FUNCTIONAL ANALYSIS

CHI-WAI LEUNG

Abstract. This is a note on the course MATH 4010: Functional analysis in 2024-25, 1st term.

1. Normed spaces
Definition 1.1. Let X be a vector space over a field K, where K = R or C. A function ∥·∥ : X → R
is called a norm on X if it satisfies the following conditions.
(i) ∥x∥ ≥ 0 for all x ∈ X.
(ii) ∥x∥ = 0 if and only if x = 0.
(iii) ∥αx∥ = |α|∥x∥ for x ∈ X and α ∈ K.
(iii) (Triangle inequality) ∥x − y∥ ≤ ∥x − z∥ + ∥z − y∥ for all x, y, z ∈ X.
In this case, the pair (X, ∥ · ∥) is called a normed space.

Example 1.2. The following are important examples of finite dimensional normed spaces.
(n)
(i) Let ℓ∞ = {(x1 , ..., xn ) : xi ∈ K, i = 1, 2..., n}. Put ∥(x1 , ..., xn )∥∞ = max{|xi | : i = 1, .., n}.
(n)
(ii) Let ℓp = {(x1 , ...., xn ) : xi ∈ K, i = 1, 2..., n}. Put ∥(x1 , ..., xn )∥p = ( ni=1 |xi |p )1/p for
P
1 ≤ p < ∞.

Proposition 1.3. If X is a normed space, then the addition (x, y) ∈ X × X 7→ x + y ∈ X and the
scalar multiplication (α, x) ∈ K × X 7→ αx ∈ X both are continuous maps.

Notation 1.4. From now on, (X, ∥ · ∥) always denotes a normed space over a field K.
For r > 0 and x ∈ X, let
(i) B(x, r) := {y ∈ X : ∥x − y∥ < r} (called an open ball with the center at x of radius r) and
B ∗ (x, r) := {y ∈ X : 0 < ∥x − y∥ < r}
(ii) B(x, r) := {y ∈ X : ∥x − y∥ ≤ r} (called a closed ball with the center at x of radius r).
Put BX := {x ∈ X : ∥x∥ ≤ 1} and SX := {x ∈ X : ∥x∥ = 1} the closed unit ball and the unit
sphere of X respectively.

Definition 1.5. We say that a sequence (xn ) in X converges to an element a ∈ X if lim ∥xn −a∥ =
0, i.e., for any ε > 0, there is N ∈ N such that ∥xn − a∥ < ε for all n ≥ N .
In this case, (xn ) is said to be convergent and a is called a limit of the sequence (xn ).

Definition 1.6. Let A be a subset of X.

(i) A point z ∈ X is called a limit point of A if for any ε > 0, there is an element a ∈ A such
that 0 < ∥z − a∥ < ε, that is, B ∗ (z, ε) ∩ A ̸= ∅ for all ε > 0.
Furthermore, if A contains the set of all its limit points, then A is said to be closed in X.
(ii) The closure of A, denoted by A, is defined by
A := A ∪ {z ∈ X : z is a limit point of A}.
1
2 CHI-WAI LEUNG

Remark 1.7. Using the notations as above, a point z ∈ A if and only if B(z, r) ∩ A ̸= ∅ for all
r > 0. This is equivalent to saying that there is a sequence (xn ) in A such that xn → a. In fact,
this can be shown by considering r = n1 for n = 1, 2....

Proposition 1.8. Using the notations as before, we have the following assertions.
(i) A is closed in X if and only if its complement X \ A is open in X.
(ii) The closure A is the smallest closed subset of X containing A. The ”smallest” in here
means that if F is a closed subset containing A, then A ⊆ F .
Consequently, A is closed if and only if A = A.
Proof. If A is empty, then the assertions (i) and (ii) both are obvious. Now assume that A ̸= ∅.
For part (i), let C = X \ A and b ∈ C. Suppose that A is closed in X. If there exists an element
b ∈ C \ int(C), then B(b, r) ⫅̸ C for all r > 0. This implies that B(b, r) ∩ A ̸= ∅ for all r > 0 and
hence, b is a limit point of A since b ∈ / A. It contradicts to the closeness of A. Thus, C = int(C)
and thus, C is open.
For the converse of (i), assume that C is open in X. Assume that A has a limit point z but z ∈ / A.
Since z ∈ / A, z ∈ C = int(C) because C is open. Hence, we can find r > 0 such that B(z, r) ⊆ C.
This gives B(z, r) ∩ A = ∅. This contradicts to the assumption of z being a limit point of A. Thus,
A must contain all of its limit points and hence, it is closed.
For part (ii), we first claim that A is closed. Let z be a limit point of A. Let r > 0. Then there
is w ∈ B ∗ (z, r) ∩ A. Choose 0 < r1 < r small enough such that B(w, r1 ) ⊆ B ∗ (z, r). Since w is a
limit point of A, we have ∅ = ̸ B ∗ (w, r1 ) ∩ A ⊆ B ∗ (z, r) ∩ A. Hence, z is a limit point of A. Thus,
z ∈ A as required. This implies that A is closed.
It is clear that A is the smallest closed set containing A.
The last assertion follows from the minimality of the closed sets containing A immediately.
The proof is complete. □

A sequence (xn ) in X is called a Cauchy sequence if for any ε > 0, there is N ∈ N such that
∥xm − xn ∥ < ε for all m, n ≥ N . We have the following simple observation.

Lemma 1.9. Every convergent sequence in X is a Cauchy sequence.

The following notation plays an important role in mathematics.

Definition 1.10. A normed space X is called a Banach space if it is a complete normed

space, i.e., every Cauchy sequence in X is convergent.

Proposition 1.11. Let X be a normed space. Then the following assertions are equivalent.
(i) X is a Banach
P∞ space. P∞
P∞ n=1 xn is absolutely convergent in X, i.e., n=1 ∥xn ∥ < ∞, implies that the
(ii) If a series
series n=1 xn converges in the norm.
Proof. (i) ⇒ (ii) is obvious.
Now suppose that Part (ii) holds. Let (yn ) be a Cauchy sequence in X. It suffices to show that
(yn ) has a convergent subsequence. In fact, by the definition of a Cauchy sequence, there is a
subsequence (ynk ) such that ∥ynk+1 − ynk ∥ < 21k for all k = 1, 2.... By the assumption, the series
P∞
k=1 (ynk+1 − ynk ) converges in the norm, and hence the sequence (ynk ) is convergent in X. The
proof is complete. □
 3

Throughout the note, we write a sequence of numbers as a function x : {1, 2, ...} → K.

The following examples are important classes in the study of functional analysis.
Example 1.12. Put
c0 := {(x(i)) : x(i) ∈ K, lim |x(i)| = 0} (the null sequence space);
ℓ∞ := {(x(i)) : x(i) ∈ K, sup x(i) < ∞ (the bounded sequence space);
i
and
c00 := {(x(i)) : there are only finitly many x(i)’s are non-zero} (the finite sequence space).
The sup-norm ∥ · ∥∞ on ℓ∞ is defined by ∥x∥∞ := supi |x(i)| for x ∈ ℓ∞ . Then ℓ∞ is a Banach
space.
Now if c00 is endowed with the sup-norm defined above, then c00 is dense in c0 , i.e., c00 = c0 .
Consequently, c0 is a closed subspace of ℓ∞ . In particular, c0 is Banach space too.
Proof. We first claim that c00 ⊆ c0 . Let z ∈ ℓ∞ . It suffices to show that if z ∈ c00 , then z ∈ c0 , i.e.,
lim z(i) = 0. Let ε > 0. Then there is x ∈ B(z, ε) ∩ c00 and hence, we have |x(i) − z(i)| < ε for
i→∞
all i = 1, 2..... Since x ∈ c00 , there is i0 ∈ N such that x(i) = 0 for all i ≥ i0 . Therefore, we have
|z(i)| = |z(i) − x(i)| < ε for all i ≥ i0 . Therefore, z ∈ c0 is as desired.
For the reverse inclusion, let w ∈ c0 . We need to show that B(w, r) ∩ c00 ̸= ∅ for all r > 0. Let
r > 0. Since w ∈ c0 , there is i0 such that |w(i)| < r for all i ≥ i0 . If we let x(i) = w(i) for 1 ≤ i < i0
and x(i) = 0 for i ≥ i0 , then x ∈ c00 and ∥x − w∥∞ := sup |x(i) − w(i)| < r is as required. □
i=1,2...

Example 1.13. For 1 ≤ p < ∞. Put

∞
X
ℓp := {(x(i)) : x(i) ∈ K, |x(i)|p < ∞}.
i=1
∞
X 1
In addition, ℓp is equipped with the norm ∥x∥p := ( |x(i)|p ) p for x ∈ ℓp . Then ℓp becomes a
i=1
Banach space under the norm ∥ · ∥p .

Example 1.14. Let X be a locally compact Hausdorff space, for example, K. Let C0 (X) be the
space of all continuous K-valued functions f on X which are vanish at infinity, i.e., for every ε > 0,
there is a compact subset D of X such that |f (x)| < ε for all x ∈ X \ D. Now C0 (X) is endowed
with the sup-norm, i.e.,
∥f ∥∞ = sup |f (x)|
x∈X
for every f ∈ C0 (X). Then C0 (X) is a Banach space. (Try to prove this fact for the case
X = R. Just use the knowledge from MATH 2060 !!!)

Proposition 1.15. Let (X, ∥ · ∥) be a normed space. Then there is a normed space (X0 , ∥ · ∥0 ),
together with a linear map i : X → X0 , satisfies the following conditions.
(i) X0 is a Banach space.
(ii) The map i is an isometry, that is, ∥i(x)∥0 = ∥x∥ for all x ∈ X.
(iii) the image i(X) is dense in X0 , that is, i(X) = X0 .
Moreover, such pair (X0 , i) is unique up to isometric isomorphism in the following sense.
If (W, ∥ · ∥1 ) is a Banach space and an isometry j : X → W is an isometry such that j(X) = W ,
then there is an isometric isomorphism ψ from X0 onto W such that
j = ψ ◦ i : X → X0 → W.
4 CHI-WAI LEUNG

In this case, the pair (X0 , i) is called the completion of X.

Example 1.16. Proposition 1.15 cannot give an explicit form of the completion of a given normed
space. The following examples are basically due to the uniqueness of the completion.
(i) If X is a Banach space, then the completion of X is itself.
(ii) The completion of the finite sequence space c00 is the null sequence space c0 .
(iii) The completion of Cc (R) is C0 (R).

2. Finite Dimensional Normed Spaces

Throughout this section, let (X, ∥ · ∥) is a normed space. Put SX the unit sphere of X, i.e.,
SX = {x ∈ X : ∥x∥ = 1}.

Definition 2.1. Two norms ∥·∥ and ∥·∥′ on a vector space X are equivalent, denoted by ∥·∥ ∼ ∥·∥′ ,
if there are positive numbers c1 and c2 such that c1 ∥ · ∥ ≤ ∥ · ∥′ ≤ c2 ∥ · ∥ on X.

Example 2.2. Consider the norms ∥·∥1 and ∥·∥∞ on ℓ1 . We want to show that ∥·∥1 and ∥·∥∞ are
not equivalent. In fact, if we put xn (i) := (1, 1/2, ..., 1/n, 0, 0, ....) for n, i = 1, 2.... Then xn ∈ ℓ1
for all n. Note that (xn ) is a Cauchy sequence with respect to the norm ∥ · ∥∞ but it is not a Cauchy
sequence with respect to the norm ∥ · ∥1 . Hence ∥ · ∥1 ≁ ∥ · ∥∞ on ℓ1 .

Proposition 2.3. All norms on a finite dimensional vector space are equivalent.
Proof. LetPX be a finite dimensional vector space and let {e1 , ..., en } be a vector basis of X. For
each x = ni=1 αi ei for αi ∈ K, define ∥x∥0 = maxni=1 |αi |. Then ∥ · ∥0 is a norm X. The result is
obtained by showing that all norms ∥ · ∥ on X are equivalent
X to ∥ · ∥0 .
Pn
Note that for each x = i=1 αi ei ∈ X, we have ∥x∥ ≤ ( ∥ei ∥)∥x∥0 . It remains to find c > 0
1≤i≤n
such that c∥ · ∥0 ≤ ∥ · ∥. In fact, let SX := {x ∈ X : ∥x∥0 = 1} be the unit sphere of X with respect
to the norm ∥ · ∥0 . Note that by using the Weierstrass Theorem on K, we see that SX is compact
with respect to the norm ∥ · ∥0 .
Define a real-valued function f on the unit sphere SX of X by
f : x ∈ SX 7→ ∥x∥.
Note that f > 0 and f is continuous with respect to the norm ∥ · ∥0 . Hence, there is c > 0 such
that f (x) ≥ c > 0 for all x ∈ SX . This gives ∥x∥ ≥ c∥x∥0 for all x ∈ X as desired. The proof is
complete. □

Corollary 2.4. We have the following assertions.

(i) All finite dimensional normed spaces are Banach spaces. Consequently, any finite dimen-
sional subspace of a normed space is closed.
(ii) The closed unit ball of any finite dimensional normed space is compact.
Proof. Let (X, ∥ · ∥) be a finite dimensional normed space. Using the notations as in the proof of
Proposition 2.3 above, we see that ∥ · ∥ must be equivalent to the norm ∥ · ∥0 . X is clearly complete
with respect to the norm ∥ · ∥0 and so is complete in the original norm ∥ · ∥. The Part (i) follows.
For Part (ii), it is clear that the compactness of the closed unit ball of X is equivalent to saying that
any closed and bounded subset is compact. Therefore, Part (ii) follows from the simple observation
that any closed and bounded subset of X with respect to the norm ∥ · ∥0 is compact. The proof is
complete. □
 5

In the remainder of this section, we want to show that the converse of Corollary 2.4(ii) holds.
Before this result, we need the following useful result.

Lemma 2.5. Riesz’s Lemma: Let Y be a closed proper subspace of a normed space X. Then for
each θ ∈ (0, 1), there is an element x0 ∈ SX such that d(x0 , Y ) := inf{∥x0 − y∥ : y ∈ Y } ≥ θ.
Proof. Let u ∈ X − Y and d := inf{∥u − y∥ : y ∈ Y }. Note that since Y is closed, d > 0 and
hence we have 0 < d < dθ because 0 < θ < 1. This implies that there is y0 ∈ Y such that
u−y0
0 < d ≤ ∥u − y0 ∥ < dθ . Now put x0 := ∥u−y 0∥
∈ SX . We are going to show that x0 is as desired.
Indeed, let y ∈ Y . Since y0 + ∥u − y0 ∥y ∈ Y , we have
1
∥x0 − y∥ = ∥u − (y0 + ∥u − y0 ∥y)∥ ≥ d/∥u − y0 ∥ > θ.
∥u − y0 ∥
Thus, d(x0 , Y ) ≥ θ. □

Remark 2.6. The Riesz’s lemma does not hold when θ = 1. The following example can be found
in the Diestel’s interesting book without proof (see [6, Chapter 1 Ex.3(i)]).
R1
Let X = {x ∈ C([0, 1], R) : x(0) = 0} and Y = {y ∈ X : 0 y(t)dt = 0}. Both X and Y are
endowed with the sup-norm. Note that Y is a closed proper subspace of X. We are going to show
that for any x ∈ SX , there is y ∈ Y such that ∥x − y∥∞ < 1. Thus, the Riesz’s Lemma does not
hold as θ = 1 in this case.
In fact, let x ∈ SX . Since x(0) = 0 with ∥x∥∞ = 1, we can find 0 < a < 1/4 such that |x(t)| ≤ 1/4
for all t ∈ [0, a].
We fix 0 < ε < 1/4 first. Since x is uniform continuous on [a, 1], we can find a partitions a = t0 <
· · · < tn = 1 on [a, 1] such that sup{|x(t) − x(t′ )| : t, t′ ∈ [tk−1 , tk ]} < ε/4. Now for each (tk−1 , tk ),
if sup{x(t) : t ∈ [tk−1 , tk ]} > ε, then we set ϕ(t) = ε. In addition, if inf{x(t) : t ∈ [tk−1 , tk ]} < −ε,
then we set ϕ(t) = −ε. From this, one can construct a continuous function ϕ on [a, 1] such that
R1
∥ϕ − x|[a,1] ∥∞ < 1 and |ϕ(x)| < 2ε for all x ∈ [a, 1]. Hence, we have | a ϕ(t)dt| ≤ 2ε(1 − a).
As |x(t)| < 1/4 on [0, a], so if we choose ε small enough such that (1 − a)(2ε) < a/4, then we can
find a continuous function y1 on [0, a] such that |y1 (t)| < 1/4 on [0, a] with y1 (0) = 0; y1 (a) = x(a)
Ra R1
and 0 y1 (t)dt = − a ϕ(t)dt. Now we define y = y1 on [0, a] and y = ϕ on [a, 1]. Then ∥y −x∥∞ < 1
and y ∈ Y is as desired.

