Você está na página 1de 33

Introduction to Functional Analysis

Goetz Grammel
February 4, 2005

Contents
1 Introduction

2 The One-dimensional Laplace Equation

3 Basic Functional Analysis


3.1 Linear Vector Spaces . .
3.2 Linear Forms . . . . . .
3.3 Inner Products . . . . .
3.4 Norms . . . . . . . . . .
4 Convergence, Continuity
4.1 Convergence . . . . . .
4.2 Continuity . . . . . . .
4.3 Completeness . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

4
4
7
9
10

and Completeness
13
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

5 Sobolev Spaces and Variational Equalities


19
5.1 Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5.2 Variational Equalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.3 Application to Linear Second Order Boundary Value Problems . . . . . . . 23
6 Approximation of Solutions
6.1 Taylor Series . . . . . . . .
6.2 Splines . . . . . . . . . . .
6.3 Fourier Series . . . . . . .
6.4 Homogenization . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

Bibliography

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

27
27
28
29
29
33

Introduction

In computational mechanics one solves problems involving partial differential equations.


Frequently those differential equations cannot be solved analytically and one has to find
approximate solutions. The standard way to compute approximate solutions is the socalled finite element method. It relies on characterizing the solutions of partial differential
equations by related minimization problems, which in turn can be described by variational
equalities. Then the approximate solutions are minimizers within a prescribed finitedimensional subspace. The purpose of this course is to develop the mathematical tools
and skills needed to understand the abstract framework related to the finite element
method.

The One-dimensional Laplace Equation

Given two real numbers a and b, find a twice continuously differentiable function u :
[0, 1] R with the boundary conditions u(0) = a, u(1) = b, such that the Laplace
equation
u00 (x) = 0
is fulfilled for all x (0, 1). Multiplying both sides of the Laplace equation with a
continuously differentiable function w : [0, 1] R and integrating leads to the variational
system
Z
1

u00 (x)w(x)dx = 0.

Integration by parts yields


Z 1

u0 (x)w0 (x)dx + u0 (1)w(1) u0 (0)w(0) = 0.


0

If we only take those functions w, whose values at the boundary points vanish, we finally
obtain the variational system
Z 1
u0 (x)w0 (x)dx = 0.
0

We claim that the solution to the Laplace equation minimizes the functional
Z 1
F (u) =
u0 (x)2 dx.
0

To this end we note that all function of the type u + w satisfy the boundary conditions.
We calculate
Z 1
Z 1
Z 1
Z 1
Z 1
0
0
2
0
2
0
0
0
2
(u (x) + w (x)) dx =
u (x) dx + 2
u (x)w (x)dx +
w (x) dx
u0 (x)2 dx
0

for all continuously differentiable functions w : [0, 1] R with w(0) = w(1) = 0.

Basic Functional Analysis

In this chapter we focus on the algebraic side of functional analysis and present the
algebraic structures needed to formulate finite element methods in a mathematical way.

3.1

Linear Vector Spaces

Definition 3.1 (Linear spaces) A real linear space (vector space, linear vector space)
is a set V together with two operations
(i) Addition + : V V V , (u, v) 7 u + v,
(ii) Scalar Multiplication : R V V , (, u) 7 u,
with the following properties.
u+v = v+u
u + (v + w) = (u + v) + w
( + ) u = u + u
(u + v) = u + v
() u = ( u)
1u = u
for all u, v, w V and , R. Furthermore, there is an element 0 V with u + 0 = u
for all u V , and for all u V there exists an element w V with u + w = 0. We write
w = u.
Remark 3.2 In any real vector space V , the equality
0u=0
is valid for all u V . Indeed, we can calculate
0 u = (0 + 0) u = 0 u + 0 u
and adding the expression (0 u) to both sides, we have 0 = 0 u. Consequently, we can
write
u = (1) u
for all u V , since
(1) u + u = (1) u + 1 u = (1 + 1) u = 0 u = 0.
Example 3.3 The smallest linear space is given by V = {0}.

Example 3.4 For any natural number n, the set Rn with the addition
u + v = (u1 , u2 , ..., un ) + (v1 , v2 , ..., vn ) = (u1 + v1 , u2 + v2 , ..., un + vn )
and the scalar multiplication
u = (u1 , u2 , ..., un ) = (u1 , u2 , ..., un )
is a real linear vector space.
Example 3.5 Let Rn . We write A(; R) for the set of all functions from to R.
We define the addition of two functions f, g A(; R) by
(f + g)(x) = f (x) + g(x)
for all x . A scalar multiplication is obtained by setting
( f )(x) = f (x)
for all x . Using the fact that the range R is a linear space, it is easy to show that
A(; R) is a linear space as well.
Lemma 3.6 Let V be a linear space and W V . Then W is a linear space (a subspace
of V ), if and only if the following two conditions are satisfied.
(i) w1 + w2 W for all w1 , w2 W ,
(ii) w W for all R, w W .
Example 3.7 The following sets are subspaces of A(; R).
B(; R), the set of all bounded functions from to R.
C(; R), the set of all continuous functions from to R.
C k (; R), the set of all k-times continuously differentiable functions from to R.
L1 (; R), the set of all integrable functions from to R.
Example 3.8 Let L2 (; R) be the set of all square integrable functions, i.e. of all functions f : R with
Z
f (x)2 dx < .

L (; R) is a linear subspace of A(; R). Recall that for all real numbers a, b R, the
inequality
1
1
ab a2 + b2
2
2
is valid. This follows from
0 (a b)2 = a2 2ab + b2 .
5

Let f, g L2 (; R). We show that f + g L2 (; R). To this end we apply the inequality
above to a = |f (x)|, b = |g(x)|. Then we obtain
Z
Z
Z
Z
2
2
2
(f (x) + g(x)) dx
f (x) dx + g(x) dx + 2 |f (x)||g(x)|dx

Z
2 f (x)2 dx + 2 g(x)2 dx <

Example 3.9 The set of polynomials over R is a linear space. For 0 n m let
n
X

p=

ai x ,

q=

m
X

i=0

bi x i

i=0

be two real polynomials. We define the addition by


p+q =

n
X

(ai + bi )x +

i=0

m
X

bi x i

i=n+1

and the scalar multiplication by


n
X

p=

(ai )xi .

i=0

Example 3.10 For fixed n 0, the set of real polynomials of the form
p=

n
X

ai x i

i=0

is a linear space as well.


