1 Normed Spaces

Functional analysis is concerned with normed spaces and with operators between normed
spaces.
As an example we consider the initial value problem on I = [t0 , t1 ],

y 0 = f (t, y) , y(t0 ) = y0 , (1.1)

with the right side f : I × Rn → Rn . We transform (1.1) into the integral equation
Z t
y(t) = y0 + f (s, y(s)) ds . (1.2)
t0

We define an operator
Z t
n n
T : C(I; R ) → C(I; R ) , (T y)(t) = y0 + f (s, y(s)) ds . (1.3)
t0

Now (1.2) is equivalent to the operator equation

y = Ty . (1.4)

In this manner we have transformed our original problem into an equation whose unknown
variable is a function (in this case the function y), not a number nor a vector with finitely
many components. In this equation there appears a mapping (here T ) between function
spaces (here C(I; Rn )). Such mappings are usually called “operators”. The function
spaces are typically infinite-dimensional Banach or Hilbert spaces.
In the following we write K for R or C.

Definition 1.1 (Norm, normed space)

Let X be a vector space over K. A mapping k · k : X → [0, ∞) is called a norm on X, if

kxk = 0 ⇔ x = 0, (1.5)

kαxk = |α| kxk for all α ∈ K, x ∈ X, (1.6)

kx + yk ≤ kxk + kyk for all x, y ∈ X. (1.7)
The pair (X, k · k) is called a normed space. (Often, one just writes X.)

We repeat some basic notions from the first year analysis course. (See, for example, my
lecture notes.)
Let X be a normed space. A sequence (xn )n∈N in X is called convergent to the limit
x ∈ X, written as
lim xn = x , (1.8)
n→∞

if
lim kxn − xk = 0 . (1.9)
n→∞

2
The sequence is called a Cauchy sequence, if for every ε > 0 there exists an N ∈ N
such that
kxn − xm k ≤ ε , for all n, m ≥ N . (1.10)
X is called complete, if every Cauchy sequence in X has a limit (which by the definition
above has to be an element of X. If X is complete, X is called a Banach space. For
x ∈ X and ε > 0 we call

B(x, ε) = {y : y ∈ X, ky − xk < ε} , (1.11)

K(x, ε) = {y : y ∈ X, ky − xk ≤ ε} , (1.12)

the open and closed ε-ball around x, respectively. A subset O of X is called open in
X, if for every x ∈ O there exists an ε > 0 such that B(x, ε) ⊂ O. A subset A of X is
called closed in X, if X \ A is open in X, or equivalently, if for every convergent sequence
(xn )n∈N in X whose elements xn all belong to A, its limit also belongs to A. The closure
Y , the interior int (Y ) and the boundary ∂Y of a subset Y of X are given by
\ [
Y = A , int (Y ) = O , ∂Y = Y \ int (Y ) . (1.13)
A⊃Y O⊂Y
A closed O open

A subspace U of the vector space X becomes itself a normed space, if we define the norm
on U to be the restriction of the given norm on X. If U is complete then U is closed in
X; if X itself is complete, the converse also holds. The closure U of a subspace U of X
is also a subspace of X.
In the basic analysis lectures we have already encountered some Banach spaces. The
spaces (Kn , k · kp ), 1 ≤ p ≤ ∞, are Banach spaces with

n
! p1
X
kxkp = |xi |p , 1 ≤ p < ∞, kxk∞ = max |xi | . (1.14)
1≤i≤n
i=1

Let D be an arbitrary set. Then (B(D; K), k · k∞ ),

B(D; K) = {f | f : D → K, f bounded} , kf k∞ = sup |f (x)| , (1.15)

x∈D

is a Banach space. If moreover D is a compact metric space (for example, if D is a closed

and bounded subset of Kn ), then (C(D; K), k · k∞ ),

C(D; K) = {f | f : D → K, f stetig} , (1.16)

is a closed subspace of (B(D; K), k · k∞ ) and therefore itself a Banach space. If D is a

Lebesgue measurable subset of Rn (this is the case, for example, if D is open or closed),
then for 1 ≤ p < ∞ the space Lp (D; K) of functions which are Lebesgue integrable to the
p-th power, that is, those measurable functions for which
Z  p1
p
kf kp = |f (x)| dx < ∞, (1.17)
D

becomes a Banach space for this norm k · kp , if we “identify” functions which are equal
almost everywhere. (“Identify” means that we pass to the quotient space formed by

3
equivalence classes of functions which are equal almost everywhere.) This is discussed in
detail in the course on measure and integration theory.
Sequence spaces. Let us set D = N. Then B(D; K) coincides with the space of all
bounded sequences, denoted as `∞ (K),

`∞ (K) = {x : x = (xk )k∈N , xk ∈ K, sup |xk | < ∞} , kxk∞ = sup |xk | . (1.18)
k∈N k∈N

Because this is a special case of (1.15), the space `∞ (K) is a Banach space. Let us consider
the subsets

c(K) = {x : x = (xk )k∈N is a convergent sequence in K} , (1.19)

c0 (K) = {x : x = (xk )k∈N converges to 0 in K} . (1.20)

Proposition 1.2 We have c0 (K) ⊂ c(K) ⊂ `∞ (K). Endowed with the supremum norm,
the spaces c0 (K) are c(K) Banach spaces.

