MA40057 Functional Analysis spring 2016

1 Normed Vector Spaces, Banach Spaces and Hilbert Spaces.

This chapter collects together some familiar things, mainly from MA40043, without proofs, and adds some
related new things, with proofs. Throughout this course, we consider vector spaces over the fields R or C.
When it can be either, we write F for the underlying field; so F = R or C.

Definition. Let X be a vector space over F. A norm on X is a function k · k : X → R such that for all x, y ∈ X
and α ∈ F,
(i) kxk ≥ 0, with equality if, and only if, x = 0,
(ii) kαxk = |α|kxk, and
(iii) kx + yk ≤ kxk + kyk (triangle inequality).
The pair (X, k · k) is then called a normed vector space or simply a normed space.

Definition. If k · k is a norm on the vector space X, then it induces a metric d(x, y) = kx − yk on X.

Recall that a sequence (xn )n∈N in X is said to converge if there exists a point x0 ∈ X such that d(xn , x0 ) → 0
as n → ∞.
The sequence is called a Cauchy sequence if d(xm , xn ) → 0 as m, n → ∞.
A metric space is called complete if every Cauchy sequence converges.
A normed vector space is called a Banach space if it is complete.
A metric space is said to be separable if it contains a countable dense subset.
q
Example. For N ∈ N, the space RN , with the Euclidean norm |x| = x21 + · · · + x2N , is a Banach space.
p
Example. The space CN , regarded as a complex vector space, with the Euclidean norm |z| = |z1 |2 + · · · + |zN |2 ,
is a Banach space.
Example. If we equip RN with the norm |x|∞ = max{|x1 |, . . . , |xN |}, then it is still a Banach space.
Example. Let a < b and consider the space C([a, b]), comprising all continuous functions f : [a, b] → F. Define

kf kC([a,b]) = sup |f (x)|.

x∈[a,b]

This is a norm that makes C([a, b]) a Banach space (cf. MA40043). [Last year’s notation was C 0 ([a, b]).]
Example. On the other hand, if C([a, b]) is equipped with the norm
ˆ b
kf kL1 (a,b) = |f (x)| dx,
a

then it is not a Banach space, as it is not complete (cf. MA40043).

Some Basic Properties of Normed Spaces

We now explore some connections between the algebraic structure of a normed vector space (or Banach space)
and concepts related to their metric structure such as closed sets and convergence.

1
Lemma 1.1. Let X be a normed vector space and x, y ∈ X. Then

|kxk − kyk| ≤ kx − yk.

Proof. We have x = (x − y) + y. So by the triangle inequality,

kxk − kyk ≤ kx − yk.

The inequality
kyk − kxk ≤ kx − yk
is proved similarly, and both inequalities together imply the claim.

Theorem 1.2. Let X be a normed vector space. Suppose that (xn )n∈N and (yn )n∈N are sequences in X with
xn → x0 and yn → y0 as n → ∞. Furthermore, suppose that (αn )n∈N is a sequence in F with αn → α0 as
n → ∞. Then
(i) limn→∞ kxn k = kx0 k,
(ii) limn→∞ (xn + yn ) = x0 + y0 , and
(iii) limn→∞ (αn xn ) = α0 x0 .

Remark. This means that the maps

k·k: X → F
(the norm),
+: X × X → X
(vector addition), and
·: F×X →X
(multiplication with a scalar) are continuous in each argument.

Proof. (i) By the triangle inequality

|kxn k − kx0 k| ≤ kxn − x0 k → 0 as n → ∞.

(ii) By the triangle inequality,

k(xn + yn ) − (x0 + y0 )k ≤ kxn − x0 k + kyn − y0 k → 0.

(iii) We have

kαn xn − α0 x0 k = k(αn − α0 )xn + α0 (xn − x0 )k

≤ |αn − α0 |kxn k + |α0 |kxn − x0 k.

Furthermore, we have |αn − α0 | → 0 and kxn k → kx0 k by (i), and kxn − x0 k → 0. Hence

kαn xn − α0 x0 k → 0 as n → ∞.

Lemma 1.3. Let X be a normed vector space and Y ⊂ X a linear subspace. Then the closure Y is a linear
subspace as well.

2
Proof. We have y ∈ Y if, and only if, there exists a sequence (yn )n∈N in Y with y = limn→∞ yn . Now let
x, y ∈ Y and α ∈ F. Choose sequences (xn )n∈N and (yn )n∈N in Y converging to x and y, respectively. Then

x + y = lim (xn + yn ) ∈ Y
n→∞

and
αx = lim (αxn ) ∈ Y ,
n→∞

as xn + yn ∈ Y and αxn ∈ Y for every n ∈ N.

Remark. If Y ⊂ X is a linear subspace of a normed vector space with norm k · k, then the restriction of k · k
to Y is a norm on Y .

Theorem 1.4. Let X be a Banach space and Y ⊂ X a linear subspace. Then Y is a Banach space if, and only
if, it is closed.

Proof. From MA30041, a subspace of a metric space is complete if and only if it is closed.

Definition. Let X be a vector space over F. An inner product (or scalar product) on X is a function h · , · i : X ×
X → F such that for all x, y, z ∈ X and α ∈ F,
(i) hx + y, zi = hx, zi + hy, zi,
(ii) hαx, yi = αhx, yi,
(iii) hx, yi = hy, xi, and
(iv) hx, xi ≥ 0, with equality if, and only if, x = 0.
The pair (X, h · , · i) is then called an inner product space.

Remarks. In the above, z̄ denotes the complex conjugate for a number z ∈ C. If z ∈ R, then we have of
course z̄ = z. Thus if F = R, then the complex conjugate can be ignored in (iii).
If F = R, then (i)-(iii) imply that ( · , · ) is bilinear. That is, we also have

hx, y + zi = hx, yi + hx, zi

and
hx, αyi = αhx, yi.
If F = C, then the scalar product is still additive in the second variable but the homogeneity needs to be
replaced by
hx, αyi = ᾱhx, yi.
In the complex case the inner product is sometimes said to be sesquilinear (“1 12 -times linear”).
An inner product induces a norm on X, defined by
p
kxk = hx, xi

for x ∈ X. This is a norm. In particular it satisfies the triangle inequality by the following result.

3
Theorem 1.5 (Cauchy-Schwarz inequality and triangle inequality). Let X be a vector space with inner
product h · , · i. Then for x, y ∈ X, the Cauchy-Schwarz inequality

|hx, yi| ≤ kxkkyk

and the triangle inequality

kx + yk ≤ kxk + kyk
hold.

Proof. This was proved in MA40043.

Definition. An inner product space is called a Hilbert space if it is complete with respect to the induced metric.
(In particular, a Hilbert space is a Banach space.)

Example. The space RN with the Euclidean inner product

N
X
hx, yi = xi yi
i=1

is a Hilbert space over R.

Example. The space CN with the Euclidean inner product
N
X
hz, wi = zi w̄i
i=1

is a Hilbert space over C.

Example. Let `2 denote the space of all sequences x = (xn )n∈N in F such that
∞
X
|xn |2 < ∞.
n=1

Consider the function h · , · i`2 : `2 × `2 → F, defined by

∞
X
hx, yi`2 = xn ȳn
n=1

for x = (xn )n∈N and y = (yn )n∈N . This is a Hilbert space. (For F = R, this was shown in MA40043. For F = C,
the proof is similar.)
Orthogonality in Inner Product Spaces

Definition. Let X be a vector space with inner product h · , · i and Y ⊂ X a linear subspace. Then

Y ⊥ = {x ∈ X : hx, yi = 0 for all y ∈ Y }

is called the orthogonal complement of Y .

Theorem 1.6 (Orthogonal Decomposition, or Projection Theorem). If H is a Hilbert space and Y ⊂ H

a closed linear subspace, then H = Y ⊕ Y ⊥

Proof. This is a result from MA40043.

4
Theorem 1.7 (Double Complement). Let H be a Hilbert space and Y ⊂ H a closed linear subspace. Then
(Y ⊥ )⊥ = Y .

Proof. Let y ∈ Y . Then hy, xi = 0 for every x ∈ Y ⊥ , so y ∈ (Y ⊥ )⊥ . This proves Y ⊂ (Y ⊥ )⊥ .

Now let y ∈ (Y ⊥ )⊥ . Decompose it with the help of Theorem 1.6 into

y = y1 + y2

with y1 ∈ Y and y2 ∈ Y ⊥ . Then hy, y2 i = 0 (because y2 ∈ Y ⊥ and y ∈ (Y ⊥ )⊥ ) and hy1 , y2 i = 0 (because y1 ∈ Y

and y2 ∈ Y ⊥ ). Hence
hy2 , y2 i = hy − y1 , y2 i = 0.
This implies y2 = 0, which means y = y1 ∈ Y . Therefore, we have (Y ⊥ )⊥ ⊂ Y .

Definition. For an index set A, a family (xα )α∈A in a Hilbert space H is called an orthonormal system if
(i) kxα k = 1 for every α ∈ A, and
(ii) hxα , xβ i = 0 for α, β ∈ A with α 6= β.
An orthonormal system is called complete if for every y ∈ H\{0}, there exists an α ∈ A such that (xα , y) 6= 0.
If the index set is N, then we also speak of an orthonormal sequence.

Remark. The condition for a complete orthonormal system can also be expressed as

(span {xα : α ∈ A})⊥ = {0}.

Theorem 1.8. Let (un )m∈N be an orthonormal sequence in a Hilbert space H and (αn )n∈N a sequence in F.
Then ∞
P P∞ 2
n=1 αn un is convergent if, and only if, n=1 |αn | < ∞. If so, then

∞ 2 ∞
X X
αn un = |αn |2 .
n=1 n=1

Proof. Let
n
X n
X
sn = αk uk and σn = |αk |2
k=1 k=1
be the partial sums. Then for m < n,
n 2 n
X X
2
ksn − sm k = αk uk = αk ᾱ` huk , u` i
k=m+1 k,`=m+1
Xn
= |αk |2 = σn − σm .
k=m+1

Thus (sn )n∈N is a Cauchy sequence in H if, and only if, the sequence (σn )n∈N is Cauchy in R. The first
statement is now clear.
Furthermore, by similar computations,
∞ 2 ∞
X X
αn un = lim ksn k2 = lim σn = |αn |2 .
n→∞ n→∞
n=1 n=1

5
Theorem 1.9 (Bessel’s inequality). Let (un )n∈N be an orthonormal sequence in a Hilbert space H. Then
for any x ∈ H,
X∞
|hx, un i|2 ≤ kxk2 .
n=1

Proof. This is another result from MA40043.

Remarks. Let (un )∞ n=1 be an orthonormal sequence in a Hilbert space H.

√
(i) kun k = 1 for every n and kun − um k = 2 if m 6= n so (un )∞n=1 is a bounded sequence having no convergent
subsequence.
(ii) From Bessel’s inequality, if x ∈ H then hx, un i → 0 as n → ∞.

Theorem 1.10. Let (un )n∈N be an orthonormal sequence in a Hilbert space H. Then the formula
∞
X
x= hx, un iun (1)
n=1

holds for all x ∈ span {un : n ∈ N} and for no other points in H. In particular, if the orthonormal sequence is
complete, then (1) holds for every x ∈ H.

Proof. If (1) holds, then it is obvious that x ∈ span {un : n ∈ N}.

Conversely, suppose that x ∈ span {un : n ∈ N}. Then by Bessel’s inequality 1.9 and Theorem 1.8, the series
in (1) is convergent. Define
∞
X
y =x− hx, un iun .
n=1

Then for any m ∈ N,

∞
X
hy, um i = hx, um i − hx, un ihun , um i = 0.
n=1

So  ⊥
y ∈ (span {un : n ∈ N})⊥ = span {un : n ∈ N} .

(The last identity is verified in Problem Sheet 3 Q2(a).) Clearly we also have y ∈ span {un : n ∈ N}. Thus y
belongs both to span {un : n ∈ N} and to its orthogonal complement, hence y = 0, as required.
If we have a complete orthonormal sequence, then
 ⊥
span {un : n ∈ N} = (span {un : n ∈ N})⊥ = {0}.

Hence span {un : n ∈ N} = H and (1) holds for every x ∈ H.

Definition. In an infinite-dimensional Hilbert space H, a complete orthonormal sequence is called an orthonor-

mal basis for H.

Remark. The orthonormal bases defined here are not bases in the algebraic sense. The latter are sometimes
called Hamel bases in order to make the distinction clear.