Remark 3.11 Depending on the domain , some of the subspaces of A(; R) are contained in others. For instance, for = [0, 1] we have the inclusions
C k ([0, 1], R) C([0, 1], R) B([0, 1], R) L2 ([0, 1], R) L1 ([0, 1], R) A([0, 1], R).
Definition 3.12 (Basis) Let V be a vector space. A subset B V is called a basis of
V , if it has the following two properties.
(i) B is a generating system for V , i.e. for all v V there are k N, {b1 , . . . , bk } B,
{a1 , . . . , ak } R such that we can represent v as a linear combination
v=

k
X

ai b i .

i=1

(ii) B consists of linearly independent vectors, i.e. the representation


0=

k
X
i=1

implies that a1 = a2 = = ak = 0.
6

ai b i

Definition 3.13 (Dimension) Let V be a linear space and B be a basis of V . The


cardinality of B is called the dimension of V .
Example 3.14 The smallest possible linear space {0} has dimension dim({0}) = 0, since
its basis is the empty set.
Example 3.15 The Euclidean space R3 has the basis
B = {(1, 0, 0), (0, 1, 0), (0, 0, 1)}.
Accordingly, dim(R3 ) = 3.
Example 3.16 The Euclidean space Rn has the basis
B = {(1, 0, . . . , 0), (0, 1, 0, . . . , 0), . . . , (0, . . . , 0, 1)}.
Accordingly, dim(Rn ) = n.
Example 3.17 Let P be the linear space of polynomials (of arbitrary degree). Then
B = {1, x1 , x2 , . . .}
is a basis of P . Accordingly dim(P ) = .

3.2

Linear Forms

Definition 3.18 (Mapping) Given two sets X, Y , a mapping f : X Y assigns to any


x X an element f (x) Y . The set X is called the domain of definition, the set Y is
the range of f .
Definition 3.19 (Linear mapping) Let V , W be a real linear vector spaces. A mapping L : V W is called a linear if
L(u + v) = L(u) + L(v)
L( u) = L(u)
for all u, v V and R.
Example 3.20 The simplest linear mapping L : V W is the zero mapping given by
L(v) = 0 for all v V .
Example 3.21 For the particular case that V = W = R, any linear mapping L : V W
is of the form L(v) = v, where R.

Remark 3.22 L(0) = 0 for any linear mapping L on V . Indeed, we can calculate
L(0) = L(0 0) = 0 L(0) = 0.
Remark 3.23 A linear mapping L : V W is completely determined by its values on a
basis B of V , since we can write
!
k
k
k
X
X
X
L(v) = L
i bi =
L(i bi ) =
i L(bi )
i=1

for any v V , where v =


v.

Pk

i=1

i=1

i=1

i bi is the unique (w.r.t. a fixed basis) representation of

Definition 3.24 (Kernel, image) Let L : V W be linear. The set


Ker(L) := {v V : L(v) = 0}
is called the kernel of L. The set
Im(L) := {w W : w = L(v) for a v V }
is called the image of L.
Remark 3.25 The smaller the kernel, the larger the image. For instance, if V = Rn one
can show that
dim(Ker(L)) + dim(Im(L)) = n = dim(Rn )
for any linear mapping L : V W .
Definition 3.26 (Linear forms) Let V be a real linear vector space. A linear mapping
L : V R is called a linear form.
Example 3.27 L : C([0, 1], R) R defined by L(f ) = f (0) is a linear form on V =
C([0, 1], R).
Example 3.28 L : C([0, 1], R) R defined by
Z 1
L(f ) =
f (x)dx
0

is a linear form on C([0, 1], R).


Example 3.29 L : C 1 ([0, 1], R) R defined by L(f ) = f 0 ( 21 ) is a linear form on
V = C 1 ([0, 1], R).
Example 3.30 Let y = (y1 , y2 , ..., yn ) Rn be fixed. Ly : Rn R defined by
Ly (x) =

n
X
i=0

is a linear form on V = Rn .
8

xi yi

3.3

Inner Products

Definition 3.31 (Bilinear forms) Let V and W be two linear spaces. A mapping a :
V W R is a bilinear form if the mappings
aw : V R,

u 7 aw (u) := a(u, w),

av : W R,

u 7 av (u) := a(v, u),

and
are linear forms for all v V and w W .
Example 3.32 Let V = R3 and W = R2 . Then
a(v, w) := v1 w1 + v2 w2
is a bilinear form.
Definition 3.33 (Inner products) Let V be a linear vector space. An inner product
on V is a bilinear form a : V V R such that
a(v, v) 0
for all v V (positivity), a(v, v) = 0 if and only if v = 0 (definiteness),
a(v, w) = a(w, v)
for all v, w V (symmetry).
A linear space which is endowed with an inner product is called an inner product space.
Remark 3.34 For inner products on V we write a(v, w) = hv, wi for u, v V .
Example 3.35 On the linear vector space V = Rn , the bilinear form a(v, w) = hv, wi =
Pn
i=0 vi wi is called the Euclidean inner product.
Example 3.36 On the linear vector space V = C([0, 1]; R), the bilinear form
Z 1
f (x)g(x)dx
a(f, g) :=
0

defines an inner product.


Remark 3.37 Let V be an inner product space. Then the equality hv, 0i = 0 is valid for
all v V . Indeed, we can calculate
hv, 0i = hv, 0 vi = 0hv, vi = 0
9

Theorem 3.38 (Schwarz inequality) Let V be an inner product space. Then the inequality
hu, vi2 hu, uihv, vi
is valid for all u, v V .
Proof. For v = 0, the claim is trivial. Let v 6= 0. Then we can calculate


hu, vi
hu, vi
0
u
v, u
v
hv, vi
hv, vi

 

hu, vi
hu, vi
hu, vi2 hv, vi
= hu, ui u,
v
v, u +
hv, vi
hv, vi
hv, vi2
hu, vi
hu, vi2
hu, vi +
hv, vi
hv, vi
2
hu, vi
= hu, ui
hv, vi
= hu, ui 2

and the claim easily follows.

3.4

Norms

Definition 3.39 (Norms) Let V be a linear vector space. A function


k k : V R,

v 7 kvk

is called a norm on V if
kvk 0 for all v V (positivity),
kvk = 0 if and only if v = 0 (definiteness),
k vk = ||kvk for all R and v V (positive homogeneity),
ku + vk kuk + kvk for all u, v V (triangle inequality).
A linear space which is endowed with a norm is called a normed space.
Example 3.40 The absolute value function x 7 |x| defines a norm in R.
Example 3.41 On Rn a norm is given by
kx|1 := |x1 | + + |xn |.
Example 3.42 On Rn a norm is given by
kx|2 :=

|x1 |2 + + |xn |2 .

