a
Lipschitz boundary.
Theorem 1.25 (Density of smooth functions). It holds that

W k,p (Ω) = Wk,p (Ω), W0k,p (Ω) = Wk,p

0 (Ω). (1.41)

A generalization of Sobolev spaces for k ≥ 0 a non-integer is often useful for

various finer investigations:
 
|∇[k] u(x) − ∇[k] u(ξ)|p
W k,p (Ω) := u ∈ W [k],p (Ω); dxdξ < +∞ , (1.42)
Ω×Ω |x − ξ|n+p(k−[k])

where [k] denotes the integer part of k. For k non-integer, W k,p (Ω) is called the
Sobolev-Slobodeckiı̆ space. They are Banach spaces if normed by the norm
  1/p
|∇[k] u(x) − ∇[k] u(ξ)|p
u W k,p (Ω) := u W [k],p(Ω) +
p
dxdξ . (1.43)
Ω×Ω |x − ξ|n+p(k−[k])
In fact, Corollary 1.22 holds for k ≥ 0 non-integer, too. A complete interpolation
theory holds for the Hilbertian case p = 2 for general k, k1 , and k2 non-negative:
   λ  1−λ
k = λk1 + (1−λ)k2 =⇒ uW k,p (Ω) ≤ C uW k1 ,p (Ω) uW k2 ,p (Ω) (1.44)

with some C = C(k1 , k2 , λ), while for p = 2, the inequality (1.44) holds provided
k is not integer and k1 , k2 ∈ N.56 Like in (1.24),
   λ  1−λ
A ≤ C A A
L (W k,p l,p
(Ω),W (Ω)) L (W k1 ,p (Ω),W l 1 ,p (Ω)) L (W k2 ,p (Ω),W l1 ,p (Ω))
(1.45)
56 For k ∈ N the Sobolev space W k,p (Ω) is to be replaced by a so-called Besov space.
1.3. Nemytskiı̆ mappings 19
 
for some constant C = C(k1 , k2 , l1 , l2 , λ) provided A ∈ L W k1 ,2 (Ω), W l1 ,2 (Ω) ∩
 k ,2 
L W 2 (Ω), W l2 ,2 (Ω) , k is as in (1.44), and l−1 = λl1−1 + (1−λ)l2−1 .
The Sobolev-Slobodeckiı̆ spaces are sometimes well fitted with some sophis-
ticated nonlinear estimates, e.g.:
Theorem 1.26 (Málek, Pražák, Steinhauer [269]57 ). If p ≥ 2 and  > 0, then
there is c > 0 such that
   1/p

  p
c u W 2/p−,p (Ω) ≤ u Lp (Ω) + |u|p−2 |∇u|2 dx . (1.46)
Ω

1.3 Nemytskiı̆ mappings

Considering integers j, m0 , m1 , . . . , mj , we say that a mapping a : Ω× Rm1 × · · ·×
Rmj → Rm0 is a Carathéodory mapping if a(·, r1 , . . . , rj ) : Ω → Rm0 is measurable
for all (r1 , . . . , rj ) ∈ Rm1 × · · · × Rmj and a(x, ·) : Rm1 × · · · × Rmj → Rm0
is continuous for a.a. x ∈ Ω. Then the so-called Nemytskiı̆ mappings Na map
functions ui : Ω → Rmi , i = 1, . . . , j, to a function Na (u1 , . . . , uj ) : Ω → Rm0
defined by
! "  
Na (u1 , . . . , uj ) (x) = a x, u1 (x), . . . , uj (x) . (1.47)