Proof: Exercise. It suffices to show that c(K) is a closed subspace of `∞ (K), and that
c0 (K) is a closed subspace of c(K). 2
Moreover, we consider the space ce (K) of all finite sequences,

ce (K) = {x : x = (xk )k∈N , there exists N ∈ N with xk = 0 for all k ≥ N } . (1.21)

The space ce (K) is a subspace of c0 (K); it is not closed in c0 (K). Thus, it is not a Banach
space; we have (exercise)
ce (K) = c0 (K) . (1.22)
Let x = (xk )k∈N be a sequence in K. We define

∞
! p1
X
kxkp = |xk |p , 1 ≤ p < ∞, (1.23)
k=1

and
`p (K) = {x : x = (xk )k∈N , xk ∈ K, kxkp < ∞} , (1.24)
the space of sequences which are summable to the p-th power.

Proposition 1.3 The space (`p (K), k · kp ), 1 ≤ p < ∞, is a Banach space.

Proof: For x ∈ `p (K) we have

∞
X
kxkp = 0 ⇔ |xk |p = 0 ⇔ |xk | = 0 for all k ⇔ x = 0.
k=1

If x ∈ `p (K), α ∈ K, then

∞
! p1 ∞
! p1
X X
kαxkp = |αxk |p = |α| |xk |p = |α|kxkp .
k=1 k=1

4
In order to prove the triangle inequality, let x, y ∈ `p (K). For arbitrary N ∈ N we have

N
! p1 N
! p1 N
! p1
X X X
|xk + yk |p ≤ |xk |p + |yk |p ,
k=1 k=1 k=1

this is just Minkowski’s inequality in KN . It follows that

N
X
|xk + yk |p ≤ (kxkp + kykp )p , for all N ∈ N.
k=1

Passing to the limit N → ∞ yields

∞
X
|xk + yk |p ≤ (kxkp + kykp )p ,
k=1

so
kx + ykp ≤ kxkp + kykp .
Thus, `p (K) is a normed space. In order to show that it is complete, let (xn )n∈N be a
Cauchy sequence in `p (K). Then for all k, n, m ∈ N we have
∞
X
|xnk − xm
k |
p
≤ |xnj − xm p n m p
j | = kx − x kp .
j=1

Therefore, (xnk )n∈N is a Cauchy sequence in K for all k. Since K is complete, there exists

x∞ n
k = lim xk , k ∈ N.
n→∞

For all k, n, m, N we have (Minkowski)

N
! p1 N
! p1 N
! p1
X X X
|xnk − x∞
k |
p
≤ |xnk − xm
k |
p
+ |xm ∞ p
k − xk |
k=1 k=1 k=1
N
! p1
X
∞ p
≤ kxn − xm kp + |xm
k − xk | . (1.25)
k=1

Let ε > 0 be arbitrary. We choose M large enough such that

kxn − xm kp ≤ ε , for all n, m ≥ M . (1.26)

Next, for every N ∈ N we choose m(N ) ∈ N such that m(N ) ≥ M and

N
! p1
X m(N )
|xk − x∞
k |
p
≤ ε. (1.27)
k=1

We set m = m(N ) in (1.25). Then it follows from (1.25) – (1.27) that

N
! p1
X
|xnk − x∞
k |
p
≤ 2ε , for all n ≥ M and all N ∈ N.
k=1

5
Passing to the limit N → ∞ we obtain

kxn − x∞ kp ≤ 2ε , for all n ≥ M .

This implies x∞ = (x∞ − xn ) + xn ∈ `p (K) and xn → x∞ in `p (K). 2

Linear continuous mappings. Let X and Y be normed spaces. By definition, a
mapping f : X → Y is continuous on X if and only if

f ( lim xn ) = lim f (xn ) (1.28)

n→∞ n→∞

holds for all convergent sequences (xn )n∈N in X. From the analysis course we know that
the assertions

f −1 (O) is open for every open set O ⊂ Y ,

and

f −1 (A) is closed for every closed set A ⊂ Y

both are equivalent to (1.28).

f is continuous at a point x ∈ X, if (1.28) holds for all sequences (xn )n∈N which converge
to x.

