Chapter 2: Banach Spaces

1. Banach spaces

Definition 2.1.1 (Normed space). A normed space is a complex vector

space which a norm ∥·∥ is defined.

Definition 2.1.2 (Banach space). A Banach space (X, ∥·∥) is a complete

normed space.

Example 2.1.3. An easy one is Cd , with p-norm, p ∈ [1, ∞):

d
X 1/p
p
∥x∥p = |xk | .
k=1

For p = ∞, we set

∥x∥∞ = max |xk |.

1≤k≤d

For p = 2, it is the standard Euclidean norm, which gives a Hilbert space.

For finite-dimensional case, ∥·∥p , p ∈ [1, ∞) is equivalent to ∥·∥2 .

1
Example 2.1.4 (The space of continuous functions). Let Ω be a topo-
logical space, a subset of Rd , and let C(Ω) be the set of all continuous func-
tions f : Ω −→ C. The uniform/supremum norm on C(Ω) is given by

∥f ∥ = sup|f (x)|.
x∈Ω

Then (C(Ω), ∥·∥) is a Banach space. This is clearly that C(Ω) is a vector
space and the norm is easy to verify. Therefore, it suffices to show that C(Ω)
is complete. Let fn be a Cauchy sequence defined on ∥·∥, then

ϵ > ∥fn − fm ∥ = sup|fn (x) − fm (x)| ≥ |fn (x) − fm (x)|

x∈Ω

Hence the limit of fn exists and f (x) = limn→∞ fn (x). Fix some point x0 ∈ Ω,
since fn (x) is continuous at x0 , for every ϵ > 0, there exist δ > 0 such that
if 0 < |x − x0 | < δ, then

|fn (x) − fn (x0 )| < ϵ/3.

Then we can see that

|f (x) − f (x0 )| = |f (x) − fn (x) + fn (x) − fn (x0 ) + fn (x0 ) − f (x0 )|

≤ |f (x) − fn (x)| + |fn (x) − fn (x0 )| + |fn (x0 ) − f (x0 )|

< ϵ/3 + ϵ/3 + ϵ/3 = ϵ.

This shows that f is continuous and hence f ∈ C(Ω).

If we take Ω = Rd , then C(Rd ) is a Banach space. Denote CK (Rd ), the

space of all compactly supported continuous functions on Rd −→ C. The
term compactly supported means we may find a compact set K ⊂ Rd such
that f (x) = 0 for all x ̸∈ K. (CK (Rd ), ∥·∥) is a normed space, but it is not

2
complete w.r.t. the supremum norm ∥·∥. The space C0 (Rd ) is the space of
all continuous functions vanishing at ∞. The latter means:

lim |f (x)| = 0
|x|→∞

which also equivalent to say that, for every ϵ > 0, there is a compact set
K such that |f (x)| < ϵ for all x ̸∈ K. The space (C0 (Rd ), ∥·∥) is a Banach
space. In total,

CK (Rd ) ⊂ C0 (Rd ) ⊂ C(Rd ).

Exercise 2.1.5. Why CK (Rd ) is not complete, but C0 (Rd ) is complete w.r.t.
the supremum norm ∥·∥?

Solutions:

Define the set

CK (R) = {f : R −→ C : f is continuous with compact support}

with supremum norm ∥f ∥ = sup |f (x)|. To prove CK (R) is not complete, it

x∈Rd
suffices to show that it is not closed. Given a compact set K = [−kπ, kπ]
and consider a sequence function

sin x
x ∈ K − {0},



 x


fk (x) = 1 x = 0,




0 x ̸∈ K,


that converges to

 sin x

x ̸= 0,
f (x) = x
1

x=0

3
as k → ∞. Then limx→±∞ f (x) = 0. We see that fk (x) ∈ CK (R) but
f ̸∈ CK (R), therefore CK (R) is not complete. Note that f ∈ C0 (R).
For given ϵ > 0, to prove the set

C0 (R) = {f : R −→ C : |f (x)| < ϵ, ∀x ̸∈ K, K is compact},

the space of all continuous function vanishing at infinity is complete, we

consider a sequence of function fn (x) ∈ C0 (R), then |fn (x)| < ϵ/2 for all
x ̸∈ K. Obviously, for every ϵ > 0, there exists N > 0 such that for n, m ≥ N
we have |fn − fm | ≤ |fn | + |fm | < ϵ, so fn is indeed a Cauchy and has a limit
f (x) = limn→∞ fn (x). To prove f (x) ∈ C0 (R), it is easy to see that

|f (x)| = |f (x) − fn (x) + fn (x)| ≤ |f (x) − fn (x)| + |fn (x)| < ϵ/2 + ϵ/2 = ϵ

for all x ̸∈ K. This shows that C0 (R) is complete.