Theorem 1.27 (Nemytskiı̆ mappings in Lebesgue spaces58 ). If a : Ω × Rm1 ×

· · · × Rmj → Rm0 is a Carathéodory mapping and the functions ui : Ω → Rmi ,
i = 1, . . . , j, are measurable, then Na (u1 , . . . , uj ) is measurable. Moreover, if a
satisfies the growth condition

  
j
 pi /p0
a(x, r1 , . . . , rj ) ≤ γ(x) + C ri  for some γ ∈ Lp0 (Ω), (1.48)
i=1

with 1 ≤ pi < +∞, 1 ≤ p0 < +∞, then Na is a bounded continuous mapping

Lp1 (Ω; Rm1 ) × · · · × Lpj (Ω; Rmj ) → Lp0 (Ω; Rm0 ). If some pi = +∞, i = 1, . . . , j,
the same holds if the respective term |·|pi /p0 is replaced by any continuous function.
The case p0 = +∞ has to be excluded, cf. Exercise 2.61. Also, it should be
emphasized that Na cannot be weakly continuous unless a(x, ·) is affine for a.a. x ∈
Ω; cf. Exercise 2.62. A “canonical” example is a sequence uk (x) := sign(sin(kπx)),
Ω = (0, 1) ⊂ R1 , which converges weakly (or weakly*) to 0 in any Lp (0, 1), but for
a(x, r) = |r| one obviously gets (cf. Figure 3) that w-lim a(uk ) = lim 1 = 1 =
  k→∞ k→∞
0 = a(0) = a w-lim uk .
k→∞

57 In fact, [269] works with periodic functions having zero means. The proof relies on the

so-called Nikolskiı̆ spaces. Here, we present an obvious modification involving the Lp -norm in
(1.46).
58 Cf. Exercise 2.60 below.
20 Chapter 1. Preliminary general material

uk 1/k
Nemystkiı̆ |uk |
mapping
uk → |uk |
weak (here even
weak strong) convergence
u convergence
Nemystkiı̆
mapping
0 1 u → |u| 0 1

Fig. 3. A counterexample for weak continuity of nonlinear Nemytskiı̆ mappings.

Sometimes, it is useful to consider Nemytskiı̆ mappings in a Sobolev space.
Here we confine ourselves to a special case of a one-argument Nemytskiı̆ mapping
with an integrand independent of x; in fact, it is then a usual superposition.
Proposition 1.28 (Superposition operator in Sobolev spaces59 ). If a : R →
R is Lipschitz continuous, then a(u) ∈ W 1,p (Ω) for any u ∈ W 1,p (Ω), p ∈ [1, +∞],
and moreover it holds that
∇a(u) = a (u)∇u (a.e. on Ω). (1.49)
Denoting u+ = max(u, 0) and u− = min(u, 0), Proposition 1.28 yields60
∇u if u > 0, 0 if u ≥ 0,
∇(u+ ) = ∇(u− ) = (a.e. on Ω). (1.50)
0 if u ≤ 0, ∇u if u < 0,
Another useful assertion addresses the change of the order of an integra-
tion and a differentiation, and is based on the Lebesgue dominated convergence
Theorem 1.14.
Theorem 1.29 (Change of integration and differentiation). Let (x, r) →
ϕ : Ω×R → R and its partial derivative ∂r ∂
ϕ : Ω×R → R be Carathéodory
functions, |ϕ(·, 0)| ∈ L (Ω) and the collection
1
 { ∂r
∂
ϕ(·, r)}|r|≤ε have a common
integrable majorant. Then Φ : r → Ω ϕ(x, r) dx is differentiable at 0 and
d
 ∂
dt Φ(0) = Ω ∂r ϕ(x, 0) dx.

1.4 Green formula and some inequalities

Let us define the unit outward normal ν = ν(x) ∈ Rn to the boundary Γ at a
point x ∈ Γ as the vector
 
∂gi ∂gi
Ai ,..., ,1
∂ξ1 ∂ξn−1
ν(x) = #   ∂g 2 (1.51)
 ∂gi 2  i 
  + ···+   +1
∂ξ1 ∂ξn−1
59 See, e.g., monographs by Cazenave and Haraux [90, Prop. 1.3.5] or Ziemer [430, Theo-

rem 2.1.11]. Original results are, in particular, due to Marcus and Mizel [273, 274].
60 See, e.g., Cazenave and Haraux [90, Corollary 1.3.6] or Ziemer [430, Corollary 2.1.8].
1.4. Green formula and some inequalities 21
 