Proposition 1.4 Let (X, k · kX ), (Y, k · kY ) be normed spaces, let T : X → Y be linear.

Then the following are equivalent:

(i) T is continuous on X.

(ii) T is continuous in 0.

(iii) There exists a C > 0 such that

kT (x)kY ≤ CkxkX , for all x ∈ X. (1.29)

(iv) T is Lipschitz continuous on X with Lipschitz constant C.

Proof: “(iii)⇒(iv)”: For all x, y ∈ X we have

kT (x) − T (y)kY = kT (x − y)kY ≤ Ckx − ykX .

“(iv)⇒(i)⇒(ii)”: obvious.
“(ii)⇒(iii)”: Contraposition. Assume that (iii) does not hold. We choose a sequence
(xn )n∈N in X satisfying
kT (xn )kY > nkxn kX . (1.30)
We set
1
zn = xn ,
nkxn kX

6
this is possible, because xn 6= 0 due to (1.30). It follows that zn → 0, but kT (zn )kY > 1
and therefore T (zn ) does not converge to 0. Consequently, (ii) does not hold. 2
Not all linear mappings are continuous. Here is an example: the unit vectors {ek : k ∈ N}
form a basis of ce (K). We define a linear mapping T : ce (K) → K by T (ek ) = k. Then we
have kek k∞ = 1 and |T (ek )| = k, so (iii) in Proposition 1.4 is not satisfied.
In the following we will write kxk instead of kxkX if it is obvious which norm is meant.
We will also write T x instead of T (x).

Definition 1.5 (Isomorphism)

Let X, Y be normed spaces. A mapping T : X → Y which is bijective, linear and continu-
ous is called an isomorphism between X and Y , if T −1 , too, is continuous. If moreover
kT xk = kxk for all x ∈ X, T is called an isometric isomorphism. X and Y are
called (isometrically) isomorphic, if there exists an (isometric) isomorphism between
X and Y . In this case we write X ' Y (X ∼
= Y ). 2

Obviously we have (and the same for “∼

=”)

X'Y , Y 'Z ⇒ X 'Z. (1.31)

Let T : X → Y be an isomorphism. According to Proposition 1.4 there exist constants

C1 and C2 such that

kT xk ≤ C1 kxk , kxk = kT −1 T xk ≤ C2 kT xk , for all x ∈ X. (1.32)

If k · k1 and k · k2 are two different norms on X, the identity mapping is an isomorphism

between (X, k · k1 ) and (X, k · k2 ) if and only if there exist constants C1 and C2 such that

kxk1 ≤ C1 kxk2 , kxk2 ≤ C2 kxk1 , for all x ∈ X. (1.33)

In this case the norms k · k1 und k · k2 are called equivalent.

We already know from the basic analysis course that on Kn all norms are equivalent. An
immediate generalization of this result is the following.

Proposition 1.6 Let X, Y be finite dimensional normed spaces with dim(X) = dim(Y ).
Then X ' Y .

Proof: Let dim(X) = n, let {v1 , . . . , vn } be a basis of X, let e1 , . . . , en be the unit vectors
in Kn . We define
n
X n
X
n
T :K →X, Tx = xi vi if x= xi ei .
i=1 i=1

One can verify immediately that

kxkX = kT xk
defines a norm on Kn ; T then becomes an isometric isomorphism. For Y , we proceed
analogously. Then
X∼ = (Kn , k · kX ) ' (Kn , k · kY ) ∼
=Y .

7
 2
If X, Y are finite dimensional vector spaces with Y ⊂ X, Y 6= X, then dim(Y ) < dim(X),
and X und Y are not isomorphic, since bijective linear mappings leave the dimension
invariant. In the infinite dimensional case, the situation is not that simple. For example,
c0 (K) ⊂ c(K) and c0 (K) 6= c(K), but we have

c0 (K) ' c(K) . (1.34)

Indeed, we claim that an isomorphism T : c(K) → c0 (K) is given by

(T x)1 = lim xj , (T x)k = xk−1 − lim xj , k ≥ 2 , (1.35)

j→∞ j→∞

and that the mapping S : c0 (K) → c(K),

(Sy)k = yk+1 + y1 , (1.36)

is the inverse of T . One may compute directly that T ◦ S and S ◦ T are equal to the
identity on c0 (K) and c(K) resp., and that kT xk∞ ≤ 2kxk∞ as well as kSyk∞ ≤ 2kyk∞
hold for all x ∈ c(K) and all y ∈ c0 (K).