Theorem 1.11. An infinite-dimensional Hilbert space is separable if, and only if, it has an orthonormal basis.

6
Proof. Suppose that H is an infinite-dimensional Hilbert space with an orthonormal basis (un )n∈N . Then

H = span {un : n ∈ N}.

If F = R, consider the set

N
( )
X
A= αn un : N ∈ N, α1 , . . . , αN ∈ Q .
n=1

This is countable and we still have A = H. Hence H is separable. If F = C, then we replace Q by Q + iQ.
Now suppose that H is separable, i.e., there exists a dense sequence (xn )n∈N in H. Then similarly to
MA40043, we can use the Gram-Schmidt process to construct an orthonormal basis from (xn )n∈N .

The Spaces `p
Fix a p with 1 ≤ p ≤ ∞ (i.e., either p ∈ R with p ≥ 1 or p = ∞). Let x = (xn )n∈N be a sequence in F. Then
we define !1
X∞ p
p
kxk` =p |xn | if 1 ≤ p < ∞, kxk`∞ = sup |xn |.
n=1 n∈N

Let `p be the set of all sequences x in F with kxk`p < ∞. We want to show that (`p , k · k`p ) is a Banach space
(with the vector space structure given by

x + y = (x1 + y1 , x2 + y2 , x3 + y3 , . . .), αx = (αx1 , αx2 , αx3 , . . .)

for x, y ∈ `p and α ∈ F). But we need some preparation first.

Definition. If p, q ∈ (1, ∞) with

1 1
+ = 1,
p q
or if p = 1, q = ∞ or p = ∞, q = 1, then we say that p and q are conjugate exponents. (In this context we
1
typically define ∞ = 0, so that we can express all three cases by one equation.)

Lemma 1.12 (Young’s inequality). Let p, q ∈ (1, ∞) be conjugate exponents. Then for all a, b ≥ 0,

ap bq
ab ≤ + .
p q
Proof. The inequality is trivial for a = 0 or b = 0. Hence we assume that a, b > 0.
Fix b > 0 and consider the function
ap
f (a) = − ab, a > 0.
p
q
We have to show that f (a) ≥ − bq for all a > 0.
We have f 0 (a) = ap−1 − b. So f 0 (a) < 0 for a < b1/(p−1) and f 0 (a) > 0 for a > b1/(p−1) . Thus there is a
global minimum at a = b1/(p−1) . Hence for all a > 0,

bp/(p−1) bq
 
1/(p−1) 1/(p−1)+1 p/(p−1) 1
f (a) ≥ f (b )= −b =b −1 =− ,
p p q
p
as q = p−1 .

7
Lemma 1.13 (Hölder’s inequality). Let p, q be conjugate exponents. If x ∈ `p and y ∈ `q , then
∞
X
|xn yn | ≤ kxk`p kyk`q
n=1

Proof. We distinguish three cases.

Case 1: p = 1 and q = ∞. Then we note that |yn | ≤ kyk`∞ for every n ∈ N. Hence
∞
X ∞
X
|xn yn | ≤ kyk`∞ |xn | = kxk`1 kyk`∞ .
n=1 n=1

Case 2: p = ∞ and q = 1. This case is analogous.

Case 3: p, q ∈ (1, ∞). The inequality is clear for x = 0 or y = 0. Otherwise, define
x y
x0 = and y 0 = .
kxk`p kyk`q

Then kx0 k`p = 1 and ky 0 k`q = 1. By Young’s inequality,

∞ ∞ ∞  0 p
|yn0 |q

X X X |xn |
|x0n yn0 | = |x0n ||yn0 | ≤ +
p q
n=1 n=1 n=1
∞ ∞
1 X 1 X kx0 kp`p ky 0 kq`q
= |x0n |p + |yn0 |q = + = 1.
p q p q
n=1 n=1

Therefore,
∞
X ∞
X
|xn yn | = kxk`p kyk`q |x0n yn0 | ≤ kxk`p kyk`q .
n=1 n=1

Lemma 1.14 (Minkowski’s inequality). Let p ∈ [1, ∞] and x, y ∈ `p . Then x + y ∈ `p and

kx + yk`p ≤ kxk`p + kyk`p .

Proof. The cases p = 1 and p = ∞ are straightforward, so we assume that p ∈ (1, ∞). Choose N ∈ N. Then
N
X N
X
|xn + yn |p ≤ (|xn | + |yn |)|xn + yn |p−1
n=1 n=1
(2)
XN N
X
= |xn ||xn + yn |p−1 + |yn ||xn + yn |p−1 .
n=1 n=1

p
Let q = p−1 , so that p and q are conjugate exponents. Then

N N
! p1 N
! 1q
X X X
|xn ||xn + yn |p−1 ≤ |xn |p |xn + yn |q(p−1)
n=1 n=1 n=1
!1− p1 (3)
N
X
≤ kxk`p |xn + yn |p
n=1

8
by Hölder’s inequality (applied to truncated sequences). Similarly,

N N
!1− p1
X X
|yn ||xn + yn |p−1 ≤ kyk`p |xn + yn |p . (4)
n=1 n=1

Plugging (3) and (4) into (2), we obtain

N N
!1− p1
X X
p p
|xn + yn | ≤ (kxk`p + kyk`p ) |xn + yn | ,
n=1 n=1

whence
N
! p1
X
|xn + yn |p ≤ kxk`p + kyk`p .
n=1

Letting N → ∞, we obtain kx + yk`p < ∞, i.e., x + y ∈ `p , and furthermore,

kx + yk`p ≤ kxk`p + kyk`p .

Remark. Hölder’s and Minkowski’s inequalities also hold for FN with the p-norms | · |p , from the above results
applied to elements of `p whose coordinates after the N -th vanish.

Theorem 1.15. For 1 ≤ p ≤ ∞, the space `p is a Banach space.

Proof. Minkowski’s inequality shows that `p is closed under vector addition and the triangle inequality holds.
All the other properties required of a vector space and a norm are easy to verify. Thus `p is a normed vector
space. It remains to show that it is complete.
Let (x(n) )n∈N be a Cauchy sequence in `p . Fix k ∈ N. Then we have
(m) (n)
|xk − xk | ≤ kx(m) − x(n) k`p → 0 as m, n → ∞.
(n)
Hence (xk )n∈N is a Cauchy sequence in F. We know that F is complete, so there exists a limit
(n)
xk = lim xk .
n→∞

Let x = (xk )k∈N . We claim that x(n) → x in `p .

Firstly consider the case p < ∞. Let ε > 0. Then there exists a number N ∈ N such that

kx(m) − x(n) k`p ≤ ε

for m, n ≥ N . Thus for any fixed K ∈ N,

K
!1/p
(m) (n)
X
|xk − xk |p ≤ kx(m) − x(n) k`p ≤ ε.
k=1

Letting m → ∞, we obtain
K
!1/p
(n)
X
|xk − xk |p ≤ ε,
k=1

9
as long as n ≥ N . Now let K → ∞. We deduce
∞
!1/p
(n)
X
|xk − xk |p ≤ε
k=1

for n ≥ N . That is, we have kx − x(n) k`p ≤ ε. By Lemma 1.14, it follows that x ∈ `p . Furthermore, as ε was
chosen arbitrarily, we have shown that x(n) → x in `p as n → ∞.
For p = ∞, the arguments are similar. Let ε > 0 and choose N ∈ N such that kx(m) − x(n) k`∞ ≤ ε for
(m) (n) (n)
m, n ≥ N . Then |xk − xk | ≤ ε for every k ∈ N. Hence |xk − xk | ≤ ε for every k ∈ N, which means

kx − x(n) k`∞ ≤ ε.

It follows that x ∈ `∞ and x(n) → x in `∞ as n → ∞.

Theorem 1.16 (Riesz’s lemma). Suppose that X is a normed vector space and Y ⊂ X a closed linear
subspace. If Y 6= X and θ ∈ (0, 1), then there exists a vector u ∈ X with kuk = 1 and

inf ku − yk ≥ θ.
y∈Y

Proof. Choose x0 ∈ X\Y . Define

ρ = inf {ky − x0 k : y ∈ Y } .
Since Y is closed, we have ρ > 0. Choose y0 ∈ Y with ky0 − x0 k ≤ ρ/θ. Set
x0 − y0
u= .
kx0 − y0 k

Obviously, we have kuk = 1. Moreover, for every y ∈ Y ,

x0 − y0 kx0 − (y0 + kx0 − y0 ky)k θ

ku − yk = −y = ≥ ρ = θ,
kx0 − y0 k kx0 − y0 k ρ

since y0 + kx0 − y0 ky ∈ Y .

Theorem 1.17. Let X be a normed vector space and

B = {x ∈ X : kxk ≤ 1}

the closed unit ball in X. Then B is compact if, and only if, X is finite-dimensional.

Proof. It follows from Exercise 2.3 and the Heine-Borel theorem that B is compact in a finite-dimensional
space.
Suppose that X is infinite-dimensional. Choose a unit vector x1 ∈ X. Consider the one-dimensional
subspace X1 = span{x1 }. According to Exercise 2.4, a finite-dimensional subspace is always closed. So we
can apply Riesz’s Lemma 1.16 to X1 and construct an x2 ∈ X with kx2 k = 1 and kx2 − x1 k ≥ 21 . Define
X2 = span{x1 , x2 }. Use Riesz’s Lemma 1.16 again to find a point x3 ∈ X with kx3 k = 1 and
1 1
kx3 − x1 k ≥ and kx3 − x2 k ≥ .
2 2
1
If we continue indefinitely, we obtain a sequence (xn )n∈N of unit vectors, such that kxm − xn k ≥ 2 for m 6= n.
But this means that we cannot have a convergent subsequence, so B is not compact.

10
2 Bounded Linear Operators
The Space B(X, Y )
In this chapter we study linear operators (also called linear maps or linear transformations) between normed
spaces, especially operators with the following additional property.
Definition. Let X, Y be normed vector spaces and T : X → Y a linear operator. We say that T is bounded if
there exists a number c > 0 such that
kT xkY ≤ ckxkX
for all x ∈ X.
Theorem 2.1. Let X, Y be normed vector spaces and T : X → Y linear. Then the following are equivalent.
(i) T is bounded.
(ii) T is uniformly continuous.
(iii) T is continuous at 0.
Proof. (i) ⇒ (ii). Suppose that there is a number c > 0 satisfying kT xkY ≤ ckxkX for all x ∈ X. Let ε > 0
and let x, y ∈ X.
kT x − T ykY = kT (x − y)kY ≤ ckx − ykX < ε
provided that kx − ykX < ε/c, so T is uniformly continuous.
(ii) ⇒ (iii). This is trivial.
(iii) ⇒ (i). Choose δ > 0 such that
kxkX ≤ δ ⇒ kT xkY ≤ 1.
Then for any x ∈ X\{0},  
kxkX δx kxkX
kT xkY = T ≤ .
δ kxkX Y δ
Hence T is bounded.

Definition. Let X, Y be normed vector spaces. Then B(X, Y ) is the space comprising all bounded linear
operators T : X → Y . For T ∈ B(X, Y ), we define
kT k = sup kT xkY = sup kT xkY ,
x∈X x∈BX
kxkX ≤1

where BX = {x ∈ X | kxkX ≤ 1}. We call kT k the operator norm of T . We also write B(X) = B(X, X).
Sometimes we write kT kB(X,Y ) or kT kB(X) for the operator norm in order to distinguish it from other
norms. Note that the linearity implies the identities
kT xkY
kT k = sup kT xkY = sup ,
x∈X x∈X\{0} kxkX
kxkX =1

unless X = {0}.
Theorem 2.2. (i) If X, Y are normed vector spaces, then B(X, Y ), equipped with the operator norm, is a
normed vector space.
(ii) If X is a normed vector space and Y is a Banach space, then B(X, Y ) is a Banach space.
(iii) Let X, Y, Z be normed vector spaces and suppose that S ∈ B(X, Y ) and T ∈ B(Y, Z). Then T S ∈ B(X, Z)
with
kT Sk ≤ kT kkSk.