where x = Ai ξ = Ai ξ1 , . . . , ξn−1 , gi (ξ1 , . . . , ξn−1 ) as in (1.32). It is again natural
to assume Ω a Lipschitz domain. Then the functions gi are Lipschitz continuous
and, by the so-called Rademacher theorem [342], they possess derivatives almost
everywhere on Γi ⊂ Rn−1 , cf. (1.32), hence ν is defined almost everywhere on
Γ and does not depend on the concrete covering of Γ by the coordinate  systems
(Gi , Ai ). For a function f : Γ → R, we can define the surface integral Γ f (x) dS
through (n − 1)-dimensional Lebesgue measure as
   
f (x) dS = f Ai (ξ1 , . . . , ξn−1 , gi (ξ1 , . . . , ξn−1 )
Γ i G̃i
$
 ∂g 2  ∂g 2
 i  i 
×   + ···+   + 1 d(ξ1 , . . . , ξn−1 ) (1.52)
∂ξ1 ∂ξn−1

where G̃i ⊂ Gi are chosen suitably to realize, instead of the overlapping

 covering
as on Figure 2, a disjoint covering of Γ. Again, the value Γ f (x) dS does not
depend on the particular coordinate systems (G̃i , Ai ).
Theorem 1.30 (Multidimensional by-part integration). If v ∈ W 1,p (Ω) and

z ∈ W 1,p (Ω), then
  
∂z ∂v
∀i = 1, . . . , n : v + z dx = vz νi dS. (1.53)
Ω ∂xi ∂xi Γ

Considering z = (z1 , . . . , zn ), writing (1.53) for zi instead

 of z, and eventually
summing it over i = 1, . . . , n with abbreviating div z := ni=1 ∂x ∂
i
zi the divergence
of the vector field z, we arrive atthe formula which we will often use:
Theorem 1.31 (Green’s formula [188]61 ). For any v ∈ W 1,p (Ω) and z ∈

W 1,p (Ω; Rn ), the following formula holds:
 
 
v (div z) + z ·∇v dx = v (z ·ν) dS. (1.54)
Ω Γ

Theorem 1.32 (Poincaré-type inequalities [339]). Let 1 ≤ q ≤ p∗ . Then

there is CP < +∞ such that
 
u W 1,p (Ω) ≤ CP ∇u Lp(Ω;Rn ) + u Lq (Ω) . (1.55)

Let 1≤q≤p# , let Ω be connected 62 , and let ΓD , ΓN ⊂ Γ be such that measn−1 (ΓD ) >
0 and measn−1 (ΓN ) > 0. Then there is CP < +∞ such that
   
u W 1,p (Ω) ≤ CP ∇u Lp(Ω;Rn ) + u|ΓN Lq (Γ ) (1.56)
N

 
61 Putting z = ∇u into (1.54), we get Ω vΔu + ∇v · ∇u dx = ∂
Γ v ∂ν u dSderived in [188]. In
fact, (1.54) holds, by continuous extension, even under weaker assumptions, cf. Nečas [302].
62 This means that, for any x , x ∈ Ω, there is z : [0, 1] → Ω continuous with z(0) = x and
0 1 0
z(1) = x1 . It has the consequence that functions with zero gradient must be constant on Ω.
22 Chapter 1. Preliminary general material

and
u|ΓD = 0 ⇒ u W 1,p (Ω) ≤ CP ∇u Lp(Ω;Rn ) , (1.57)
and also   
     
u 1,p  
≤ CP ∇u Lp (Ω;Rn ) +  u dx . (1.58)
W (Ω)
Ω

A special case of (1.56) with ΓD = Γ and p = q = 2, is sometimes also called

Friedrichs’ inequality [157].
Theorem 1.33 (Korn inequality [237]63 ). Let 1 < p < +∞. There is a constant
CK = CK (p, Ω) such that for any v ∈ W01,p (Ω) it holds that