Proposition 1.7 Let (X1 , k·k1 ), . . . , (Xm , k·km ) be normed spaces. On the product space
m
Y
X= X i = X1 × · · · × Xm (1.37)
i=1

the expressions
m
! p1
X
kxk∞ = max kxi ki , kxkp = kxi kpi , (1.38)
1≤i≤m
i=1

where x = (x1 , . . . , xm ) ∈ X, define norms k · kp , 1 ≤ p ≤ ∞. All these norms are

equivalent. A sequence (xn )n∈N in X converges to an x = (x1 , . . . , xm ) ∈ X if and only if
all component sequences (xni )n∈N converge to xi . X is complete if and only if all Xi are
complete.

Proof: Exercise. 2

Corollary 1.8 Let X be a normed space. The addition + : X × X → X and the scalar
multiplication · : K × X → X are continuous.

Proof: If xn → x and yn → y, then

0 ≤ k(xn + yn ) − (x + y)k ≤ kxn − xk + kyn − yk → 0 .

If αn → α and xn → x, then

0 ≤ kαn xn − αxk ≤ |αn | kxn − xk + |αn − α| kxk → 0 .

8
Definition 1.9 (Space of operators, dual space)
Let X, Y be normed spaces. We define

L(X; Y ) = {T | T : X → Y, T is linear and continuous} . (1.39)

The space L(X; K) is called the dual space of X, denoted X ∗ . The elements of X ∗ are
called functionals. 2

From linear algebra and analysis it is known that L(X; Y ) is a vector space.

Proposition 1.10 (Operator norm)

Let X, Y be normed spaces. Then

kT xk
kT k = sup (1.40)
x∈X, x6=0 kxk

defines a norm on L(X, Y ), called the operator norm. We have

kT xk ≤ kT k kxk , for all x ∈ X, (1.41)

and
kT k = sup kT xk = sup kT xk , (1.42)
x∈X, kxk≤1 x∈X, kxk=1

as well as
kT k = inf{C : C > 0, kT xk ≤ Ckxk for all x ∈ X} . (1.43)
If Y is a Banach space, then so is L(X, Y ).

Proof: Let T ∈ L(X; Y ), let C > 0 with kT xk ≤ Ckxk for all x ∈ X. Dividing by kxk
we see that kT k ≤ C, so kT k ∈ R+ , and (1.41) as well as (1.43) hold. From (1.41) we
obtain that
sup kT xk ≤ sup kT xk ≤ kT k .
x∈X, kxk=1 x∈X, kxk≤1

Since
kT xk x
= T( ) ,
kxk kxk
(1.42) follows. We have

kT k = 0 ⇔ kT xk = 0 for all x ∈ X ⇔ T x = 0 for all x ∈ X ⇔ T = 0.

For α ∈ K we get

kαT k = sup kαT xk = |α| sup kT xk = |α| kT k ,

kxk=1 kxk=1

and for S, T ∈ L(X; Y ) we get

kS + T k = sup kSx + T xk ≤ sup kSxk + sup kT xk = kSk + kT k .

kxk=1 kxk=1 kxk=1

9
Thus, the properties of a norm are satisfied. We now prove that L(X, Y ) is complete. Let
(Tn )n∈N be a Cauchy sequence in L(X; Y ). Since
kTn x − Tm xk = k(Tn − Tm )(x)k ≤ kTn − Tm k kxk
the sequence (Tn x)n∈N is a Cauchy sequence in Y for every x ∈ X. Then (as Y is complete)
T x = lim Tn x
n→∞

defines a mapping T : X → Y . Let x, z ∈ X, α, β ∈ K, then

αT x + βT z = α lim Tn x + β lim Tn z = lim (αTn x + βTn z)
n→∞ n→∞ n→∞
= lim Tn (αx + βz) = T (αx + βz) .
n→∞

Therefore, T is linear. Since (as the norm is continuous)

kT xk = k lim Tn xk = lim kTn xk ≤ (sup kTn k)kxk ,
n→∞ n→∞ n∈N

T is continuous. It remains to prove that kTn − T k → 0. Let ε > 0. We choose N ∈ N

such that kTn − Tm k ≤ ε for all n, m ≥ N . For arbitrary x ∈ X with kxk = 1 we get
k(Tn − T )xk ≤ k(Tn − Tm )xk + k(Tm − T )xk ≤ kTn − Tm kkxk + kTm x − T xk .
Therefore, for n ≥ N it follows, choosing m sufficiently large,
k(Tn − T )xk ≤ 2ε ,
and thus kTn − T k ≤ 2ε if n ≥ N . 2

Example 1.11

1. D compact metric space, X = C(D; K) with the supremum norm, a ∈ D, Ta : X →

K, Ta x = x(a). Ta is linear, and |Ta x| = |x(a)| ≤ kxk∞ with equality, if x is a
constant function. Therefore, Ta is continuous, and kTa k = 1. The functional Ta is
called the Dirac functional in a.
2. X = C([a, b]; R) with the supremum norm, T : X → R,
Z b
Tx = x(t) dt .
a

T is linear, Z b
|T x| = x(t) dt ≤ (b − a)kxk∞
a
with equality if x is constant. Therefore, T is continuous and kT k = b − a.
3. X = L1 ([a, b]; R) with the L1 norm, T as before, then
Z b Z b
|T x| = x(t) dt ≤ |x(t)| dt = kxk1
a a

with equality if x is constant. Therefore, T is continuous and kT k = 1.