Theorem 2.7. X is a finite dimensional normed space if and only if the closed unit ball BX of X
is compact.
Proof. The necessary condition has been shown by Proposition 2.4(ii).
Now assume that X is of infinite dimension. Fix an element x1 ∈ SX . Let Y1 = Kx1 . Then
Y1 is a proper closed subspace of X. The Riesz’s lemma gives an element x2 ∈ SX such that
∥x1 − x2 ∥ ≥ 1/2. Now consider Y2 = span{x1 , x2 }. Then Y2 is a proper closed subspace of X since
dim X = ∞. To apply the Riesz’s Lemma again, there is x3 ∈ SX such that ∥x3 − xk ∥ ≥ 1/2 for
k = 1, 2. To repeat the same step, there is a sequence (xn ) ∈ SX such that ∥xm − xn ∥ ≥ 1/2 for
all n ̸= m. Thus, (xn ) is a bounded sequence without any convergence subsequence. Hence, BX is
not compact. The proof is complete. □
Recall that a metric space Z is said to be locally compact if for any point z ∈ Z, there is a
compact neighborhood of z. Theorem 2.7 implies the following corollary immediately.
Corollary 2.8. Let X be a normed space. Then X is locally compact if and only if dim X < ∞.

In view of the proof of Theorem 2.7, we call a subset A of a normed space X a r-discrete set if
∥x − y∥ > r whenever x, y ∈ A with x ̸= y.
6 CHI-WAI LEUNG

Proposition 2.9. Let X be an infinite dimensional normed space and r > 0. If A := {xi }i∈I is
an infinite bounded r-discrete subset of X, then |I| ≤ dim X. Consequently, if {Bi : i ∈ I} is a
collection of pairwise disjoint r/2-balls in a bounded subset of X, then |I| ≤ dim X.
Proof. Let C be the collection of linearly independent subsets of A. Then the Zorn’s Lemma gives
a maximal element {xi : i ∈ J} in C for a subset J of I. We are going to show that |J| = |I|.
Suppose not. We assume that |J| < |I|. Now for each finite subset F of J, put VF the linear span
of the set F . Since VF is of finite dimension, the set VF ∩ A has finite volume with respect to the
Lebesgue measure in VF . Then VF ∩ A is a finite set because it is a r-discrete set. Therefore, if we
let F(J) := {F ⊆ J : |F | < ∞}, then we have
[
| VF ∩ A| = |F(J)| = |J|.
F ∈F(J)
[
Hence, if |J| < |I|, then VF ∩ A ⊊ A. Hence, there is an element xi0 ∈ A so that xi0 ∈
/ VF ∩ A
F ∈F(J)
for all F ∈ F(J). This implies that xi0 is linearly independent to the set (xi )i∈J . This contradicts
to the maximality of (xi )i∈J . Hence, |I| = |J| ≤ dim X by the definition of dimension. □

3. Bounded Linear Operators

Proposition 3.1. Let T be a linear operator from a normed space X into a normed space Y . Then
the following statements are equivalent.
(i) T is continuous on X.
(ii) T is continuous at 0 ∈ X.
(iii) sup{∥T x∥ : x ∈ BX } < ∞.
In this case, let ∥T ∥ = sup{∥T x∥ : x ∈ BX } and T is said to be bounded.
Proof. (i) ⇒ (ii) is obvious.
For (ii) ⇒ (i), suppose that T is continuous at 0. Let x0 ∈ X. Let ε > 0. Then there is δ > 0 such
that ∥T w∥ < ε for all w ∈ X with ∥w∥ < δ. Therefore, we have ∥T x − T x0 ∥ = ∥T (x − x0 )∥ < ε for
any x ∈ X with ∥x − x0 ∥ < δ. Part (i) follows.
For (ii) ⇒ (iii), since T is continuous at 0, there is δ > 0 such that ∥T x∥ < 1 for any x ∈ X with
∥x∥ < δ. Now for any x ∈ BX with x ̸= 0, we have ∥ 2δ x∥ < δ. Therefore, we see have ∥T ( 2δ x)∥ < 1
and hence, we have ∥T x∥ < 2/δ. Part (iii) follows.
Finally, we need to show (iii) ⇒ (ii). Note that by the assumption of (iii), there is M > 0 such
that ∥T x∥ ≤ M for all x ∈ BX . Thus, for each x ∈ X, we have ∥T x∥ ≤ M ∥x∥. This implies that
T is continuous at 0. The proof is complete. □

Corollary 3.2. Let T : X → Y be a bounded linear map. Then we have

sup{∥T x∥ : x ∈ BX } = sup{∥T x∥ : x ∈ SX } = inf{M > 0 : ∥T x∥ ≤ M ∥x∥, ∀x ∈ X}.
Proof. Let a = sup{∥T x∥ : x ∈ BX }, b = sup{∥T x∥ : x ∈ SX } and c = inf{M > 0 : ∥T x∥ ≤
M ∥x∥, ∀x ∈ X}.
Clearly, we have b ≤ a. Now for each x ∈ BX with x ̸= 0, then we have b ≥ ∥T (x/∥x∥)∥ =
(1/∥x∥)∥T x∥ ≥ ∥T x∥. Thus, we have b ≥ a and thus, a = b.
Now if M > 0 satisfies ∥T x∥ ≤ M ∥x∥, ∀x ∈ X, then we have ∥T w∥ ≤ M for all w ∈ SX . Hence,
we have b ≤ M for all such M , and so we have b ≤ c. Finally, it remains to show c ≤ b. Note that
by the definition of b, we have ∥T x∥ ≤ b∥x∥ for all x ∈ X. Thus, c ≤ b. □
 7

Proposition 3.3. Let X and Y be normed spaces. Let B(X, Y ) be the set of all bounded linear
maps from X into Y . For each element T ∈ B(X, Y ), let
∥T ∥ = sup{∥T x∥ : x ∈ BX }.
be defined as in Proposition 3.1.
Then (B(X, Y ), ∥ · ∥) becomes a normed space.
Furthermore, if Y is a Banach space, then so is B(X, Y ).
In particular, if Y = K, then B(X, K) is a Banach space. In this case, put X ∗ := B(X, K) and call
it the dual space of X.
Proof. We can directly check that B(X, Y ) is a normed space (Do It By Yourself !).
We want to show that B(X, Y ) is complete if Y is a Banach space. Let (Tn ) be a Cauchy sequence in
B(X, Y ). Then for each x ∈ X, it is easy to see that (Tn x) is a Cauchy sequence in Y . Thus, lim Tn x
exists in Y for each x ∈ X because Y is complete. Hence, we can define a map T x := lim Tn x ∈ Y
for each x ∈ X. Clearly, T is a linear map from X into Y .
We need show that T ∈ B(X, Y ) and ∥T − Tn ∥ → 0 as n → ∞. Let ε > 0. Since (Tn ) is a Cauchy
sequence in B(X, Y ), there is a positive integer N such that ∥Tm − Tn ∥ < ε for all m, n ≥ N .
Hence, we have ∥(Tm − Tn )(x)∥ < ε for all x ∈ BX and m, n ≥ N . Taking m → ∞, we have
∥T x − Tn x∥ ≤ ε for all n ≥ N and x ∈ BX . Therefore, we have ∥T − Tn ∥ ≤ ε for all n ≥ N . From
this, we see that T − TN ∈ B(X, Y ) and thus, T = TN + (T − TN ) ∈ B(X, Y ) and ∥T − Tn ∥ → 0
as n → ∞. Therefore, limn Tn = T exists in B(X, Y ). □

Remark 3.4. By using Proposition 3.1, we can show that if f : X → K is any linear functional
defined on a vector space X, then X can be endowed with a norm so that f is bounded. P
In fact, if we fix a vector basis (ei )i∈I for X and put ∥x∥∞ := maxi∈I |ai | as x = i∈I ai ei ∈ X,
(note that it is a finite sum), where ai ∈ K, then the function ∥ · ∥∞ is a norm on X. Now for each
x ∈ X, set
∥x∥1 := |f (x)| + ∥x∥∞ .
Clearly, the function ∥ · ∥1 is a norm on X. In addition, we have |f (x)| ≤ ∥x∥1 for all x ∈ X.
Hence, f is bounded on X with respect to the norm ∥ · ∥1 as required.

Proposition 3.5. Let X and Y be normed spaces. Suppose that X is of finite dimension n. Then
we have the following assertions.
(i) Any linear operator from X into Y must be bounded.
(ii) If Tk : X → Y is a sequence of linear operators such that Tk x → 0 for all x ∈ X, then
∥Tk ∥ → 0.
Proof. Using Proposition 2.3 and the notations as in the proof, then there is c > 0 such that
Xn n
X
|αi | ≤ c∥ αi ei ∥
i=1 i=1
for all scalars α1 , ..., αn . Therefore, for any linear map T from X to Y , we have


∥T x∥ ≤ max ∥T ei ∥ c∥x∥
1≤i≤n

for all x ∈ X. This gives the assertions (i) and (ii) immediately. □

Remark 3.6. The assumption of X of finite dimension in Proposition 3.5 cannot be removed. For
example, if for each positive integer k, we define fk : c0 → R by fk (x) := x(k), then fk is bounded
for each k and
lim fk (x) = lim x(k) = 0
k→∞ k→∞
for all x ∈ c0 . However fk ↛ 0 because ∥fk ∥ ≡ 1 for every k.
8 CHI-WAI LEUNG

Proposition 3.7. Let Y be a closed subspace of X and X/Y be the quotient space. For each
element x ∈ X, put x̄ := x + Y ∈ X/Y the corresponding element in X/Y . Define

(3.1) ∥x̄∥ = inf{∥x + y∥ : y ∈ Y }.

If we let π : X → X/Y be the natural projection, i.e., π(x) = x̄ for all x ∈ X, then (X/Y, ∥ · ∥)
is a normed space and π is bounded with ∥π∥ ≤ 1. In particular, ∥π∥ = 1 as Y is a proper closed
subspace.
Furthermore, if X is a Banach space, then so is X/Y .
In this case, we call ∥ · ∥ in (3.1) the quotient norm on X/Y .

Proof. Note that since Y is closed, we can directly check that ∥x̄∥ = 0 if and only is x ∈ Y , i.e.,
x̄ = 0̄ ∈ X/Y . It is easy to check the other conditions of the definition of a norm. Thus, X/Y is a
normed space. Moreover, π is clearly bounded with ∥π∥ ≤ 1 by the definition of the quotient norm
on X/Y .
Furthermore, if Y ⊊ X, then by using the Riesz’s Lemma 2.5, we see that ∥π∥ = 1.
We show the last assertion. Suppose that X is a Banach space. Let (x̄n ) be a Cauchy sequence in
X/Y . It suffices to show that (x̄n ) has a convergent subsequence in X/Y .
Indeed, since (x̄n ) is a Cauchy sequence, we can find a subsequence (x̄nk ) of (x̄n ) such that

∥x̄nk+1 − x̄nk ∥ < 1/2k

for all k = 1, 2.... Then by the definition of quotient norm, there is an element y1 ∈ Y such that
∥xn2 − xn1 + y1 ∥ < 1/2. Note that we have, xn1 − y1 = x̄n1 in X/Y . Thus, there is y2 ∈ Y such
that ∥xn2 − y2 − (xn1 − y1 )∥ < 1/2 by the definition of quotient norm again. In addition, we have
xn2 − y2 = x̄n2 . Then we also have an element y3 ∈ Y such that ∥xn3 − y3 − (xn2 − y2 )∥ < 1/22 .
To repeat the same step, we can obtain a sequence (yk ) in Y such that

∥xnk+1 − yk+1 − (xnk − yk )∥ < 1/2k

for all k = 1, 2.... Therefore, (xnk − yk ) is a Cauchy sequence in X and thus, limk (xnk − yk ) exists
in X while X is a Banach space. Set x = limk (xnk − yk ). On the other hand, note that we have
π(xnk − yk ) = π(xnk ) for all k = 1, 2, , ,. This tells us that limk π(xnk ) = limk π(xnk − yk ) = π(x) ∈
X/Y since π is bounded. Therefore, (x̄nk ) is a convergent subsequence of (x̄n ) in X/Y . The proof
is complete. □

Corollary 3.8. Let T : X → Y be a linear map. Suppose that Y is of finite dimension. Then T
is bounded if and only if ker T := {x ∈ X : T x = 0} is closed.

Proof. The necessary part is clear.

Now assume that ker T is closed. Then by Proposition 3.7, X/ ker T becomes a normed space.
Morover, it is known that there is a linear injection Te : X/ ker T → Y such that T = Te ◦ π, where
π : X → X/ ker T is the natural projection. Since dim Y < ∞ and Te is injective, dim X/ ker T < ∞.
This implies that Te is bounded by Proposition 3.5. Hence T is bounded because T = Te ◦ π and π
is bounded. □

Remark 3.9. The converse of PCorollary 3.8 does not hold when Y is of infinite dimension. For
∞
example, let X := {x ∈ ℓ : n=1 n |x(n)|2 < ∞} (note that X is a vector space Why?) and
2 2

Y = ℓ2 . Both X and Y are endowed with ∥ · ∥2 -norm.

Define T : X → Y by T x(n) = nx(n) for x ∈ X and n = 1, 2.... Then T is an unbounded
operator(Check !!). Note that ker T = {0} and hence, ker T is closed. Hence, the closeness of
ker T does not imply the boundedness of T in general.
 9

Two normed spaces X and Y are said to be isomorphic (resp. isometric isomorphic) if there is
a bi-continuous linear isomorphism (resp. isometric) between X and Y . We write X = Y if X and
Y are isometric isomorphic.

Remark 3.10. Note that the inverse of a bounded linear isomorphism need not be bounded.
Example 3.11. Let X : {f ∈ C ∞ (−1, 1) : f (n) ∈ C b (−1, 1) for all n = 0, 1, 2...} and Y := {f ∈
X : f (0) = 0}. In addition, X and Y both are equipped with the sup-norm ∥ · ∥∞ . Define an
operator S : X → Y by Z x
Sf (x) := f (t)dt
0
for f ∈ X and x ∈ (−1, 1). Then S is a bounded linear isomorphism but its inverse S −1 is
unbounded. In fact, the inverse S −1 : Y → X is given by
S −1 g := g ′
for g ∈ Y .

A metric space is said to be separable if there is a countable dense subset, for example, the base
field K is separable. Moreover, it is easy to see that a normed space is separable if and only if it is
the closed linear span of a countable dense subset.

Definition 3.12. A sequence of element (en )∞ n=1 in a normed space X is called a Schauder basis
for X if for each element x ∈ X, there is a unique sequence of scalars (αn ) such that
∞
X
(3.2) x= αn e n .
n=1
Note: The expression in Eq. 3.2 depends on the order of en ’s.

Remark 3.13. Note that if X has a Schauder basis, then X must be separable. The following
natural question was first raised by Banach (1932).
The basis problem: Does every separable Banach space have a Schauder basis?
The answer is “No′′ !
This problem was completely solved by P. Enflo in 1973.

Example 3.14. We have the following assertions.

(i) The space ℓ∞ is non-separable under the sup-norm ∥·∥∞ . Consequently, ℓ∞ has no Schauder
basis.
(ii) The spaces c0 and ℓp for 1 ≤ p < ∞ have Schauder bases.
Proof. For Part (i) let D = {x ∈ ℓ∞ : x(i) = 0 or 1}. Then D is an uncountable set and
∥x − y∥∞ = 1 for x ̸= y. Therefore {B(x, 1/4) : x ∈ D} is an uncountable family of disjoint open
balls. Therefore, ℓ∞ has no countable dense subset.
For each n = 1, 2..., let en (i) = 1 if n = i, otherwise, is equal to 0.
In addition, (en ) is a Schauder basis for the space c0 and ℓp for 1 ≤ p < ∞. □

In the rest of this section, we are going to investigate some concrete examples of dual spaces.
Example 3.15. Let X = KN . Consider the usual Euclidean norm on X, i.e., ∥(x1 , ..., xN )∥ :=
|x1 |2 + · · · |xN |2 . Define θ : KN → (KN )∗ by θx(y) = x1 y1 + · · · + xN yN for x = (x1 , ..., xN )
p

and y = (y1 , ..., yN ) ∈ KN . Note that θx(y) = ⟨x, y⟩, the usual inner product on KN . Then by
the Cauchy-Schwarz inequality, it is easy to see that θ is an isometric isomorphism. Therefore, we
have KN = (KN )∗ .
10 CHI-WAI LEUNG

Example 3.16. Define a map T : ℓ1 → c∗0 by

∞
X
(T x)(η) = x(i)η(i)
i=1

for x ∈ ℓ1 and η ∈ c0 .
Then T is isometric isomorphism and hence, c∗0 = ℓ1 .