    ∇v + (∇v)
v  1,p ≤ CK e(v)Lp (Ω;Rn×n ) where e(v) = . (1.59)
W 0 (Ω;Rn ) 2

1.5 Bochner spaces

We will now define spaces of abstract functions on a bounded interval I ⊂ R
valued in a Banach space V , invented by Bochner [59], which is a basic tool on
the abstract level for Part II. We say that u : I → V is simple if it takes only
finite number of values vi ∈ V and Ai := u−1 (vi ) is Lebesgue measurable; then
T 
0 u(t) dt := finite meas1 (Ai )vi . We say that u : I → V is Bochner measurable
if it is a point-wise limit (in the strong topology) of V of a sequence {uk }k∈N
of simple functions; i.e. uk (t) → u(t) for a.a. t ∈ I. We say that u : I → V is
K
absolutely continuous64 if, for each ε > 0, there is δ > 0 such that k=1 u(tk ) −
u(sk ) V ≤ ε whenever tk−1 ≤ sk ≤ tk ≤ T for k = 1, . . . , K ∈ N, t0 = 0,
K
and k=1 tk − sk ≤ δ. A point t ∈ I is called a Lebesgue point of u : I → V if
 h/2
limh0 h1 −h/2 u(t+ϑ) − u(t) dϑ = 0. Analogously, a right Lebesgue point t ∈ I
h
means limh0 h1 0 u(t+ϑ) − u(t) dϑ = 0.
Theorem 1.34 (Pettis [334]65 ). If V is separable, then u is Bochner measurable
if and only if it is weakly measurable in the sense: v ∗ , u(·) is Lebesgue measurable
for any v ∗ ∈ V ∗ .
Considering simple functions {uk }k∈N as above, we call u : I → V
T T
Bochner integrable if limk→∞ 0 u(t) − uk (t) V dt = 0. Then 0 u(t) dt :=
T
limk→∞ 0 uk (t) dt; this limit exists and is independent of the particular choice
63 See Nečas [307] or monographs [130, 308, 427]. A counterexample for p=1 is by Ornstein

[318].
64 If V = R1 , this definition naturally coincides with the absolute-continuity with respect to

the Lebesgue measure on I used on p. 12.

65 In fact, [334] works with a general Banach space, showing equivalence of the Bochner mea-

surable mappings with a.e. separably valued weakly measurable ones. See Example 1.42 for a
weakly* continuous function u valued in V = L∞ (0, 1) which is not Bochner measurable.
1.5. Bochner spaces 23

of the sequence {uk }k∈N . Moreover, if V is separable, then a Bochner measur-

able function u is Bochner integrable if and only if u(·) V is Lebesgue inte-
T T
grable. Then also  0 u(t) dt V ≤ 0 u(t) V dt. From this, we can see that
 h/2
limh0 h1 −h/2 u(t+ϑ)dϑ → u(t) at each Lebesgue point t ∈ I.66

Theorem 1.35. If u ∈ L1 (I; V ), then a.e. t ∈ I is a Lebesgue point for u.

An analogous assertion holds for right Lebesgue points.

An example for Bochner integrable functions is any u ∈ C(I; (V, weak)).67
However, some classical “scalar-valued” results need not hold in vectorial cases68 .
For 1 ≤ p < ∞, a Bochner space Lp (I; V ) is the linear space (of classes with
respect to equivalence a.e.) of Bochner integrable functions u : I → V satisfying
T
0
u(t) pV dt < +∞. This space is a Banach space if endowed with the norm

 T 1/p

u Lp(I;V ) := u(t)p dt . (1.60)
V
0

For p = ∞, we modify it standardly, i.e. u L∞ (I;V ) := ess supt∈I u(t) V . Later,

we will often use partition of I to subintervals of the length τ := 2−K T , K ∈ N.

Proposition 1.36 (Density of piecewise constant functions). If 1 ≤ p <

+∞, then the set {v : I → V ; ∃K ∈ N : ∀1 ≤ k ≤ 2K : v|((k−1)τ,kτ ) is constant,
τ = 2−k T } is dense in Lp (I; V ). In particular, if p ∈ [1, +∞) and V is separable,
Lp (I; V ) is separable too.