Example 2.1.6 (Lp -spaces). Let (Ω, M, µ) be a measure space and p ∈

[1, ∞). We denote
 Z 
p p
L (Ω, µ) = f : Ω −→ C : ∥f ∥Lp = |f (x)| dµ(x) < ∞
Ω

For p = ∞, we define

L∞ (Ω, µ) = {f : Ω −→ C : ∥f ∥L∞ < ∞},

∥f ∥L∞ = ess sup|f (x)| = inf{c > 0 : |f (x)| ≤ cµ − a.s}.
x∈Ω

Then Lp (Ω, µ) is a Banach space.

4
Exercise 2.1.7. Write down the following inequality/facts from Lp -spaces:
(1). Minkowski’s inequality;
(2). Hölder’s inequality;
(3). Riesz-Fisher theorem.

Solutions:

(1). ∥f + g∥p ≤ ∥f ∥p + ∥g∥p

(2). ∥f g∥L1 ≤ ∥f ∥Lp ∥g∥Lq , 1/p + 1/q = 1. Note that p = 2 is Cauchy-
schwartz.
(3). Lp ’s are complete.

Example 2.1.8 (Sequence spaces). Consider a = {ak }∞

k=1 , a sequence of

complex numbers. We denote for p ∈ [1, ∞]:

ℓ∞ = {a : ∥a∥ℓ∞ = supn |an | < ∞};

c0 = {a : limn→∞ |an | = 0};
ℓp = {a : ∥a∥ℓp = ( ∞ p 1/p
P
n=1 |ak | ) < ∞};
s = {a : limn→∞ np an = 0};
f = {a : an = 0 for except finitely many terms}.

Then f ⊂ s ⊂ ℓp ⊂ c0 ⊂ ℓ∞ . They are all Banach spaces, with appropriate

norms, except f .

Exercise 2.1.9. What is the correspondent norm for c0 , s so that it is a

Banach space?

Solutions:

c define norm ∥·∥ℓ∞ , s define norm ∥·∥ℓp .

5
Space of Bounded Operators

Definition 2.1.10.
(a). Let X, Y be two normed spaces with norm ∥·∥X , ∥·∥Y . A linear map
T : X −→ Y is bounded if there exists C ≥ 0 such that

∥T x∥Y ≤ C ∥x∥X .

(b). The operator norm ∥T ∥op of T is defined as the smallest constant C ≥ 0

such that the above holds. This is equivalent to
∥T x∥Y
∥T ∥op = sup .
x∈X,x̸=0 ∥x∥X

(c). We denote L(X, Y ) or B(X, Y ) the space of all bounded operators from
X −→ Y . If X = Y , we write L(X) or B(X).

In fact, L(X, Y ) is a Banach space, with the norm ∥·∥op . In fact, X needs
not to be complete.

Theorem 2.1.11. Let X be a normed space and Y be a Banach space. Then

L(X, Y ) is a Banach space w.r.t the operator norm ∥·∥op

Proof. The fact that L(X, Y ) is a vector space. To prove ∥·∥op is a norm,
without loss of generality we set ∥x∥X = 1. We see that

∥αT ∥op = sup ∥αT x∥Y = ∥αT ∥Y = |α| ∥T ∥Y = |α| ∥T ∥op .

{∥x∥X =1}

If ∥T ∥op = 0, then clearly that T = 0. To prove triangle inequality,

∥T + S∥op = sup ∥(T + S)x∥Y ≤ sup (∥T x∥Y + ∥Sx∥Y )

{∥x∥=1} {∥x∥=1}

= sup ∥T x∥Y + sup ∥Sx∥Y

{∥x∥=1} {∥x∥=1}

= ∥T ∥op + ∥S∥op .

6
To prove completeness, let {Tn } be a Cauchy sequence in ∥·∥op . Given x ∈ X,
to show that {Tn x} is Cauchy, set ∥x∥X = 1, then

ϵ > ∥Tn − Tm ∥op = sup ∥Tn x − Tm x∥Y ≥ ∥Tn x − Tm x∥Y

{∥x∥X =1}

This shows that {Tn x} is Cauchy in Y . Since Y is Banach, then the limit
of Tn x is in Y , so T x = limn→∞ Tn x. To prove T x is in L(X, Y ), we need
to show that T x is linear and bounded. Obviously T x is linear. To prove
bounded, we see that

∥T x∥Y = lim sup ∥Tn x∥op ∥x∥X ≤ c ∥x∥X

n→∞

since Tn x is Cauchy so is bounded by c. This shows that T x is bounded.

Now we need to show that Tn → T in operator norm, we see that

∥(Tn − T )x∥Y = lim ∥(Tn − Tm )x∥Y ≤ lim ∥(Tn − Tm )∥op ∥x∥X ≤ ϵ ∥x∥X .
m→∞ m→∞

This implies that

∥(Tn − T )x∥Y
≤ϵ
∥x∥X
and the supremum is
∥(Tn − T )x∥Y
∥(Tn − T )∥op = sup ≤ ϵ.
x∈X,x̸=0 ∥x∥X
Hence Tn − T in operator norm.