10
4. X as before, f ∈ C([a, b]; R), T : X → R,
Z b
Tx = f (t)x(t) dt .
a

T is linear, and
Z b Z b
|T x| ≤ |f (t)| |x(t)| dt ≤ kf k∞ |x(t)| dt = kf k∞ kxk1 .
a a

Therefore, T is continuous and kT k ≤ kf k∞ . In order to prove that actually equality

holds, we choose t∗ ∈ [a, b] with f (t∗ ) = kf k∞ (if the maximum of the absolute value
is attained at a point where f is negative, we instead consider −f resp. −T ). Let
now ε > 0, ε < kf k∞ . We choose an interval I with t∗ ∈ I ⊂ [a, b], so that
f (t) ≥ kf k∞ − ε holds for all t ∈ I, and set (|I| denotes the length of I)
(
1
1 |I|
, t∈I,
x= 1I , so x(t) =
|I| 0 , sonst .

Then kxk1 = 1 and

Z b Z Z
1 1
|T x| = f (t)x(t) dt = f (t) dt ≥ (kf k∞ − ε) dt = kf k∞ − ε .
a I |I| I |I|

It follows that kT k ≥ kf k∞ − ε and therefore kT k = kf k∞ .

5. Let D = (0, 1) × (0, 1), k ∈ L2 (D; R). On X = L2 ((0, 1); R) we want to define an
integral operator T : X → X by
Z 1
(T x)(s) = k(s, t)x(t) dt , s ∈ (0, 1) .
0

Once we have proved that this integral is well defined, it is clear that T is linear.
We have (the integrals are well defined as a consequence of Fubini’s and Tonelli’s
theorem, the second inequality follows from Hölder’s inequality)
Z 1 Z 1 2 Z 1 Z 1 2
kT xk2L2 ((0,1);R) = k(s, t)x(t) dt ds ≤ |k(s, t)| |x(t)| dt ds
0 0 0 0
Z 1 Z 1  Z 1 
2 2
≤ |k(s, t)| dt · |x(t)| dt ds
0 0 0
Z 1Z 1
2
= kxkL2 ((0,1);R) |k(s, t)|2 dt ds .
0 0

Therefore, T is continuous and

Z 1 Z 1  12
2
kT k ≤ |k(s, t)| dt ds = kkkL2 (D;R) .
0 0

11
 6. Let X, Y be normed spaces with dim(X) < ∞. Then every linear mapping T : X →
Y is continuous: Let {v1 , . . . , vn } be a basis of X. Then for
n
X
x= xi vi ∈ X , xi ∈ K ,
i=1

we have that
n
X n
X n
X
kT xk = xi T vi ≤ |xi | kT vi k ≤ max kT vi k |xi | .
1≤i≤n
i=1 i=1 i=1

Therefore,
n
X
kT k ≤ max kT vi k , if kxk = |xi | .
1≤i≤n
i=1

7. Let T : Rn → Rm be linear, let T (x) = Ax for the matrix A ∈ R(m,n) . Via this
correspondence, we can identify the space L(Rn ; Rm ) with the space R(m,n) of all
m × n-matrices. The operator norm then becomes a so-called matrix norm whose
explicit form depends on the choice of the norms in Rn and Rm . Matrix norms
play an important role in the construction and analysis of numerical algorithms, in
particular in numerical linear algebra. 2

Lemma 1.12 Let X, Y, Z be normed spaces, let T : X → Y and S : Y → Z be linear

and continuous. Then S ◦ T : X → Z is linear and continuous, and
kS ◦ T k ≤ kSk kT k . (1.44)

Proof: For all x ∈ X we have k(S ◦ T )xk ≤ kSk kT xk ≤ kSk kT k kxk. The assertion now
follows from Proposition 1.10. 2
Seminorms and quotient spaces.

Definition 1.13 (Seminorm)

Let X be a vector space over K. A mapping p : X → [0, ∞) is called a seminorm on X,
if
p(αx) = |α| p(x) for all α ∈ K, x ∈ X, (1.45)
p(x + y) ≤ p(x) + p(y) for all x, y ∈ X. (1.46)
In this case, (X, p) is called a seminormed space.