Proposition 1.37 (Uniform convexity). If V is uniformly convex and 1 < p <

+∞, then Lp (I; V ) is uniformly convex, too.

Proposition 1.38 (Dual space). If p ∈ [1, +∞), the dual space to Lp (I; V ) always

contains Lp (I; V ∗ ) and the equality holds if V ∗ is separable, the duality pairing
being given by the formula

% & T% &
f, u Lp(I;V ∗ )×Lp (I;V ) := f (t), u(t) V ∗ ×V dt . (1.61)
0

Thus, if p ∈ (1, +∞) and V is reflexive and separable, then Lp (I; V ) is reflexive.

 h/2  h/2
66 This follows simply from u(t) − 1
h −h/2
u(t+ϑ)dϑV =  h1 −h/2
u(t+ϑ) − u(t)dϑV ≤
 h/2
1
h −h/2
u(t+ϑ) − u(t)V dϑ.
67 The weak* continuity would not be sufficient, however, as Example 1.42 shows.
68 E.g., u ∈ C(I; (V, weak∗)) need not be Bochner integrable, as Example 1.42 shows. Also, a

 mapping u : I → V need not be represented by an integrable

σ-additive absolutely continuous
function v in the sense u(t) = 0t v(ϑ)dϑ; for the counterexample see Yosida [425, Sect.V.5].
24 Chapter 1. Preliminary general material

Theorem 1.39 (Komura [235]). If V is reflexive and u : I → V is absolutely

t
continuous, then u is (strongly) differentiable a.e.69 and u(t) = u(0) + 0 u , u ∈
L1 (I; V ).70
Convention 1.40. Often, we will use V = Lq (Ω; Rm ). Considering p, q ∈ [1, +∞],
it is then natural to identify Lp (I; Lq (Ω; Rm )) with
 T 
 q p/q
ũ : I × Ω → R ;
m  
ũ(t, x) dx dt < +∞ (1.62)
0 Ω

through the natural isomorphism ũ → u: [u(t)](x) := ũ(t, x). Likewise,

Lp (I; W 1,q (Ω)) is identified via this isomorphism with functions ũ : I × Ω → R for
 T  p/q
Ω |ũ(t, x)| + |∇ũ(t, x)| dx
q q
which 0 dt < +∞.
Proposition 1.41 (Interpolation of Lp1 (I; Lq1 (Ω)) and Lp2 (I; Lq2 (Ω)) 71 ). Let
p1 , p2 , q1 , q2 ∈ [1, +∞], λ ∈ [0, 1], and v ∈ Lp1 (I; Lq1 (Ω)) ∩ Lp2 (I; Lq2 (Ω)). Then
1 λ 1−λ 1 λ 1−λ
= + and = +
p p1 p2 q q1 q2
   λ  1−λ
=⇒ v  
≤ v p q
v  p . (1.63)
Lp (I;Lq (Ω)) L 1 (I;L 1 (Ω)) L 2 (I;Lq2 (Ω))

Example 1.42. (The space L∞ (I; L∞ (Ω)) is not L∞ (Q).72 ) Using the isomorphism
from Convention 1.40, we can identify abstract functions from L∞ (I; L∞ (Ω)) with
functions on Q := I × Ω. However, L∞ (I; L∞ (Ω)) = L∞ (Q). For Ω = I, the
function u(t) = χ[0,t] induces a function ũ(t, x) = 1 if x ≤ t and = 0 if x > t which
obviously belongs to L∞ (Q). Even u is weakly* continuous but it is not Bochner
measurable, hence u ∈ L∞ (I; L∞ (Ω)).
Considering Banach spaces V0 , V1 , . . . , Vk and a : I × V1 × · · · × Vk → V0 ,
let us still define the Nemytskiı̆ mappings Na again by the formula (1.47). The
following generalization of Theorem 1.27 holds:
Theorem 1.43 (Nemytskiı̆ mappings in Bochner spaces [265]73 ). Let V0 , V1 ,
. . . , Vk be separable Banach spaces, a : I × V1 × · · · × Vk → V0 a Carathéodory
mapping74 and the growth condition
  
k
 pi /p0
a(t, r1 , . . . , rk ) ≤ γ(t) + C ri  for some γ ∈ Lp0 (I), (1.64)
V0 Vi
i=1