11
Proof. (i) is routine.
(ii) Suppose that Y is complete. Let (Tn )n∈N be a Cauchy sequence in B(X, Y ). Then for any fixed x ∈ X, we
have
kTm x − Tn xkY ≤ kTm − Tn kkxkX → 0 as m, n → ∞.
Hence (Tn x)n∈N is a Cauchy sequence in Y and the limit

T x := lim Tn x
n→∞

exists. This gives rise to a map T : X → Y . It is linear, because for x, y ∈ X and α ∈ F, we have

T (αx + y) = lim Tn (αx + y) = α lim Tn x + lim Tn y = αT x + T y.

n→∞ n→∞ n→∞

It is bounded, because  
kT xkY = lim kTn xkY ≤ lim kTn k kxkX .
n→∞ n→∞

(Note limn→∞ kTn k exists because (kTn k)n∈N is a Cauchy sequence in F.)
It remains to prove that Tn → T in B(X, Y ) (which is stronger than pointwise convergence). Let ε > 0 and
choose N ∈ N such that kTm − Tn k ≤ ε for m, n ≥ N . Then for every x ∈ X,

kTm x − Tn xkY ≤ εkxkX .

Letting m → ∞, we obtain
kT x − Tn xkY ≤ εkxkX .
Thus kT − Tn k ≤ ε, as required.
(iii) Let x ∈ X. Then
kT SxkZ ≤ kT kkSxkY ≤ kT kkSkkxkX .
This implies the desired inequality.

Definition. If X is a normed vector space over F, then X ∗ = B(X, F) is called the dual space of X. Its
elements are called bounded linear functionals on X.

Example. Let a < b and c ∈ [a, b]. Define T : C([a, b]) → F by T f = f (c). This is a linear operator and we
have
|T f | = |f (c)| ≤ kf kC([a,b]) .
So T ∈ (C([a, b]))∗ and kT k ≤ 1. Since T 1 = 1, we have in fact kT k = 1.
Example. Let g ∈ C([a, b]) and define T : C([a, b]) → C([a, b]) by T f = gf . Then

kT f kC([a,b]) = kgf kC([a,b]) ≤ kgkC([a,b]) kf kC([a,b]) .

Hence T ∈ B(C([a, b])) and kT k ≤ kgkC([a,b]) . Since T 1 = g, we have in fact kT k = kgkC([a,b]) .

Example. Let p ∈ [1, ∞] and consider the operators

R : `p → `p , (x1 , x2 , x3 , . . .) 7→ (0, x1 , x2 , . . .),

L : `p → `p , (x1 , x2 , x3 , . . .) 7→ (x2 , x3 , x4 , . . .).

For x = (x1 , x2 , x3 , . . .) ∈ `p , we clearly have

kRxk`p = kxk`p and kLxk`p ≤ kxk`p .

12
So R, L ∈ B(`p ). It also follows immediately that kRk = 1 and kLk ≤ 1. In fact we have kLk = 1, because L
maps (0, 1, 0, 0, . . .) (with norm 1) to (1, 0, 0, . . .) (still with norm 1).
Example. Let p, q be conjugate exponents. Fix y = (y1 , y2 , y3 , . . .) ∈ `q . Define T : `p → F by
∞
X
Tx = x n yn .
n=1

Then by the Hölder inequality,

|T x| ≤ kxk`p kyk`q
for every x ∈ `p . So T ∈ (`p )∗ with kT k ≤ kyk`q . We will see later that kT k = kyk`q .

The Uniform Boundedness Principle

It is not so easy to come up with an example of a linear operator between Banach spaces that is not bounded.
Nevertheless, boundedness is a severe restriction and has some deep consequences. We now discuss some of
them. Recall

Theorem 2.3 (Baire Category Theorem). Let (X, d) be a nonempty complete metric space and let {Fn }∞
n=1
be a sequence of closed sets that covers X. Then Fn has nonempty interior for some n ∈ N.

Proof. Exercise: convince yourself this is the same as the Baire category theorem from MA40043!

As a consequence we obtain a result on bounded linear operators.

Theorem 2.4 (Banach-Steinhaus Uniform Boundedness Principle). Let X be a Banach space and Y
a normed vector space. Let T ⊂ B(X, Y ) be such that for all x ∈ X, the set {T x : T ∈ T } is bounded in Y .
Then T is bounded in B(X, Y ).

In other words, if we have pointwise bounds, then we automatically get a bound in the norm of B(X, Y ) as
well. The second property appears much stronger and implies in particular that we have uniform bounds on
bounded subsets of X. That is, if A ⊂ X with

sup kxk < ∞,

x∈A

then   
sup sup kT xkY ≤ sup kT k sup kxkX < ∞.
T ∈T x∈A T ∈T x∈A

Proof. For n ∈ N, define

\
Xn = {x ∈ X : kT xkY ≤ n for all T ∈ T } = {x ∈ X : kT xkY ≤ n}.
T ∈T

Then Xn is closed in X. By the hypothesis, we have

∞
[
X= Xn .
n=1

By the Baire Category Theorem 2.3, there exists an n ∈ N such that Xn contains a ball

B0 = {x ∈ X : kx − x0 k ≤ r}

13
for some x0 ∈ X and r > 0. That is, for x ∈ B0 and T ∈ T , we have kT xkY ≤ n.
Fix T ∈ T and suppose that x ∈ X with kxkX ≤ 1. Define y = rx + x0 . Then y ∈ B0 . So
 
y − x0 1
kT xkY = T = (T y − T x0 )
r Y r Y
1 2n
≤ (kT ykY + kT x0 kY ) ≤ .
r r
2n
Hence kT k ≤ r . The right-hand side is independent of T , so T is bounded in B(X, Y ).

Corollary 2.5. Let X be a Banach space and Y a normed vector space. Suppose that (Tn )n∈N is a sequence in
B(X, Y ) such that (Tn x)n∈N is convergent for every x ∈ X. Then there exists a unique operator T ∈ B(X, Y )
such that
T x = lim Tn x
n→∞
for all x ∈ X.

Proof. It is clear (by the assumption and the uniqueness of limits) that the formula defines a unique map
T : X → Y . It is also clear that T is linear. We need to show that T is bounded.
Let T = {Tn : n ∈ N}. For every x ∈ X, the sequence (Tn x)n∈N is convergent by the hypothesis; in
particular it is bounded. Therefore, we can apply the Uniform Boundedness Theorem 2.4 to T . It follows that
there exists a number M ≥ 0 with the property that kTn k ≤ M for every n ∈ N. Now for any x ∈ X,

kT xkY = lim kTn xkY ≤ M kxkX .

n→∞

So kT k ≤ M and T ∈ B(X, Y ).

Example. Let (an )n∈N be a sequence in R such that for every x ∈ `2 , the sequence (a1 x1 , a2 x2 , a3 x3 , . . .) also
belongs to `2 . Then the linear operator

T : `2 → `2 , (x1 , x2 , x3 , . . .) 7→ (a1 x1 , a2 x2 , a3 x3 , . . .),

is bounded.
To prove this statement, define Tn : `2 → `2 with

Tn (x1 , x2 , x3 , . . .) = (a1 x1 , . . . , an xn , 0, 0, . . .).

Then Tn ∈ B(`2 ) with kTn k ≤ max{|a1 |, . . . , |an |}. Moreover, for every x ∈ `2 ,

T x = lim Tn x.
n→∞

Thus the claim follows from Corollary 2.5.

The Open Mapping Theorem and Closed Graph Theorem

Recall that a continuous map can be characterised by the property that the inverse image of any open set is
open. We now consider the reverse property, which is useful, e.g., when we study the continuity of the inverse
of a bijective operator.

Definition. Let X, Y be metric spaces and f : X → Y a map. We say that f is open if for every open set
U ⊂ X, the image f (U ) = {f (x) : x ∈ U } is open in Y .

14
Theorem 2.6 (Open Mapping Theorem). Let X, Y be Banach spaces and T ∈ B(X, Y ). If T is surjective,
then T is open.

Proof. Consider the closed unit balls

BX = {x ∈ X : kxkX ≤ 1} and BY = {y ∈ Y : kykY ≤ 1}.

For R > 0 we use the notation RBX for the ball of centre 0 and radius R in X. Since T is surjective, we have
∞
[
Y = T (nBX ).
n=1

By the Baire category theorem there exists an n ∈ N such that T (nBX ) contains a ball, say with centre y1 and
radius 2r, so we may choose any y0 ∈ (y1 + rBY ) ∩ T (nBX) and we then have

y0 + rBY ⊂ y1 + 2rBY ⊂ T (nBX ).

Let x0 ∈ X be such that T (x0 ) = y0 . Then

BY ⊂ −r−1 y0 + r−1 T (nBX ) = T (−r−1 x0 + r−1 nBX ) ⊂ T (RBX )

where R = r−1 kx0 k + r−1 n.

Next we claim that
BY ⊂ T (2RBX ).
In order to verify this, fix y ∈ BY . Then there exists an x1 ∈ RBX with

ky − T x1 kY ≤ 12 .

Hence
(y − T x1 ) ∈ 21 BY ⊂ T 1


2 RBX .
Therefore, there exists a point x2 ∈ 12 RBX such that

k(y − T x1 ) − T x2 kY ≤ 14 ,

so
ky − T (x1 + x2 )kY ≤ 14 .
Now we recursively define x3 , x4 , . . . such that xk ∈ ( 12 )k−1 RBX and

1 k


ky − T (x1 + · · · + xk )kY ≤ 2 .

Then
∞
X ∞
X
kxk kX ≤ R ( 21 )k−1 = 2R.
k=1 k=1
Therefore,
∞
X
x= xk
k=1

exists (homework exercise) and belongs to 2RBX . Moreover,

ky − T xkY = lim ky − T (x1 + · · · + xk )kY = 0.

k→∞

15
So y = T x. That is, we have proved BY ⊂ T (2RBX ).
Finally, let U ⊂ X be an open set and let V = T (U ). Fix y0 ∈ V . There exists an x0 ∈ U with T x0 = y0 .
As U is open, there exists a number ρ > 0 such that x0 + ρBX ⊂ U . Then
ρ ρ
V ⊃ T (x0 + ρBX ) = y0 + T (2RBX ) ⊃ y0 + BY .
2R 2R
Hence V is open.

Corollary 2.7. Let X, Y be Banach spaces and T ∈ B(X, Y ). If T is bijective, then T −1 ∈ B(Y, X).

Proof. We have T −1 ∈ B(X, Y ) if, and only if, the inverse T −1 is continuous by Theorem 2.1. A map is
continuous if, and only if, inverse images of open sets are open. For an inverse map T −1 , this is equivalent to
T being open. So the claim follows directly from the Open Mapping Theorem 2.6.

If X, Y are normed spaces, then we can equip the Cartesian product X × Y with the norm

k(x, y)kX×Y = kxkX + kykY .

If X and Y are Banach spaces, then this makes X × Y a Banach space as well.
We now have the following criterion for boundedness of an operator between Banach spaces.

Theorem 2.8 (Closed Graph Theorem). Let X, Y be Banach spaces and T : X → Y a linear operator.
Then T ∈ B(X, Y ) if, and only if, the graph G = {(x, T x) : x ∈ X} is closed in X × Y .

Proof. Suppose that T ∈ B(X, Y ). Then G is closed since G = L−1 (0), where L : X × Y → Y is the linear
operator L(x, y) = y − T x, which is bounded because

kL(x, y)kY ≤ kykY + kT kkxkX ≤ (1 + kT k)k(x, y)kX×Y .

Now suppose that G is closed. Consider the projections PX : X × Y → X and PY : X × Y → Y with

PX (x, y) = x and PY (x, y) = y. We have

kPX (x, y)kX = kxkX ≤ k(x, y)kX×Y .

So PX ∈ B(X × Y, X) with kPX k ≤ 1. Similarly, we see that PY ∈ B(X × Y, Y ) with kPY k ≤ 1.

Now G is a Banach space because G is a closed linear subspace of the Banach space X × Y . The restriction
PX |G of PX to G belongs to B(G, X) and is bijective. By Corollary 2.7, we have

(PX |G )−1 ∈ B(X, G).

Finally, we note that

T = PY ◦ (PX |G )−1 .
Hence T ∈ B(X, Y ).

16
3 Dual Spaces
Recall that for a normed vector space X over F, we define X ∗ = B(X, F). This is always a Banach space, even
if X is incomplete. Recall that the norm k kX ∗ (or simply k k∗ ) on X ∗ is the operator norm:

kf kX ∗ = sup |f (x)|, f ∈ X ∗.
x∈BX

The Dual of a Hilbert Space

If we have a vector space H with an inner product h · , · iH , then for every fixed y ∈ H, we obtain a linear
functional fy ∈ H ∗ , defined by fy (x) = hx, yiH . We have

|fy (x)| = |hx, yiH | ≤ kxkH kykH

by the Cauchy-Schwarz inequality. Moreover,

|fy (y)| = kyk2H .