This norm is called the Euclidean norm.


10

Example 3.43 On C([0, 1]; R) we obtain a norm by setting


kf k := max |f (x)|.
x[0,1]

This norm is called the maximum norm.


Example 3.44 On C 1 ([0, 1]; R) we obtain a norm by setting
kf kC 1 := max |f (x)| + max |f 0 (x)|.
x[0,1]

x[0,1]

Theorem 3.45 (Natural norms) Let V be a linear vector space with an inner product
h, i. Then we obtain a norm on V by setting
1

kvk := hv, vi 2 .
This norm is called the natural norm on an inner product space.
1

Proof. As for the positivity we immediately obtain kvk = hv, vi 2 0, since hv, vi 0
for all v V . The definiteness follows from the definiteness of the inner product. The
positive homogeneity is obtained as follows.
1
1
1
1
k vk = h v, vi 2 = 2 hv, vi 2 = (2 ) 2 hv, vi 2 = || kvk.
As for the triangle inequality, we calculate
ku + vk2 = hu + v, u + vi
= hu, ui + hu, vi + hv, uihv, vi
= hu, ui + 2hu, vi + hv, vi,
and using the Schwarz inequality we obtain
ku + vk2 kuk2 + 2kukkvk + kvk2
= (kuk + kvk)2 .
2

Hence ku + vk kuk + kvk.

Theorem 3.46 (Parallelogram law) Let V be an inner product space. Then the parallelogram law
ku + vk2 + ku vk2 = 2kuk2 + 2kvk2
is valid for all u, v V .
Conversely, If V is a normed space such that the parallelogram law is valid for all u, v V ,
then the norm is a natural norm of an inner product, i. e. there is an inner product on
1
V such that kuk = (hu, ui) 2 for all u V .
11

Example 3.47 The Euclidean norm on Rn given by


v
u n
uX
kxk := t
x2i
i=1

is the natural norm corresponding to the Euclidean inner product on Rn .


Example 3.48 (Energy norms) We set V := {u C([0, 1]; R) : u(0) = u(1) = 0}.
Obviously V is a linear vector space. On V we define a bilinear form by
Z 1
a : V V R, a(u, v) =
u0 (x)v 0 (x)dx.
0

The bilinear form a is an inner product on V . Indeed, we immediately obtain the positivity
Z 1
a(u, u) =
(u0 (x))2 dx 0
0

R1
for all u V . As for the definiteness we note that a(u, u) = 0 if and only if 0 (u0 (x))2 dx =
0. Since u0 C([0, 1]; R), we conclude that u0 (x) = 0 for all x [0, 1]. Since the domain
[0, 1] is an interval, there is a constant c R with u(x) = c for all x [0, 1]. By the
definition of V , we have c = u(0) = 0, and hence u = 0. The symmetry is obvious. The
natural norm
 12
Z 1
0
0
kuka :=
u (x)u (x)dx
0

induced by a is called the energy norm on V .


Example 3.49 L2 ([0, 1]; R) is an inner product space with inner product
Z 1
hf, gi :=
f (x)g(x)dx.
0

The corresponding natural norm defined by


Z 1
1/2
2
kf kL2 :=
|f (x)| dx
0

is called the L2 -norm. Here, the bilinearity, positivity and symmetry are easy to show. In
contrast to the space of continuous functions, definiteness, in the usual sense, does not
hold, since there are many square integrable functions f with
Z 1
f (x)2 dx = 0.
0

However, there is an artificial mathematical procedure to obtain definiteness. We put all


those functions with square integral equal to zero into one class, the zero-class, and moreover say that two square integrable functions f, g are equivalent, if their difference f g is
an element of this zero-class. In this way we obtain equivalence classes of functions. So,
we do not talk about single functions any more, but about equivalence classes of functions.
A single function can be viewed as a representative of its equivalence class.
12

Convergence, Continuity and Completeness

In this chapter, we present the notions that are necessary in order to understand the
analytical aspects of functional analysis.

4.1

Convergence

Definition 4.1 (Convergence) Let V be a normed space and (vn )nN be a sequence in
V . We say that the sequence vn converges to a limit v V , and write
vn v,

as n ,

if
kvn vk 0,

as n .

Hence, the norm allows us to reduce the convergence in a vector space to a convergence
of real numbers. However, note that the convergence of a given sequence in a linear space
depends on the norm we have chosen. For instance, since we have the inclusions
C([0, 1]; R) L2 ([0, 1]; R) L1 ([0, 1]; R),
we can equip the linear space of real valued continuous functions on [0, 1] with at least
three different norms. The following theorem tells us that the maximum norm is nicely
adapted to C([0, 1]; R).
Theorem 4.2 Let (fn )nN be a sequence in C([0, 1]; R) and f B([0, 1]; R) with
kfn f k 0,

as n .

Then the function f is continuous as well.


Proof. Let x [0, 1] be arbitrary (but fixed) and  > 0. We take an arbitrary sequence
(xk )kN in [0, 1] with xk x, as k . We choose n N large enough such that

kfn f k < ,
3
and, for this particular n N, we choose k0 N large enough such that

|fn (xk ) fn (x)| <
3
for all k k0 . Then we obtain (using the triangle inequality) the estimation
|f (xk ) f (x)| |f (xk ) fn (xk )| + |fn (xk ) fn (x)| + |fn (x) f (x)|



<
+ + = ,
3 3 3
for all k k0 . In other words, we have the convergence f (xk ) f (x), as k , and
f is continuous at x [0, 1]. Since x [0, 1] was arbitrarily chosen, the function f is
continuous on [0, 1].
2

13

Example 4.3 Consider the sequence fn C([0, 2]; R) given by


(
xn for 0 x < 1
fn (x) :=
.
1 for 1 x 2
For any fixed x [0, 2] we obtain a real sequence fn (x), which converges to
(
0 for 0 x < 1
f (x) :=
.
1 for 1 x 2
Note that the function x 7 f (x) is not continuous on [0, 2], since it is not continuous at
x = 1. But f is square integrable. We easily show that
kfn f kL2 0,
Indeed, we calculate
Z 2
Z
2
|fn (x) f (x)| dx =
0

as n .

x2n dx =

1
0,
2n + 1

as n .