Remark. The space of bounded operators on X −→ X is more than just a

Banach space. We are allowed to do multiplication, that is, if T, S ∈ L(X),
then T S ∈ L(X). It is a Banach algebra.

7
Exercise 2.1.12. Show that if T, S ∈ L(X), then T S ∈ L(X). Moreover,
∥T S∥op ≤ ∥T ∥op ∥S∥op .

Proof. Note that L(X) is the space of all bounded operators from X to X.
If T, S ∈ L(X), then

∥T x∥X ≤ C ∥x∥X ∥Sx∥X ≤ K ∥x∥X .

Now we see that

∥(T S)x∥X = ∥T ◦ (Sx)∥X ≤ C ∥Sx∥X ≤ CK ∥x∥X .

This shows that T S ∈ L(X). To prove second part, by definition, we have

∥T S∥op = sup ∥T ◦ (Sx)∥X ≤ sup ∥T x∥X sup ∥Sx∥X = ∥T ∥op ∥S∥op .

{∥x∥X =1} {∥x∥X =1} {∥x∥X =1}

Isometry and equivalence of norms

Definition 2.1.13. Let ∥·∥1 , ∥·∥2 be two norms defined on a vector space X.
(a). ∥·∥1 is stronger than ∥·∥2 if there exists a M ≥ 1 such that for all x ∈ X,

∥x∥2 ≤ M ∥x∥1 .

(b). ∥·∥1 is weaker than ∥·∥2 if ∥·∥2 is stronger than ∥·∥1 .

(c). ∥·∥1 and ∥·∥2 are equivalent if (a), (b) hold, That is, there exists M ≥ 1
such that

M −1 ∥x∥X ≤ ∥x∥Y ≤ M ∥x∥X , x ∈ X.

Remark. The definition 2.14 (a) asserts that the convergence sequence in
norm 1 is less than norm 2.

8
Exercise 2.1.14. Denote ∥·∥1 ≺ ∥·∥2 if latter is strong than the former, and
∥·∥1 ∼ ∥·∥2 if the two norms are equivalent.
(a). Show that ≺ is a partial order.
(b). Show that ∼ is an equivalence relation.

Proof. (a). Reflexive: ∥x∥1 ≺ ∥x∥1 since ∥x∥1 ≤ M ∥x∥1 .

Anti-symmetry: If ∥x∥1 ≺ ∥x∥2 and ∥x∥2 ≺ ∥x∥1 , then by definition

∥x∥1 ≤ M ∥x∥2 and ∥x∥2 ≤ M ∥x∥1

implies that M −1 ∥x∥1 ≤ ∥x∥2 ≤ M ∥x∥1 . Thus, ∥x∥1 is equivalent to ∥x∥2 .

Transitive: If ∥x∥1 ≺ ∥x∥2 and ∥x∥2 ≺ ∥x∥3 , then by definition

∥x∥1 ≤ M ∥x∥2 and ∥x∥2 ≤ N ∥x∥3 .

implies that ∥x∥1 ≤ M N ∥x∥3 so that ∥x∥1 ≺ ∥x∥3 . Thus, ≺ is a partial

order.
(b).
Reflexive: This is clearly that ∥x∥1 ∼ ∥x∥1 since

M −1 ∥x∥1 ≤ ∥x∥1 ≤ M ∥x∥1 .

Symmetry: If ∥x∥1 ∼ ∥x∥2 , then

M −1 ∥x∥1 ≤ ∥x∥2 ≤ M ∥x∥1 .

This is equivalent to M −1 ∥x∥2 ≤ ∥x∥1 ≤ M ∥x∥2 and hence ∥x∥2 ∼ ∥x∥1 .

Transitive: If ∥x∥1 ∼ ∥x∥2 and ∥x∥2 ∼ ∥x∥3 , then

M −1 ∥x∥1 ≤ ∥x∥2 ≤ M ∥x∥1 ,

N −1 ∥x∥2 ≤ ∥x∥3 ≤ N ∥x∥2 .

9
This implies that M −1 N −1 ∥x∥1 ≤ N −1 ∥x∥2 ≤ ∥x∥3 ≤ N ∥x∥2 ≤ M N ∥x∥1 .
This shows that ∥x∥1 ∼ ∥x∥3 . Hence ∼ is an equivalence relation.

Here is a stronger form of equivalence.

Definition 2.1.15. We say the two Banach spaces X, Y are isomorphic if

there exists a bounded bijective linear map T : X −→ Y with bounded inverse.
We say X, Y are isometry if additionally it satisfies ∥T x∥Y = ∥x∥X for all
x.

Exercise 2.1.16. Show that if two norms ∥·∥1 , ∥·∥2 over the space X are
equivalent, then (X, ∥·∥1 ), (X, ∥·∥2 ) are isomorphic.