Obviously,(1.45) implies p(0) = 0, but p(x) = 0 does not imply x = 0. Every norm is
a seminorm. If (Y, q) is a seminormed space, then every linear mapping T : X → Y
generates a seminorm on X by
p(x) = q(T x) . (1.47)
Here are some examples of seminorms:
X = Rn , p(x) = |x1 | , (1.48)
X = B(D; K) , a ∈ D , p(x) = |x(a)| , (1.49)
Z 1
1
X = L ((0, 1); R) , p(x) = x(t) dt , (1.50)
0
X = C 1 ([0, 1]; R) , p(x) = kẋk∞ . (1.51)

12
From linear algebra, we recall the concept of a quotient space. Let X be a vector space,
let U be a subspace of X. Then

x ∼U z ⇔ x−z ∈U (1.52)

defines an equivalence relation on X. Let

[x] = {z : z ∈ X, x ∼U z} (1.53)

denote the equivalence class of x ∈ X. The quotient space X/U is defined by

X/U = {[x] : x ∈ X} . (1.54)

With the addition and scalar multiplication on X/U defined by

[x] + [y] = [x + y] , α[x] = [αx] , (1.55)

X/U becomes a vector space, and the mapping

Q : X → X/U , Qx = [x] , (1.56)

is linear and surjective. We have 0 = [x] = Qx if and only if x ∈ U .

Proposition 1.14 (Quotient norm) Let U be a subspace of a normed space X.

(i) The formula

p([x]) = dist (x, U ) = inf kx − zk (1.57)
z∈U

defines a seminorm on X/U which satisfies

p([x]) ≤ kxk , for all x ∈ X. (1.58)

(ii) If U is closed, then p is a norm.

(iii) If moreover X is a Banach space and U is closed, then (X/U, p) is a Banach space.

Proof: Part (i) is an exercise. Concerning (ii): Let

0 = p([x]) = inf kx − zk .
z∈U

There exists a sequence (zn )n∈N in U satisfying kx − zn k → 0, thus zn → x and therefore

(if U is closed) x ∈ U , so finally [x] = 0. Concerning (iii): We begin by proving: If
x, y ∈ X, then there exists a ỹ ∈ [y] with

kỹ − xk ≤ 2p([y − x]) . (1.59)

We obtain such a ỹ by first choosing a z ∈ U with

ky − x − zk ≤ 2p([y − x])

13
and then setting ỹ = y − z. Let now ([xn ])n∈N be a Cauchy sequence in X/U . Passing to
a subsequence if necessary we may assume that

p([xn+1 ] − [xn ]) ≤ 2−n .

We now choose according to (1.59) an x̃1 ∈ [x1 ] and for n > 1 an x̃n ∈ [xn ] mit

kx̃n − x̃n−1 k ≤ 2p([xn ] − [xn−1 ]) . (1.60)

Then for all n, p ∈ N we have

p p
X X
kx̃n+p − x̃n k ≤ kx̃n+j − x̃n+j−1 k ≤ 2 · 2−n−j+1 ≤ 2−n+2 .
j=1 j=1

Therefore, (x̃n )n∈N is a Cauchy sequence in X. As X is complete, there exists x =

limn∈N x̃n . From (1.58) it follows that

0 ≤ p([xn ] − [x]) = p([x̃n ] − [x]) ≤ kx̃n − xk → 0 ,

so [xn ] → [x] in X/U . 2

Here is an example: X = C([0, 1]), U = {x : x ∈ X, x(0) = 0}. We have

x ∼U y ⇔ x(0) = y(0) .

One directly checks that

p([x]) = |x(0)| , X/U ∼
= R.
The mapping T : X/U → R, T ([x]) = x(0), is an isometric isomorphism
Dense subsets. A subset A of a metric space (X, d) is called dense in X, if

A=X (1.61)

holds. If A is dense in (X, d) and B is dense in (A, dA ), then B is dense in (X, d).

Definition 1.15 (Separable space)

A metric space (X, d) is called separable if there exists a finite or countably infinite
subset A of X which is dense in X. 2

Example: Qn is a dense subset of Rn , therefore Rn is separable. Analogous for Cn .

Proposition 1.16 The space (C([a, b]; K), k · k∞ ) is separable.

Proof: By the approximation theorem of Weierstra”s, the set P of all polynomials is dense
in C([a, b]; K). Moreover, the set of all polynomials with rational coefficients is countable.
It is dense in P , and therefore in C([a, b]; K), too. 2

Proposition 1.17 Let X be a normed space, let (xn )n∈N be a sequence in X with X =
span {xn : n ∈ N}. Then X is separable.