Hence, if we chose the L2 -norm in C([0, 2]; R), the sequence fn is convergent. But if we
chose the maximum norm in C([0, 2]; R), the sequence fn does not converge, because the
limit function f is not continuous.
Note that in the previous example the linear space considered is infinite-dimensional. In
contrast, in finite-dimensional linear vector spaces the situation is less complex. Namely,
if a sequence in a finite-dimensional space does converge with respect to one particular
norm, then it is convergent with respect to any other norm as well. This very useful fact
is an immediate consequence of the following theorem.
Definition 4.4 Let V be a linear space and k k1 , k k2 be two norms on V . We say that
the norm k k1 is stronger than the norm k k2 , if there is a real constant C > 0 such that
kvk2 Ckvk1
for all v V .
When the norm k k1 is stronger than the norm k k2 , a sequence vn in V converges w.r.t.
the norm k k2 whenever it is convergent w.r.t. the norm k k1 .
Definition 4.5 (Equivalence of norms) Let k k1 and k k2 be two norms on a linear
space X. We say that both norms are equivalent, if there is a real constant C > 0 such
that we can estimate
kxk1 Ckxk2 and kxk2 Ckxk1
for all x X.
14

Remark 4.6 On the n-dimensional space Rn all norms are equivalent.


We are not going to prove this result, but present a simple example.
Example 4.7 Consider the two-dimensional space R2 with the two norms
p
kxk1 = |x1 | + |x2 | and kxk2 = x21 + x22 .
A short look at the corresponding unit spheres tells us that

kxk2 kxk1 and kxk1 2kxk2

for all x R2 . Hence, we can chose C = 2 and see that both norms are equivalent.

4.2

Continuity

Definition 4.8 (Continuity) Let V, W be normed spaces and V . A function f :


W is continuous at v , if f (vn ) f (v), as n , for any sequence (vn )nN in
with vn v, as n . A mapping f : W is continuous, if it is continuous at
any v .
Proposition 4.9 (Continuity of linear mappings) Let V, W be normed linear spaces
and L : V W be a linear mapping. Then L is continuous if and only if it is continuous
at 0 V .
Proof. The only if part is obviuos, we only show the if part. Let L be continuous
at 0 V and let v V . We take a sequence (vn )nN in V with vn v, as n . Then
we can write
kLvn Lvk = kL(vn v)k 0, as n ,
2

since (vn v) 0, as n .

Example 4.10 Let V = C 1 ([0, 1]; R) and W = C([0, 1]; R). We define a mapping D :
V W by f 7 Df = f 0 . Clearly, D is a linear mapping by the rules of differentiation. It
turns out that the continuity of D strongly depends on the norms we choose. For instance,
D is not continuous, if we equip both spaces V, W with the maximum norm. This can be
seen as follows. Consider the sequence (fn )nN in V given by
fn (x) =

sin(nx)
n

for all x [0, 1]. Then we immediately calculate that kfn k 0, as n , but
kDfn k = max0x1 |cos(nx)| = 1 for all n N.
In contrast, D becomes continuous, if we equip only W with the maximum norm and take
for f V the stronger norm
kf kC 1 = kf k + kf 0 k .
15

Indeed, we easily show that D is continuous at 0 V . Let fn be a sequence in V with


kfn kC 1 0, as n inf ty. Then we have kDfn k = kfn0 k kfn k + kfn0 k = kfn kC 1
for all n N, and hence Dfn 0 in W , as n .
Proposition 4.11 (Continuity of the inner product) Let V be an inner product space.
Then the mapping
P : V V R,

(v1 , v1 ) 7 P (v1 , v2 ) := hv1 , v2 i

is continuous, where the product space V V is equipped with the norm


k(v1 , v2 )kV V =

p
kv1 k2 + kv2 k2 .

Proof. For v1 , v2 , w1 , w2 V we can estimate


|P (v1 , v2 ) P (w1 , w2 )| = |hv1 , v2 i hw1 , w2 i|
|hv1 , v2 i hw1 , v2 i| + |hw1 , v2 i hw1 , w2 i|
= |hv1 w1 , v2 i| + |hw1 , v2 w2 i|
kv1 w1 kkv2 k + kw1 kkv2 w2 k
L (kv1 w1 k + kv2 w2 k)
p
2L kv1 w1 k2 + kv2 w2 k2
= 2Lk(v1 , v2 ) (w1 , w2 )kV V ,
2

where L = max (kv2 k, kw1 k).

Definition 4.12 (Bounded linear mappings) Let V, W be normed linear spaces with
norms k kV , k kW . We say that a linear mapping L : V W is bounded, if there exists
a constant C 0 such that
kL(v)kW CkvkV
for all v V .
Proposition 4.13 Let V, W be normed linear spaces and L : V W be a linear mapping.
Then L is continuous if and only if it is bounded.
Proof. Let L be bounded. We show that L is continuous at the origin. Let vn be a
sequence in V with vn 0, as n . Then we can estimate
kL(vn )kW Ckvn kV
for all n N. Hence, L(vn ) 0, as n , and L is continuous at 0 V .
16

As for the converse, let L be continuous. Assume that L is not bounded, then there is a
sequence vn in V \ {0} with
kL(vn )kW
>n
kvn kV
for all n N. Hence, for wn :=

vn
nkvn kV

we obtain wn 0, as n , but

kL(wn )kW > 1


for all n N, a contradiction to the continuity of L. Thus, our assumption was wrong
and L is bounded.
2

Proposition 4.14 Let V, W be normed linear spaces and a : V W R be a bilinear


mapping. We equip the product space V W with the norm
k(v, w)kV W = kvkV + kwkW .
Then a is continuous if and only if there is a constant C 0 such that the inequality
|a(v, w)| CkvkV kwkW
it valid for all v V , w W .

4.3

Completeness

Definition 4.15 (Cauchy sequence) Let V be a normed space. A sequence (vn )nN is
called a Cauchy sequence, if for any real  > 0 there is a number n0 N such that
kvn vm k < 
for all n, m n0 .
Proposition 4.16 Let V be a normed space and (vn )nN a convergent sequence in V with
limit v V . Then (vn )nN is a Cauchy sequence.
Proof. For  > 0 there is an n0 N such that
kvn vk <


2

for all n n0 . Hence we can estimate


kvn vm k kvn vk + kv vm k <



+ =
2 2
2

for all n, m n0 .