Proof. The proof is divided into the following steps.

Step 1. T x ∈ c∗0 for all x ∈ ℓ1 .
In fact, let η ∈ c0 . Then
∞
X ∞
X
|T x(η)| ≤ | x(i)η(i)| ≤ |x(i)||η(i)| ≤ ∥x∥1 ∥η∥∞ .
i=1 i=1

Step 1 follows.
Step 2. T is an isometry.
Note that by Step 1, we have ∥T x∥ ≤ ∥x∥1 for all x ∈ ℓ1 . We need to show that ∥T x∥ ≥ ∥x∥1 for
all x ∈ ℓ1 . Fix x ∈ ℓ1 . Now for each k = 1, 2.., consider the polar form x(k) = |x(k)|eiθk . Note that
ηn := (e−iθ1 , ..., e−iθn , 0, 0, ....) ∈ c0 for all n = 1, 2.... Then we have
n
X n
X
|x(k)| = x(k)ηn (k) = T x(ηn ) = |T x(ηn )| ≤ ∥T x∥
k=1 k=1

for all n = 1, 2.... Hence, we have ∥x∥1 ≤ ∥T x∥.

Step 3. T is a surjection.
Let ϕ ∈ c∗0 and let ek ∈ c0 be given by ek (j) = 1 if j = k, otherwise, is equal to 0. Put x(k) := ϕ(ek )
for k = 1, 2... and consider the polar form x(k) = |x(k)|eiθk as above. Then we have
n
X n
X n
X
|x(k)| = ϕ( e−iθk ek ) ≤ ∥ϕ∥∥ e−iθk ek ∥∞ = ∥ϕ∥
k=1 k=1 k=1

for all n = 1, 2.... Therefore, x ∈ ℓ1 .

Finally, we need to show that T x = ϕ and thus, T is surjective. In fact, if η = ∞
P
k=1 η(k)ek ∈ c0 ,
then we have
∞
X ∞
X
ϕ(η) = η(k)ϕ(ek ) = η(k)x(k) = T x(η).
k=1 k=1
The proof is complete by the Steps 1-3 above. □

Example 3.17. We have the other important examples of the dual spaces.
(i) (ℓ1 )∗ = ℓ∞ .
(ii) For 1 < p < ∞, (ℓp )∗ = ℓq , where p1 + 1q = 1.
(iii) For a locally compact Hausdorff space X, C0 (X)∗ = M (X), where M (X) denotes the space
of all regular Borel measures on X.
Parts (i) and (ii) can be obtained by the similar argument as in Example 3.16 (see also in [?,
Chapter 8]). Part (iii) is known as the Riesz representation Theorem which is referred to [?,
Section 21.5] for the details.

Example 3.18. Let C[a, b] be the space of all continuous R-valued functions defined on a closed
and bounded interval [a, b]. Moreover, the space C[a, b] is endowed with the sup-norm, i.e., ∥f ∥∞ :=
 11

sup{|f (x)| : x ∈ [a, b] for f ∈ C[a, b].

A function ρ : [a, b] → R is said to be a bounded variation if it satisfies the condition:
Xn
V (ρ) := sup{ |ρ(xk ) − ρ(xk−1 )| : a = x0 < x1 < · · · < xn = b} < ∞.
k=1

Let BV ([a, b]) denote the space of all bounded variations on [a, b] and let ∥ρ∥ := V (ρ) for ρ ∈
BV ([a, b]). Then BV ([a, b]) becomes a Banach space.
Besides, for f ∈ C[a, b], the Riemann-Stieltjes integral of f with respect to a bounded variation ρ
on [a, b] is defined by
Z b n
X
f (x)dρ(x) := lim f (ξk )(ρ(xk ) − ρ(xk−1 )),
a P
k=1

where P : a = x0 < x1 < · · · < xn = b and ξk ∈ [xk−1 , xk ] (Fact: the Riemann-Stieltjes

integral of a continuous function always exists).
Define a mapping T : BV ([a, b]) → C[a, b]∗ by
Z b
T (ρ)(f ) := f (x)dρ(x)
a
for ρ ∈ BV ([a, b]) and f ∈ C[a, b]. Then T is an isometric isomorphism, and hence, we have
C[a, b]∗ = BV ([a, b]).

4. Hahn-Banach Theorem
A real valued function p : X → R defined on a vector space X is called a positively homogeneous
sub-additive if the following conditions hold:
(i) p(αx) = αp(x) for all x ∈ X and α ≥ 0.
(ii) p(x + y) ≤ p(x) + p(y) for all x, y ∈ X.
Lemma 4.1. Let X be a real vector space and Y be a subspace of X. Assume that there is an
element v ∈ X \ Y such that X = Y ⊕ Rv, i.e., the space X is the linear span of Y and v. Let p be
a positive homogeneous sub-additive function defined on X. Suppose that f is real linear functional
defined on Y satisfying f (y) ≤ p(y) for all y ∈ Y . Then there is a real linear extension F of f
defined on X so that
F (x) ≤ p(x) for all x ∈ X.
Proof. It is noted that if F is a linear extension of f on X and γ := F (v) which satisfies
F (y + tv) = f (y) + tγ ≤ p(y + tv) for all y ∈ Y and for all t ∈ R,
then it suffices to saying that the following inequalities hold:
(4.1) f (y1 ) + γ ≤ p(y1 + v) and f (y2 ) − γ ≤ p(y2 − v)
for all y1 , y2 ∈ Y . Thus, we need to determine γ := F (v) so that the following holds:

(4.2) f (y1 ) − p(y1 − v) ≤ γ ≤ −f (y2 ) + p(y2 + v) for all y1 , y2 ∈ Y .

Note that if we fix y1 , y2 ∈ Y , we see that
f (y1 ) + f (y2 ) = f (y1 + y2 ) ≤ p(y1 + y2 ) ≤ p(y1 − v) + p(y2 + v).
This implies that we have
f (y1 ) − p(y1 − v) ≤ −f (y2 ) + p(y2 + v)
12 CHI-WAI LEUNG

for all y1 , y2 ∈ Y . Therefore, it gives

a := sup{f (y1 ) − p(y1 − v) : y1 ∈ Y } ≤ b := inf{−f (y2 ) + p(y2 + v) : y2 ∈ Y }.
Now if we choose a real number γ so that a ≤ γ ≤ b, then the Inequality 4.2 holds. The proof is
complete. □

Remark 4.2. Before completing the proof of the Hahn-Banach Theorem, Let us first recall one
of super important results in mathematics, called Zorn’s Lemma, a very humble name. Every
mathematics student should know it.
Zorn’s Lemma: Let X be a non-empty set with a partially order “ ≤ ”. Assume that every totally
order subset C of X has an upper bound, i.e. there is an element z ∈ X such that c ≤ z for all c ∈ C.
Then X must contain a maximal element m, that is, if m ≤ x for some x ∈ X, then m = x.

The following is the typical argument of applying the Zorn’s Lemma.

Theorem 4.3. Hahn-Banach Theorem : Let X be a vector space ( not necessary to be a normed
space) over R and let Y be a subspace of X. Let p be a positive homogeneous sub-additive function
defined on X. Suppose that f is a real linear functional defined on Y satisfying f (y) ≤ p(y) for all
y ∈ Y . Then there is a real linear extension F of f defined on X so that
F (x) ≤ p(x) for all x ∈ X.
Proof. Let X be the collection of the pairs (Y1 , f1 ), where Y ⊆ Y1 is a subspace of X and f1 is a
linear extension of f defined on Y1 such that and f1 ≤ p on Y1 . Define a partial order ≤ on X by
(Y1 , f1 ) ≤ (Y2 , f2 ) if Y1 ⊆ Y2 and f2 |Y1 = f1 . Then by the Zorn’s lemma, there is a maximal element
(Ye , F ) in X. The maximality of (Ye , F ) and Lemma 4.1 give Ye = X. The proof is complete. □

Definition 4.4. Let D be a convex subset of a normed space X, i.e., tx + (1 − t)y ∈ D for all
x, y ∈ D and t ∈ (0, 1). Suppose that 0 is an interior point of D. Define
µD (x) := inf{t > 0 : x ∈ tD}
for x ∈ X. In addition, set µD (x) = ∞ if {t > 0 : x ∈ tD} = ∅.
The function µD is called the Minkowski functional with respect to D.

Lemma 4.5. Let D be a convex subset of a normed space X. Suppose that 0 is an interior point
of D. Then the Minkowski functional µ := µD : X → [0, ∞) is positively homogeneous and sub-
additive on D.
In addition, we have {x ∈ X : µ(x) < 1} ⊆ D ⊆ {x ∈ X : µ(x) ≤ 1}.
Proof. It is noted that since 0 ∈ int(D), the set {t > 0 : x ∈ tD} = ̸ ∅ for all x ∈ X. Thus, the
function µ : X → [0, ∞) is defined.
Clearly, if we fix t > 0 and x ∈ X, then we have µ(tx) ≤ s if and only if tµ(x) ≤ s. Hence, the
function µ is positively homogeneous.
Next, we show the subadditivity of µ. Let ε > 0. For x, y ∈ X, we choose s, t > 0 such that x ∈ sD
and y ∈ tD satisfying s < µ(x) + ε and t < µ(y) + ε. Then x = sd1 and y = td2 for some d1 , d2 ∈ D.
Since D is convex, we have
s t
x + y = sd1 + td2 = (s + t)( d1 + d2 ) ∈ (s + t)D.
s+t s+t
Thus, µ(x + y) ≤ s + t and so, µ(x + y) < µ(x) + µ(y) + 2ε. Therefore, µ is sub-additive. The last
assertion is clear by the definition of µ. □
 13

Proposition 4.6. Let C be a closed convex subset of a real vector space X. Let d and A be the
positive constants. If x0 is an element in X with ∥x0 ∥ ≤ A such that 0 < d < dist(x0 , C), then
there is an element F1 ∈ BX ∗ such that
(4.3) F1 (y) + α < F1 (x0 ).
d d −1
for all y ∈ C, whenever 0 < α < − 2A
2 (1 ) .
Consequently, for any closed convex subset C1 of X and x′0 ∈
/ C1 , then there is an element g ∈ X ∗
with ∥g∥ ≤ 1 such that
(4.4) g(z) + β < g(x′0 ).
for all z ∈ C1 , whenever 0 < β < dist(x′0 , C1 ).
Proof. For showing the first assertion, we first note that if there is x0 ∈ X such that ∥x0 ∥ ≤ A and
d
d(x0 , C) > d. then 2A < 1. To see this, we have A ≥ ∥x0 ∥ ≥ d(x0 , C) > d2 because 0 ∈ C.
Now since 0 < d < dist(x0 , C), we have (x0 + B(0, d)) ∩ C = ∅. Thus, we have (x0 + B(0, 12 d)) ∩
(C + B(0, 21 d)) = ∅. Put D := C + B(0, 21 d). Notice that D is a convex subset of X and x0 ∈ / D.
Moreover, we have 0 ∈ int(D). Let µ := µD be the Minkowski functional corresponding to D.
Then µ is positive homogeneous and sub-additive on X by Lemma 4.5.
Put Y := Rx0 and define f : Y → R by f (αx0 ) := αµ(x0 ) for α ∈ R. Then f (y) ≤ µ(y) for all
y ∈ Y since µ ≥ 0 and positive homogenous. The Hahn-Banch Theorem 4.3 implies that there is a
linear extension F defined on X satisfying F (x) ≤ µ(x) for all x ∈ X. We want to show that the
linear functional F1 := d2 F ∈ BX ∗ is as required.
We first notice that F is bounded because we have |F (y)| ≤ µ(y) ≤ 1 for all y ∈ B(0, 12 d) ⊆ D and
so, ∥F1 ∥ = ∥ d2 F ∥ ≤ 1. Note that µ(x) ≤ 1 for all x ∈ C because C ⊆ D. Thus, sup F (C) ≤ 1.
On the other hand, since x0 ∈ / D, we have F (x0 ) = µ(x0 ) ≥ 1. Now if µ(x0 ) = 1, then there
is a decreasing sequence of positive numbers (λn ) with λn ↓ 1 and λ1n x0 ∈ D. This implies
that x0 ∈ D. It contradicts to the fact that (x0 + B(0, 21 d)) ∩ D is empty. Hence, we have
F (y) ≤ 1 < F (x0 ) = µ(x0 ) for all y ∈ D.
Next, we are going to show that the Inequality 4.3 holds. In fact, for λ > 0, we see that x0 ∈ λD
if and only if λ1 x0 ∈ D. Hence, In this case, we have 1 < µ(x0 ) ≤ λ. Also, we have λ1 x0 = y + z for
some y ∈ C and for some z ∈ B(0, 12 d). Then we have
1 1 1 1
d ≤ ∥x0 − y∥ = ∥x0 − x0 − z∥ = |1 − |∥x0 ∥ + ∥z∥ < |1 − |A + d.
λ λ λ 2
1 d d −1
This implies that 1 − λ > 2A because λ ≥ µ(x0 ) > 1. This gives 1 < (1 − 2A ) < λ whenever
d −1 d −1
λ > 0 with x0 ∈ λD and hence, µ(x0 ) ≥ (1 − 2A ) . Now if we put 0 < α1 := (1 − 2A ) − 1, then
we have
d −1
F (y) + α1 ≤ 1 + α1 < (1 − ) ≤ µ(x0 ) = F (x0 )
2A
for all y ∈ C. Therefore, if 0 < α < d2 α1 , then the element F1 := d2 F ∈ BX ∗ satisfies the inequality
4.3 as desired.
td0 td0 −1
For showing the last assertion, let d0 := dist(x′0 , C1 ) > 0. Clearly, we have lim (1− ) = d0 .
t→1− 2 2d0
Thus if 0 < β < d0 , there is 0 < t1 < 1 such that β < t12d0 (1 − t2d 1 d0 −1
0
) . Now if we put d := t1 d0 ,
then we have 0 < β < 2 (1 − 2d0 ) . From this, we choose ε > 0 such that β < d2 (1 − 2(d0d+ε) )−1 .
d d −1

Now we fix a point x1 ∈ C1 so that ∥x′0 − x1 ∥ < d0 + ε. Put x0 := x′0 − x1 ; C := C1 − x1 and

A := d0 + ε into the first assertion. Then there is an element g ∈ BX ∗ such that the inequality 4.4
holds immediately.
We finish the proof. □

The following result is also referred to the Hahn-Banach Theorem.