Proof. We consider the identity map

I : (X, ∥·∥1 ) −→ (X, ∥·∥2 )

which is clearly linear and bijective. Since ∥·∥1 and ∥·∥2 are equivalent, then
by definition

M −1 ∥x∥1 ≤ ∥Ix∥2 = ∥x∥2 ≤ M ∥x∥1

implies that I is bounded. Also ∥x∥1 ≤ M ∥Ix∥2 which is bounded inverse.

Hence I is isomorphism.

10
Example 2.1.17. Let ∥·∥p be the p-norm over Cd , with d < ∞. Then ∥·∥p
are equivalent with each other for all p ∈ [1, ∞]. Particularly, given x ∈ Cd ,
we have
d
X 1/p
p
∥x∥∞ ≤ ∥x∥p ≤ d ∥x∥∞ ⇐⇒ max|xk | ≤ |xk | ≤ d max|xk |.
k k
k=1

Hence, ∥·∥p ∼ ∥·∥∞ . The equivalence relation over ∼ shows every norm are
equivalent. However, the space (Cd , ∥·∥p ) are not isometry to any different p.
See Exercise below. Note that also the bound becomes worse when d → ∞.

Exercise 2.1.18. Show that (R2 , ∥·∥1 ) are not isometric to (R2 , ∥·∥2 ). Par-
ticularly, the unit ball of ∥·∥1 is a square, and that of ∥·∥2 is a unit disk. It
is possible to bijectively map a square to a disk via linear transform?

Proof. Define a linear map T by T (x, y) = (x + y, x). Then

p p
∥T x∥2 = (x + y)2 + x2 = 2x2 + 2xy + y 2 ,
∥x∥1 = |x| + |y|

which is clearly they are not equal. Hence T is not isometry between (R2 , ∥·∥1 )
and (R2 , ∥·∥2 ). It is not possible.

Exercise 2.1.19.
(a). If (Ω, µ) is bounded measure space, that is µ(Ω) = 1, which norm is
stronger? ∥·∥L1 or ∥·∥L2 .
(b). Show that the two norms on L1 (Rd ) and L2 (Rd ) are not equivalent.

Proof. (a). From Example 2.1.6,

Z Z
∥f ∥L1 = |f (x)| dµ(x) ∥f ∥L2 = |f (x)|2 dµ(x).
Ω Ω

11
Then, since f is integrable, then f is bounded, so |f (x)| ≤ M . Hence,
Z Z
2
∥f ∥L2 = |f (x)| dµ(x) ≤ M |f (x)|dµ(x) = M ∥f ∥L1 .
Ω Ω

This shows that ∥f ∥L1 is stronger than ∥f ∥L2 .

(b). Suppose contrary that two norms on L1 (Rd ) and L2 (Rd ) are equivalent,
then by definition

M −1 ∥f ∥L1 ≤ ∥f ∥L2 ≤ M ∥f ∥L1 .

The right inequality holds, but the left inequality does not hold, a contradic-
tion.

2. Duals and Double Duals

Definition 2.2.1. Let X be a Banach space. The dual space X ∗ is the vector
space of all bounded linear functionals ℓ : X −→ C. Boundedness is defined
as: there exists M ≥ 0 such that

|ℓ(x)| ≤ M ∥x∥X , x ∈ X.

The norm of ℓ is defined as the smallest M > 0 such that the above holds,
that is, the operator norm of ℓ:
|ℓ(x)|
∥ℓ∥X ∗ = sup = sup |ℓ(x)|.
x:x̸=0 ∥x∥ x:∥x∥=1

Theorem 2.2.2. X ∗ equipped with the norm ∥·∥X ∗ is a Banach Space.

Proof. Free from Theorem 2.1.11.

12
A hint to generalize the inner product on Hilbert spaces to Banach is Riesz
representation theorem, which says inner product x 7→ (v, x) is in fact a
linear functional ℓv , and vice versa. That is we may view (·, ·) : H × H −→ C
as the map ⟨·, ·⟩ : H∗ × H −→ C, where

(v, x) = ℓv (x) = ⟨ℓv , x⟩.

That is, taking the inner product between v, x is equivalent to test x against
a linear functional ℓv ∈ H∗ . Hence, to define inner product for a Banach
space X, we have

⟨·, ·⟩ : X ∗ × X −→ C,

where the bilinear map ⟨·, ·⟩ takes a linear functional ℓ over X and a member
x ∈ X to produce a complex number ⟨ℓ, x⟩ = ℓ(x) ∈ C.

Example 2.2.3. If H is a Hilbert space, then H∗ ∼

= H, by Riesz representa-
tion theorem

Theorem 2.2.4. Let p ∈ [1, ∞). Then the dual space of Lp (Ω, µ) (for some
measure space (Ω, µ)) is isomorphic to Lq (X, µ), where p, q are Hölder dual:
p−1 + q −1 = 1.