17

Remark 4.17 The converse of the Proposition 4.16 above is wrong, i.e. it may happen
that a Cauchy sequence does not converge towards an element of the normed space considered. Recall, for instance, Example 4.3. There, the sequence fn converges w.r.t. the
L2 -norm to an element of L2 ((0, 2); R). According to Proposition 4.16, fn necessarily is a
Cauchy sequence w.r.t. the L2 -norm. However, the sequence fn does not converge to an
element of the linear space C([0, 2]; R), which has been originally considered.
Definition 4.18 (Completeness) Let V be a normed space. V is called complete, if
any Cauchy sequence in V is convergent in V , i.e. if any Cauchy sequence in V has a
limit in V .
According to our considerations above, the linear space C([0, 2]; R) cannot be complete
w.r.t. the L2 -norm. One could also say that it is not large enough to contain all limits of
Cauchy sequences.
Normally, one tries to verify the convergence of a sequence by showing that the distance
to the limit is tending to zero. However, this procedure requires that the limit is already
known. So, convergence within complete spaces is much easier to show, since one only has
to check whether the distance between the elements of the sequence is tending to zero,
withput referring to a limit. Fortunately, many important infinite-dimensional function
spaces are complete, as well as all finite-dimensional spaces.
Definition 4.19 (Banach space, Hilbert space) A complete normed space is called a
Banach space. An inner product space, which is complete w.r.t. the natural norm induced
by the inner product, is called a Hilbert space.
Example 4.20 The finite-dimensional Euclidean space Rn is a Hilbert space with the
P
inner product hx, yi = ni=1 xi yi .
Example 4.21 The linear space of continuous functions, C([a, b]; R), endowed with the
maximum norm k k is a Banach space.
Example 4.22 The linear space of continuously differentiable functions, C 1 ([a, b]; R),
endowed with the C 1 -norm k kC 1 is a Banach space.
Example 4.23 The linear space of integrable functions, L1 ([a, b]; R), endowed with the
L1 -norm k kL1 is a Banach space.
Example 4.24 The linear space of square integrable functions, L2 ([a, b]; R), endowed with
the L2 -norm k kL2 is a Hilbert space.

18

5
5.1

Sobolev Spaces and Variational Equalities


Sobolev Spaces

The linear space C([a, b]; R) is not complete w.r.t. the L2 -norm. But there is a larger linear
space, which is a Hilbert space w.r.t. the L2 -norm, namely the linear space L2 ([a, b]; R).
Actually, one can show that L2 ([a, b]; R) is a minimal complete extension of C([a, b]; R).
However, the space L2 ([a, b]; R) is not suitable for solving boundary value problems, since
even the variational equalities constructed in the second section require some kind of differentiability, let alone the original differential equation. In this section, we present an appropriate notion of differentiability which is applicable to certain elements of L2 ([a, b]; R).
First of all, we introduce the inner product, which is related to variational equalities.
Proposition 5.1 We consider the linear space V := C 1 ([a, b]; R). Then the mapping
h, iH 1 : V V R
defined by
Z

hf1 , f2 iH 1 :=

Z
f1 (x)f2 (x)dx +

f10 (x)f20 (x)dx

is an inner product on V .
Proof. This is an easy exercise. Note that the first term represents the L2 -inner product,
whereas the second term defines a positive bilinear form.
2
C 1 ([a, b]; R) is not a Hilbert space w.r.t. this inner product. In the sequel we construct a
minimal complete extension of C 1 ([a, b]; R). To this end recall the rule of integration by
parts. For continuously differentiable functions f it yields the variational equality
Z b
Z b
0
f (x)v(x)dx +
f (x)v 0 (x)dx = 0
a

for all v C01 ([a, b]; R), where we set


C01 ([a, b]; R) := {v C 1 ([a, b]; R) : v(a) = v(b) = 0}.
The variational equality above is used to generalize the notion of differentiability.
Definition 5.2 Let f L2 ([a, b]; R). We say that g L2 ([a, b]; R) is the generalized
derivative of f , if
Z b
Z b
f (x)v 0 (x)dx = 0
g(x)v(x)dx +
a

for all v

C01 ([a, b]; R).

In this case we write g = f 0

19

Example 5.3 In L2 ([1, 1]; R) we consider the function f given by


f (x) := |x|.
Surely, f is not contained in C 1 ([1, 1]; R), since it is not differentiable at x = 0. However, f does possess a generalized derivative, namely the function g L2 ([1, 1]; R) given
by
(
1 for 1 x 0
g(x) =
1 for 0 < x 1.
This can be seen as follows. We calculate
Z 1
g(x)v(x)dx
1
Z 0
Z 1
=
g(x)v(x)dx +
g(x)v(x)dx
1
0
Z 0
Z 1
=
v(x)dx +
v(x)dx
1
0
Z 0
Z 1
0
=
f (x)v(x)dx +
f 0 (x)v(x)dx
1
0
Z 0
Z 1
0
=
f (x)v (x)dx + f (0)v(0) f (1)v(1) +
f (x)v 0 (x)dx + f (1)v(1) f (0)v(0)
1
0
Z 1
f (x)v 0 (x)dx
=
1

for all v C01 ([1, 1]; R). Here, we make use of the fact that f is continuously differentiable on the intervals (1, 0) and (0, 1).
Definition 5.4 (Sobolev space) The Sobolev space H 1 ([a, b]; R) is defined as the set of
all functions f L2 ([a, b]; R), which possess a generalized derivative f 0 L2 ([a, b]; R).
Since
C 1 ([a, b]; R) H 1 ([a, b]; R),
we know that H 1 ([a, b]; R) 6= . On the other hand, the example above shows that
C 1 ([a, b]; R) 6= H 1 ([a, b]; R).
Remark 5.5 The definition of the Sobolev space H 1 ([a, b]; R) given above easily can be
generalized to functions defined on finite dimensional domains Rn . Let us note
that for one-dimensional domains, i.e. for intervals [a, b], the Sobolev space H 1 ([a, b]; R)
equivalently can be regarded as the space of all continuous functions f C([a, b]; R) that
can be written as
Z x
g(y)dy,
f (x) = f (a) +
a
2

where g L ([a, b]; R). In this case g is the generalized derivative of f . Hence, we can
apply the fundemental theorem of calculus.
20

Theorem 5.6 The Sobolev space H 1 ([a, b]; R) is a Hilbert space with the inner product
Z b
Z b
hf1 , f2 iH 1 :=
f1 (x)f2 (x)dx +
f10 (x)f20 (x)dx.
a

Actually, one can show that the Sobolev space H 1 ([a, b]; R) is the smallest complete
extension of C 1 ([a, b]; R) w.r.t. the inner product h, iH 1 . Hence the Sobolev functions
can be approximated by continuously differentiable functions w.r.t. the inner product
h, iH 1 . However, the Sobolev space H 1 ([a, b]; R) is still too large for finding the solutions
to boundary value problems, since we did not prescribe boundary values. For this reason
we finally define the subspace H01 ([a, b]; R) H 1 ([a, b]; R), which consists of those Sobolev
functions f : [a, b] R with f (a) = f (b) = 0.
Definition 5.7 The Sobolev space H01 ([a, b]; R) consists of those functions f H 1 ([a, b]; R),
that can be approximated w.r.t. the inner product h, iH 1 by functions taken from C01 ([a, b]; R).