Hence kfy kH ∗ = kykH . It turns out that all elements of H ∗ are of this form if H is a Hilbert space:

Theorem 3.1 (Riesz Representation Theorem). Let H be a Hilbert space and f ∈ H ∗ . Then there exists
a unique y ∈ H such that f (x) = hx, yiH for all x ∈ H. Furthermore, kykH = kf kH ∗ .

Proof. We first prove existence. If f = 0, then we simply choose y = 0. Otherwise, we set

V = ker f = {x ∈ H : f (x) = 0}.

This is a closed linear subspace of H. Hence H = V ⊕ V ⊥ by Theorem 1.6. As f 6= 0, we have V 6= H. Hence

there exists a vector u ∈ V ⊥ with kukH = 1. We show that a suitable multiple of u provides the required y.
For an arbitrary x ∈ H, consider
z = f (u)x − f (x)u.
We have
f (z) = f (u)f (x) − f (x)f (u) = 0.
So z ∈ V . It follows that hz, uiH = 0, i.e.,

f (u)hx, uiH − f (x)hu, uiH = 0

and since hu, u, iH = 1 we obtain

f (x) = f (u)hx, uiH = hx, f (u)uiH
so we can take y = f (u)u.
Next we show uniqueness. Suppose that y1 , y2 ∈ H satisfy hx, y1 iH = hx, y2 iH for all x ∈ H. Then in
particular for x = y1 − y2 ,

0 = hy1 − y2 , y1 iH − hy1 − y2 , y2 iH = ky1 − y2 k2H .

Therefore, we have y1 = y2 .
Finally, the identity kykH = kf kH ∗ has been verified at the beginning of this subsection.

17
 The theorem implies that there is a bijection Φ : H → H ∗ that maps a point y ∈ H to the functional
fy =: Φ(y), where
fy (x) = hx, yiH , x ∈ H.
If F = R, then can be checked by a direct calculation that Φ is linear. If F = C, then it is not linear. Instead,
for x, y, z ∈ H and α ∈ F, we have

Φ(αx + y)(z) = hz, αx + yiH = ᾱhz, xiH + hz, yiH = ᾱΦ(x)(z) + Φ(y)(z).

That is, we have Φ(αx + y) = ᾱΦ(x) + Φ(y). Nevertheless, the map Φ induces a Hilbert space structure on H ∗ :

Theorem 3.2. Let H be a Hilbert space and consider the above map Φ : H → H ∗ . Then H ∗ , with the inner
product
hf, giH ∗ = hΦ−1 (g), Φ−1 (f )iH ,
is a Hilbert space. Furthermore, Φ is an isometry between H and H ∗ .

Proof. The proof is routine and is not required. Note the orders in which the two inner products are taken.

The Dual Space of `p

Recall that `2 is a Hilbert space, so we can identify `2 with (`2 )∗ . For `p with p 6= 2, this is not possible, but
we have a similar property.
Let p, q ∈ [1, ∞] be conjugate exponents. Let y ∈ `q and consider the linear functional
∞
X
fy (x) = x n yn , x ∈ `p .
n=1

By Hölder’s inequality, the series is convergent, so fy : `p → F is well-defined. Moreover, we have |fy (x)| ≤
kxk`p kyk`q . Thus fy ∈ (`p )∗ . Therefore, we obtain a linear map Φp : `q → (`p )∗ , setting Φp y = fy . i.e.,
∞
X
(Φp y)(x) = x n yn , x ∈ `p .
n=1

Theorem 3.3. The operator Φp is an isometry, i.e., kΦp yk(`p )∗ = kyk`q for every y ∈ `p . If p < ∞, then Φp
is bijective.

Proof. Let y ∈ `q and set fy = Φp y. We have seen that kfy k(`p )∗ ≤ kyk`q by Hölder’s inequality. To obtain the
reverse inequality, we distinguish three cases.
Case 1. If 1 < q < ∞, define (
|yn |q−2 ȳn if yn 6= 0,
xn =
0 if yn = 0.
Then for x = (xn )n∈N , we have
∞
|yn |q = kykq`q .
X
fy (x) =
n=1

Note also that !1/p

∞
q/p
= kyk`q = kyk`q−1
X
kxk`p = |yn |p(q−1) q ,
n=1

18
 q
as p = q−1 . Therefore,
fy (x) = kykq−1
`q kyk` = kxk` kyk` .
q p q

So kfy k(`p )∗ = kyk`q .

Case 2. If q = 1, define x ∈ `∞ as before, that is,
(
|yn |−1 ȳn if yn 6= 0,
xn =
0 if yn = 0.

Except for the trivial case y = 0, we have kxk`∞ = 1. We still compute

fy (x) = kyk`1 ,

so kfy k(`∞ )∗ = kyk`1 .

Case 3. If q = ∞, fix k ∈ N and define x(k) ∈ `1 by
(
0 if n 6= k,
x(k)
n =
1 if n = k.

Then kx(k) k`1 = 1, and we have

|fy (x(k) )| = |yk |.
So
kfy k(`1 )∗ = sup |fy (x)| ≥ sup |fy (x(k) )| = sup |yk | = kyk`∞ .
kxk(`1 )∗ =1 k∈N k∈N

Thus we have kfy k(`1 )∗ = kyk`∞ in this case as well, and this concludes the first part of the proof.
Note also that an isometry is necessarily injective, so for the second part, it suffices to show that Φp is
surjective. Suppose that p < ∞. Let f ∈ (`p )∗ . Consider the sequences

e1 = (1, 0, 0, . . .),
e2 = (0, 1, 0, 0, . . .),
e3 = (0, 0, 1, 0, 0, . . .),
..
.

Define yn = f (en ) for n ∈ N and consider the sequence y = (yn )n∈N . We claim that y ∈ `q . This is clear if
q = ∞, as |yn | = |f (en )| ≤ kf k(`1 )∗ . If q < ∞, fix N ∈ N and define
N
X
x(N ) = |yn |q−2 ȳn en .
n=1

Then x(N ) ∈ `p with

N
! p1 N
!1− 1q
X X
kx(N ) k`p = |yn |(q−1)p = |yn |q .
n=1 n=1
Moreover,
N
X N
X
|yn |q = |yn |q−2 ȳn f (en ) = f (x(N ) )
n=1 n=1
N
!1− 1q
X
≤ kf k(`p )∗ kx(N ) k`p = kf k(`p )∗ |yn |q .
n=1

19
So
N
! 1q
X
|yn |q ≤ kf k(`p )∗ .
n=1
Letting N → ∞, we obtain kyk`q ≤ kf k(`p )∗ . In particular y ∈ `q .
Next we claim that f = Φp y. Let x ∈ `p . Then we have
∞
X
x= xn e n .
n=1

Hence
∞
X ∞
X
f (x) = xn f (en ) = xn yn = Φp y(x),
n=1 n=1
as required.

The Hahn-Banach Theorem for real vector spaces

In this section we will see an important tool for the construction of bounded linear functionals with specific
properties. But we need some preparation first.

Definition. A partial order ≤ on a set S is a relation that satisfies:

(i) a ≤ a for every a ∈ S,
(ii) if a, b ∈ S with a ≤ b and b ≤ a, then a = b, and
(iii) if a, b, c ∈ S with a ≤ b and b ≤ c, then a ≤ c.
A subset T ⊂ S is totally ordered if for all a, b ∈ T , either a ≤ b or b ≤ a. We say that a0 ∈ S is an upper bound
for T ⊂ S if a ≤ a0 for every a ∈ T . We say that a0 is a maximal element of T if a0 ∈ T and

∀a ∈ T : a0 ≤ a ⇒ a = a0 .

We also need a lemma from set theory. The proof is outside the scope of this course. However, it is known
that the lemma is equivalent to the so-called axiom of choice (see P.R. Halmos, Naive Set Theory, for a proof),
so for our purposes, the statement can be taken as an axiom.

Lemma 3.4 (Zorn’s Lemma). Let S 6= ∅ be a partially ordered set. If every totally ordered subset of S has
an upper bound, then S has a maximal element.

Before we state the main result of this section, we prove another statement that relies on similar arguments.

Theorem 3.5. Let V be a vector space over an arbitrary field k. Then V contains a maximal linearly inde-
pendent set S. Moreover, this set satisfies V = span S.

Remark. The expression ‘maximal’ refers to the partial order ⊂ on the power set of V ; thus, if S 0 is a linearly
independent subset of V and S ⊂ S 0 , then S = S 0 .
A set with these properties is a basis of V in the sense of linear algebra (i.e., a Hamel basis). The statement
can thus be summarised as follows: every vector space has a Hamel basis.

Proof. Let A comprise all linearly independent subsets of V . Then A is non-empty, as ∅ ∈ A. The inclusion
relation ⊂ gives a partial order of A. We want to show that A satisfies the hypotheses of Zorn’s lemma.
Suppose that U ⊂ A is totally ordered. Define
[
T = U.
U ∈U

20
We claim that T is linearly independent. To see this, let v1 , . . . , vn ∈ T be distinct vectors and let α1 , . . . , αn ∈ k
such that
α1 v1 + · · · + αn vn = 0.
Then for every i = 1, . . . , n, there exists a Ui ∈ U with vi ∈ Ui . Since U is totally ordered, there exists an
i0 ∈ {1, . . . , n} such that Ui ⊂ Ui0 for i = 1, . . . , n. Thus v1 , . . . , vn ∈ Ui0 . Since Ui0 is linearly independent,
this means that
α1 = · · · = αn = 0.
So T is linearly independent, and therefore T ∈ A. Obviously, T is an upper bound for U.
By Zorn’s lemma, there exists an maximal element S of A, as required. It remains to show that span S = V .
Let v0 ∈ V \S. Then S ∪ {v0 } is not linearly independent. Hence there exist certain vectors v1 , . . . , vn ∈ S
and scalars α0 , . . . , αn ∈ k with (α0 , . . . , αn ) 6= (0, . . . , 0) such that

α0 v0 + α1 v1 + · · · + αn vn = 0.

Since S is linearly independent, we must have α0 6= 0. Thus

α1 αn
v0 = − v1 − · · · − vn .
α0 α0
In particular, we have v0 ∈ span S.

Definition. Let V be a real vector space. We say that a functional p : V → R is sublinear if for all x, y ∈ V
and α ≥ 0,
(i) p(x + y) ≤ p(x) + p(y), and
(ii) p(αx) = αp(x).

Property (i) in the definition is called subadditivity and (ii) is called positive homogeneity.
Example. If p is linear, then it is also sublinear.
Example. If k · k is a norm on V and λ > 0, then p(x) = λkxk is sublinear.

Theorem 3.6 (Hahn-Banach). Let V be a real vector space and p : V → R a sublinear functional. Suppose
that L ⊂ V is a linear subspace and f : L → R a linear functional such that f (x) ≤ p(x) for all x ∈ L. Then
there exists a linear extension F : V → R such that
(i) F (x) = f (x) for all x ∈ L, and
(ii) F (x) ≤ p(x) for all x ∈ V .

Remark. From (ii) it also follows that

F (x) = −F (−x) ≥ −p(−x).

If p is even (for example if p is a norm) we then conclude that |F (x)| ≤ p(x).

In the rest of this subsection, we prove Theorem 3.6. We divide the proof into three steps.
Step 1: set-up.
Consider the set S, consisting of all pairs (U, φ), where U is a linear subspace of V and φ : U → R is linear
with φ ≤ p. Define a partial order ≤ on S as follows: for (U, φ), (U 0 , φ0 ) ∈ S, we have

(U, φ) ≤ (U 0 , φ0 )

21
if, and only if, U ⊂ U 0 and φ0 |U = φ. We note that the pair (L, f ) belongs to S. Let

S0 = {(U, φ) ∈ S : (L, f ) ≤ (U, φ)}.

Then S0 6= ∅ as (L, f ) ∈ S0 . We want to apply Zorn’s lemma to S0 .

Step 2: verify the hypotheses of Zorn’s lemma.
Consider a totally ordered subset T ⊂ S0 . If T = ∅, then (L, f ) is an upper bound. Otherwise, set
[
W = U.
(U,φ)∈T

Define ψ : W → R as follows: if (U, φ) ∈ T and x ∈ U , set ψ(x) = φ(x).

Claim: The functional ψ is well-defined.