Proof. Set X = Lp and denote

 Z 
p p
L (Ω, µ) = f : Ω −→ C : ∥f ∥Lp = |f (x)| dµ(x) < ∞
Ω

Then Lq ⊂ X ∗ , that is, every Lq function is secretly a bounded linear func-

tional on X, is straightforward. Particularly given g ∈ Lq , we define
Z
ℓg (f ) = g(x)f (x)dµ(x).
Ω

13
Clearly, it defines a linear functional. Moreover, by Hölder’s inequality,

|ℓg (f )| ≤ ∥g∥Lq ∥f ∥Lp .

Now we may define T : Lq −→ X ∗ by

Z
⟨T (g), f ⟩ = ℓg (f ) = g(x)f (x)dµ(x).
Ω

One shows: T is linear, injective, surjective, bounded, norm preserving (that

is, ∥T g∥X ∗ = ∥g∥Lq ). Most of these are not hard, except surjectivity. In fact,
norm-preserving is not hard, Hölder’s inequality shows ∥T (g)∥X ∗ ≤ ∥g∥Lq .
To show it equals one just needs to find f ∈ Lp such that ℓg (f ) = ∥g∥Lq .
Choose f of the form
eiϕ(x) |g(x)|q−1
f (x) = ,
∥g∥Lq−1
q

where ϕ(x) is the phase angle such that g(x)f (x) is real. One have to verify
also g ∈ Lp (using p−1 + q −1 = 1).
Surjective of T is technical. One needs to show every linear functional ℓ
arisen as ℓg for some g. We may seek help from more powerful representation
theorem, such has Riesz representation theorem for L2 , or an even powerful
one, Radon-Nykodym theorem.

Example 2.2.5. The dual space of sequence spaces can be easily identified,
because sequence spaces are essentially Rd for d = ∞. Some techniques from
Rd may apply in this case. For instance,
(i). (ℓp )∗ ∼
= ℓq for p ∈ [1, ∞) and p−1 + q −1 = 1;
(ii). (ℓ1 )∗ ∼
= ℓ∞ as above. We also have c∗0 ∼= ℓ1 , where c0 ⊂ ℓ∞ is the space
of sequences that vanishes as the term → ∞.

14
Definition 2.2.6 (Double dual). The double dual X ∗∗ of a Banach space
is the dual space of its dual space X ∗ , that is, X ∗∗ = (X ∗ )∗ .

The double dual space X ∗∗ of X is larger than X, see theorem below. But
in the case when X ∗∗ ∼ = X, we say the space X is reflexive. For instance, a
= H∗ ∼
Hilbert space is reflexive since H∗∗ ∼ = H, and so are Lp for p ∈ (1, ∞)
= (Lq )∗ ∼
since (Lp )∗∗ ∼ = Lp . The space L1 (Ω), C0 (Ω) are not.

Theorem 2.2.7. A Banach space X is isometrically embedded in its double

dual X ∗∗ . That is, there exists a linear injection ι : X −→ X ∗∗ such that
∥ι(X)∥X ∗∗ = ∥x∥X .

Proof. For each x ∈ X, let ℓ̂(·) be the linear functional on X ∗ which assigns
to each λ ∈ X ∗ the number λ(x). Since

|ℓ̂(λ)| = |λ(x)| ≤ ∥λ∥X ∗ ∥x∥X

ℓ̂ is a bounded linear functional on X ∗ with norm ℓ̂ ≤ ∥x∥X . It follows

X ∗∗
from Hann-Banach Theorem that, given x, we can find a λ ∈ X ∗ such that

∥λ∥X ∗ = 1 and λ(x) = ∥x∥X .

This shows that

ℓ̂ = sup |ℓ̂(λ)| ≥ ∥x∥X

X ∗∗ λ∈X ∗ ,∥λ∥≤1

which implies that

ℓ̂ = ∥x∥X
X ∗∗

Thus, ι is an isometry of X into X ∗∗ .

15
3. Hahn-Banach theorem
Let us work with a simple example to show how Hann-Banach can be useful.
Consider a Hilbert space H and let S ⊂ H be a subspace. Our goal here is
to find an annihilator y ∈ H, y ̸= 0 such that

(y, x) = 0, for all x ∈ H.

We simply choose any y ∈ S ⊥ , and it will do the work so long as S ⊥ ̸= {0}.

Let us extend this problem to Banach space. In a Banach space the concept
of inner product is replaced by linear functional. So whether there is some
nonzero linear functional ℓ ∈ X such that ℓ(x) = 0 for all x ∈ S ⊂ X. We can
also call such ℓ an annihilator. The set of all annihilators can be called as the
orthogonal complement of S, which is different from the Hilbert space setting.

Such a problem is usually handled by Hann-Banach theorem. First, let X be

a Banach space, and p : X −→ R+ be a sublinear function, that is,
1. p(α)x = |α|p(x);
2. p(x + y) ≤ p(x) + p(y).
For example, ∥·∥X : X −→ R+ is sublinear.