14
Proof: Exercise. 2

Proposition 1.18 The space `p (K) is separable for 1 ≤ p < ∞. The space `∞ (K) is not
separable.

Proof: For p < ∞ we have `p (K) = span {en : n ∈ N}. The assertion then follows from
1.17. Let now p = ∞. For arbitrary M ⊂ N we define xM ∈ `∞ (K) by
(
M 1, k ∈ M ,
xk =
0, k ∈ /M.

If M, N ⊂ N with M 6= N , then kxM − xN k∞ = 1. Therefore,

1
{B(xM , ) : M ⊂ N}
2
is an uncountable set whose elements are disjoint open balls. If A is a dense subset of
`∞ (K), it has to be dense in each of those balls. Therefore A cannot be countable. 2
Analogously it holds that, for D ⊂ Rn open, the space (Lp (D; K), k · kp ) is separable
for 1 ≤ p < ∞, but not separable for p = ∞. If p < ∞ and moreover D is bounded,
this follows from the result that C(D) is dense in (Lp (D), k · kp ) and that (in analogy to
Proposition 1.16) the polynomials with rational coefficients form a dense subset of C(D).
For D unbounded the claim follows from the representation
[
Lp (D; K) = Lp (D ∩ Bn ; K)
n∈N

with Bn = {x : kxkp < n}.

Proposition 1.19 Let X be a normed space, Y a Banach space, U a dense subspace of X

and S : U → Y linear and continuous. Then there exists a unique linear and continuous
mapping T : X → Y with T |U = S, and kT k = kSk.

Proof: Given x ∈ X, we choose a sequence in (xn )n∈N in U with xn → x and define

T x = limn→∞ Sxn . The limit exists since (xn )n∈N and consequently (Sxn )n∈N are Cauchy
sequences, and since Y is complete. The assertions now follow directly from the definitions
and elementary properties of the operator norm and of convergent sequences; we do not
write down the details. 2
Dual spaces. Let a normed space X be given. According to Proposition 1.10, the dual
space X ∗ = L(X; K) is a Banach space.

Proposition 1.20 Let p, q ∈ (1, ∞), let

1 1
+ = 1. (1.62)
p q
Then
(`p (K))∗ ∼
= `q (K) . (1.63)

15
Proof: We want to define an isometric isomorphism T : `q (K) → (`p (K))∗ by
∞
X
(T x)(y) = xk y k , x ∈ `q (K) , y ∈ `p (K) . (1.64)
k=1

By virtue of Hölder’s inequality,

N N
! 1q N
! p1
X X X
|xk | |yk | ≤ |xk |q · |yk |p ≤ kxkq kykp .
k=1 k=1 k=1
P
Therefore, the series k xk yk converges absolutely, and
∞
X
|(T x)(y)| ≤ |xk yk | ≤ kxkq kykp .
k=1

Thus, (1.64) defines for any given x ∈ `q (K) a linear continuous mapping T x : `p (K) → K
which satisfies
kT xk ≤ kxkq . (1.65)
Therefore T : `q (K) → (`p (K))∗ is well-defined. It follows from (1.64) that T is linear
and from (1.65) that T is continuous. T is injective, as (T x)(ek ) = xk , and consequently
T x = 0 implies that xk = 0 for all k. We now prove that T is surjective. Let y ∗ ∈ (`p (K))∗
be arbitrary. We want to find an x = (xk )k∈N in `q (K) such that T x = y ∗ . Since
(T x)(ek ) = xk for such an x, we have to define

xk = y ∗ (ek ) . (1.66)

Let y ∈ ce (K),
N
X
y= yk ek , yk ∈ K , N ∈ N, (1.67)
k=1

be given. We have
N
! N N
X X X
∗ ∗ ∗
y (y) = y yk ek = yk y (ek ) = y k xk . (1.68)
k=1 k=1 k=1

We choose
yk = |xk |q−1 sign (xk ) , 1≤k≤N. (1.69)
Then
|yk |p = |xk |p(q−1) = |xk |q = xk yk ,
and it follows from (1.68) that

N N N
! p1 N
! p1
X X X X
|xk |q = xk yk = y ∗ (y) ≤ ky ∗ k kykp = ky ∗ k |yk |p = ky ∗ k |xk |q .
k=1 k=1 k=1 k=1

Thus,
N
! 1q
X
|xk |q ≤ ky ∗ k . (1.70)
k=1

16
Passing to the limit N → ∞ yields x ∈ `q (K) and

kxkq ≤ ky ∗ k . (1.71)

From (1.64) and (1.68) it follows that

(T x)(y) = y ∗ (y) , for all y ∈ ce (K).