5.2

Variational Equalities

The following result describes the strong relation between variational equalities and quadratic
minimization problems, which underlies the finite element method.
Theorem 5.8 Let V be a linear space, a : V V R be a symmetric, positive definite
bilinear form and L : V R be a linear form. Then a vector u V minimizes the
functional F : V R given by
1
F (v) := a(v, v) L(v).
2
if and only if u V solves the variational equality
a(u, v) = L(v) v V.
Proof. For t R and v V we calculate
1
F (u + tv) = F (u) + t(a(u, v) L(v)) + t2 a(v, v).
2
If u V minimizes the functional F , we obtain
F (u + tv) F (u)
= a(u, v) L(v)
t0
t

0 = lim

for all v V . As for the converse, let u V satisfy the variational equality
a(u, v) = L(v)
for all v V . Then setting t = 1, we obtain
1
F (u + v) = F (u) + a(v, v) > F (u)
2
21

For all v V \ {0} and the proof is finished.

Note that the theorem above does not say anything about the existence of a solution
to the problems involved. Indeed, one needs more structure in order to guarantee the
solvability of the variational equality (and hence of the minimization problem). The
following fundamental theorem is the corner stone for obtaining various existence results
for boundary value problems.
Theorem 5.9 (Riesz representation theorem) Let V be a Hilbert space and L : V
R be a continuous linear form. Then there exists a unique vector u V such that
hu, vi = L(v)
for all v V .
Many versions of this basic result running under the title Lax-Milgram theorem can be
found in the literature. Now we are nearly in a position to solve the variational equalities
related to linear second order boundary value problems. We only need one more property
of the bilinear form involved, that makes sure that the norm induced by the bilinear form
and by the original inner product are equivalent.
Definition 5.10 Let V be an inner product space. A bilinear form a : V V R is
called coercive, if there is a constant c > 0 such that
|a(v, v)| chv, vi
for all v V .
Theorem 5.11 Let V be a Hilbert space and L : V R be a continuous linear form.
Moreover, let a : V V R be a coercive, positive definite, symmetric, continuous
bilinear form. Then there exists a unique vector u V solving the variational equality
a(u, v) = L(v) v V.
Moreover, we can estimate
2C
,
c
where c > 0 is the coercivity constant of the bilinear form a and C 0 is the boundedness
constant of the linear form L.
kuk

Proof. The bilinear form a defines a new inner product on the linear space V . Hence,
according to the representation theorem of Riesz, we only have to show that V is complete
and L is continuous w.r.t. the norm
p
kvka := a(v, v).
22

To this end, we calculate


kvka =

p
p

a(v, v) Ckvkkvk = Ckvk

and
kvk =

hv, vi

ca(v, v) =

ckvka

for all v V . Hence the norms k k and k ka are equivalent. As for the estimation, we
observe that u minimizes the corresponding functional F (v) = 21 a(v, v) L(v), hence we
have
F (u) F (0) = 0
and

1
F (u) ckuk2 Ckuk,
2

which gives us
1
ckuk2 Ckuk 0
2
2

and the proof is finished.

5.3

Application to Linear Second Order Boundary Value Problems

In this part, we finally apply the abstract framework developped to linear second order
boundary value problems. For a given real number R and f L2 ([0, 1]; R) consider
the following differential equation
u00 (x) + f (x) = u(x),
where x [0, 1]. We are interested in the unique existence of a solution u H01 ([0, 1]; R)
for prescribed boundary values
u(0) = u(1) = 0.
Multiplying both sides of the differential equation by test functions v H01 ([0, 1]; R) and
integrating by parts leads to the variational equation
Z 1
Z 1
Z 1
0
0
u (x)v (x)dx +
u(x)v(x)dx =
f (x)v(x)dx v H01 ([0, 1]; R).
0

In order to apply our abstact existence result, Theorem 5.11, we have to check three points.
(I) The bilinear form given by
Z
a(w, v) :=

1
0

w(x)v(x)dx

w (x)v (x)dx +
0

23

should be positive definite, coercive and continuous on H01 ([0, 1]; R).
(II) The linear form given by
Z

L(v) :=

f (x)v(x)dx
0

should be continuous on H01 ([0, 1]; R).


(III) The space H01 ([0, 1]; R) should be a Hilbert space w.r.t. the inner product h, iH 1 .
As for (I), we calculate using the fundamental theorem of calculus and Hoelders inequality
2
2
Z 1
Z 1/2 Z x
Z 1  Z 1
2
0
0
v(x) dx =
v (y)dy dx +

v (y)dy dx
0

0
1/2

v 0 (y)2 dydx +

x
0

1
8
1
8

1/2

v 0 (y)2 dydx

x
1

1/2

1/2

v 0 (y)2 dy +

1
8

(1 x)

v (y) dydx +
0

(1 x)

1/2

Z
x

0
1/2

1/2
x

v 0 (y)2 dydx

1/2

v 0 (y)2 dy

1/2

v 0 (y)2 dy

C01 ([0, 1]; R),

for all v
and hence for all v H01 ([0, 1]; R). This estimation is of importance in order to obtain coercivity. For 8 < 1, it yields
Z 1
Z 1
0
2
a(v, v) =
v (x) dx +
v(x)2 dx
0
0
Z 1

Z 1
Z 1
8+
8+
0
2
0
2
=
v (x) dx + 1
v (x) dx +
v(x)2 dx
9
9
0
 Z0 1

 0Z 1

Z 1
8+
8+

v 0 (x)2 dx + 8 1
v(x)2 dx +
v(x)2 dx
9
9
0
0

 Z0 1

Z 1
8+
8+
=
v 0 (x)2 dx +
v(x)2 dx
9
9
0

 0
8+
=
hv, viH 1
9
for all v H01 ([0, 1]; R). For 1, we immediately obtain
Z 1
Z 1
Z 1
Z
0
2
2
0
2
a(v, v) =
v (x) dx +
v(x) dx
v (x) dx +
for all v
write

0
1
H0 ([0, 1]; R).

v(x)2 dx = hv, viH 1

The continuity of the bilinear form a is obvious, since we can

Z

|a(v, w)| (1 + ||)

1
0

v (x)w (x)dx +

24



v(x)w(x)dx (1 + ||)kvkH 1 kwkH 1

for all v, w H01 ([0, 1]; R).