Theorem 2.3.1 (Hann-Banach theorem). Let S ⊂ X be a subspace and

ℓ : S −→ C be a linear functional on S such that

|ℓ(x)| ≤ p(x), x ∈ S,

where p : X −→ R+ is a sublinear function. Then there is an extension

ℓ : X −→ C such that
(1). ℓ(x) = ℓ(x) for all x ∈ S;
(2). |ℓ(x)| ≤ p(x) for all x ∈ X.

16
Proof. The finite dimension X is easy, given S ⊂ X, and ℓ pre-defines on
S, we add a vector x1 ∈ X \ S, and we extend ℓ to S1 = span(S ∪ {x1 }) so
that the extension ℓ1 still satisfies the bound |ℓ1 (x)| ≤ p(x) for all x in the
extended subspace. Then we add another vector x2 ∈ X \ S1 . We continuous
this until the vector space is exhausted. The final extension will be what we
need.
This argument cannot be worked out for the case of infinite dimensions. We
need some argument which says if things can be done for finite case, it can
also be done on the infinite case. Therefore, we need Zorn’s lemma.
We first prove the case of real Banach spaces, then the complex case follows
by complexification.
Step 1: We show we can always extend ℓ : S −→ R from S ̸= X to ℓ : S ′ −→
R, where S ′ has one more (dimension) than S. As mentioned earlier, we take
v ∈ X \ S, and we define S ′ = span(S ∪ {v}) and ℓ : S ′ −→ R by

ℓ(x + hv) = ℓ(x) + hα0 , h ∈ R,

where α0 ∈ R. It is trivial that ℓ is an extension. The critical part is: how

to choose α0 ∈ R so that the bound |ℓ(x)| ≤ p(x) is still true. So we need a
choice of α0 such that

|ℓ(x + hv)| = |ℓ(x) + hα0 | ≤ p(x + hv), for all x ∈ S.

Set h = −1, the above is equivalent to

|ℓ(x) − α0 | ≤ p(x − v), for all x ∈ S.

To verify the existence of such α0 , we find observe

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

17
Rearranging terms: for all x, y ∈ S,

ℓ(x) − p(x − v) ≤ ℓ(y) + p(y − v).

If we set f (x) = ℓ(x)−p(x−v) and g(y) = ℓ(y)+p(y−v). We find f (x) ≤ g(y)

for all x, y ∈ S, namely, supx∈S f (x) ≤ inf y∈S g(y). Set α0 to be a number in
between the sup and inf. Then we have

ℓ(x) − p(x − v) ≤ α0 ≤ ℓ(y) + p(y − v), x, y ∈ S.

Set x = y and rearranging:

−p(x − v) ≤ α0 − ℓ(x) ≤ p(x − v) =⇒ |α0 − ℓ(x)| ≤ p(x − v).

Step 2: Consider

P = {(M, λ) : S ⊂ M ⊂ a subspace, λ : M −→ R is linear and |ℓ(x)| ≤

p(x)}.

On the set P we have a partial order (M, λ) ≺ (M ′ , λ′ ) if M ⊂ M ′ and λ′

is an extension of λ. Given a chain C, an upper bound exists, namely, set
M+ to be the subspace spanned by the union of all subspaces in the chain,
and λ+ by λ+ (x) = λ(x) if x ∈ M and (M, λ) ∈ C. By Zorn’s lemma, we
find a maximal element (M, Λ) ∈ P. Note M = X. If not we may find
an extension using Step 2 above. Also Λ is a linear functional satisfying
|Λ(x)| ≤ p(x) as (M, Λ) is a member of P.
To get the complex version of the theorem, given ℓ : S −→ C with |ℓ(x)| ≤
p(x). Re(ℓ) defines a real linear functional on S with |Re(ℓ)(x)| ≤ p(x).
The real Hann-Banach theorem implies the existence of Λ : X −→ R with
|Λ(x)| ≤ p(x). Now we choose

ℓ(x) = Λ(x) − iΛ(ix).

18
ℓ is a desired extension of ℓ.

Corollary 2.3.2. Let X be a normed space and Y ⊂ X a subspace. Suppose

λ ∈ Y ∗ (a bounded linear functional on Y ). Then there exists an extension
Λ ∈ X ∗ satisfying ∥Λ∥X ∗ = ∥λ∥Y ∗ .

Proof. Choose p(x) = ∥λ∥Y ∗ ∥x∥ and apply Theorem 2.3.1.

Corollary 2.3.3. Let y ∈ X, a normed space. Then there exists a nonzero

Λ ∈ X ∗ such that

Λ(y) = ∥Λ∥X ∗ ∥y∥X .

Proof. Let Y be the subspace consisting of all scalar multiples of y and define
λ(ay) = a ∥y∥. By using Corollary 2.3.2, we can construct Λ with ∥Λ∥ = ∥λ∥
extending λ to all of X. But, since Λ(y) = ∥y∥, ∥Λ∥ = 1 and therefore

Λ(y) = ∥Λ∥X ∗ ∥y∥.