ε→0 ε u(t + ε) − ε u(t) does exist for a.a. t ∈ I. The reflexivity of X is indeed
69 This means lim 1 1

necessary, e.g. u : I → X := L∞ (0, 1) defined by [u(t)](x) := x sin(t/x) is Lipschitz continuous

but nowhere differentiable.
70 If V is a uniformly convex space, this formula has been proved already by Clarkson [97].
71 Cf. Exercise 8.60 on p.266.
72 This observation is due to Fattorini [144, Example 5.0.10].
73 Besides the original paper [265] by Lucchetti and Patrone, we refer also, e.g., to Hu and

Papageorgiou [209, Part I, Sect.9.1].

74 Here, it means that, for a.a. t ∈ I, a(t, ·) : V ×· · ·×V → V is to be (norm,norm)-continuous
1 k 0
and, as in Section 1.3, a(·, r1 , . . . , rk ) : I → V0 is to be Bochner measurable.
1.6. Some ordinary differential equations 25

hold with p0 , p1 , . . . , pk as for (1.48). Then Na maps Lp1 (I; V1 ) × · · · × Lpk (I; Vk )
continuously into Lp0 (I; V0 ).

1.6 Some ordinary differential equations

As an auxiliary problem, we will often use the initial-value problem for the system
of k ordinary differential equations in the form:

du  
= f t, u(t) for a.a. t ∈ I, u(0) = u0 , (1.65)
dt

where f : (0, +∞) × Rk → Rk is a Carathéodory mapping. By a solution on a time

interval [0, T ] we will understand an absolutely continuous mapping u : [0, T ] → Rk
such that the equation (1.65) holds a.e. on [0, T ] and u(0) = u0 holds, too.

Theorem 1.44 (Local-in-time existence). Let f : (0, +∞) × Rk → Rk be a

Carathéodory mapping. Then there is T > 0 (not given a-priori) such that the
initial-value problem has a solution on the interval [0, T ].

The main ingredient for estimation of evolution systems in general is the

so-called Gronwall inequality,75 which we will also often use. In the general form,
this inequality says that, for all t ≥ 0,
  t θ t
y(t) ≤ C+ b(θ)e− 0
a(ϑ)dϑ
dθ e 0 a(θ)dθ (1.66)
0

t
whenever we know that y(t) ≤ C + 0 (a(ϑ)y(ϑ) + b(ϑ))dϑ for some a, b ≥ 0 inte-
t
grable.76 For a ≥ 0 constant, (1.66) simplifies to y(t) ≤ eat (C + 0 b(ϑ)e−aϑ dϑ) ≤
t
(C + 0 b(ϑ)dϑ)eaT for t ∈ I.

Theorem 1.45 (Existence and uniqueness). Let T be fixed and f : I ×Rk → Rk

be a Carathéodory mapping satisfying the growth condition |f (t, r)| ≤ γ(t) + C|r|
with some γ ∈ L1 (I). Then:
(i) The initial-value problem (1.65) has a solution u ∈ W 1,1 (I; Rk ) on the interval
I = [0, T ].
(ii) If f (t, ·) is also Lipschitz continuous in the sense |f (t, r1 )−f (t, r2 )| ≤ (t)|r1 −
r2 | with some ∈ L1 (I), then the solution is unique.