As for (II), we calculate
|L(v)| kf kL2 kvkL2 kf kL2 kvkH 1
for all v H01 ([0, 1]; R).
As for (III), let vn be a Cauchy sequence in H01 ([0, 1]; R). Since H 1 ([0, 1]; R) is a Hilbert
space, there is a v H 1 ([0, 1]; R) with vn v, as n . It remains to show that
v H01 ([0, 1]; R). To this end let wn C01 ([0, 1]; R) with kvn wn kH 1 1/n for all n N.
Then we can estimate
kv wn kH 1 kv vn kH 1 + kvn wn kH 1 ,
where the right-hand side tends to zero, as n . Hence, v H01 ([0, 1]; R).
Overall, we can conclude from the existence theorem, Theorem 5.11, that the variational
equation has a unique solution u H01 ([0, 1]; R), at least for parameters > 8. For
parameters 8, we cannot prove the coercivity of the bilinear form involved. Actually
it is not surprising that negative parameters < 0 might cause difficulties. Consider, for
instance, the homogeneous differential equation
u00 (x) = u(x),
where x [0, 1] and < 0. It possesses a two-dimensional space of solutions, which can
be represented by
p
p
u(x) = sin( ||x) + cos( ||x),
where (, ) R2 . In order to fulfil the boundary condition u(0) = 0, we have to set
p
= 0. Accordingly, for || = k, k N, the boundary conditions are fulfiled by the
solution
p
u(x) = sin( ||x).
Hence, for = k 2 2 , k N, the boundary problem has more than one solution, i.e.
uniqueness is destroyed. This happens, in particular, for = 2 , which is slightly
smaller than our deduced lower bound for the parameters .
Remark 5.12 The inequality
Z
0

1
v(x) dx
8
2

25

Z
0

v 0 (x)2 dx

derived above is called Poincare-inequailty and is valid in a more general setting. Namely,
one can show that for any bounded domain Rn there is a constant C = C() such
that we can estimate
Z
Z
2
v(x) dx C hgradv(x), gradv(x)idx

for all v C01 (; R).

26

Approximation of Solutions

Definition 6.1 Let (V, k k) be a normed space and X V be a subset. We say that X
is dense in V , if for any v V and any  > 0 there is an x X with
kx vk .
Density of a subset means that any element of the larger space can be approximated by
elements of the smaller subset.
Example 6.2 The set of rational numbers Q is dense in the linear space of real numbers
(R, | |).
In connection with finite element methods, one is interested in the approximation of
functions.
Theorem 6.3 (Weierstrass approximation theorem) The space of real polynomials
P := {p C([a, b]; R) : p(x) = a0 + a1 x + . . . + an xn , n N, a0 , . . . , an R} is dense in
the space of continuous functions (C([a, b]; R), k k ).
Theorem 6.4 The space of continuous functions C([a, b]; R) is dense in the space of
square integrable functions (L2 ([a, b]; R), k kL2 ).
Theorem 6.5 The space of continuously differentiable functions C 1 ([a, b]; R) is dense in
the space of Sobolev functions (H 1 ([a, b]; R), k kH 1 ).
All these statements are quite interesting from a theoretical point of view. But for practical considerations one needs more information on the rate of approximation.

6.1

Taylor Series

Theorem 6.6 (Taylor) Let f C N +1 ([a, b]; R). Then the N -th Taylor polynomial of f
at a
TfN (x) := f (a) + Df (a)(x a) + D2 f (a)

(x a)2
(x a)N
+ + DN (a)
2!
N!

fulfils the estimation


kf TfN k kDN +1 f k

(b a)N +1
.
(N + 1)!

For f C ([a, b]; R) the Taylor series is given by


Tf (x) :=

Dk f (a)

k=0

(x a)k
.
k!

Note that there are functions for which the Taylor series does not converge at any x [a, b]
or for which the Taylor series does converge for all x [a, b] but its limit has nothing in
common with the original function f .
27

6.2

Splines

We consider linear splines on the interval [0, 1]. To this end we introduce a decomposition
[0, 1] =

N
[

[xk1 , xk ],

k=1

where

k
(k = 0, . . . , N ).
N
For k = 0, . . . , N we define the continuous mapping bk : [0, 1] R by

1 + (x xk )N for xk N x xk
bk (x) :=
1 (x xk )N for xk x xk + N1

0
for
|x xk | N1 .
xk :=

The subspace given by the basis BN = {b0 , . . . , bN } C([0, 1]; R) is called the space
1
1
of linear splines and denoted by SN
([0, 1]; R). In other words, SN
([0, 1]; R) consists of
continous functions in C([0, 1]; R) which are piecewise affine linear. Naturally we have
1
dim(SN
([0, 1]; R)) = N + 1.
Theorem 6.7 (Approximation with linear splines) For any N N and any f
1
C 2 ([0, 1]; R) the linear spline s SN
([0, 1]; R) given by
s(x) =

N
X

f (xk )bk (x)

k=0

fulfills the estimation

1 kf 00 k
.
N2 2
This approximation theorem relies on the Taylor approximation and often is used for
numerical integration.
kf sk

Theorem 6.8 (Trapezoid rule for integration) For any N N and any f C 2 ([0, 1]; R)
we can estimate
Z 1



1 kf 00 k
N

f
(x)dx

,
f

N2 2
0

where
IfN :=

f (0) + 2f (x1 ) + + 2f (xN 1 ) + f (1)


.
N

Proof. We can estimate


Z 1
Z


f (x)dx

0

Z

s(x)dx

kf sk dx

1 kf 00 k
.
N2 2

On the other hand, it is obvious that


IfN

s(x)dx
0

and the proof is finished.

28

6.3

Fourier Series

Here we consider the Hilbert space L2 ([0, 2]; R) together with particular functions of the
form
1
cos(kx)
sin(kx)
x 7 , x 7
, x 7
(k = 1, 2, . . .).