Example 2.3.4. Let x ∈ X be Hilbert spaces, a dual element x∗ ∈ X ∗ is

such that

⟨x∗ , x⟩ = ℓ(x) = ∥x∥.

If X is Hilbert space, then x∗ = x/ ∥x∥. Let L1 (Rd ) be L1 -spaces and f ∈

L1 (Rd ). There is a dual element f ∗ ∈ L∞ (Rd ) such that
Z Z
∗ ∗
⟨f , f ⟩ = f (x)f (x)dx = ∥f ∥L1 = |f (x)|dx
Rd Rd

which is f ∗ (x) = sgnf (x).

19
4. Axiom of Choices and Baire Category Theorem

Let X = {Eα }α be a set of nonempty sets. A choice function is a map

S
f : X −→ α Eα such that f (Eα ) is an element in Eα . The axiom of choice
(AC) states: under the assumptions above, there exists a choice function.

The axiom of choice seems to be very intuitive. In fact, there are quite
some counter intuitive (often wild) consequence follows.
1. Zorn’s lemma-it is proved Zorn’s lemma is equivalent to AC;
2. Well-ordering theorem-every set can be well-ordered. It is also proved this
is equivalent to AC.
3. Banach-Tarski theorem-one can construct two balls of the same weight
from one using piecewise rotation and translation;
4. Existence of non-measurable set;
5. Tarski’s theorem-A has the same cardinality as A × A.

5. Baire’s category theorem and its consequences

Definition 2.5.1. Let X be a topological space, and E ⊂ X.

(i). E is nowhere dense if (E)0 = ∅ (the interior of the closure is empty).
(ii). E is of first category if E is the countable union of nowhere dense sets.
(iii). E is of second category if E is not of first.

Example 2.5.2. In R, Z is nowhere dense, and so is ϵZ for every ϵ > 0. Q is

S
not nowhere dense, but it is of first category because Q = i {ri } is countably
union of rational point, where the single rational point {ri } is nowhere dense.
In fact, any countable sets is of first category. Cantor set is second category
since this is complete metric space.

20
Theorem 2.5.3 (Baire’s category). A complete metric space X is not of
first category.
S
Proof. If X is of first category, we write X = k Fk , where Fk ’s are nowhere
dense. We assume that Fk are closed (Note: If Fk is nowhere dense, so
is its closure). Construct a sequence {xk }k as follows: Take x1 such that
Br1 (x1 ) ∈ X \ F1 (because X \ F1 is open). Take x2 ∈ Br1 (x1 ) \ F2 . The set
is nonempty because F2 is nowhere dense (otherwise Br1 (x1 ) ⊂ F2 , and F2
cannot be nowhere dense). Since Br1 (x1 )\F2 is open, we have some r2 ≤ r1 /2
such that Br (x2 ) ⊂ Br1 (x1 ) \ F2 . Repeat the construction we get a sequence
S
{xk }k with the property that xk ̸∈ j≤k Fj . The sequence is Cauchy from the
construction (because {xk }k≥n ∈ Br/2n (xn )). The completeness leads to the
existence of some element x ∈ X which is not in any Fk ’s, a contradiction.

Uniform boundedness principle

If we have a family of bounded operators {Tα } that has a pointwise bound

∥Tα x∥ ≤ Mx (which is a weak bound), then this family of operators is uni-
formly bounded (a strong bound). The proof relies on Baire category theo-
rem.

Theorem 2.5.4. Let T : X −→ Y be a linear operator, with X, Y two

normed spaces. Then T is bounded if and only if T −1 (B1Y (0)) contains an
interior point.

Proof. If T is bounded (that is, also continuous by Problem 1.4), then the
preimage of B1Y (0) is open, and hence T −1 (B1Y (0)) has an interior point. Con-
versely, if T −1 (B1Y (0)) has an interior point, namely, Bϵ (x0 ) ⊂ T −1 (B1Y (0)).

21
Pick x ∈ Bϵ (0), then

∥T x∥Y = ∥T (x + x0 ) − T (x0 )∥Y ≤ ∥T (x + x0 )∥Y + ∥T x0 ∥Y ≤ 1 + C.

The right hand side does not depend on x. Scaling x by ϵ−1 we find for all
x ∈ B1X (0) we have

∥T x∥Y ≤ (1 + C)/ϵ.

Hence T is bounded.

Theorem 2.5.5 (Uniform boundedness principle). Let X be a Banach

space and F = {Tα }α be a collection of bounded operators X −→ Y (where
Y is a norm space). Suppose for all x ∈ X,

sup ∥Tα x∥ < ∞,

α

that is, ∥Tα x∥ ≤ Mx , then we have

sup ∥Tα ∥op < ∞.