14 CHI-WAI LEUNG

Theorem 4.7. Let X be a normed space and let Y be a subspace of X. If f ∈ Y ∗ , then there exists
a linear extension F ∈ X ∗ of f such that ∥F ∥ = ∥f ∥.
Proof. W.L.O.G, we may assume that ∥f ∥ = 1. We first show the case when X is normed space
over R. It is noted that the norm function p(·) := ∥ · ∥ is positively homogeneous and sub-additive
on X. Since ∥f ∥ = 1, we have f (y) ≤ p(y) for all y ∈ Y . Then by the Hahn-Banach Theorem 4.3,
there is a linear extension F of f on X such that F (x) ≤ p(x) for all x ∈ X. This implies that
∥F ∥ = 1 as required.
Now for the complex case, let h = Ref and g = Imf . Then f = h + ig and f, g both are real
linear on Y with ∥h∥ ≤ 1. Note that since f (iy) = if (y) for all y ∈ Y , we have g(y) = −h(iy)
for all y ∈ Y . This gives f (·) = h(·) − ih(i·) on Y . Then by the real case above, there is a real
linear extension H on X such that ∥H∥ = ∥h∥. Now define F : X −→ C by F (·) := H(·) − iH(i·).
Then F ∈ X ∗ and F |Y = f . Thus it remains to show that ∥F ∥ = ∥f ∥ = 1. We need to show that
|F (z)| ≤ ∥z∥ for all z ∈ X. For z ∈ X, consider the polar form F (z) = reiθ . Then F (e−iθ z) = r ∈ R
and thus F (e−iθ z) = H(e−iθ z). This yields that
|F (z)| = r = |F (e−iθ z)| = |H(e−iθ z)| ≤ ∥H∥∥e−iθ z∥ ≤ ∥z∥.
The proof is complete. □

Proposition 4.8. Let X be a normed space and x0 ∈ X. Then there is f ∈ X ∗ with ∥f ∥ = 1 such
that f (x0 ) = ∥x0 ∥. Consequently, we have
∥x0 ∥ = sup{|g(x)| : g ∈ BX ∗ }.
In addition, if x, y ∈ X with x ̸= y, then there exists f ∈ X ∗ such that f (x) ̸= f (y).
Proof. Let Y = Kx0 . Define f0 : Y → K by f0 (αx0 ) := α∥x0 ∥ for α ∈ K. Then f0 ∈ Y ∗ with
∥f0 ∥ = ∥x0 ∥. The result follows immediately from the Hahn-Banach Theorem. □

Remark 4.9. Proposition 4.8 tells us that the dual space X ∗ of X must be non-zero. Indeed, the
dual space X ∗ is very “Large′′ so that it can separate any pair of distinct points in X.
Furthermore, for any normed space Y and any pair of points x1 , x2 ∈ X with x1 ̸= x2 , we can find
an element T ∈ B(X, Y ) such that T x1 ̸= T x2 . In fact, fix a non-zero element y ∈ Y . Then by
Proposition 4.8, there is f ∈ X ∗ such that f (x1 ) ̸= f (x2 ). Thus, if we define T x = f (x)y, then
T ∈ B(X, Y ).

Proposition 4.10. Using the notations as above, if M is closed subspace and v ∈ X \ M , then
there is f ∈ X ∗ such that f (M ) ≡ 0 and f (v) ̸= 0.
Proof. Since M is a closed subspace of X, we can consider the quotient space X/M . Let π : X →
X/M be the natural projection. Note that v̄ := π(v) ̸= 0 ∈ X/M because v̄ ∈ X \ M . Then by
Corollary 4.8, there is a non-zero element f¯ ∈ (X/M )∗ such that f¯(v̄) ̸= 0. Therefore, the linear
functional f := f¯ ◦ π ∈ X ∗ is as desired. □

Proposition 4.11. Using the notations as above, if X ∗ is separable, then X is separable.

Proof. Let F := {f1 , f2 ....} be a dense subset of X ∗ . Then there is a sequence (xn ) in X with
∥xn ∥ = 1 and |fn (xn )| ≥ 1/2∥fn ∥ for all n. Now let M be the closed linear span of xn ’s. Then M
is a separable closed subspace of X. We are going to show that M = X. Suppose that M ̸= X and
hence Proposition 4.10 gives us a non-zero element f ∈ X ∗ such that f (M ) ≡ 0. Since {f1 , f2 ....}
is dense in X ∗ , we have B(f, r) ∩ F ̸= ∅ for all r > 0. Therefore, if B(f, r) ∩ F ̸= ∅ is finite for some
r > 0, then f = fm for some fm ∈ F . This implies that ∥f ∥ = ∥fm ∥ ≤ 2|fm (xm )| = 2|f (xm )| = 0
and thus, f = 0 which contradicts to f ̸= 0.
 15

Therefore, B(f, r) ∩ F is infinite for all r > 0. In this case, there is a subsequence (fnk ) such that
∥fnk − f ∥ → 0. This gives
1
∥fnk ∥ ≤ |fnk (xnk )| = |fnk (xnk ) − f (xnk )| ≤ ∥fnk − f ∥ → 0
2
because f (M ) ≡ 0. Thus∥fnk ∥ → 0 and hence f = 0. It leads to a contradiction again. Thus, we
can conclude that M = X as desired. □

Remark 4.12. The converse of Proposition 4.11 does not hold. For example, consider X = ℓ1 .
Then ℓ1 is separable but the dual space (ℓ1 )∗ = ℓ∞ is not.

Proposition 4.13. Let X and Y be normed spaces. For each element T ∈ B(X, Y ), define a linear
operator T ∗ : Y ∗ → X ∗ by
T ∗ y ∗ (x) := y ∗ (T x)
for y ∗ ∈ Y ∗ and x ∈ X. Then T ∗ ∈ B(Y ∗ , X ∗ ) and ∥T ∗ ∥ = ∥T ∥. In this case, T ∗ is called the
adjoint operator of T .
Proof. We first claim that ∥T ∗ ∥ ≤ ∥T ∥ and hence, ∥T ∗ ∥ is bounded.
In fact, for any y ∗ ∈ Y ∗ and x ∈ X, we have |T ∗ y ∗ (x)| = |y ∗ (T x)| ≤ ∥y ∗ ∥∥T ∥∥x∥. Hence,
∥T ∗ y ∗ ∥ ≤ ∥T ∥∥y ∗ ∥ for all y ∗ ∈ Y ∗ . Thus, ∥T ∗ ∥ ≤ ∥T ∥.
We need to show ∥T ∥ ≤ ∥T ∗ ∥. Let x ∈ BX . Then by Proposition 4.8, there is y ∗ ∈ SX ∗ such that
∥T x∥ = |y ∗ (T x)| = |T ∗ y ∗ (x)| ≤ ∥T ∗ y ∗ ∥ ≤ ∥T ∗ ∥. This implies that ∥T ∥ ≤ ∥T ∗ ∥. □

Example 4.14. Let X and Y be the finite dimensional normed spaces. Let (ei )ni=1 and (fj )m j=1 be
the bases for X and Y respectively. Let θX : X → X ∗ and θY : X → Y ∗ be the identifications as
in Example 3.15. Let e∗i := θX ei ∈ X ∗ and fj∗ := θY fj ∈ Y ∗ . Then e∗i (el ) = δil and fj∗ (fl ) = δjl ,
where, δil = 1 if i = l; otherwise is 0.
Now if T ∈ B(X, Y ) and (aij )m×n is the representative matrix of T corresponding to the bases
(ei )ni=1 and (fj )m ∗ ∗ ∗ ′
j=1 respectively, then akl = fk (T el ) = T fk (el ). Therefore, if (alk )n×m is the
representative matrix of T ∗ corresponding to the bases (fj∗ ) and (e∗i ), then akl = a′lk . Hence the
transpose (akl )t is the the representative matrix of T ∗ .

Proposition 4.15. Let Y be a closed subspace of a normed space X. Let i : Y → X be the natural
inclusion and π : X → X/Y the natural projection. Then
(i) the adjoint operator i∗∗ : Y ∗∗ → X ∗∗ is an isometry.
(ii) the adjoint operator π ∗ : (X/Y )∗ → X ∗ is an isometry.
Consequently, Y ∗∗ and (X/Y )∗ can be viewed as the closed subspaces of X ∗∗ and X ∗ respectively.
Proof. For Part (i), we first note that for any x∗ ∈ X ∗ , the image i∗ x∗ in Y ∗ is just the restriction
of x∗ on Y , denoted by x∗ |Y . Now let ϕ ∈ Y ∗∗ . Then for any x∗ ∈ X ∗ , we have
|i∗∗ ϕ(x∗ )| = |ϕ(i∗ x∗ )| = |ϕ(x∗ |Y )| ≤ ∥ϕ∥∥x∗ |Y ∥Y ∗ ≤ ∥ϕ∥∥x∗ ∥X ∗ .
Thus, ∥i∗∗ ϕ∥ ≤ ∥ϕ∥. WE need to show the inverse inequality. Now for each y ∗ ∈ Y ∗ , the Hahn-
Banach Theorem gives an element x∗ ∈ X ∗ such that ∥x∗ ∥X ∗ = ∥y ∗ ∥Y ∗ and x∗ |Y = y ∗ and hence,
i∗ x∗ = y ∗ . Then we have
|ϕ(y ∗ )| = |ϕ(x∗ |Y )| = |ϕ(i∗ x∗ )| = |(i∗∗ ◦ ϕ)(x∗ )| ≤ ∥i∗∗ ϕ∥∥x∗ ∥X ∗ = ∥i∗∗ ϕ∥∥y ∗ ∥Y ∗
for all y ∗ ∈ Y ∗ . Therefore, we have ∥i∗∗ ϕ∥ = ∥ϕ∥.
For Part (ii), let ψ ∈ (X/Y )∗ . Note that since ∥π ∗ ∥ = ∥π∥ ≤ 1, we have ∥π ∗ ψ∥ ≤ ∥ψ∥. On the
other hand, for each x̄ := π(x) ∈ X/Y with ∥x̄∥ < 1, we can choose an element m ∈ Y such that
∥x + m∥ < 1. Therefore, we have
|ψ(x̄)| = |ψ ◦ π(x)| = |ψ ◦ π(x + m)∥ ≤ ∥ψ ◦ π∥ = ∥π ∗ (ψ)∥.
16 CHI-WAI LEUNG

Therefore, we have ∥ψ∥ ≤ ∥π ∗ (ψ)∥. The proof is complete. □

Remark 4.16. By using Proposition 4.15, we can give an alternative proof of the Riesz’s Lemma
2.5.
Using the notations as in Proposition 4.15, if Y ⊊ X, then we have ∥π∥ = ∥π ∗ ∥ = 1 because π ∗ is an
isometry by Proposition 4.15(ii). Thus we have ∥π∥ = sup{∥π(x)∥ : x ∈ X, ∥x∥ = 1} = 1. Hence,
for any 0 < θ < 1, we can find element z ∈ X with ∥z∥ = 1 such that θ < ∥π(z)∥ = inf{∥z + y∥ :
y ∈ Y }. The Riesz’s Lemma follows.

5. Reflexive Spaces
Proposition 5.1. For a normed space X, let Q : X −→ X ∗∗ be the canonical map, that is,
Qx(x∗ ) := x∗ (x) for x∗ ∈ X ∗ and x ∈ X. Then Q is an isometry.
Proof. Note that for x ∈ X and x∗ ∈ BX ∗ , we have |Q(x)(x∗ )| = |x∗ (x)| ≤ ∥x∥. Then ∥Q(x)∥ ≤
∥x∥.
We need to show that ∥x∥ ≤ ∥Q(x)∥ for all x ∈ X. In fact, for x ∈ X, there is x∗ ∈ X ∗ with
∥x∗ ∥ = 1 such that ∥x∥ = |x∗ (x)| = |Q(x)(x∗ )| by Proposition 4.8. Thus we have ∥x∥ ≤ ∥Q(x)∥.
The proof is complete. □

Remark 5.2. Let T : X → Y be a bounded linear operator and T ∗∗ : X ∗∗ → Y ∗∗ the second dual
operator induced by the adjoint operator of T . Using notations as in Proposition 5.1 above, the
following diagram commutes.
T
X −−−−→ Y
 

QX y
Q
y Y
T ∗∗
X ∗∗ −−−−→ Y ∗∗

Definition 5.3. A normed space X is said to be reflexive if the canonical map Q : X −→ X ∗∗ is

surjective. (Note that every reflexive space must be a Banach space.)

Example 5.4. We have the following examples.

(i) : Every finite dimensional normed space X is reflexive.
(ii) : ℓp is reflexive for 1 < p < ∞.
(iii) : c0 and ℓ1 are not reflexive.
Proof. For Part (i), if dim X < ∞, then dim X = dim X ∗∗ . Hence, the canonical map Q : X → X ∗∗
must be surjective.
Part (ii) follows from (ℓp )∗ = ℓq for 1 < p < ∞, p1 + 1q = 1.
For Part (iii), note that c∗∗ 1 ∗ ∞ ∞
0 = (ℓ ) = ℓ . Since ℓ is non-separable but c0 is separable. Therefore,
∗∗ ∞
the canonical map Q from c0 to c0 = ℓ must not be surjective.
For the case of ℓ1 , we have (ℓ1 )∗∗ = (ℓ∞ )∗ . Since ℓ∞ is non-separable, the dual space (ℓ∞ )∗ is
non-separable by Proposition 4.11. Therefore, ℓ1 ̸= (ℓ1 )∗∗ . □

Proposition 5.5. Every closed subspace of a reflexive space is reflexive.

Proof. Let Y be a closed subspace of a reflexive space X. Let QY : Y → Y ∗∗ and QX : X → X ∗∗ be
the canonical maps as before. Let y0∗∗ ∈ Y ∗∗ . We define an element ϕ ∈ X ∗∗ by ϕ(x∗ ) := y0∗∗ (x∗ |Y )
for x∗ ∈ X ∗ . Since X is reflexive, there is x0 ∈ X such that QX x0 = ϕ. Suppose x0 ∈ / Y . Then
∗ ∗ ∗ ∗
by Proposition 4.10, there is x0 ∈ X such that x0 (x0 ) ̸= 0 but x0 (Y ) ≡ 0. Note that we have
x∗0 (x0 ) = QX x0 (x∗0 ) = ϕ(x∗0 ) = y0∗∗ (x∗0 |Y ) = 0. It leads to a contradiction, and so x0 ∈ Y . The
 17

proof is complete if we have QY (x0 ) = y0∗∗ .

In fact, for each y ∗ ∈ Y ∗ , then by the Hahn-Banach Theorem, y ∗ has a continuous extension x∗ in
X ∗ . Then we have
QY (x0 )(y ∗ ) = y ∗ (x0 ) = x∗ (x0 ) = QX (x0 )(x∗ ) = ϕ(x∗ ) = y0∗∗ (x∗ |Y ) = y0∗∗ (y ∗ ).
□

Example 5.6. By using Proposition 5.5, we immediately see that the space ℓ∞ is not reflexive
because it contains a non-reflexive closed subspace c0 .

Proposition 5.7. Let X be a Banach space. Then we have the following assertions.
(i) X is reflexive if and only if the dual space X ∗ is reflexive.
(ii) If X is reflexive, then so is every quotient of X.
Proof. For Part (i), suppose that X is reflexive first. Let ze ∈ X ∗∗∗ . Then the restriction z := ze|X ∈
X ∗ . Then one can directly check that Qz = z on X ∗∗ since X ∗∗ = X.
For the converse, assume that X ∗ is reflexive but X is not. Therefore, X is a proper closed subspace
of X ∗∗ . Then by using the Hahn-Banach Theorem, we can find a non-zero element ϕ ∈ X ∗∗∗ such
that ϕ(X) ≡ 0. However, since X ∗∗∗ is reflexive, we have ϕ ∈ X ∗ and hence, ϕ = 0 which leads to
a contradiction.
For Part (ii), we assume that X is reflexive. Let M be a closed subspace of X and π : X → X/M
the natural projection. Note that the adjoint operator π ∗ : (X/M )∗ → X ∗ is an isometry (Check
!). Thus, (X/M )∗ can be viewed as a closed subspace of X ∗ . By Part (i) and Proposition 5.5, we
see that (X/M )∗ is reflexive. Then X/M is reflexive by using Part (i) again.
The proof is complete. □

Lemma 5.8. Let M be a closed subspace of a normed space X. Let r : X ∗ → M ∗ be the restriction
map, that is x∗ ∈ X ∗ 7→ x∗ |M ∈ M ∗ . Put M ⊥ := ker r := {x∗ ∈ X ∗ : x∗ (M ) ≡ 0}. Then the
canonical linear isomorphism re : X ∗ /M ⊥ → M ∗ induced by r is an isometric isomorphism.
Proof. We first note that r is surjective by using the Hahn-Banach Theorem. We need to show
that re is an isometry. Note that re(x∗ + M ⊥ ) = x∗ |M for all x∗ ∈ X ∗ . Now for any x∗ ∈ X ∗ , we
have ∥x∗ + y ∗ ∥X ∗ ≥ ∥x∗ + y ∗ ∥M ∗ = ∥x∗ |M ∥M ∗ for all y ∗ ∈ M ⊥ . Thus, we have ∥e
r(x∗ + M ⊥ )∥ =
∗ ∗ ⊥
∥x |M ∥M ∗ ≤ ∥x + M ∥. We need to show the reverse inequality.
Now for any x∗ ∈ X ∗ , then by the Hahn-Banach Theorem again, there is z ∗ ∈ X ∗ such that
z ∗ |M = x∗ |M and ∥z ∗ ∥ = ∥x∗ |M ∥M ∗ . Then x∗ − z ∗ ∈ M ⊥ and hence, we have x∗ + M ⊥ = z ∗ + M ⊥ .
This implies that
∥x∗ + M ⊥ ∥ = ∥z ∗ + M ⊥ ∥ ≤ ∥z ∗ ∥ = ∥x∗ |M ∥M ∗ = ∥e
r(x∗ + M ⊥ )∥.
The proof is complete. □

Proposition 5.9. (Three-space property): Let M be a closed subspace of a normed space X.