2
Those functions form a so-called orthogonal system, since they are pairwise orthogonal
w.r.t. the standard inner product on L2 ([0, 2]; R). Furthermore those functions are unit
vectors w.r.t. the L2 -norm. For a given function f L2 ([0, 2]; R) we define the Fourier
coefficients by
Z 2
Z 2
Z 2
sin(kx)
f (x)
cos(kx)
dx, ak :=
dx (k = 1, 2, . . .).
dx, bk :=
a0 :=
f (x)
f (x)

2
0
0
0
Theorem 6.9 The trigonometric polynomial of order N




1
cos(x)
sin(x)
cos(N x)
sin(N x)
N
Ff := a0 + a1
+ b1
+ + (aN
+ bN

2
fulfils the estimation
kf

FfN k2L2

(a2k + b2k ).

k=N +1

The Fourier series is given by




X
1
cos(kx)
sin(kx)
Ff (x) := a0 +
ak
+ bk
.

2 k=1
Note that kf k2L2 = a20 +
f w.r.t. the L2 -norm.

6.4

2
k=1 (ak

+ b2k ) < . Hence, the Fourier series does converge to

Homogenization

Sometimes it is advisable to approximate the solution of differential equation by a solution


of a simpler differential equation before starting the numerical calculation. For instance, if
the material considered consitsts of different thin layers that are ordered in a regular, say
periodic way. A direct numerical calculation has to take into account the fine structure of
the material and hence leads to fine discretization grids. For this reason one simplifies the
equation before starting discretizations, a procedure which is often called homogenization.
We consider the linear second order differential equation
d
dx


d
a (x) u (x) = f (x)
dx

on the interval x (0, 1) with boundary conditions


29

(1)

u (0) = u (1) = 0.
Here, we set
a (x) := a

x

,

where the mapping a : R R is bounded and periodic. In particular, we assume that
there are constants 0 < < and Y > 0 such that
0 < a(y) < and a(y) = a(y + Y )
for all x R. As for the right-hand side we assume that
f L2 ((0, 1); R).
The corresponding variational equation becomes
Z 1
Z 1
0
0
a (x)u (x)v (x)dx +
f (x)v(x)dx = 0 v H01 ([0, 1]; Rn ).
0

With the Poincare inequality one easily obtains the coercivity of the bilinear form
Z 1
a(u, v) :=
a (x)u0 (x)v 0 (x)dx
0

w.r.t. the H 1 -norm. The homogenized differential equation is given by


d
dx


d
a0 u0 (x) = f (x),
dx

where the real constant a0 can be calculated via averaging


Z
1
1 Y 1
=
dy.
a0
Y 0 a(y)
Theorem 6.10 The functions u , u0 are continuous on [0, 1] and we can estimate
ku u0 k  3Y

2 2
kf kL2
2

for all  > 0.


Proof. We set

d
u (x).
dx
Then the differential equation (1) is equivalent to
 (x) := a (x)

d
 (x) = f (x),
dx

d
 (x)
u (x) =
.
dx
a (x)
30

(2)

Hence,  is absolutely continuous and we can write


Z x
Z
 (x) =  (0) +
f (w)dw, u (x) =
0

Rz
 (0) + 0 f (w)dw
dz
a (z)

for all  > 0. Similarly, we set


d
u0 (x).
dx
Then the homogenized differential equation (2) is equivalent to
0 (x) := a0

d
0 (x) = f (x),
dx

d
0 (x)
u0 (x) =
.
dx
a0

Hence, 0 is absolutely continuouswe can write


Z x
Z
0 (x) = 0 (0) +
f (w)dw, u0 (x) =
0

0 (0) +

Rz
0

f (w)dw

a0

dz.

We conclude that
 (x) 0 (x) =  (0) 0 (0)
and
u (x) u0 (x)
Z x
Z x
 (0) + F (z)
0 (0) + F (z)
=
dz
dz
a (z)
a0
0
0




Z x
Z x
Z x
 (0) 0 (0)
1
1
1
1
dz +
0 (0)

dz +
F (z)

dz
=
a (z)
a (z) a0
a (z) a0
0
0
0
Z x
x Z x
z 
x
 (0) 0 (0)
=
dz + 0 (0)A

f (z)A
dz + F (x)A
,
a (z)



0
0
where

Z
F (z) :=

Z
f (w)dw,

A(y) :=

1
1

a(q) a0


dq.

The boundary condition 0 = u (1) = u0 (1) implies that


  Z 1
 
Z 1
z 
 (0) 0 (0)
1
1
dz = 0 (0)A
+
f (z)A
dz F (1)A
a (z)



0
0
and that accordingly

  Z

1
| (0) 0 (0)| 0 (0)A
+



  

1
f (z)A
dz + F (1)A
.


z 

Overall, we obtain
 
  Z 1
  
 z 

1
1



|u (x) u0 (x)|
+1
0 (0)A
+
f (z)A
dz + F (1)A




0



+ 1  (|0 (0)| + 2kf kL2 ) max |A(y)|.


yR

31

By the periodicity of a : R R and A(0) = A(Y ) = 0 we easily obtain the estimation




1
1
max |A(y)| = max |A(y)| Y

0yY
yR

Furthermore, the boundary condition u0 (1) = 0 gives us
|0 (0)| kf kL2
2

and the proof is finished.

32

References
[1] T.J.R. Hughes, The finite element method. Linear static and dynamic finite element
analysis. Englewood Cliffs, New Jersey: Prentice-Hall International, Inc. (1987).
[2] J.L. Nowinski, Applications of functional analysis in engineering. Mathematical Concepts and Methods in Science and Engineering, Vol. 22. New York and London:
Plenum Press (1981).
[3] M. Pedersen, Functional analysis in applied mathematics and engineering. Studies in
Advanced Mathematics. Boca Raton, FL: Chapman & Hall/CRC (2000).
[4] B. Daya Reddy, Introductory functional analysis. With applications to boundary
value problems and finite elements. Texts in Applied Mathematics. 27. New York,
NY: Springer (1998).
[5] W. Rudin, Principles of mathematical analysis. 3rd ed. International Series in Pure
and Applied Mathematics. Dsseldorf etc.: McGraw-Hill Book Company (1976).
[6] W. Rudin, Real and complex analysis. 3rd ed. New York, NY: McGraw-Hill (1987).
[7] K. Yosida, Functional analysis. Repr. of the 6th ed. Berlin: Springer (1994).

33

Você também pode gostar