Proof. If (U, φ), (U 0 , φ0 ) ∈ T with x ∈ U ∩ U 0 , then either (U, φ) ≤ (U 0 , φ0 ) or (U 0 , φ0 ) ≤ (U, φ) by the total
ordering. In both cases, we have φ(x) = φ0 (x) by the definition of ≤. So the value of φ(x) does not depend on
the choice of (U, φ).

Claim: The set W is a linear subspace and ψ is linear.

Proof. Let x1 , x2 ∈ W and α ∈ R. There exist (U1 , φ1 ), (U2 , φ2 ) ∈ T with x1 ∈ U1 and x2 ∈ U2 . By the total
ordering, we have U1 ⊂ U2 or U2 ⊂ U1 . Say U1 ⊂ U2 . Then x1 , x2 ∈ U2 and thus αx1 + x2 ∈ W . Moreover,

ψ(αx1 + x2 ) = φ2 (αx1 + x2 ) = αφ2 (x1 ) + φ2 (x2 ) = αψ(x1 ) + ψ(x2 ).

Finally, we have ψ ≤ p in W by construction and also (L, f ) ≤ (W, ψ). Hence (W, ψ) ∈ S0 . Again by
construction, (W, ψ) is an upper bound for T .
Step 3: construct the extension.
Zorn’s lemma now implies that there exists a maximal element (X, F ) of S0 .
Claim: X = V .

Proof. Assume for contradiction that X 6= V . Choose v ∈ V \X. Then for all x, y ∈ X, we have

F (x) + F (y) = F (x + y) ≤ p(x + y) ≤ p(x − v) + p(y + v).

Hence
F (x) − p(x − v) ≤ p(y + v) − F (y).
Set
κ = sup (F (z) − p(z − v)),
z∈X

so that, for all x, y ∈ X,

F (x) − p(x − v) ≤ κ ≤ p(y + v) − F (y);
i.e.,
F (x) − κ ≤ p(x − v) and F (y) + κ ≤ p(y + v). (5)
Now define
G(x + αv) = F (x) + ακ, x ∈ X, α ∈ R.

22
Then G : X ⊕ span{v} → R is linear and G|X = F . Moreover from (5), for α > 0 we have
 x  x 
G(x + αv) = F (x) + ακ = α F + κ ≤ αp + v = p(x + αv),
α α
and for α < 0,
   
x
G(x + αv) = F (x) − |α|κ = |α| F −κ
|α|
 
x
≤ |α|p − v = p(x + αv).
|α|
Therefore, we have (X ⊕ span{v}, G) ∈ S0 and (X, F ) ≤ (X ⊕ span{v}, G). But we do not have equality, which
contradicts the maximality of (X, F ). Hence X = V .

Now by construction, F has both the required properties (i) and (ii). This completes the proof of the
Hahn-Banach Theorem.

Corollaries of the Hahn-Banach Theorem

Here we collect a few useful results that follow almost immediately from the Hahn-Banach Theorem 3.6.

Corollary 3.7. Let X be a normed vector space and V ⊂ X a linear subspace. Suppose that f ∈ V ∗ . Then
there exists an extension F ∈ X ∗ with F (x) = f (x) for all x ∈ V and kF kX ∗ = kf kV ∗ .

Proof. First assume that F = R. Apply the Hahn-Banach Theorem 3.6 with

p(x) = kf kV ∗ kxkX .

We obtain a linear functional F : X → R with F |V = f and

F (x) ≤ p(x) = kf kV ∗ kxkX

for all x ∈ X. Moreover,

−F (x) = F (−x) ≤ p(−x) = kf kV ∗ kxkX ,
so that we deduce F ∈ X ∗ and kF kX ∗ ≤ kf kV ∗ . On the other hand,

kF kX ∗ = sup |F (x)| ≥ sup |F (x)| = kf kV ∗ .

x∈X x∈V
kxkX ≤1 kxkX ≤1

Thus kF kX ∗ = kf kV ∗ .
Now we consider the case F = C. Then we use the fact that a complex linear functional is uniquely
determined by its real part. Define g(x) = Re f (x). Regarding X as a real vector space, we can apply the
previous arguments to g. Thus we construct an R-linear functional G : X → R with G|V = g = Re f . Now
define
F (x) = G(x) − iG(ix).
This is a C-linear functional by Exercise Sheet 7. For x ∈ V , we have

Re F (x) = g(x) = Re f (x),

and therefore also

Im F (x) = − Re(iF (x)) = − Re F (ix) = − Re f (ix) = Im f (x).
Hence F is an extension of f . Furthermore, we have kF kX ∗ = kf kV ∗ by Exercise Sheet 7.

23
Corollary 3.8. Let L be a closed linear subspace of a normed vector space X. Suppose that x0 ∈ X\L. Then
there exists an F ∈ X ∗ with F = 0 on L and F (x0 ) = 1.

Proof. Define V = span{x0 } ⊕ L and let f : V → F be the functional with

f (λx0 + w) = λ, λ ∈ F, w ∈ L.

This is well-defined and linear. Clearly f = 0 on L and f (x0 ) = 1. If we can show that f ∈ V ∗ , then the claim
follows from Corollary 3.7.
Assume for contradiction that f is unbounded. Then there exists a sequence of points λn x0 + wn ∈ V (i.e.,
λn ∈ F and wn ∈ L), such that
kλn x0 + wn k ≤ 1
for every n ∈ N, but |λn | → ∞ as n → ∞. Then
wn
x0 + →0 as n → ∞.
λn
So
wn
x0 = − lim ∈ L,
n→∞ λn
as L is closed. But this contradicts the assumptions.

Corollary 3.9. Let X be a normed vector space and x0 ∈ X\{0}. Then there exists an f ∈ X ∗ with kf kX ∗ = 1
and f (x0 ) = kx0 kX .

Proof. This follows from Corollary 3.7. A proof directly from the Hahn-Banach Theorem, when F = R, is set
on Probem Sheet 6.

Corollary 3.10. Let X be a normed vector space and x1 , x2 ∈ X with x1 6= x2 . Then there exists an f ∈ X ∗
with f (x1 ) 6= f (x2 ).

Proof. This follows from Corollary 3.8.

Corollary 3.11. Let X be a normed vector space. Then for every x ∈ X,

kxkX = sup |f (x)|.

f ∈X ∗
kf kX ∗ ≤1

Proof. If f ∈ X ∗ with kf kX ∗ ≤ 1, then it is clear that |f (x)| ≤ kxkX . On the other hand, there exists an
f ∈ X ∗ with kf kX ∗ ≤ 1 and |f (x)| = kxkX (by Corollary 3.9 if x 6= 0 and trivially if x = 0). The result
follows.

Reflexivity
For conjugate exponents p, q ∈ (1, ∞), we have seen that we can identify (`p )∗ with `q and (`q )∗ with `p . So
(`p )∗∗ has the same structure as `p . This is not the case for other spaces in general; indeed, the spaces `1 and
`∞ provide counterexamples. But a normed vector space X can always be embedded in X ∗∗ .

Theorem 3.12. Let X be a normed vector space and for x ∈ X define

(JX x)(f ) = f (x) ∀f ∈ X ∗ .

Then JX is a linear isometry from X onto a linear subspace of X ∗∗ .

24
Proof. The verification of linearity of JX x for each x is routine. For f ∈ X ∗ , we have

|Fx (f )| = |f (x)| ≤ kf kX ∗ kxkX .

It follows that Fx ∈ X ∗∗ and kFx kX ∗∗ ≤ kxkX . By Corollary 3.9, there exists a g ∈ X ∗ with kgkX ∗ = 1 and
g(x) = kxkX . So |Fx (g)| = kxkX , and therefore kFx kX ∗∗ = kxkX .
Finally the linearity of x 7→ JX x is routine.

Definition. For a normed vector space X, the operator JX : X → X ∗∗ , defined in Theorem 3.12 is called the
canonical embedding of X in X ∗∗ .
The notation x
b := JX (x) is frequently used.
If JX (X) = X ∗∗ then X is said to be reflexive.
A linear isometric bijection is called an isometric isomorphism.

Remarks.
• A linear isometry is necessarily injective, i.e., we have ker JX = {0}.
• If X is reflexive then X is complete since X ∗∗ , being the dual space of X ∗ , is complete.
• Reflexivity is not quite the same as saying that X and X ∗∗ are isometrically isomorphic. For reflexivity, the
isometric isomorphism must be given by the specific operator JX .
Examples.
• Any finite-dimensional normed vector space is reflexive.
• It follows from the Riesz representation theorem 3.1 that any Hilbert space is reflexive. (However, a bit of
additional work is necessary to show that the resulting map H → H ∗∗ coincides with JH .)
• It follows from Theorem 3.3 that `p is reflexive for 1 < p < ∞ (see Exercise Sheet 9 for the details).
• On the other hand, both `1 and `∞ are not reflexive (cf. Exercise Sheet 9).

Theorem 3.13. If X is reflexive and Y ⊂ X is a closed linear subspace, then Y is reflexive.

Proof. Let F ∈ Y ∗∗ . Then we obtain an element F0 ∈ X ∗∗ with

F0 (f ) = F (f |Y ), f ∈ X ∗.

As X is reflexive, there exists an x0 ∈ X with F0 = JX x0 . We claim that x0 ∈ Y .

Assume for contradiction that x0 6∈ Y . Then by Corollary 3.8, there exists a g ∈ X ∗ with g = 0 on Y and
g(x0 ) = 1. So
1 = g(x0 ) = (JX x0 )(g) = F0 (g) = F (g|Y ) = 0,
which is the desired contradiction. Hence x0 ∈ Y .
Next we claim that F = JY x0 . In order to prove this, let φ ∈ Y ∗ . According to Corollary 3.7, there exists
an extension f ∈ X ∗ of φ. Now
F (φ) = F0 (f ) = f (x0 ) = φ(x0 ).
So F = JY x0 .

25
Convexity
Definition. Let V be a (real or complex) vector space. A subset C ⊂ V is called convex if for all x, y ∈ C and
λ ∈ (0, 1), the point λx + (1 − λ)y belongs to C as well.

Examples.
• A linear subspace is convex.
• In a normed vector space X, consider the open unit ball B ◦ = {x ∈ X : kxk < 1}. Let x, y ∈ B ◦ and λ ∈ (0, 1).
Then
kλx + (1 − λ)yk ≤ λkxk + (1 − λ)kyk < 1.
So B ◦ is convex. Similarly, any open or closed ball in X is convex.
• Suppose that V is a vector space, f : V → F is a linear functional, and D ⊂ F is convex. Then

C = f −1 (D) = {x ∈ V : f (x) ∈ D}

is convex, as for x, y ∈ C and λ ∈ (0, 1),

f (λx + (1 − λ)y) = λf (x) + (1 − λ)f (y) ∈ D.

Definition. Let C be a subset of a real normed vector space X with 0 ∈ C ◦ . The Minkowski functional h of
C is defined by
h(x) = inf{α > 0 : x ∈ αC}, x ∈ X.

Lemma 3.14. Let X be a real normed vector space and C ⊂ X a convex set with 0 ∈ C ◦ . Then the Minkowski
functional h of C satisfies
(i) h(λx) = λh(x) for all x ∈ X and λ ≥ 0,
(ii) h(x + y) ≤ h(x) + h(y) for all x, y ∈ X,
(iii) h(x) < 1 if, and only if, x ∈ C ◦ ,
(iv) h(x) ≤ 1 if, and only if, x ∈ C, and
(v) there exists a number A such that h(x) ≤ AkxkX for all x ∈ X.

Proof. (i) For λ = 0, this is clear, as h(0) = 0. For λ > 0,

 
β
h(λx) = inf{β > 0 : λx ∈ βC} = inf β > 0 : x ∈ C
λ
= inf{λγ : γ > 0, x ∈ γC} = λh(x).

(ii) Let x, y ∈ X. Fix ε > 0 and choose α, β > 0 such that x ∈ αC, y ∈ βC, and

α ≤ h(x) + ε, β ≤ h(y) + ε.

Set x0 = x
α and y 0 = βy . Then x0 ∈ C and y 0 ∈ C. Hence
 
0 0 α β
x + y = αx + βy = (α + β) x0 + y0 ∈ (α + β)C.
α+β α+β

Thus h(x + y) ≤ α + β ≤ h(x) + h(y) + 2ε. Now let ε → 0.

(iii) If x ∈ C ◦ , then there exists an ε > 0 such that (1 + ε)x ∈ C. So h(x) ≤ 1
1+ε < 1.