If M and the quotient space X/M both are reflexive, then so is X.
Proof. Let π : X → X/M be the natural projection. Let ψ ∈ X ∗∗ . We going to show that
ψ ∈ im(QX ). Since π ∗∗ (ψ) ∈ (X/M )∗∗ , there exists x0 ∈ X such that π ∗∗ (ϕ) = QX/M (x0 + M )
because X/M is reflexive. Thus we have
π ∗∗ (ψ)(x̄∗ ) = QX/M (x0 + M )(x̄∗ )
for all x̄∗ ∈ (X/M )∗ . This implies that
ψ(x̄∗ ◦ π) = ψ(π ∗ x̄∗ ) = π ∗∗ (ψ)(x̄∗ ) = QX/M (x0 + M )(x̄∗ ) = x̄∗ (x0 + M ) = QX x0 (x̄∗ ◦ π)
18 CHI-WAI LEUNG

for all x̄∗ ∈ (X/M )∗ . Therefore, we have

ψ = QX x0 on M ⊥.
Therefore, we have ψ −QX (x0 ) ∈ (X ∗ /M ⊥ )∗ . Let f : M ∗ → X ∗ /M ⊥ be the inverse of the isometric
isomorphism re which is defined as in Lemma 5.8. Then the composite (ψ − QX x0 ) ◦ f : M ∗ →
X ∗ /M ⊥ → K lies in M ∗∗ . Then by the reflexivity of M , there is an element m0 ∈ M such that
(ψ − QX x0 ) ◦ f = QM (m0 ) ∈ M ∗∗ .
Notice that for each x∗ ∈ X ∗ , we can find an element m∗ ∈ M ∗ such that f (m∗ ) = x∗ + M ⊥ ∈
X ∗ /M ⊥ because f is surjective. Moreover, by the construction of re in Lemma 5.8, we see that
x∗ |M = m∗ . This gives
ψ(x∗ ) − x∗ (x0 ) = (ψ − QX x0 )(m∗ ) ◦ f = QM (m0 )(m∗ ) = m∗ (m0 ) = x∗ (m0 ).
Thus, we have ψ(x∗ ) = x∗ (x0 +m0 ) for all x∗ ∈ X ∗ . From this we have ψ = QX (x0 +m0 ) ∈ im(QX )
as desired. The proof is complete. □

Remark 5.10. In view of the definition of a reflexive space, it is naturally raised the question
that whether a Banach space X is reflexive whenever it is isometrically isomorphic to its second
dual. The answer is negative. A counter example was given by R.C. James in 1951 (see [8]).

6. Weakly convergent and Weak∗ convergent

Definition 6.1. Let X be a normed space. A sequence (xn ) is said to be weakly convergent if there
is x ∈ X such that f (xn ) → f (x) for all f ∈ X ∗ . In this case, x is called a weak limit of (xn ).

Proposition 6.2. A weak limit of a sequence is unique if it exists. In this case, if (xn ) weakly
w
converges to x, denoted by x = w-lim xn or xn −
→ x.
n

Proof. The uniqueness follows immediately from the Hahn-Banach Theorem. □

w
Remark 6.3. Clearly, if a sequence (xn ) converges to x ∈ X in norm, then xn −
→ x. However,
the weakly convergence of a sequence does not imply the norm convergence.
For example, consider X = c0 and (en ). Then f (en ) → 0 for all f ∈ c∗0 = ℓ1 but (en ) is not
convergent in c0 .

Proposition 6.4. Suppose that X is finite dimensional. A sequence (xn ) in X is norm convergent
if and only if it is weakly convergent.
Proof. Suppose that (xn ) weakly converges to x. Let B := {e1 , .., eN } be a basis for X and let fk be
the k-th coordinate functional corresponding to the basis B, i.e., v = N
P
k=1 fk (v)ek for all v ∈ X.
Since dim X < ∞, we have fk in X ∗ for all k = 1, ..., N . Therefore, we have limn fk (xn ) = fk (x)
for all k = 1, ..., N . Thus, we have ∥xn − x∥ → 0. □
Definition 6.5. Let X be a normed space. A sequence (fn ) in X ∗ is said to be weak∗ convergent
if there is f ∈ X ∗ such that limn fn (x) = f (x) for all x ∈ X, that is fn point-wise converges to f .
w∗
In this case, f is called the weak∗ limit of (fn ). Write f = w∗ -limn fn or fn −−→ f .

Remark 6.6. In the dual space X ∗ of a normed space X, we always have the following implications:
“Norm Convergent” =⇒ “Weakly Convergent” =⇒ “Weak∗ Convergent”.
However, the converse of each implication does not hold.
 19

Example 6.7. Remark 6.3 has shown that the w-convergence does not imply ∥ · ∥-convergence.
We now claim that the w∗ -convergence also Does Not imply the w-convergence.
Consider X = c0 . Then c∗0 = ℓ1 and c∗∗ 1 ∗ ∞ ∗ 1 ∗
0 = (ℓ ) = ℓ . Let en = (0, ...0, 1, 0...) ∈ ℓ = c0 , where
w∗
the n-th coordinate is 1. Then e∗n −−→ 0 but e∗n ↛ 0 weakly because e∗∗ (e∗n ) ≡ 1 for all n, where
e∗∗ := (1, 1, ...) ∈ ℓ∞ = c∗∗ ∗
0 . Hence the w -convergence does not imply the w-convergence.

w
Proposition 6.8. Let (fn ) be a sequence in X ∗ . Suppose that X is reflexive. Then fn −
→ f if and
w∗
only if fn −−→ f .
In particular, if dim X < ∞, then the followings are equivalent:
∥·∥
(i) : fn −−→ f ;
w
(ii) : fn −→ f;
w∗
(iii) : fn −−→ f .

Theorem 6.9. (Banach) : Let X be a separable normed space. If (fn ) is a bounded sequence in
X ∗ , then it has a w∗ -convergent subsequence.
Proof. Let D := {x1 , x2 , ...} be a countable dense subset of X. Note that since (fn )∞ n=1 is bounded,
(fn (x1 )) is a bounded sequence in K. Then (fn (x1 )) has a convergent subsequence, say (f1,k (x1 ))∞k=1
in K. Let c1 := limk f1,k (x1 ). Now consider the bounded sequence (f1,k (x2 )). Then there is
convergent subsequence, say (f2,k (x2 )), of (f1,k (x2 )). Put c2 := limk f2,k (x2 ). Note that we still
have c1 = limk f2,k (x1 ). To repeat the same step, if we define (m, k) ≤ (m′ , k ′ ) if m < m′ ; or
m = m′ with k ≤ k ′ , we can find a sequence (fm,k )m,k in X ∗ such that
(i) : (fm+1,k )∞ ∞
k=1 is a subsequence of (fm,k )k=1 for m = 0, 1, .., where f0,k := fk .
(ii) : ci = limk fm,k (xi ) exists for all 1 ≤ i ≤ m.
Now put hk := fk,k . Then (hk ) is a subsequence of (fn ). Note that for each i, we have limk hk (xi ) =
limk fi,k (xi ) = ci by the construction (ii) above. Since (∥hk ∥) is bounded and D is dense in X, we
have h(x) := limk hk (x) exists for all x ∈ X and h ∈ X ∗ . That is h = w∗ -limk hk . The proof is
complete. □
Remark 6.10. Theorem 6.9 does not hold if the separability of X is removed.
For example, consider X = ℓ∞ and δn the n-th coordinate functional on ℓ∞ . Then δn ∈ (ℓ∞ )∗
with ∥δn ∥(ℓ∞ )∗ = 1 for all n. Suppose that (δn ) has a w∗ -convergent subsequence (δnk )∞
k=1 . Define
x ∈ ℓ∞ by 
0
 if m ̸= nk ;
x(m) = 1 if m = n2k ;

−1 if m = n2k+1 .


Hence we have |δni (x) − δni+1 (x)| = 2 for all i = 1, 2, ... It leads to a contradiction. Thus(δn ) has
no w∗ -convergent subsequence.

Corollary 6.11. Let X be a separable space. Assume that a sequence in X ∗ is w∗ -convergent if

and only if it is norm convergent. Then dim X < ∞.
Proof. We need to show that the closed unit ball BX ∗ in X ∗ is compact in norm. Let (fn ) be a
sequence in BX ∗ . By using Theorem 6.9, (fn ) has a w∗ -convergent subsequence (fnk ). Then by
the assumption, (fnk ) is norm convergent. Note that if lim fnk = f in norm, then f ∈ BX ∗ . Thus
k
BX ∗ is compact and thus dim X ∗ < ∞. Thus dim X ∗∗ < ∞ that gives dim X is finite because
X ⊆ X ∗∗ . □
20 CHI-WAI LEUNG

Corollary 6.12. Suppose that X is a separable. If X is reflexive space, then the closed unit ball
BX of X is sequentially weakly compact, i.e. it is equivalent to saying that any bounded sequence
in X has a weakly convergent subsequence.
Proof. Let Q : X → X ∗∗ be the canonical map as before. Let (xn ) be a bounded sequence in X.
Hence, (Qxn ) is a bounded sequence in X ∗∗ . We first note that since X is reflexive and separable,
X ∗ is also separable by Proposition 4.11. We can apply Theorem 6.9, (Qxn ) has a w∗ -convergent
subsequence (QxnK ) in X ∗∗ = Q(X) and hence, (xnk ) is weakly convergent in X. □

Remark 6.13. In fact, the converse of Corollary 6.12 also holds (see Appendix 7 below). The
assumption of separability of X can be removed. We have the following stronger result which was
shown by R. C. James (see [10, §1.13]).

Theorem 6.14. Let X be a Banach space. Then the following are equivalent.
(i) X is reflexive.
(ii) Every bounded sequence in X has a weakly convergent subsequence.
(iii) The closed unit ball BX of X is weakly compact, that is, BX is compact in the weak topology.

7. Appendix: Sequentially weakly compactness and reflexivity for a separable

space
This section is devoted to show the theorem that a separable Banach space is reflexive if and
only if its closed unit ball is weakly sequentially compact. This results was obtained by Banach [2,
Chapter VIII and Chapter XI]. For simply, throughout this section, all Banach spaces are assumed
to be over R.

Some set theory

Before showing the main theorem, we need some basic knowledge of the set theory which can be
found in the Halmos’s classic book [7].
Recall that a partially order set S is called a well ordered set if for any non-empty subset A of S
contains the least element, that is, there is an element x0 ∈ A such that x0 ≤ x for all x ∈ A. In
particular, S is automatically a totally order set. The well-ordering theorem tells us that every
set can be equipped with a well-ordering.
Two well ordered sets A and B are said to have the same ordinal number or ordinal for simply
if there is an order preserving bijection from A onto B. In particular, each ordinal can be viewed
as a well ordered set. More precisely, an ordinal number α is a well ordered set such that for any
element η ∈ α, we have η = {ξ ∈ α : ξ < η}, thus, we have η ⊆ α whenever η ∈ α. This definition
was due to von Neumann.

Put ω the least infinite countable ordinal, that is, ω := {0, 1, 2...} and is endowed with the usual
order.

On the other hand, it is naturally led to define an order on the class of ordinals as the following.
(Warning: We DO NOT HAVE a statement about ” the set of All ordinals numbers” !!!! )
Definition 7.1. Let α and β be two ordinals. We say that
(i) α = β if there is an order preserving bijection from α onto β.
(ii) α ≤ β if there is an order preserving injection from α to β.
(iii) α < β if α ≤ β but α ̸= β.
From the von Neumann’s definition, we have (see [7, Section 20])
 21

Lemma 7.2. Every nonempty set of ordinal numbers is a well ordered set.

Next we need the following definition for comparing the size of two given sets.
Definition 7.3. Two sets A and B are said to have the same cardinality, write A ≈ B, if there is
a bijection from A onto B.
The cardinal number of A, write |A|, is defined by an ordinal number given by
|A| := min{α : α is an ordinal number such that α ≈ A}
(Notice that the well ordering theorem and Lemma 7.2 assure the existence of |A|. )

Definition 7.4. We will use the following terminologies later.

(i) For an ordinal θ, put θ+ := θ ∪ {θ} and define ξ ≤ θ for all elements ξ ∈ θ. Then θ+
becomes a well ordered set and θ < θ+ . In this case, θ+ is called the successor ordinal of θ.
Notice that there is no ordinal γ such that θ < γ < θ+ .
(ii) An ordinal β is called a limit ordinal if there is no ordinal γ such that γ + = β. In this case,
we are given any ordinal η with η < β, there is an ordinal ξ such that η < ξ < β, write
ξ → β. For example, ω1 :=the cardinal of R and ω := N both are limit ordinals.
(iii) Let X be a non-empty set. A function x : [0, θ) → X is called a transfinite sequence in X
(or ”θ-transfinite sequence”) for some ordinal θ, where [0, θ) denotes the set of all ordinals
ξ satisfying ξ < θ. We also write (xξ )ξ<θ for a transfinite sequence. Note that ω-transfinite
sequences are the usual sequences defined as before.
In this case, if θ is a limit ordinal, one can naturally define the similar notation lim xξ ;
ξ→θ
lim xξ and lim xξ in any metric space as in the usual sequences case.
ξ→θ ξ→θ
Clearly, one can show that if (tξ )ξ<θ is a bounded θ-sequence in R, then
lim tξ := lim sup tξ = inf sup tξ
ξ→θ η→θ ξ≥η η<θ ξ≥η

exists.
(iv) Let θ be a limit ordinal. A subset M of θ is said to be cofinal if for every ordinal µ < θ,
there is ν ∈ M such that µ < ν < θ. In this case, if lim xξ exists, then so does lim xν
ξ→θ ν∈M :ν→θ
and they are the same.

Lemma 7.5. Every infinite cardinal is a limit ordinal.

Proof. If γ is an infinite cardinal and γ = θ+ for some ordinal θ. Then by the definition of the
successor ordinal, we have θ ≈ θ+ ≈ γ. Hence, γ = θ+ is not a cardinal because θ < θ+ . □

Lemma 7.6. Let X be a Banach space. If (fξ )ξ<θ is a norm bounded θ-sequence in X ∗ , then there
is an element f ∈ X ∗ such that ∥f ∥ ≤ supξ<θ ∥fξ ∥ and

(7.1) f (x) ≤ lim fξ (x)

ξ→θ

for all x ∈ X. Consequently, we have

(7.2) lim fξ (x) ≤ f (x) ≤ lim fξ (x)
ξ→θ ξ→θ

for all x ∈ X.
In this case, f is called a transfinite limit of (fξ ) ( note that f may not be unique).
22 CHI-WAI LEUNG

Proof. Let M := supξ<θ ∥fξ ∥. We first notice that since (fξ )ξ<θ is bounded, lim fξ (x) exists for all
ξ→θ
x ∈ X. Hence, one can define a function p : X → R by
p(x) := lim fξ (x)
ξ→θ

for x ∈ X. Clearly, p is a positively homogenous and sub-additive function. We may assume that
p(x0 ) > 0 for some x0 ∈ X. To see this, if p(x0 ) < 0, then p(−x0 ) = limfξ (−x0 ) ≥ limfξ (−x0 ) =
−limfξ (x0 ) > 0 as desired. Now if we define a linear map f0 on Rx0 by f0 (tx0 ) := tp(x0 ), then
f0 (tx0 ) ≤ p(tx0 ) for all t ∈ R. Then by the Hahn-Banach Theorem 4.3, there is a linear extension
f of f0 defined on X such that f (x) ≤ p(x) for all x ∈ X. Notice that since |fξ (x)| ≤ M ∥x∥ for
all ξ < θ and for all x ∈ X, we have p(x) ≤ M ∥x∥. Thus, we have ∥f ∥ ≤ M as desired. The last
assertion is obtained by putting −x ∈ X into Eq 7.1. The proof is complete. □

Definition 7.7. A normed subspace Γ of X ∗ is said to be transfinitely closed if for every norm
bounded transfinite θ-sequence (fξ )ξ<θ in Γ for some limit ordinal θ, one can find an element f ∈ Γ
satisfying the Eq 7.1 above, that is,
f (x) ≤ lim fξ (x)
ξ→θ
for all x ∈ X.
Clearly, every transfinitely closed subspace is norm closed by considering θ = ω in Eq 7.2 above.