We recall that ce (K) is dense in (`p (K), k · kp ). It follows from Proposition 1.19 that

T x = y∗ . (1.72)

Therefore T is surjective, and from (1.65) and (1.71) we obtain that kT xk = kxkq . There-
fore, T is an isometric isomorphism. 2
We mention some other results concerning the representation of dual spaces. We have

c0 (K)∗ ∼
= `1 (K) , `1 (K)∗ ∼
= `∞ (K) . (1.73)

We do not present the proof here.

For function spaces, analogous results hold. For example, let D ⊂ Rn be open. Then
1 1
Lp (D; K)∗ ∼
= Lq (D; K) , 1 < p < ∞, + = 1. (1.74)
p q

For the case p = 2, (1.74) is a consequence of a general result for Hilbert spaces, which
will be treated in the next chapter. We sketch the proof for arbitrary p ∈ (1, ∞); its
general structure is the same as that for the sequence space `p (K). Setting
Z
(T x)(y) = x(t)y(t) dt , x ∈ Lq (D; K), y ∈ Lp (D; K) , (1.75)
D

one obtains a linear continuous mapping T : Lq (D; K) → Lp (D; K)∗ with kT xk = kxkq .
Indeed, by virtue of Hölder’s inequality,
Z Z  1q Z  p1
q p
|(T x)(y)| ≤ |x(t)| |y(t)| dt ≤ |x(t)| dt |y(t)| dt = kxkq kykp , (1.76)
D D D

and for
y(t) = sign (x(t))|x(t)|q−1
we have |y(t)|p = |x(t)|q and therefore
Z Z
q−1
(T x)(y) = x(t)sign (x(t))|x(t)| dt = |x(t)|q dt
D D
Z  1q Z  p1
q q
= |x(t)| dt |x(t)| dt = kxkq kykp .
D D

In order to prove that T is surjective, for a given y ∗ ∈ Lp (D; K)∗ one constructs an
x ∈ Lq (D; K) satisfying T x = y ∗ by employing the Radon-Nikodym theorem from measure
and integration theory.

17
One also has the result that

L1 (D; K)∗ ∼
= L∞ (D; K) . (1.77)

The representation theorem of Riesz states that, for compact sets D ⊂ Rn , the space
C(D; R)∗ is isometrically isomorphic to the space of all signed regular measures on the
Borel σ-algebra on D. In particular, for every y ∗ ∈ C(D; R)∗ there exists a measure µ
such that Z
∗
y (y) = y dµ . (1.78)
D

Definition 1.21 (Series in normed spaces)

Let X be a normed space, let (xk )k∈N be a sequence in X. If the sequence
n
X
sn = xk (1.79)
k=1

of the partial sums converges to an element s ∈ X, we say that the corresponding series
P ∞
k=1 xk converges, and we define
X∞
xk = s . (1.80)
k=1
P∞
The series k=1 xk is called absolutely convergent if
∞
X
kxk k < ∞ . (1.81)
k=1

Because addition and scalar multiplication in normed spaces are continuous operations,
we have the rules
∞
X ∞
X ∞
X ∞
X ∞
X
(xk + yk ) = xk + yk , αxk = α xk , (1.82)
k=1 k=1 k=0 k=1 k=1

α ∈ K. They are valid when the limits on the right side exist.
P∞
Proposition 1.22 Let X be a Banach space, assume that the series k=1 xk converges
absolutely. Then it also converges, and
∞
X ∞
X
xk ≤ kxk k . (1.83)
k=1 k=1

Moreover, every reordering of the series converges, and the limits are identical.

Proof: Let n
X
σn = kxk k .
k=1

18
For the partial sums defined in (1.79) we have, if n > m,
n
X
ksn − sm k ≤ kxk k = |σn − σm | .
k=m+1

Since (σn ) is a Cauchy sequence in K, also (sn ) is a Cauchy sequence in X, hence conver-
gent to some s ∈ X. Due to ksn k ≤ |σn |, (1.83) follows from
∞
X
ksk = lim ksn k ≤ lim |σn | = kxk k .
n→∞ n→∞
k=1

Let ∞
X
x̃k , x̃k = xπ(k) , π : N → N bijective,
k=1
P
be a reordering of xk with the partial sums
n
X
s̃n = x̃k .
k=1

Let ε > 0. We choose M large enough such that

∞
X
kxk k ≤ ε .
k=M +1

Next, we choose N such that N ≥ M and π({1, . . . , N }) ⊃ {1, . . . , M }. Then we have

for all n > N ∞
X
ks̃n − sn k ≤ kxk k ≤ ε ,
k=M +1

therefore ks̃n − sn k → 0. 2

19