26
 Conversely, suppose that h(x) < 1. Then there exists an ε > 0 such that y := (1 + ε)x ∈ C. Since 0 ∈ C ◦ ,
there exists an open ball rB ◦ ⊂ C, where
B ◦ = {x ∈ X : kxkX < 1}.
Since C is convex,
εr y ε
x+ B◦ = + rB ◦ ⊂ C.
1+ε 1+ε 1+ε
So x ∈ C ◦ .
(v) Choose r > 0 such that rB ◦ ⊂ C ◦ . Then it follows from (iii) that h < 1 on rB ◦ . Hence by (i)
 
2kxkX rx 2
h(x) = h ≤ kxkX
r 2kxkX r
for all x ∈ X\{0}. Moreover, we know that h(0) = 0.
(iv) Firstly note that (ii) and (v) imply that h is continuous, since for z, w ∈ X,
h(z) − h(w) = h((z − w) + w) − h(w) ≤ h(z − w) ≤ Akz − wkX .
By definition, we have h ≤ 1 on C. Hence h ≤ 1 on C.
Conversely, suppose that h(x) ≤ 1. If h(x) < 1, then x ∈ C ◦ ⊂ C by (iii). If h(x) = 1, then there exist
αn > 0 with αn−1 x ∈ C for every n ∈ N and limn→∞ αn = 1. Hence x = limn→∞ αn−1 x ∈ C.

Theorem 3.15. Let X be a real normed vector space and C ⊂ X a convex set with 0 ∈ C ◦ . Then for any
x0 ∈ X\C, there exists an f ∈ X ∗ such that
f (x0 ) > sup f (x).
x∈C

Proof. Let h be the Minkowski functional of C. Let L = span{x0 } and define φ : L → R by

φ(αx0 ) = αh(x0 ), α ∈ R.
This is a linear functional, and we have φ ≤ h on L. Furthermore, h is sublinear by Lemma 3.14. Hence
according to the Hahn-Banach theorem, there exists a linear extension f : X → R of φ with
f (x) ≤ h(x) ≤ AkxkX
and
−f (x) = f (−x) ≤ h(−x) ≤ AkxkX
for all x ∈ X, where A is the constant from Lemma 3.14. In particular, we have f ∈ X ∗ . Moreover, by
statement (iv) of Lemma 3.14, we have f ≤ 1 in C and f (x0 ) = h(x0 ) > 1, as x0 6∈ C.

Corollary 3.16. Let X be a real normed vector space and C ⊂ X non-empty, closed, and convex. Let
x0 ∈ X\C. Then there exists an f ∈ X ∗ with
f (x0 ) > sup f (x).
x∈C
Remarks. In contrast to Theorem 3.15, it is not required here that 0 ∈ C ◦ .

Proof. Fix y0 ∈ C. Then it suffices to prove the statement for C − y0 instead of C and for the point x0 − y0
instead of x0 . So we can assume without loss of generality that 0 ∈ C.
Let B ◦ be the open unit ball in X and
Cr = C + rB ◦ = {x + ry : x ∈ C, y ∈ B ◦ }
for r > 0. Then Cr is convex. For r sufficiently small, we have x0 6∈ Cr . Hence we can apply Theorem 3.15 and
the claim follows.

27
4 Weak and Weak* Convergence
Basic Properties
We have seen that closed, bounded sets need not be compact in a Banach space. In fact, according to Theorem
1.17, the closed unit ball is never compact in an infinite-dimensional space. However, compactness is an
important and useful concept in analysis. A way to recover some sort of compactness is to change our notion
of convergence.

Definition. Let X be a normed vector space. A sequence (xn )n∈N converges weakly to x0 ∈ X if for all f ∈ X ∗ ,

f (x0 ) = lim f (xn ).

n→∞

A sequence (fn )n∈N in X ∗ converges weak* to f0 ∈ X ∗ if for all x ∈ X,

f0 (x) = lim fn (x).

n→∞

Remarks.
∗
• We use the notation xn * x0 for weak convergence and fn * f0 for weak* convergence.
• We sometimes use the expression strong convergence for convergence with respect to the norm on X to avoid
confusion with weak convergence.
• Note that weak* convergence is just pointwise convergence of functionals.

Theorem 4.1. If X is a normed vector space, then weak limits in X and weak* limits in X ∗ are unique.

Proof. Suppose that (xn )n∈N is a sequence in X that converges weakly to x0 ∈ X. Let y0 ∈ X with y0 6= x0 .
Then by Corollary 3.10, there exists an f ∈ X ∗ with f (x0 ) 6= f (y0 ). Hence

lim f (xn ) = f (x0 ) 6= f (y0 ).

n→∞

Thus we cannot have weak convergence to y0 as well.

∗ ∗
Now suppose that (fn )n∈N is a sequence in X ∗ with fn * f0 and fn * g0 weak* in X ∗ . Then for every
x ∈ X,
f0 (x) = lim fn (x) = g0 (x).
n→∞
Thus f0 = g0 .

Theorem 4.2. Let X be a normed vector space.

(i) Every weakly convergent sequence in X is bounded.
(ii) If X is complete, then every weak* convergent sequence in X ∗ is bounded.
∗
Proof. (ii) Suppose that fn * f0 weak* in X ∗ . Then for every x ∈ X, the sequence (fn (x))n∈N is convergent
and thus bounded. According to the Uniform Boundedness Theorem 2.4, the sequence (fn )n∈N is bounded in
X ∗ (since X is assumed complete).
(i) Consider the canonical embedding JX : X → X ∗∗ with (JX x)(f ) = f (x) for x ∈ X and f ∈ X ∗ . Let
(xn )n∈N be a sequence in X with xn * x0 weakly. Define x̂n = JX xn and x̂0 = JX x0 . Then for every f ∈ X ∗ ,

x̂n (f ) = f (xn ) → f (x0 ) = x̂0 (f ).

∗
That is, we have x̂n * x̂0 weak* in X ∗∗ = (X ∗ )∗ . Since X ∗ is complete (Theorem 2.2) it follows from (ii)
that the sequence (x̂n )n∈N is bounded in X ∗∗ . But we have seen in Theorem 3.12 that kxn kX = kx̂n kX ∗∗ . So
(xn )n∈N is bounded.

28
Theorem 4.3. Suppose that C is a convex, closed subset of a normed vector space X. If (xn )n∈N is a sequence
in C with weak limit x0 , then x0 ∈ C.

Proof. We assume for contradiction that x0 6∈ C.

If F = R, then by Corollary 3.16, there exists an f ∈ X ∗ with f (x0 ) > supx∈C f (x). Therefore,

f (x0 ) = lim f (xn ) ≤ sup f (x) < f (x0 ),

n→∞ x∈C

which is impossible.
If F = C, then we first regard X as a real vector space. We choose bounded real linear f as above and
define g(x) = f (x) − if (ix), so that g ∈ X ∗ (cf. Corollary 3.7). Then we have the contradiction

Re g(x0 ) = lim Re g(xn ) < Re g(x0 ).

n→∞

Theorem 4.4. Let X be a normed vector space.

(i) If xn * x0 weakly in X, then
kx0 kX ≤ lim inf kxn kX .
n→∞
∗
(ii) If fn * f0 weak* in X ∗ , then
kf0 kX ∗ ≤ lim inf kfn kX ∗ .
n→∞

Proof. (i) It is enough to consider the case when x0 6= 0. According to Corollary 3.9 we can choose f ∈ X ∗
such that kf k∗ = 1 and f (x0 ) = kx0 k. Then

kx0 k = |f (x0 )| = | lim f (xn )| = lim |f (xn )| ≤ lim inf kf k∗ kxn k = lim inf kxn k.
n→∞ n→∞ n→∞ n→∞

(ii) Let ε > 0 and from the definition of k k∗ choose x ∈ X with kxk ≤ 1 and |f0 (x)| ≥ kf0 k∗ − ε. Then

kf0 k∗ ≤ |f0 (x)| + ε = lim |fn (x)| + ε ≤ lim inf kfn k∗ kxk + ε ≤ lim inf kfn k∗ + ε.
n→∞ n→∞ n→∞

This holds for all ε > 0 hence the result.

It follows that the closed unit ball B in X is closed under weak convergence. That is, the weak limit of a
sequence in B belongs to B, too. (This statement can also be verified using Theorem 4.3.) Similarly, the closed
unit ball in X ∗ is closed under weak* convergence. However, we will see in the next subsection that we have in
fact a much stronger statement.

Lemma 4.5. Let X be a normed vector space. If X ∗ is separable, then X is separable.

Proof. Let (fn )n∈N be a dense sequence in X ∗ . For each n ∈ N, choose a point xn ∈ X with kxn kX ≤ 1 and
|fn (xn )| ≥ 21 kfn kX ∗ . Set
Y = span{xn : n ∈ N}.
We claim that Y = X.
Assume that Y 6= X for contradiction. Choose a unit vector x0 ∈ X \ Y . Then by Corollary 3.8, there
exists an F ∈ X ∗ with F = 0 on Y , but F (x0 ) = 1. Let ε > 0 and choose n ∈ N such that kfn − F kX ∗ < ε.
Now kF k∗ ≥ 1 so
kfn k∗ ≥ kF k∗ − kF − fn k∗ > 1 − ε.

29
Then

|F (xn )| ≥ |fn (xn )| − |(F − fn )(xn )|

1
≥ kfn k∗ − kF − fn k∗ kxn kX
2
1 1 3ε
> (1 − ε) − ε = − .
2 2 2
But if we choose ε = 14 then this contradicts the condition that F = 0 on Y , since xn ∈ Y .
So span{xn : n ∈ N} is dense in X. Then the set of all linear combinations of {xn : n ∈ N} with coefficients
in Q or Q + iQ is also dense, and this set is countable.

Weak and Weak* Sequential Compactness

Recall that in a metric space, compactness is equivalent to sequential compactness. In other words, a set is
compact if, and only if, every sequence in the set has a convergent subsequence. The notions of weak or weak*
convergence do not correspond to any metric, but we can still ask the question whether or not a given sequence
has a weakly or weak* convergent subsequence. That is, the notion of weak or weak* sequential compactness
is meaningful.

Theorem 4.6 (Banach). Let X be a separable normed vector space. Then every sequence in the closed unit
ball B ∗ = {f ∈ X ∗ : kf kX ∗ ≤ 1} has a subsequence converging weak* to an element of B ∗∗ .

Proof. As X is separable, there exists a dense sequence (xk )k∈N in X. Now consider a sequence (fn )n∈N in B ∗ .
Since |fn (x1 )| ≤ kfn kX ∗ kx1 kX ≤ kx1 kX for every n ∈ N, the sequence (fn (x1 ))n∈N in F is bounded. Therefore,
there exists a convergent subsequence. That is, there exists an infinite set Λ1 ⊂ N such that the subsequence
(fn (x1 ))n∈Λ1 is convergent.
Next we consider the sequence (fn (x2 ))n∈Λ1 . It is bounded as well, hence there exists an infinite set Λ2 ⊂ Λ1
such that the subsequence (fn (x2 ))n∈Λ2 is convergent.
Similarly, we construct Λ3 , Λ4 , Λ5 , . . . such that each Λk is an infinite subset of Λk−1 for k = 2, 3, 4, . . . and
(fn (xk ))n∈Λk is convergent. We now want to construct an increasing sequence (n` )`∈N in N such that we have
simultaneous convergence of (fn` (xk ))`∈N for all k ∈ N. This works with a ‘diagonal sequence argument’ as
follows.
Let n1 = min Λ1 . Now define n2 , n3 , . . . recursively. Assuming that n` has been chosen, let

n`+1 = min{m ∈ Λ`+1 : m > n` }.

Then for every k, we have nk , nk+1 , nk+2 , . . . ∈ Λk . Therefore, (fn` (xk ))`∈N is convergent.
We claim that (fn` (x))`∈N is convergent for all x ∈ X. In order to prove this, fix x ∈ X. Let ε > 0 and
choose k ∈ N such that kx − xk kX ≤ ε. We know that (fn` (xk ))`∈N is convergent and thus Cauchy. Therefore,
there exists an L ∈ N such that for `, m ≥ L,

|fn` (xk ) − fnm (xk )| ≤ ε.

Thus

|fn` (x) − fnm (x)| ≤ |fn` (x) − fn` (xk )| + |fn` (xk ) − fnm (xk )|
+ |fnm (xk ) − fnm (x)|
≤ kfn` kX ∗ kx − xk k + ε + kfnm kX ∗ kx − xk k ≤ 3ε.