Lemma 7.8. Let Γ be a transfinitely closed subspace of X ∗ and f0 be an element in X ∗ \ Γ. Let

0 < c < dist(f0 , Γ). Then there is a non-empty finite subset G of X so that there is no element
f ∈ Γ satisfying the following:

(7.3) |f (x) − f0 (x)| ≤ c∥x∥ for all x ∈ G.

Proof. Let
W := {γ : γ is a cardinal such that γ ≤ |X|}.
Now for each element γ ∈ W , put P (γ) the sentence given by:
whenever G is a subset of X with |G| = γ, there exists an element fγ ∈ Γ such that |fγ (x)−f0 (x)| ≤
c∥x∥ for all x ∈ G. Let
A := {γ ∈ W : P (γ) holds}.
Clearly, the sentence P holds for the zero cardinal, that is, 0 ∈ A. One the other hand, notice that
the cardinal |X| ∈ W \ A. To see this, we can enumerate the elements in X such that (xξ )ξ<|X|
because [0, |X|) = |X|. Thus, if P (|X|) holds, there is an element f ∈ Γ so that |f (x) = f0 (x)| ≤
c∥x∥ for all x ∈ X. Hence, ∥f − f0 ∥ ≤ c which contradicts to the choice of c because c < dist(f0 , Γ).
Therefore, W \ A is a non-empty well ordered set by Lemma 7.2 and hence the set W \ A contains
the least element, say m.
We will show that if m is infinite, then it will lead to a contradiction.
Claim: P (m) holds if m is infinite.
Now let G be any subset of X with |G| = m, so, we can enumerate the elements in G as a m-
sequence, say (xξ )ξ<m . Now for any ordinal η < m, notice that if we put a := |(xξ )ξ<η |, then
a ≤ |η| ≤ η < m. The minimality of m implies that P (a) holds. Hence, there is an element fη ∈ Γ
so that
|fη (xξ ) − f0 (xξ )| ≤ c∥xξ ∥ for all ξ < η.
Now if m is infinite, then m is a limit ordinal by Lemma 7.5. Then by the definition of a transfinitely
closed set, there is an element f ∈ Γ such that
lim fη (x) ≤ f (x) ≤ lim fη (x)
η→m η→m
 23

for all x ∈ X. This gives

|f (xξ ) − f0 (xξ )| ≤ c∥xξ ∥ for all ξ < m.
Recall that G = (xξ )ξ<m . Hence, P (m) holds if m is infinite. It leads to a contradiction since m ∈
/A
by the definition of the set A.
Thus, we can conclude that m is finite.
Since m ∈/ A, then by the definition of the sentence P we can find a finite subset G of X with
|G| = m as desired for the Lemma.
You see!! It is a very clever proof, isn’t it!!!. □

Proposition 7.9. Let Γ be a transfinitely closed subspace of X ∗ and f0 be an element in X ∗ \ Γ.

Then there is an element x0 ∈ X such that f0 (x0 ) = 1 and f (x0 ) = 0 for all f ∈ Γ.
Proof. Let 0 < c < dist(f0 , Γ). Recall Lemma 7.8 that there is a non-empty finite subset G of X
such that there is no element f ∈ Γ satisfying the following condition:
(7.4) |f (tx) − f0 (tx)| ≤ c∥tx∥ for all x ∈ G and for all t > 0.
Notice that we may assume that ∥x∥ ≤ 1 for all x ∈ G because G is finite. In particular, by
considering t = 1/k, k = 1, 2.. into the Eq 7.4 above, we can find a sequence (xj ) in X with
limj xj = 0 such that there is no element f ∈ Γ satisfying the condition:
(7.5) |f (xj ) − f0 (xj )| ≤ c for all j = 1, 2.....
Now let fe0 := (f0 (xj ) and fe := (f (xj )) for f ∈ Γ. Then fe0 ∈ c0 and F := {fe : f ∈ Γ} is a subspace
of c0 . Note that the Eq 7.5 implies ∥fe0 − fe∥∞ > c for all f ∈ Γ. This gives dist(fe0 , F ) ≥ c > 0.
Then by the Hahn-Banach separation Theorem, there is an element φ = (tj ) ∈ ℓ1 = c∗0 such that
φ(fe0 ) = 1 and φ(fe) = 0 for all f ∈ Γ. In particular, we have
∞
X ∞
X
tj f0 (xj ) = 1 and tj f (xj ) = 0 for all f ∈ Γ .
j=1 j=1
P∞
Therefore, if we put x0 := j=1 tj xj ∈ X, then the element x0 is as desired. □

Recall the notation that for every element x in a Banach space X, x b denotes the element in X ∗∗
given by x(f ) := f (x) for all f ∈ X ∗ . Put A a : a ∈ A} ⊆ X ∗∗ for a subset A of X.
b := {b

Lemma 7.10. Let X be a Banach space. Assume that every normed bounded sequence in X has
a weakly convergent subsequence. Let D be a countably infinite subset of X. Then every bounded
transfinite sequence in D b has a transfinite limit in X x : x ∈ X} ⊆ X ∗∗ , that is, for every
b := {b
transfinite sequence (xξ )ξ<θ in D, there is an element z ∈ X such that
(7.6) f (z) ≤ lim f (xξ )
ξ→θ

for all f ∈ X ∗ .
Proof. Let θ be a limit ordinal and (xξ )ξ<θ be a bounded transfinite sequence in D. By the
assumption of D, we can write D = {xi : i = 0, 1, 2...}.
Case 1: there is an infinite sequence (ξn )∞
n=0 in [0, θ) such that for every ordinal η < θ, there is
N ∈ N such that η < ξn < θ for all n > N , write lim ξn = θ. In this case, put xξi := xi for
n→∞
i = 0, 1, 2.... Then by assumption, there is a weakly convergent sequence (xξik ) with the weak limit,
say z ∈ X. This implies that
zb(f ) = f (z) = lim f (xξik ) ≤ lim f (xξ ) = lim xbξ (f )
k→∞ ξ→θ ξ→θ
24 CHI-WAI LEUNG

for all f ∈ X ∗ .
The proof is complete if Eq 7.6 also holds for the following case.
Case 2: there is no sequence (ξn ) in [0, θ) such that lim ξn = θ.
n→∞
In this case, for each xi ∈ D , let Zi := {ξ < θ : xξ = xi }. We will see that Zi is confinal subset
of [0, ∞) for some i ∈ ω, that is, for any ordinal η < θ, there is ν ∈ Zi such that η < ν < θ.
To see this, if we assume that every Zi , i ∈ ω, is not a cofinal subset of [0, θ), then for each iω,
there is an ordinal µi with µi < θ such that ξ ≤ µi for all ξ ∈ Zi . Now put λ0 := µ0 and
λn := max{µ : i = 0, 1, .., n − 1}. This gives an increasing sequence (λn )n∈ω in [0, θ). Then by the
assumption of thisS case, there is λ ∈ [0, θ) such that λn ≤ λ for all n ∈ ω and hence, ξ ≤ λ for all
ξ ∈ [0, θ) because i∈ω Zi = [0, θ). This implies that there is no ordinal ξ such thhat λ < ξ < θ
that will lead to a contradiction because θ is a limit ordinal.
Now let Zi0 be a confinal subset of [0, θ) and z := xi0 . Then by the definition of Zi0 , xξ = z for all
ξ ∈ Zi0 . Thus, we have
zb(f ) = f (z) = lim f (xξ ) ≤ lim f (xξ ) = lim xbξ (f ).
ξ∈Zi0 ;ξ→θ ξ→θ ξ→θ

The proof is finished. □

We are now in a position to reach the following main result in this section.

Theorem 7.11. Let X be a separable Banach space. Then X is reflexive if and only if every
bounded sequence in X has a weakly convergent subsequence.
Proof. The necessary condition has been shown in Corollary 6.12.
We are going to show the converse statement. Let D be a countable subset of X.
Claim: The space X b := {b x : x ∈ X} is transfinitely closed in X ∗∗ , that is, for every bounded
b there is an element z ∈ X such that
transfinite sequence (xbξ )ξ<θ in X,
f (z) ≤ lim f (xξ )
ξ→θ

for all f ∈ X ∗.
Now for each n = 1, 2... and each xξ , ξ < θ, we choose an element xnξ ∈ D such that
1
∥xξ − xnξ ∥ <
.
n
From this, for each n = 1, 2, ..., we obtain a bounded a θ-transfinite sequence (xnξ )ξ<θ in D. For
each n = 1, 2.., Lemma 7.10 gives an element zn ∈ X such that
f (zn ) ≤ lim f (xnξ )
ξ→θ

for all f ∈ X ∗. In addition from this we have ∥zn ∥ ≤ 1 + supξ<θ ∥xξ ∥ for all n = 1, 2, .... Then the
necessary condition implies that (zn ) has a weak convergent subsequence (znj ). Let z be the weak
limit of (znj ). Then we have
f (z) = lim f (znj )
j→∞
n
≤ lim lim f (xξ j )
j→∞ ξ→θ
∥f ∥
≤ lim ( lim f (xξ ) + )
j→∞ ξ→θ nj
= lim f (xξ )
ξ→θ
for all f ∈ X ∗.
The Claim follows.
b ⊊ X ∗∗ , then there is an element ϕ ∈ X ∗∗ \ X.
Finally, if X b Note that X
b is a closed subspace of
X . Using Lemma 7.9, the transfinite closeness of X implies that there is an element f0 ∈ X ∗
∗∗ b
 25

b(f0 ) = 0 for all x ∈ X and so, f0 = 0 but ϕ(f0 ) = 1 that is ridiculous. Thus,
such that f0 (x) = x
Xb = X ∗∗ . The proof is complete. □

8. Appendix: w∗ -compactness
Throughout this section X always denotes a normed space. I suppose that the students have
learned a standard course of topology before.
Now for each ε > 0 and for finitely many elements x1 , ..., xm in X, let
W (x1 , .., xm ; ε) := {f ∈ X ∗ : |f (xi )| < ε; ∀i = 1, .., m}.
It is noted that 0 ∈ W (x1 , .., xm ; ε) for any ε > 0 and for all finitely many elements x1 , ..., xm in X.
Definition 8.1. The weak ∗ -topology on the dual space X ∗ is the topology generated by the collection
{h + W (x1 , .., xm ; ε) : h ∈ X ∗ ; for ε > 0 and for finitely many x1 , .., xm ∈ X}.

The following result is an important feature for a weak∗ -continuous linear functional on X ∗ .

Proposition 8.2. Let X be a normed space and let φ ∈ X ∗∗ . Then φ : X ∗ → K is w∗ -continuous

c0 (f ) := f (x0 ) for all f ∈ X ∗ .
if and only if there is a unique element x0 ∈ X such that φ(f ) = x
Proof. The converse is clear. We are going to show the necessary condition.
Assume that φ is w∗ -continuous, in particular, φ is w∗ continuous at 0. By the definition of w∗ -
open neighbourhood of 0 given above (Definition 8.1), there are finitely elements x1 , ..., xN in X
and δ > 0 such that |φ(f )| < 1 whenever f ∈ X ∗ satisfies |f (xk )| < δ for all k = 1, ..., N . Now if
we put L(f ) := max |f (xk )| for f ∈ X ∗ , then clearly we have L(tf ) = tL(f ) for all f ∈ X ∗ and
k=1,...,N
δ ∗
t > 0. From this, we see that L( 2L(f ) f ) < δ while f ∈ X satisfies L(f ) > 0. Therefore, we have

2
(8.1) |φ(f )| ≤ L(f )
δ
whenever f ∈ X ∗ with L(f ) > 0.
Claim 1: Eq 8.1 holds for all f ∈ X ∗ . To see this, it suffices to show that if L(f ) = 0, then
φ(f ) = 0. In fact, if L(f ) = 0, then L(tf ) = tL(f ) = 0 for all t > 0. In particular, we have
|(tf )(xk )| = 0 < δ for all t > 0 and for all k = 1, .., N . This implies that t|φ(f )| = |φ(tf )| < 1 for
all t > 0. Thus, we have φ(f ) = 0 as desired.
So, from Eq 8.1, we have
N
\
(8.2) ck ⊆ ker φ.
ker x
k=1

Claim 2: φ ∈ span{c xk : k = 1, ..., N } in X ∗∗ . To see this, we first notice that if xbj ∈ span{c xk :
T TN T
k ̸= j} for some j, then k̸=j ker x ck ⊆ ker xbj and so, we have k=1 ker x ck = k̸=j ker x
ck . Therefore,
we may assume that {c x1 , ..., xcN } is a linearly independent set by considering a maximal linearly
independent subset of {c x1 , ..., xcN }. In this case, {x1 , ..., xN } is a linearly independent subset of X.
Then the Hahn-Banach Theorem implies that there are x∗1 , ..., x∗N in X ∗ such that x∗i (xj ) = 1 if
i = j; otherwise is 0 for i, j = 1, ..., N . From this for any f ∈ X ∗ , we have xbj (f − N ∗
P
PN k=1 f (xk )xk ) = 0
∗ ∗
for all j = 1, ..., N . Eq 8.2 implies that φ(f − k=1 f (xk )xk ) = 0 for all f ∈ X . This gives
N
X N
X
φ(f ) = f (xk )φ(x∗k ) = φ(x∗k )c
xk (f )
k=1 k=1
26 CHI-WAI LEUNG

for all f ∈ X ∗ . Hence, Claim 2 follows from

N
X
φ= φ(x∗k )c
xk .
k=1
PN ∗
Now if we put x0 := ∈ X, then φ = x
k=1 φ(xk )xk c0 . On the other hand, the uniqueness is clear
due to the Hahn-Banach Theorem. We finish the proof. □

The following is clearly shown by the definition.

Lemma 8.3. Using the notations as above, we have
(i) The weak ∗ -topology is Hausdorff.
(ii) Let f ∈ X ∗ . Then for each open neighborhood V of f , there are ε > 0 and x1 , .., xm in X
such that f + W (x1 , .., xm ; ε) ⊆ V , that is, the collection {f + W (x1 , .., xm ; ε)} forms an
open basis at f .
(iii) A sequence (fn ) waek ∗ converges to f in X ∗ if and only if for each ε > 0 and for
finitely many elements x1 , ..., xm in X, there is a positive integer N such that fn − f ∈
W (x1 , .., xm ; ε) for all n ≥ N .

Before showing the main result in this section, let us recall that product topologies.
Let (Zi )i∈I be a collection of topological spaces. Let Z be the usual Cartesian product, that is
Y [
Z := Zi : {z : I → Zi : z(i) ∈ Zi ; ∀i ∈ I}.
i∈I i∈I
Let pi : Z → Zi be the natural projection for i ∈ I. The product topology on Z is the weakest
topology such that each projection pi is continuous. More precisely, the following collection forms
an open basis for the product topology:
\
{ p−1 i (Wi ) : J is a finite subset of I and Wi is an open subset of Zi }.
i∈J
We have the following famous result in topology.

Theorem 8.4. Tychonoff ’s Theorem: The Cartesian product of compact spaces is compact
under the product topology.

The following result is known as the Alaoglu’s Theorem.