75 In the general form presented here, which can be found, e.g., in Ioffe and Tihomirov [211,

Sect. 9.1, Lemma 3], it is also called the Bellman-Gronwall inequality according to the original
works [193] (for C = 0, a, b constant) and [41] (for C ≥ 0, b = 0, a ∈ L1 (0, T )).
76 A generalization for slightly superlinear growth in y, namely y ln(y) or y ln(y) ln(ln(y)), is

possible, too.
26 Chapter 1. Preliminary general material

Applying the so-called bootstrap argument, i.e. knowing that u ∈ W 1,1 (I; Rk )
and therefore f (t, u) ∈ W 1,1 (I; Rk ) provided f is smooth in the sense

∀ρ ∈ R+ ∃γρ ∈ L1 (I), Cρ < +∞ ∀t ∈ I, |r| ≤ ρ :

 ∂f 
 
 (t, r) ≤ γρ (t) and |f (t, r)| ≤ Cρ , (1.67)
∂t
we know immediately that d
dt u = f (t, u) ∈ W 1,1 (I; Rk ).77 Hence:
Theorem 1.46 (Regularity). If f satisfies all assumptions in Theorem 1.45 and
is also smooth in the sense (1.67), then u ∈ W 2,1 (I; Rk ).
The following discrete version of the Gronwall inequality will be often used:78
 
l−1  
l−1
yl ≤ C+τ bk eτ k=1 ak (1.68)
k=1
l−1
provided yl ≤ C + τ k=1 (ak yk + bk ) for any l ≥ 0 (of course, for l = 0 it means
yl ≤ C). We will often use ak ≡ a constant, and the condition


l
 
yl ≤ C + τ ayk + bk , (1.69)
k=1
l−1
from which we can easily derive yl ≤ (1 − aτ )−1 (C + τ b1 + τ k=1 (ayk + bk+1 )),
so that (1.68) gives
 
l 
eτ la/(1−aτ ) 1
yl ≤ C+τ bk if τ < . (1.70)
1 − aτ a
k=1

There is a variant of the discrete Gronwall inequality by Nochetto, Savaré

and Verdi [314] designed for fine estimates for variable time steps and even bet-
ter fit to discretisation of some evolutionary problems. While the condition (1.69)
is an analog of dt d
y ≤ ay + b, the modified version is based rather on the con-
tinuous condition y dt79 d
y ≤ ay+b+cy 2 : If non-negative sequences {yk }K
k=0 and
{(zk , ak , bk )}k=1 satisfy
K 80

yk − yk−1
yk + zk ≤ ak yk + bk + cyk2 for k = 1, . . . , K (1.71a)
τ
77 We use d f (t, u) = ∂ f (t, u)+ ∂ f (t, u) d u ∈ L1 (I; Rk ) because ∂
f (t, u) ∈ L1 (I; Rk ) while
dt ∂t ∂r dt ∂t
| ∂r f (t, u)| ≤  and dt
∂ d
u = f (t, u) ∈ L∞ (I; Rk ).
78 See Lees [255], Krejčı́ and Roche [241, Lemma A.1], Quarteroni and Valli [341, Lemma 1.4.2]
or Thomée [405].
79 Note that 1 d y 2 = y d y ≤ ay + b + cy 2 ≤ ( 1 a + c)y 2 + 1 a + b, from which the boundedness
2 dt dt 2 2
of y follows by the classical Gronwall inequality (1.66) provided a, b, and c are integrable.
80 More specifically, this is (up to some scaling) a special case of [314, Lemma 4.12] for c

independent of k.
1.6. Some ordinary differential equations 27

with a constant c ≥ 0, then one has both maxk=1,...,K e−kcτ /(1−cτ )yk and
K
(2τ k=1 e−((2k−1)cτ )/(1−cτ ) zk )1/2 estimated from above by


K
√ 
K
(y02 + 2τ e−(2k−1)cτ bk )1/2 + 2τ e−(k−1)cτ ak
k=1 k=1

provided 0 < τ < 1/c. A particular but well applicable case can be obtained for
K = T /τ : assuming (1.71a) and also
1
0 < τ ≤ τ0 with τ0 < , (1.71b)
c
one has, with some constant CT independent of τ ,

 T /τ 1/2   T /τ 1/2 
T /τ 
max yk + τ zk ≤ CT y0 + τ bk +τ ak . (1.72)
k=1,...,K
k=1 k=1 k=1