30
It follows that (fn` (x))`∈N is Cauchy, and since F is complete, it is convergent, as claimed.
Finally, we define
f (x) = lim fn` (x), x ∈ X.
`→∞
Then f is linear and |fn (x)| ≤ kxk for all n ∈ N and x ∈ X so

|f (x)| = lim |fn` (x)| ≤ kxk.

`→∞

∗
Hence f ∈ X ∗ . The weak* convergence fn` * f follows from the construction.

Now suppose that we want a similar statement for weak convergence. If X happens to be reflexive, then
we can identify X with X ∗∗ and weak convergence in X with weak* convergence in X ∗∗ via the canonical
embedding JX : X → X ∗∗ . We can then apply Theorem 4.6 and thus obtain sequential compactness of the
closed unit ball in X, provided that X ∗ is separable. If X ∗ is not separable, then we need to use Lemma 4.5 in
addition.

Theorem 4.7. Let X be a reflexive Banach space. Then any sequence in the closed unit ball B of X has a
subsequence converging weakly to an element of B.

Proof. Firstly assume that X is separable. Consider the canonical embedding JX : X → X ∗∗ , which is isometric
and bijective since X is reflexive. Then X ∗∗ is separable, and by Lemma 4.5, the space X ∗ is separable as well.
Let B and B ∗∗ be the closed unit balls in X and X ∗∗ , respectively.
Suppose that (xn )n∈N is a sequence in B. Defining x̂n = JX xn , we obtain a sequence in B ∗∗ . By Theorem
4.6, there exists a weak* convergent subsequence (x̂nk )k∈N . Let x̂0 ∈ B ∗∗ be the weak* limit and define
−1
x0 = JX x̂0 . Then for any f ∈ X ∗ ,

f (x0 ) = x̂0 (f ) = lim x̂nk (f ) = lim f (xnk ).

k→∞ k→∞

That is, we have xnk * x0 weakly in X.

If X is not separable, then we proceed as follows. Let (xn )n∈N be a sequence in B. Define

Y = span{xn : n ∈ N}.

This subspace is clearly separable. According to Theorem 3.13, it is reflexive. Therefore, the previous arguments
give a subsequence (xnk )k∈N that converges weakly in Y , say with weak limit x0 ∈ Y . We need to show that
xnk * x0 weakly in X as well.
Let f ∈ X ∗ . Define g = f |Y . Then g ∈ Y ∗ . Hence

f (x0 ) = g(x0 ) = lim g(xnk ) = lim f (xnk ).

k→∞ k→∞

Thus xnk * x0 weakly in X.

Remark. If p, q are conjugate exponents and 1 < q ≤ ∞ then `q is isometrically isomorphic to (`p )∗ , hence we
can, by slight misuse of terminology, talk about the weak* topology on `q . This is especially interesting in the
case q = ∞, when the dual space of `∞ is not isometrically isomorphic to `1 . More exactly, if f n ∈ (`p )∗ for all
n ≥ 0 and y n = Φ−1 n
p (f ) then

∞ ∞
∗
X X
f n * f 0 ⇔ ∀x ∈ `p Φp (y n )(x) → Φ(y 0 )(x) ⇔ ∀x ∈ `p ykn xk → yk0 xk .
n=1 n=1

31
5 Spectral Theory
For a finite-dimensional vector space X, in order to understand a linear operator T : X → X, it is essential to
know the eigenvalues and eigenvectors. For infinite-dimensional spaces, the corresponding concepts still exist,
but there are a few more subtle issues to consider, too.

The Spectrum of an Operator

For a Banach space X and an operator T ∈ B(X), we examine the solvability of an equation of the form

T x − λx = y (6)

for given y ∈ X and λ ∈ F. This leads to the question: is the operator λidX − T invertible?
There are two potential problems when we want to solve equation (6): we could have ker(λidX − T ) 6= {0}
or im(λidX − T ) 6= X. In a finite-dimensional space, the two conditions are equivalent, but not in an infinite-
dimensional space.

Definition. Let X be a Banach space and T ∈ B(X). Suppose that λ ∈ F such that ker(λidX − T ) 6= {0}.
Then λ is called an eigenvalue of T . Any vector u ∈ ker(λidX − T )\{0} is called an eigenvector of T .
The spectrum of T is the set

σ(T ) = {λ ∈ F : λidX − T is not invertible} .

The resolvent set is ρ(T ) = F\σ(T ).

Remark. By Corollary 2.6 of the Open Mapping Theorem, an invertible bounded linear operator from a
Banach space X to itself has a bounded inverse.

Theorem 5.1. Let X be a Banach space and T ∈ B(X) with kT kB(X) < 1. Then idX − T is invertible with
∞
X
−1
(idX − T ) = T n.
n=0

Proof. Firstly, since kT n kB(X) ≤ kT knB(X) , we have absolute convergence of the series. Set
n
X
Sn = T k.
k=0

Then
n
X n+1
X
(idX − T )Sn = Sn − T Sn = Tk − T k = idX − T n+1 → idX
k=0 k=1
in B(X). Moreover, setting
∞
X
S= T n,
n=0
we have
k(idX − T )Sn − (idX − T )SkB(X) ≤ kidX − T kB(X) kSn − SkB(X) → 0
as n → ∞. It follows that
(idX − T )S = lim (idX − T )Sn = idX .
n→∞
Similarly, we show that S(idX − T ) = idX .

32
Corollary 5.2. If X is a Banach space, then the set of all invertible operators in B(X) is open.

Proof. Suppose that T ∈ B(X) is invertible. If S ∈ B(X) then T − S = T (idX − T −1 S) which will be
invertible if kT −1 Sk < 1, by Theorem 5.1. Since kT −1 Sk ≤ kT −1 kkSk it now follows that T − S is invertible if
kSk < kT −1 k−1 , so T is an interior point of the set of invertible elements of B(X).

Theorem 5.3. Let X be a Banach space and T ∈ B(X). Then |λ| ≤ kT kB(X) for every λ ∈ σ(T ). Furthermore,
σ(T ) is compact.

Proof. If |λ| > kT kB(X) , then kλ−1 T kB(X) < 1, so (idX −λ−1 T )−1 exists by Theorem 5.1. Thus (λidX −T )−1 =
λ−1 (idX − λ−1 T )−1 exists as well. That is, we have λ 6∈ σ(T ).
Define the map τ : F → B(X) with τ (λ) = λidX − T . Then τ is continuous. Let G ⊂ B(X) be the set of
all invertible operators. This is open by Corollary 5.2. Hence B(X)\G is closed, and σ(T ) = τ −1 (B(X)\G) is
closed. It is bounded as well, and thus σ(T ) is compact by the Bolzano-Weierstrass theorem.

Compact Linear Operators

Definition. Let X, Y be normed vector spaces. A bounded linear operator T : X → Y is called compact if for
every bounded sequence (xn )n∈N in X, the sequence (T xn )n∈N in Y has a convergent subsequence.

Theorem 5.4. Let X be a normed vector space, let Y be a Banach space, let (Tn )n∈N be a sequence in B(X)
that converges in the operator norm to T ∈ B(X). Then T is compact.

Proof. Let ε > 0 and choose N ∈ N such that kTn − T k < ε for all n ≥ N . Since TN is compact it follows that
TN (BX ) is a compact set so we can choose y1 , . . . , ym ∈ Y such that y1 + εBY , . . . , ym + εBY cover TN (BX ).
Then y1 + 2εBY , . . . , ym + 2εBY cover T (BX ) and since their union is closed they cover T (BX ). It follows that
T (BX ) is compact so T is a compact operator.

Remarks.
• Any bounded finite rank operator (i.e. one whose image is finite-dimensional) must be compact.
• Consequently, any operator that is the limit in B(X, Y ) of a sequence of bounded finite rank operators must
be compact if Y is complete.
Example. Let (λn )n∈N be a sequence in F that converges to 0 and define T : `2 → `2 by

T (x1 , x2 , x3 , . . .) = (λ1 x1 , λ2 x2 , λ3 x3 , . . .).

Then T is compact, being the limit in the operator norm of the bounded finite rank operators

Tn : (x1 , x2 , x3 , . . .) 7→ (λ1 x1 , . . . , λn xn , 0, . . .)

since kT x − Tn xk2 ≤ supk>n |λk |kxk2 so kT − Tn k → 0 as n → ∞.

Theorem 5.5. Let X, Y, Z be normed vector spaces and S ∈ B(X, Y ) and T ∈ B(Y, Z).
(i) If S is compact, then T ◦ S is compact.
(ii) If T is compact, then T ◦ S is compact.

Proof. (i) Suppose that S is compact. Let (xn )n∈N be a bounded sequence. Then there exists a subsequence
(xnk )k∈N such that Sxnk → y0 for some y0 ∈ Y . As T is continuous, we have (T ◦ S)(xnk ) → T y0 .
(ii) Suppose that T is compact. Let (xn )n∈N be a bounded sequence in X. Then (Sxn )n∈N is bounded in
Y , hence ((T ◦ S)(xn ))n∈N has a convergent subsequence.

33
Theorem 5.6. Let X be an infinite-dimensional Banach space and K ∈ B(X) a compact operator. Then
0 ∈ σ(K).

Proof. Assume for contradiction that 0 6∈ σ(K). Then K −1 exists. By the Open Mapping Theorem (or more
precisely, Corollary 2.6), we have K −1 ∈ B(X).
By Theorem 1.17, the closed unit ball in X is not compact. Hence there exists a sequence (yn )n∈N in X
such that kyn kX ≤ 1 for every n ∈ N, but there exists no convergent subsequence. Set xn = K −1 yn . This
gives rise to a bounded sequence, but (Kxn )n∈N does not have a convergent subsequence, in contradiction to
the compactness of K.

Theorem 5.7. Let X be a Banach space and K ∈ B(X) compact. Then for any λ ∈ F\{0}, the space
ker(λidX − K) is finite-dimensional.

Proof. Let Y = ker(λidX − K). Assume for contradiction that dim Y = ∞. Then, using Theorem 1.17 again,
we construct a bounded sequence (yn )n∈N in Y with no convergent subsequence. However, as K is compact,
there exists a subsequence, say (ynk )k∈N , such that

z0 := lim Kynk
k→∞

exists. Then
1 z0
ynk = Kynk → .
λ λ
But this contradicts the construction of (yn )n∈N .

Self-adjoint Operators on Hilbert Spaces

Theorem 5.8. Let H be a Hilbert space and A ∈ B(H). Then there exists a unique A∗ ∈ B(H) such that

hAx, yi = hx, A∗ yi ∀x, y ∈ H. (7)

Moreover kA∗ kB(H) = kAkB(H) .

Proof. Fix y ∈ H and consider φ ∈ H ∗ with

φ(x) = hAx, yi, x ∈ H.

By the Riesz Representation Theorem, there exists a unique z ∈ H such that φ(x) = hx, zi for all x ∈ X. Set
A∗ y = z. Then (7) holds by construction.
We need to check that the resulting operator A∗ is linear. Let y1 , y2 ∈ H and α ∈ F. Then

hx, A∗ (αy1 + y2 )i = hAx, αy1 + y2 i

= ᾱhAx, y1 i + hAx, y2 i
= ᾱhx, A∗ y1 i + hx, A∗ y2 i
= hx, αA∗ y1 i + hx, A∗ y2 i
= hx, αA∗ y1 + A∗ y2 i.

Hence A∗ (αy1 + y2 ) = αA∗ y1 + A∗ y2 .

34
 Finally, we have

sup kA∗ ykH = sup sup |hx, A∗ yi|

y∈H y∈H x∈H
kykH ≤1 kykH ≤1 kxkH ≤1

= sup sup |hAx, yi|

y∈H x∈H
kykH ≤1 kxkH ≤1

= sup sup |hAx, yi| = kAkB(H) .

x∈H y∈H
kxkH ≤1 kykH ≤1

In particular, we have A∗ ∈ B(H) and kA∗ kB(H) = kAkB(H) .

Definition. Let H be a Hilbert space and A ∈ B(H). The operator A∗ ∈ B(H) characterised by (7) is called
the adjoint of A. If A∗ = A, then we say that A is self-adjoint.

Remark. Consider the Euclidean space Fn . An operator A ∈ B(Fn ) is represented by an (n × n)-matrix M .

T
The adjoint A∗ then corresponds to M .

Theorem 5.9. If H is a Hilbert space and A ∈ B(H), then A∗∗ = A.