Theorem 8.5. The closed unit ball BX ∗ of the dual space X is compact with respect to the weak ∗ -
topology.
Proof. For each x ∈ X, put Zx := [−∥x∥, ∥x∥] ⊆ R. Each Zx is endowed with the usual subspace
topology of R. Then Zx is a compact set for all x ∈ X. Let
Y
Z := Zx .
x∈X
Then the set Z is a compact Hausdorff space under the product topology. Define a mapping by
T : f ∈ BX ∗ 7→ T f ∈ Z; T f (x) := f (x) ∈ Zx for x ∈ X.
Then by the definitions of weak ∗ -topology and the product topology, it is clear that T is a home-
omorphism from BX ∗ onto its image T (BX ∗ ). Recall a fact that any closed subset of a compact
Hausdorff space is compact. Since Z is compact Haudorsff, it suffices to show that T (BX ∗ ) is a
closed subset of Z.
Let z ∈ T (BX ∗ ). We are going to show that there is an element f ∈ BX ∗ such that f (x) = z(x)
 27

for all x ∈ X.
Define a function f : X → K by
f (x) := z(x)
for x ∈ X.
Claim : f (x+y) = f (x)+f (y) for all x, y ∈ X. In fact if we fix x, y ∈ X and for any ε > 0, then by
the definition of product topology, there is an element g ∈ BX ∗ such that |g(x + y) − z(x + y)| < ε;
|g(x) − z(x)| < ε; and |g(y) − z(y)| < ε. Since g is linear, we have g(x + y) − g(x) − g(y) = 0. This
implies that
|z(x + y) − z(x) − z(y)| = |z(x + y) − g(x + y) − (z(x) − g(x)) − (z(y) − g(y))| < 3ε
for all ε > 0. Thus we have z(x + y) = z(x) = z(y). The Claim follows.
Similarly, we have z(αx) = αz(x) for all α ∈ K and for all x ∈ X.
Therefore, the functional f (x) := z(x) is linear on X. It remains to show f is bounded with
∥f ∥ ≤ 1. In fact, for any x ∈ X and any ε > 0, then there is an element g ∈ BX ∗ such that
g(x) − z(x)| < ε. Therefore, we have |f (x)| = |z(x)| ≤ |g(x)| + ε ≤ ∥x∥ + ε. Therefore, f is bounded
and ∥f ∥ ≤ 1 as desired. The proof is complete. □

9. Open Mapping Theorem

Let E and F be the metric spaces. A mapping f : E → F is called an open mapping if f (U ) is
an open subset of F whenever U is an open subset of E.
Clearly, a continuous bijection is a homeomorphism if and only if it is an open map.
Remark 9.1. Warning An open map need not be a closed map.
For example, let p : (x, y) ∈ R2 7→ x ∈ R. Then p is an open map but it is not a closed map. In
fact, if we let A = {(x, 1/x) : x ̸= 0}, then A is closed but p(A) = R \ {0} is not closed.

Lemma 9.2. Let X and Y be normed spaces and T : X → Y a linear map. Then T is open if and
only if 0 is an interior point of T (U ) where U is the open unit ball of X.
Proof. The necessary condition is obvious.
For the converse, let W be a non-empty subset of X and a ∈ W . Put b = T a. Since W is open,
we choose r > 0 such that BX (a, r) ⊆ W . Note that U = 1r (BX (a, r) − a) ⊆ 1r (W − a). Thus, we
have T (U ) ⊆ 1r (T (W ) − b). Then by the assumption, there is δ > 0 such that BY (0, δ) ⊆ T (U ) ⊆
1
r (T (W ) − b). This implies that b + rBY (0, δ) ⊆ T (W ) and so, T (a) = b is an interior point of
T (W ). □

Corollary 9.3. Let M be a closed subspace of a normed space X. Then the natural projection
π : X → X/M is an open map.
Proof. Put U and V the open unit balls of X and X/M respectively. Using Lemma 9.2, the result
is obtained by showing that V ⊆ π(U ). Note that if x̄ = π(x) ∈ V , then by the definition a quotient
norm, we can find an element m ∈ M such that ∥x + m∥ < 1. Hence we have x + m ∈ U and
x̄ = π(x + m) ∈ π(U ). □

Before showing the main result, we have to make use one of important properties of a metric
space which is known as the Baire Category Theorem. Recall that a subset A of a metric space E
is called a nowhere dense set if the closure A of A has no interior point.

Proposition 9.4. Let E be a complete metric space with a metric d. If E is a union of a sequence
of subsets (An ) of E, then int(AN ) ̸= ∅ for some AN . Hence, every complete metric space is not a
countable union of nowhere dense sets.
28 CHI-WAI LEUNG

Proof. Let Fn := An . Hence, E = ∞

S
n=1 Fn . Assume that each Fn has no interior points. Fix
an element x1 ∈ E. Let 0 < η1 < 1/2. Then B(x1 , η1 ) ⊈ F1 . Then there is an element x2 ∈
B(x1 , η1 ) \ F1 . Since F1 is closed, we can choose 0 < η2 < 1/22 such that B(x2 , η2 ) ∩ F1 = ∅
and B(x2 , η2 ) ⊆ B(x1 , η1 ). To repeat the same step, we have a sequence of elements (xk ) in E;
a decreasing sequence of positive of numbers (ηk ) such that for all k = 1, 2... satisfy the following
conditions:
(1) 0 < ηk < 1/2k .
(2) B(xk+1 , ηk+1 ) ⊆ B(xk , ηk ).
(3) B(xk+1 , ηk+1 ) ∩ Fk = ∅.
The completeness of E, together with conditions (1) and (2) imply that the sequence (xk ) is a
Cauchy sequence and thus, the limit l := limk xk exists with l ∈ ∞
T
k=1 B(xk , ηk ). Since E =
S ∞
F
n=1 n , the limit l ∈ F K for some K. However, it leads to a contradiction because FK ∩
B(xK , ηK ) = ∅ by the condition (3) above. □

Lemma 9.5. Let T : X −→ Y be a bounded linear surjection from a Banach space X onto a
Banach space Y . Then 0 is an interior point of T (U ), where U is the open unit ball of X, i.e.,
U := {x ∈ X : ∥x∥ < 1}.
Proof. Set U (r) := {x ∈ X : ∥x∥ < r} for r > 0 and so, U = U (1).
Claim 1 : 0 is an interior point of T (U (1)).
S∞
Note that since T is surjective, Y = n=1 T (U (n)). Then by the Baire Category Theorem, there
exists N such that int T (U (N )) ̸= ∅. Let y ′ be an interior point of T (U (N )). Then there is
η > 0 such that BY (y ′ , η) ⊆ T (U (N )). Since BY (y ′ , η) ∩ T (U (N )) ̸= ∅, we may assume that
y ′ ∈ T (U (N )). Let x′ ∈ U (N ) such that T (x′ ) = y ′ . Then we have
0 ∈ BY (y ′ , η) − y ′ ⊆ T (U (N )) − T (x′ ) ⊆ T (U (2N )) = 2N T (U (1)).
Thus, we have 0 ∈ 2N 1
(BY (y ′ , η) − y ′ ) ⊆ T (U (1)). Hence 0 is an interior point of T (U (1)). The
Claim 1 follows.
Therefore there is r > 0 such that BY (0, r) ⊆ T (U (1)). This implies that we have
(9.1) BY (0, r/2k ) ⊆ T (U (1/2k ))
for all k = 0, 1, 2....
Claim 2 : D := BY (0, r) ⊆ T (U (3)).
Let y ∈ D. By Eq 9.1, there is x1 ∈ U (1) such that ∥y − T (x1 )∥ < r/2. Then by using Eq 9.1
again, there is x2 ∈ U (1/2) such that ∥y − T (x1 ) − T (x2 )∥ < r/22 . To repeat the same steps, there
exists is a sequence (xk ) such that xk ∈ U (1/2k−1 ) and
∥y − T (x1 ) − T (x2 ) − ... − T (xk )∥ < r/2k
for all k. On the other hand, since ∞
P P∞ k−1 and X is Banach, x :=
P∞
k=1 ∥xk ∥ ≤ k=1 1/2 k=1 xk
exists in X and ∥x∥ ≤ 2. This implies that y = T (x) and ∥x∥ < 3.
Thus we the result follows. □
.

Theorem 9.6. Open Mapping Theorem : Using the notations as in Lemma 9.5, then T is an
open mapping.
Proof. The proof is complete by using Lemmas 9.2 and 9.5. □

Proposition 9.7. Let T be a bounded linear isomorphism between Banach spaces X and Y . Then
T −1 is bounded.
 29

Consequently, if ∥ · ∥ and ∥ · ∥′ both are complete norms on X such that ∥ · ∥ ≤ c∥ · ∥′ for some c > 0,
then these two norms ∥ · ∥ and ∥ · ∥′ are equivalent.
Proof. The first assertion follows immediately from the Open Mapping Theorem.
Therefore, the last assertion can be obtained by considering the identity map I : (X, ∥·∥) → (X, ∥·∥′ )
which is bounded by the assumption. □

Corollary 9.8. Let X and Y be Banach spaces and T : X → Y a bounded linear operator. Then
the followings are equivalent.
(i) The image of T is closed in Y .
(ii) There is c > 0 such that
d(x, ker T ) ≤ c∥T x∥
for all x ∈ X.
(iii) If (xn ) is a sequence in X such that ∥xn + ker T ∥ = 1 for all n, then ∥T xn ∥ ↛ 0.
Proof. Let Z be the image of T . Then the canonical map Te : X/kerT → Z induced by T is a
bounded linear isomorphism. Note that Te(x̄) = T x for all x ∈ X, where x̄ := x + ker T ∈ X/ ker T .
For (i) ⇒ (ii): suppose that Z is closed. Then Z becomes a Banach space. Then the Open
Mapping Theorem implies that the inverse of Te is also bounded. Thus, there is c > 0 such that
d(x, ker T ) = ∥x̄∥X/ ker T ≤ c∥Te(x̄)∥ = c∥T (x)∥ for all x ∈ X. The part (ii) follows.
For (ii) ⇒ (i), let (xn ) be a sequence in X such that lim T xn = y ∈ Y exists and so, (T xn ) is
a Cauchy sequence in Y . Then by the assumption, (x̄n ) is a Cauchy sequence in X/ ker T . Since
X/ ker T is complete, we can find an element x ∈ X such that lim x̄n = x̄ in X/ ker T . This gives
y = lim T (xn ) = lim Te(x̄n ) = Te(x̄) = T (x). Therefore, y ∈ Z.
(ii) ⇔ (iii) is obvious. The proof is complete. □

Proposition 9.9. Let X and Y be Banach spaces. Let T and K belong to B(X, Y ). Suppose that
T (X) is closed and K is of finite rank, then the image (T + K)(X) is also closed.
Proof. Suppose the conclusion does not hold. We write z̄ := z + ker(T + K) for z ∈ X. Then by
Corollary 9.8, there is a sequence (xn ) in X such that ∥x̄n ∥ = 1 for all n and ∥(T +K)xn ∥ → 0. Thus,
(xn ) can be chosen so that it is bounded. By passing a subsequence of (xn ) we may assume that
y := limn K(xn ) exists in Y because K is of finite rank. Therefore, we have limn T (xn ) = −y. Since
T has closed range, we have T x = −y for some x ∈ X. This gives lim T (xn − x) = 0. Note that the
natural map Te is a topological isomorphism from X/ ker T onto T (X) because T (X) is closed. We
see that ∥xn −x+ker T ∥ → 0 and thus, ∥y−K(x)+K(ker T )∥ = lim ∥K(xn )−K(x)+K(ker T )∥ = 0.
From this we have y − Kx = Ku for some u ∈ ker T . In addition, for each n, there is an element
tn ∈ ker T so that ∥xn − x + tn ∥ < 1/n. This implies that
∥K(tn − u)∥ ≤ ∥K(tn + (xn − x))∥ + ∥ − K(xn + x) − K(u)∥ ≤ ∥K∥1/n → 0.
Therefore, we have ∥tn − u + (ker T ∩ ker K)∥ → 0 because tn − u ∈ ker T and the image of K| ker T
is closed. From this we see that ∥tn − u + ker(T + K)∥ → 0.
On the other hand, since T x = −y = −Kx − Ku and u ∈ ker T , we have (T + K)x = −Ku − T u
and so, x + u ∈ ker(T + K). Then we can now conclude that
∥x̄n ∥ = ∥x̄n − (x̄ + ū)∥ ≤ ∥x̄n − x̄ − t̄n ∥ + ∥t̄n − ū∥ → 0.
It contradicts to the choice of xn such that ∥x̄n ∥ = 1 for all n. The proof is complete. □

Remark 9.10. In general, the sum of operators of closed ranges may not have a closed range.
Before looking for those examples, let us show the following simple useful lemma.
30 CHI-WAI LEUNG

Lemma 9.11. Let X be a Banach space. If T ∈ B(X) with ∥T ∥ < 1, then the operator 1 − T is
invertible, i.e., there is S ∈ B(X) such that (1 − T )S = S(1 − T ) = 1.
Proof. Note that since X is a Banach space, the set of all bounded
P operators B(X) is a Banach
space under the usual operator norm. This implies that the series ∞k=0 T k is convergent in B(X)

Xn
because ∥T ∥ < 1. On the other hand, we have 1 − T n = (1 − T )( T k ) for all n = 1, 2.... Taking
k=0
∞
X
n → ∞, we see that (1 − T )−1 exists, in fact, (1 − T )−1 = T k. □
k=0

Example 9.12. Define an operator T0 : ℓ∞ → ℓ∞ by

1
T0 (x)(k) := x(k)
k
for x ∈ ℓ∞ and k = 1, 2.... Note that T0 is injective with ∥T0 ∥ ≤ 1 and im T0 ⊆ c0 . The Open
mapping Theorem tells us that the image im T0 must not be closed. Otherwise T0 becomes an
isomorphism from ℓ∞ onto a closed subspace of c0 . It is ridiculous since ℓ∞ is nonseparable but
c0 is not. Now if we let T := 21 T0 , then ∥T ∥ < 1 and T is without closed range. Applying Lemma
9.11, we see that the operator S := 1 − T is invertible and thus, S has closed range. Then by our
construction T = 1 − S is the sum of two operators of closed ranges but T does not have closed
range as required.

10. Closed Graph Theorem

Let T : X −→ Y . The graph of T , denoted by G(T ), is defined by the set {(x, y) ∈ X × Y : y =
T (x)}.
Now the direct sum X ⊕ Y is endowed with the norm ∥ · ∥∞ , i.e., ∥x ⊕ y∥∞ := max(∥x∥X , ∥y∥Y ).
We write X ⊕∞ Y when X ⊕ Y is equipped with this norm.
An operator T : X −→ Y is said to be closed if its graph G(T ) is a closed subset of X ⊕∞ Y , i.e.,
whenever, a sequence (xn ) of X satisfies the condition ∥(xn , T xn ) − (x, y)∥∞ → 0 for some x ∈ X
and y ∈ Y , we have T (x) = y.

Theorem 10.1. Closed Graph Theorem : Let T : X −→ Y be a linear operator from a Banach
space X to a Banach Y . Then T is bounded if and only if T is closed.
Proof. The part (⇒) is clear.
Assume that T is closed, i.e., the graph G(T ) is ∥ · ∥∞ -closed. Define ∥ · ∥0 : X −→ [0, ∞) by
∥x∥0 = ∥x∥ + ∥T (x)∥
for x ∈ X. Then ∥ · ∥0 is a norm on X. Let I : (X, ∥ · ∥0 ) −→ (X, ∥ · ∥) be the identity operator. It
is clear that I is bounded since ∥ · ∥ ≤ ∥ · ∥0 .
Claim: (X, ∥ · ∥0 ) is Banach. In fact, let (xn ) be a Cauchy sequence in (X, ∥ · ∥0 ). Then (xn ) and
(T (xn )) both are Cauchy sequences in (X, ∥ · ∥) and (Y, ∥ · ∥Y ). Since X and Y are Banach spaces,
there are x ∈ X and y ∈ Y such that ∥xn − x∥X → 0 and ∥T (xn ) − y∥Y → 0. Thus y = T (x) since
the graph G(T ) is closed.
Then by Theorem 9.7, the norms ∥ · ∥ and ∥ · ∥0 are equivalent. Hence, there is c > 0 such that
∥T (·)∥ ≤ ∥ · ∥0 ≤ c∥ · ∥ and hence, T is bounded since ∥T (·)∥ ≤ ∥ · ∥0 . The proof is complete. □

P∞ 2
Example 10.2. Let D := {c = (cn ) ∈ ℓ2 : 2 2
n=1 n |cn | < ∞}. Define T : D −→ ℓ by
T (c) = (ncn ). Then T is an unbounded closed operator.
 31

Proof. Note that since ∥T en ∥ = n for all n, T is not bounded. Now we claim that T is closed.
Let (xi ) be a convergent sequence in D such that (T xi ) is also convergent in ℓ2 . Write xi = (xi,n )∞
n=1
with lim xi = x := (xn ) in D and lim T xi = y := (yn ) in ℓ2 . This implies that if we fix n0 , then
i i
lim xi,n0 = xn0 and lim n0 xi,n0 = yn0 . This gives n0 xn0 = yn0 . Thus T x = y and hence T is
i i
closed. □

Example 10.3. Let X := {f ∈ C b (0, 1) ∩ C ∞ (0, 1) : f ′ ∈ C b (0, 1)}. Define T : f ∈ X 7→ f ′ ∈

C b (0, 1). Suppose that X and C b (0, 1) both are equipped with the sup-norm. Then T is a closed
unbounded operator.
Proof. Note that if a sequence fn → f in X and fn′ → g in C b (0, 1). Then f ′ = g. Hence T is
closed. In fact, if we fix some 0 < c < 1, then by the Fundamental Theorem of Calculus, we have
Z x Z x
0 = lim(fn (x) − f (x)) = lim( (fn′ (t) − f ′ (t))dt) = (g(t) − f ′ (t))dt
n n c c
Rx Rx
for all x ∈ (0, 1). This implies that we have c g(t)dt = c f ′ (t)dt. Thus g = f ′ on (0, 1).
On the other hand, since ∥T xn ∥∞ = n for all n ∈ N. Thus T is unbounded as desired. □

Let X be a normed space and let X ∗ be its dual space. Then there is a natural bi-linear mapping
on X × X ∗ (call a dual pair) given by
⟨·, ·⟩ : X × X ∗ → K; ⟨x, f ⟩ = f (x).
Moreover, this dual pair is non-degenerate, that is, ⟨x, f ⟩ = 0 for all f ∈ X ∗ if and only if x = 0
and ⟨x, f ⟩ = 0 for all x ∈ X if and only if f = 0.