α

Proof. Fill up the space X with a collection of closed set {Bn } with
\
Bn = {x ∈ X : ∥Tα x∥Y ≤ n, ∀α} = Tα−1 (BnY (0)).
α
S
We need to show that X = n Bn . Since X is complete, a consequence of
Baire category theorem shows one of these sets, say Bn is not nowhere dense
0
for some n, that is, Bn0 = Bn ̸= ∅. That is, Bn contains an open ball. Now
we proceed with the same argument from the proof of the theorem 2.5.4.
First we have Bϵ (x0 ) ⊂ Bn . Then for all x ∈ Bϵ (0),

∥Tα x∥Y ≤ ∥Tα (x + x0 )∥Y + ∥Tα x0 ∥Y ≤ n + C,

22
where the right hand side is independent of x ∈ Bϵ (0). Scaling x by ϵ−1 we
find

∥Tα ∥op ≤ (n + C)/ϵ.

Taking the supremum over α yields the desired result.

Open mapping theorem

Theorem 2.5.6 (Open mapping theorem). Let T : X −→ Y be a linear

bounded surjective. Then T is an open map, that is, T sends open sets to
open sets.

Proof. Write X as the union of increasing open ball Bn of radius n ∈ N.

S
Surjectivity implies n T (Bn ) = T (X) = Y . Since Y is complete, Baire
category implies one of the closure T (Bn ) has nonempty interior, namely, for
some r > 0 and x,

BrY (y) ⊂ T (B1 ).

Translating y to 0, we find some R > 0:

BrY (0) = BrY (y) − y ⊂ T (Bn ) − y = T (Bn ) − T (x) = T (Bn ) − T (x)

= T (Bn − x)

= T (Bn (−x))

⊂ T (BR (0))

where x is such that T x = y. Scaling the ball BR (0) down to B1 we conclude

BϵY (0) ⊂ T (B1 ).

23
If we prove

T (B1 ) ⊂ T (B2 ),

then this proves T (B2 ) has an interior point. By scaling and translation, this
shows that T (Bϵ (x)) has an interior point. This then shows that T (U ) is
open.
It remains to prove the inclusion. Let ϵ > 0 be from above. Pick y ∈ T (B1 ),
and ∥y − T (x1 )∥Y < ϵ/2 (by the definition of closure), that is,
Y
y − T x1 ∈ Bϵ/2 ⊂ T (B1/2 ).

Pick x2 ∈ B1/2 such that

Y
y − T (x1 ) − T (x2 ) ∈ Bϵ/4 ⊂ T (B1/4 ).

Construct the sequence {xn }n recursively such that

n
X
Y
y− T (xk ) ∈ Bϵ/2n.

k=1
∗
P∞
Define x = k=1 xk . The series is summable because ∥xk ∥ ≤ 2−k . Moreover,
x∗ ∈ B2 . Passing n → ∞, we find

T (x∗ ) = y ∈ T (B2 ).

This proves T (B1 ) ⊂ T (B2 ). This concludes the proof.

Corollary 2.5.7 (Bounded inverse theorem). Let T : X −→ Y be a

bounded bijective linear map. Then the inverse T −1 : Y −→ X is bounded
linear.

24
Proof. To show T −1 is bounded, it suffices to show that T −1 is continuous.
For any open set U ⊂ X, consider T (U ) in Y . Since T is bijective bounded
linear map, then by open mapping theorem T (U ) is open, so T −1 is contin-
uous.

Definition 2.5.8. Let T be a mapping of a normed linear space X into a

normed linear space Y . The graph of T , denoted by Γ(T ), is defined as

Γ(T ) = {(x, y) : (x, y) ∈ X × Y, y = T x}.

Corollary 2.5.9 (Closed graph theorem). Let T : X −→ Y be an oper-

ator between two Banach spaces. Then T is bounded if and only if the graph
Γ(T ) = {(x, T x) : x ∈ X} is closed in X × Y w.r.t. the product topology,
that is, with the norm

∥(x, y)∥ = ∥x∥X + ∥y∥Y .

Proof. Suppose Γ(T ) is closed. Then since T is linear, Γ(T ) is a subspace of

the Banach space X × Y . By assumption Γ(T ) is closed and thus is a Banach
space in the norm

∥(x, T x)∥ = ∥x∥ + ∥T x∥.

Consider the continuous map π1 , π2 ,

π1 : (x, T x) −→ x, π2 : (x, T x) −→ T x.

π1 is a bijection so by the bounded inverse theorem, π1−1 is continuous. But

T = π2 ◦ π1−1 , so T is continuous. The converse is trivial.

25
Corollary 2.5.10 (Helinger-Toplitz). Let A : H −→ H be a linear oper-
ator, where H is a Hilbert space. Suppose A satisfies

(x, Ay) = (Ax, y) for all x, y ∈ H.

Then A is bounded.

Proof. We will prove that Γ(A) is closed. Suppose that (xn , Axn ) → (x, y).
We need only prove that (x, y) ∈ Γ(A), that is, that y = Ax. But, for any
z ∈ H,

(z, y) = lim (z, Axn ) = lim (Az, xn ) = (Az, x) = (z, Ax).

n→∞ n→∞

Thus y = Ax and Γ(A) is closed.

26