Proof. For x, y ∈ H, we have

hx, A∗∗ yi = hA∗ x, yi = hy, A∗ xi = hAy, xi = hx, Ayi.

Thus A∗∗ = A.

Theorem 5.10. Let H be a Hilbert space and A ∈ B(H) a self-adjoint operator. Then
(i) all eigenvalues of A are real, and
(ii) eigenvectors belonging to distinct eigenvalues are orthogonal.

Proof. (i) Let λ be an eigenvalue of A and choose a corresponding eigenvector u 6= 0. Then

λkuk2H = hAu, ui = hu, Aui = λ̄kuk2H .

Hence λ = λ̄, i.e., we have λ ∈ R.

(ii) Let λ, µ ∈ R be two distinct eigenvalues and u, v ∈ H\{0} corresponding eigenvectors. Then

λhu, vi = hAu, vi = hu, Avi = µhu, vi.

It follows that hu, vi = 0.

Lemma 5.11. Let H be a Hilbert space and A ∈ B(H) a self-adjoint operator. Then

kAkB(H) = sup |hAx, xi|.

x∈H
kxkH ≤1

Proof. Set
M = sup |hAx, xi|.
x∈H
kxkH ≤1

If x ∈ H is a point with kxkH ≤ 1, then

|hAx, xi| ≤ kAxkH kxkH ≤ kAkB(H) kxk2H ≤ kAkB(H) .

35
The inequality M ≤ kAkB(H) follows.
In order to prove the reverse inequality, we may assume that A 6= 0, as the statement is obvious otherwise.
Let x ∈ H with Ax 6= 0. Fix β > 0 and set v = β1 Ax. Then hAx, vi = β −1 kAxk2 which is real and so by
self-adjointness hAv, xi = β −1 kAxk2 also. Thus we obtain

4kAxk2H = 2β (hAx, vi + hAv, xi)

= hA(βx + v), βx + vi − hA(βx − v), βx − vi
≤ M kβx + vk2H + kβx − vk2H

= 2M β 2 kxk2H + kvk2H


 
1
= 2M β 2 kxk2 + 2 kAxk2H ,
β
using the parallelogram identity. For
kAxkH
β2 = ,
kxkH
this gives kAxk2H ≤ M kAxkH kxkH . Now the desired inequality follows.

Lemma 5.12. For a Hilbert space H, let T ∈ B(H). Then (im T ∗ )⊥ = ker T .

Proof. For y ∈ H, we have

y ∈ (im T ∗ )⊥ ⇔ ∀x ∈ H : hy, T ∗ xi = 0
⇔ ∀x ∈ H : hT y, xi = 0 ⇔ y ∈ ker T.

Lemma 5.13. Let A ∈ B(H) be a self-adjoint operator on a Hilbert space H. If M ⊂ H is a linear subspace
with A(M ) ⊂ M , then A(M ⊥ ) ⊂ M ⊥ .

Proof. Let y ∈ M ⊥ . Then for every x ∈ M ,

hAy, xi = hy, Axi = 0,

since Ax ∈ M . Hence Ay ∈ M ⊥ .

The Spectral Theorem

We have some information about the spectrum of a compact self-adjoint operator from Theorem 5.6 (namely,
that the spectrum contains 0) and Theorem 5.10 (the eigenvalues are real). But it turns out that we can say
much more.

Theorem 5.14. Let H be a Hilbert space and let K ∈ B(H) be a self-adjoint compact operator. Then for every
λ ∈ F\{0}, the subspace im(λidH − K) of H is closed.

Proof. Let (zn )n∈N be a convergent sequence in im(λidH − K) with limit z0 = limn→∞ zn . For n ∈ N, choose
xn ∈ H with
zn = λxn − Kxn .
Note that ker(λidH − K) is closed, so we have, by the orthogonal decomposition,

H = ker(λidH − K) ⊕ (ker(λidH − K))⊥ .

36
Consider the decomposition

xn = un + yn , un ∈ ker(λidH − K), yn ∈ (ker(λidH − K))⊥ .

Then
zn = λyn − Kyn , n ∈ N.
We claim that the sequence (yn )n∈N is bounded.
Assume for contradiction that it is unbounded. Discarding a subsequence if necessary, we can suppose that
yn 6= 0 for every n and kyn kH → ∞ as n → ∞. Set
yn
wn = .
kyn kH
Then
zn
λwn − Kwn = →0 as n → ∞. (8)
kyn kH
As K is compact, we may assume that the sequence (Kwn )n∈N is convergent (discarding another subsequence);
say Kwn → v0 . Then (8) implies
wn → λ−1 v0 .
Putting this into (8) we obtain
v0 − λ−1
n Kv0 = lim (λwn − Kwn ) = 0.
n→∞

That is, we have v0 ∈ ker(λidH − K). Since wn ∈ (ker(λidH − K))⊥ , which is closed, we also have v0 ∈
(ker(λidH − K))⊥ . Therefore, v0 = 0. On the other hand,

kv0 kH = λ lim kwn kH = λ 6= 0.

n→∞

The contradiction proves that (yn )n∈N is bounded.

By compactness again, we can discard a subsequence and then (Kyn )n∈N converges. Let x0 := limn→∞ Kyn .
Then
1 1
yn = (zn + Kyn ) → (z0 + x0 ) =: y0 .
λ λ
Now we have
z0 = lim zn = lim (λyn − Kyn ) = λy0 − Ky0 .
n→∞ n→∞

So z0 ∈ im(λidH − K).

Corollary 5.15. Let K ∈ B(H) be a self-adjoint compact operator on a Hilbert space H. Then for every
λ ∈ F\{0},
im(λidH − K) = (ker(λ̄idH − K))⊥ .

Proof. Note that

(λidH − K)∗ = λ̄idH − K.
By Lemma 5.12, this means
ker(λ̄idH − K) = (im(λidH − K))⊥ .
Theorem 5.14 tells us that im(λidH − K) is closed. Hence

im(λidH − K) = (im(λidH − K))⊥⊥ = (ker(λ̄idH − K))⊥ ,

as stated.

37
Theorem 5.16. Let K ∈ B(H) be a self-adjoint compact operator on a Hilbert space H. If λ ∈ σ(K), then
λ = 0 or λ is an eigenvalue of K.

Proof. Suppose that 0 6= λ ∈ F and that λ is not an eigenvalue of K. Then λ̄ is not an eigenvalue of K either,
for otherwise λ̄ would be real by Theorem 5.10 and then λ would be an eigenvalue.
By Corollary 5.15, we have
im(λidH − K) = (ker(λ̄idH − K))⊥ .
Since ker(λ̄idH − K) = {0} we now have im(λidH − K) = H, and since ker(λidH − K) = {0} it now follows
that λidH − K is invertible, thus λ ∈
/ σ(K).

Theorem 5.17. Let K ∈ B(H) be a self-adjoint compact operator on a Hilbert space H. Then for every ε > 0,
the set {λ ∈ σ(K) : |λ| ≥ ε} is finite.

Proof. Suppose for contradiction that there exists a sequence (λn )n∈N in σ(K) with |λn | ≥ ε for every n ∈ N and
λm 6= λn for m 6= n. Then every λn is an eigenvalue according to Theorem 5.16. So there exists a corresponding
eigenvector un ∈ H with kun kH = 1. By Theorem 5.10, we have hum , un i = 0 for m 6= n. Thus

hKum , Kun i = λm λn hum , un i = 0.

Furthermore,
kKun kH = |λn |kun kH ≥ ε.
Hence
kKum − Kun k2H = kKum k2H + kKun k2H ≥ 2ε2 .
But then the sequence (Kun )n∈N cannot have a convergent subsequence, which contradicts the compactness of
K.

Lemma 5.18. Let K ∈ B(H) be a self-adjoint compact operator on a Hilbert space H. Then at least one of
kKkB(H) and −kKkB(H) is an eigenvalue of K.

Proof. We may assume that K 6= 0. By Lemma 5.11, we have

kKkB(H) = sup |hKx, xi|.

x∈H
kxkH ≤1

Hence there exists a sequence (un )n∈N of unit vectors (i.e., kun kH = 1 for every n ∈ N), such that hKun , un i →
m, where m = ±kKkB(H) . Note that

kKun − mun k2H = kKun k2H − 2mhKun , un i + m2

≤ m2 − 2mhKun , un i + m2 → 0

as n → ∞. By compactness, there exists a convergent subsequence of hKun in∈N , say Kunj → v as j → ∞.

Then it follows that unj → v/m. Hence

Kv = m lim Kunj = mv.

j→∞

Furthermore, we have
kvkH = |m| lim kunj kH = |m| =
6 0.
j→∞

Thus v is an eigenvector with eigenvalue m.

38
Remark. If π is the orthogonal projection map on a closed linear subspace M of a Hilbert space H then

hu, yi = hu, πyi ∀u ∈ M, y ∈ H.

Theorem 5.19 (Spectral theorem for self-adjoint compact operators on Hilbert spaces). Let H be
an infinite-dimensional Hilbert space and let K ∈ B(H) be a self-adjoint compact operator. Then:
(i) the spectrum σ(K) is a subset of R, it contains 0, and σ(K)\{0} comprises either finitely many eigenvalues
or a sequence of eigenvalues converging to 0;
(ii) the eigenspace corresponding to each non-zero eigenvalue is finite-dimensional;
(iii) let (λn )n∈Λ comprise all non-zero eigenvalues of K, repeated according to the dimensions of their eigenspaces,
where Λ = N or Λ = {1, . . . , N } for some N ∈ N. Then there exists an orthonormal system (un )n∈Λ such that
un is an eigenvector corresponding to λn for each n ∈ Λ, and
X
Kx = λn hx, un iun
n∈Λ

for every x ∈ H.

Proof. The statement of (i) is covered by Theorems 5.6, 5.10, 5.16, and 5.17. Part (ii) is covered by Theorem
5.7. It remains to prove part (iii).
For each eigenvalue λ, we choose an orthonormal basis of the eigenspace ker(λidH − K). By Theorem
5.10, the union of all all of these gives rise to an orthonormal system. Furthermore, if we label the elements
appropriately, then we can represent it in the form (un )n∈Λ , where un is an eigenvector corresponding to the
eigenvalue λn for every n ∈ Λ.
Let L = span {un : n ∈ Λ}. It follows from Theorem 1.10 that for any x ∈ L,
X
x= hx, un iun .
n∈Λ

Therefore, X
Kx = λn hx, un iun .
n∈Λ

We now want to show the same identity for all x ∈ H.

Let π ⊥ be the orthogonal projection onto L⊥ . Define K ⊥ = K ◦π ⊥ . This is a compact operator by Theorem
5.5. We have K(L) ⊂ L, and Lemma 5.13 then implies K(L⊥ ) ⊂ L⊥ . Hence for x, y ∈ H,

hK ⊥ x, yi = hKπ ⊥ x, yi
= hKπ ⊥ x, π ⊥ yi = hπ ⊥ x, Kπ ⊥ yi = hx, K ⊥ yi.

That is, the operator K ⊥ is self-adjoint, and we can apply Lemma 5.18 to K ⊥ instead of K. Therefore, we
have an eigenvalue m of K ⊥ with |m| = kK ⊥ kB(H) . Let v0 be a corresponding eigenvector. Then

mv0 = K ⊥ v0 ∈ L⊥ .

So either m = 0 or v0 ∈ L⊥ . In the second case, we have

Kv0 = K ⊥ v0 = mv0 ,

so m is also an eigenvalue of K and v0 is a corresponding eigenvector. But all eigenvectors for non-zero
eigenvalues are in L, hence m = 0, which means K ⊥ = 0.

39
 Now for x ∈ H, we have the decomposition x = x1 + x2 with x1 ∈ L and x2 ∈ L⊥ , according to H = L ⊕ L⊥ .
Therefore,

Kx = Kx1 + Kx2 = Kx1 + K ⊥ x2 = Kx1

X X
= λn hx1 , un iun = λn hx, un iun .
n∈Λ n∈Λ

Corollary 5.20. Let H be a separable Hilbert space and let K ∈ B(H) be a self-adjoint compact operator.
Then H has an orthonormal basis consisting of eigenvectors of K.

Proof. For dim H < ∞, this is a result of linear algebra. For dim H = ∞, define (un )n∈Λ and L as in the
previous proof. Choose an orthonormal basis (vm )m∈M of L⊥ , where M = N or a finite index set. For every
m ∈ M , we have Kvm = 0. Hence vm is an eigenvector. Now combine (un )n∈Λ with (vm )m∈M .

40