Proposition 10.4. Let X and Y be Banach spaces. Let G : Y ∗ → X ∗ be a w∗ -w∗ continuous

linear map. Then we have the following assertions.
(i) G is bounded.
(ii) There exists a bounded linear map T ∈ B(X, Y ) such that T ∗ = G.
∥·∥ ∥·∥
Proof. For showing Part (i), let (yn∗ ) be a sequence in Y ∗ such that yn∗ −−→ y ∗ and Gyn∗ −−→ x∗ in
the norm topologies. By using the Closed Graph Theorem, we want to show Gy ∗ = x∗ , that is,
∥·∥ w∗ w∗
(Gy ∗ )(x) = x∗ (x) for all x ∈ X. In fact, yn∗ −−→ y ∗ , so yn∗ −−→ y ∗ . Thus, we have Gyn∗ −−→ Gy ∗ , so
∥·∥
(Gyn∗ )(x) → (Gy ∗ )(x) for all x ∈ X. On the other hand, since Gyn∗ −−→ x∗ , we have (Gyn∗ )(x) →
x∗ (x) for all x ∈ X. Therefore, (Gy ∗ )(x) = x∗ (x) for all x ∈ X as desired.
For Part (ii), note that for each x ∈ X, the map f ∈ Y ∗ 7→ ⟨x, Gf ⟩ is w∗ -continuous on Y ∗ . Hence,
Proposition 8.2 gives a unique element Rx ∈ Y such that
⟨Rx, f ⟩ = ⟨x, Gf ⟩
for all f ∈ Y ∗. Then by using Part (i) and the Closed Graph Theorem, R is bounded. The proof
is complete. □

11. Uniform Boundedness Theorem

Theorem 11.1. Uniform Boundedness Theorem : Let {Ti : X −→ Y : i ∈ I} be a family of
bounded linear operators from a Banach space X into a normed space Y . Suppose that for each
x ∈ X, we have sup ∥Ti (x)∥ < ∞. Then sup ∥Ti ∥ < ∞.
i∈I i∈I

Proof. For each x ∈ X, define

∥x∥0 := max(∥x∥, sup ∥Ti (x)∥).
i∈I
32 CHI-WAI LEUNG

Then ∥ · ∥0 is a norm on X and ∥ · ∥ ≤ ∥ · ∥0 on X. If (X, ∥ · ∥0 ) is complete, then by the Open

Mapping Theorem. This implies that ∥ · ∥ is equivalent to ∥ · ∥0 and thus there is c > 0 such that
∥Tj (x)∥ ≤ sup ∥Ti (x)∥ ≤ ∥x∥0 ≤ c∥x∥
i∈I
for all x ∈ X and for all j ∈ I. Thus∥Tj ∥ ≤ c for all j ∈ I is as desired.
Thus it remains to show that (X, ∥ · ∥0 ) is complete. In fact, if (xn ) is a Cauchy sequence in
(X, ∥ · ∥0 ), then it is also a Cauchy sequence with respect to the norm ∥ · ∥ on X. Write x := limn xn
with respect to the norm ∥ · ∥. For any ε > 0, there is N ∈ N such that ∥Ti (xn − xm )∥ < ε
for all m, n ≥ N and for all i ∈ I. Now fixing i ∈ I and n ≥ N and taking m → ∞, we have
∥Ti (xn − x)∥ ≤ ε and thus supi∈I ∥Ti (xn − x)∥ ≤ ε for all n ≥ N . Thus, we have ∥xn − x∥0 → 0
and hence (X, ∥ · ∥0 ) is complete. □

Remark 11.2. Consider c00 := {x = (xn ) : ∃ N, ∀ n ≥ N ; xn ≡ 0} which is endowed with ∥ · ∥∞ .

Now for each k ∈ N, if we define Tk ∈ c∗00 by Tk ((xn )) := kxk , then supk |Tk (x)| < ∞ for each
x ∈ c00 but (∥Tk ∥) is not bounded, in fact, ∥Tk ∥ = k. Thus the assumption of the completeness of
X in Theorem 11.1 is essential.

Corollary 11.3. Let X and Y be as in Theorem 11.1. Let Tk : X −→ Y be a sequence of bounded

operators. Assume that limk Tk (x) exists in Y for all x ∈ X. Then there is T ∈ B(X, Y ) such that
limk ∥(T − Tk )x∥ = 0 for all x ∈ X. Moreover, we have ∥T ∥ ≤ lim inf ∥Tk ∥.
k

Proof. Note that by the assumption, we can define a linear operator T from X to Y given by
T x := limk Tk x for x ∈ X. We need to show that T is bounded. In fact, (∥Tk ∥) is bounded by the
Uniform Boundedness Theorem since limk Tk x exists for all x ∈ X. Hence, for each x ∈ BX , there
is a positive integer K such that ∥T x∥ ≤ ∥TK x∥ + 1 ≤ (supk ∥Tk ∥) + 1. Thus, T is bounded.
Finally, it remains to show the last assertion. In fact, note that for any x ∈ BX and ε > 0, there
is N (x) ∈ N such that ∥T x∥ < ∥Tk x∥ + ε < ∥Tk ∥ + ε for all k ≥ N (x). This gives ∥T x∥ ≤
inf k≥N (x) ∥Tk ∥ + ε for all k ≥ N (x) and hence, ∥T x∥ ≤ inf k≥N (x) ∥Tk ∥ + ε ≤ supn inf k≥n ∥Tk ∥ + ε
for all x ∈ BX and ε > 0. Therefore, we have ∥T ∥ ≤ lim inf ∥Tk ∥. □
k

Corollary 11.4. Every weakly convergent sequence in a normed space must be bounded.
Proof. Let (xn ) be a weakly convergent sequence in a normed space X. If we let Q : X → X ∗∗
be the canonical isometry, then (Qxn ) is a bounded sequence in X ∗∗ . Note that (xn ) is weakly
convergent if and only if (Qxn ) is w∗ -convergent. Thus, (Qxn (x∗ )) is bounded for all x∗ ∈ X ∗ .
Note that the dual space X ∗ must be complete. We can apply the Uniform Boundedness Theorem
to see that (Qxn ) is bounded and so is (xn ). □

12. Projections on Banach Spaces

Throughout this section, let X be a Banach space. A linear operator P : X → X is called a
projection (or idempotent) if it is bounded and satisfies the condition P 2 = P .
In addition, a closed subspace E of X is said to be complemented if there is a closed subspace F
of X such that X = E ⊕ F .

Proposition 12.1. A closed subspace E of X is complemented if and only if there is a projection

Q on X with E = im Q.
Proof. We first suppose that there is a closed subspace F of X such that X = E ⊕ F . Define an
operator Q : X → X by Qx = u if x = u + v for u ∈ E and v ∈ F . It is clear that we have Q2 = Q.
For showing the boundedness of Q, by using the Closed Graph Theorem, we need to show that if
 33

(xn ) is a sequence in E such that lim xn = x and lim Qxn = u for some x, u ∈ E, then Qx = u.
Indeed, if we let xn = yn + zn where yn ∈ E and zn ∈ F , then Qxn = yn . Note that (zn ) is
a convergent sequence in F because zn = xn − yn and (xn ) and (yn ) both are convergent. Let
w = lim zn . This implies that
x = lim xn = lim(yn + zn ) = u + w.
Since E and F are closed, we have u ∈ E and w ∈ F . Therefore, we have Qx = u as desired.
The converse is clear. In fact, we have X = im Q ⊕ ker Q in this case. □

Proposition 12.2. If M is a finite dimensional subspace of a normed space X, then M is com-

plemented in X. Furthermore, if dim M = 1, then there is a bounded linear projection Q from X
onto M such that ∥Q∥ = 1.
Proof. In fact, if M is spanned by {ei : i = 1, 2.., m}, then M is closed and by the Hahn-Banach
Theorem, for eachTmi = 1, ..., m, there is e∗i ∈ X ∗ such that e∗i (ej ) = 1 if i = j, otherwise, it is equal
∗
to 0. Put N := i=1 ker ei . Then X = M ⊕ N .
To see the last assertion, let M := Kx0 for some x0 ∈ X. The Hahn-Banach Theorem gives an
element f ∈ X ∗ with ∥f ∥ = 1 such that f (x0 ) = ∥x0 ∥. Notice that we have X = ker f ⊕M . Then for
any x ∈ X we have x = y + tx0 for some t ∈ K and for some y ∈ ker f . Thus, ∥tx0 ∥ = f (x) ≤ ∥x∥.
Now if let Q be the corresponding projection from X onto M along this decomposition, then we have
∥Qx∥ = ∥tx0 ∥ ≤ ∥x∥ for all x ∈ X. Hence, ∥Q∥ ≤ 1. However, we always have ∥Q∥ = ∥Q2 ∥ ≤ ∥Q∥2
and so ∥Q∥ ≥ 1. Therefore, ∥Q∥ = 1 as desired. □

The following example can be found in [4].

Example 12.3. c0 is not complemented in ℓ∞ .

Proof. It will be shown by the contradiction. Suppose that cT ∞
0 is complemented in ℓ .
∞
Claim 1: There is a sequence (fn ) in (ℓ∞ )∗ such that c0 = n=1 ker fn .
In fact, by the assumption, there is a closed subspace F of ℓ∞ such that ℓ∞ = c0 ⊕ F . If we let P
be the projection from ℓ∞ onto F along this decomposition, then ker P = c0 and P is bounded by
the Closed Graph Theorem. Let e∗n : ℓ∞ → K be the n-th coordinate functional. Then e∗n ∈ (ℓ∞ )∗ .
∞
Thus, if we put fn = e∗n ◦ P , then fn ∈ (ℓ∞ )∗ and c0 = n=1 ker fn as desired.
T
Claim 2: For each irrational number α ∈ [0, 1], there is an infinite subset Nα of N such that
Nα ∩ Nβ is a finite set if α and β both are distinct irrational numbers in [0, 1].
In fact, we write [0, 1] ∩ Q as a sequence (rn ). Then for each irrational α in [0, 1], there is a
subsequence (rnk ) of (rn ) such that limk rnk = α. Let Nα := {nk : k = 1, 2...}. From this, we see
that Nα ∩ Nβ is a finite set whenever α, β ∈ [0, 1] ∩ Qc with α ̸= β. Claim 2 follows.
Now for each α ∈ [0, 1] ∩ Qc , define an element xα ∈ ℓ∞ by xα (k) ≡ 1 as k ∈ Nα ; otherwise,
xα (k) ≡ 0.
Claim 3: If f ∈ (ℓ∞ )∗ with c0 ⊆ ker f , then for any η > 0, the set {α ∈ [0, 1] ∩ Qc : |f (xα )| ≥ η}
is finite.
Note that by considering the decomposition f = Re(f ) + iIm(f ), it suffices to show that the set
{α ∈ [0, 1] ∩ Qc : f (xα ) ≥ η} is finite. Let α1 , ...αN in [0, S
1] ∩ Qc such that f (xαj ) ≥ η, j = 1, ..., N .
Now for each j = 1, .., N , set yj (k) ≡ 1 as k ∈ Nαj \ m̸=j Nαm ; otherwise yj ≡ 0. Note that
xαj − yj ∈ c0 since Nα ∩ Nβ is finite for α ̸= β by Claim 2. Hence, we have f (xαj ) = f (yj ) for all
j = 1, ..., N . Moreover, we have {k : yj (k) = 1} ∩ {k; yi (k) = 1} = ∅ for i, j = 1, ..., N with i ̸= j.
Thus, we have ∥y∥∞ = 1. Now we can conclude that
XN N
X
∥f ∥ ≥ f ( yj ) = f (xαj ) ≥ N η.
j=1 j=1
34 CHI-WAI LEUNG

This implies that |{α : f (α) ≥ η}| ≤ ∥f ∥/η. Claim 3 follows.

We are now going to complete the proof. Now ∞ ∗
S∞ let (fn ) cbe the sequence in (ℓ ) as found in the
Claim 1. Claim 3 implies that the set S := n=1 {α ∈ Q ∩ [0, 1] :Tfn (xα ) ̸= 0} is countable. Thus,
there exists γ ∈ [0, 1] ∩ Qc such that γ ∈/ S. Thus, we have xγT∈ ∞ n=1 ker fn . Besides, since Nγ is
an infinite set, we see that xγ ∈/ c0 . Therefore, we have c0 ⊊ ker fk which contradicts to Claim
1. □

Proposition 12.4. (Dixmier) Let X be a normed space. Let i : X → X ∗∗ and j : X ∗ → X ∗∗∗

be the natural embeddings. Then the composition Q := j ◦ i∗ : X ∗∗∗ → X ∗∗∗ is a projection with
Q(X ∗∗∗ ) = X ∗ .
Consequently, X ∗ is a complemented closed subspace of X ∗∗∗ .
Proof. Clearly, Q is bounded. Note that i∗ ◦ j = IdX ∗ : X ∗ → X ∗ . From this, we see that Q2 = Q
as desired.
We need to show that im Q = X ∗ , more precisely, im Q = j(X ∗ ). In fact, it follows from Q ◦ j = j
by using the equality i∗ ◦ j = IdX ∗ again.
The last assertion follows immediately from Proposition 12.1. □

Corollary 12.5. c0 is not isomorphic to the dual space of a normed space.

Proof. Suppose not. Let T : c0 → X ∗ be an isomorphism from c0 onto the dual space of some
normed space X. Then T ∗∗ : c∗∗ 0 = ℓ
∞ → X ∗∗∗ is an isomorphism too. Let Q : X ∗∗∗ → X ∗∗∗ be
∗
the projection with im Q = X which is found in Proposition 12.4.
Now put P := (T ∗∗ )−1 ◦ Q ◦ T ∗∗ : ℓ∞ → ℓ∞ . Then P is a projection.
On the other hand, we always have T ∗∗ |c0 = T (see Remark 5.2). This implies that im P = c0 .
Thus, c0 is complemented in ℓ∞ by Proposition 12.1 which leads to a contradiction by Example
12.3. □

Recall that a closed subspace M of a Banach space E is called an M -ideal if the space M ⊥ :=
{x∗ ∈ E ∗ : x∗ (M ) ≡ 0} is L
a ℓ1 -direct summand of E ∗ , that is, there is another closed subspace N
∗ ∗ ⊥ ∗ ∗ ∗
of E such that E = M ℓ1 N , i.e., for every element x ∈ E satisfies the condition: x = u + v
and ∥x∗ ∥ = ∥u∥ + ∥v∥ for a pair of elements u and v in M ⊥ and N respectively.

Proposition 12.6. We keep the notation as give

L in Proposition 12.4. If X is viewed as a closed
subspace of X ∗∗ and suppose that X ∗∗∗ = X ⊥ ℓ1 N for some closed subspace N of X ∗∗∗ , i.e. X
is an M -ideal of X ∗∗ , then N = X ∗ .
Proof. Let Q : X ∗∗∗ → X ∗∗∗ be the projection given in Proposition 12.4. Recall that Qz = j(z|X )
for z ∈ X ∗∗∗ and im Q = ∗ ⊥
LX .∗ Moreover ∥Q∥ ≤ 1. Note that ker Q = X := {z ∈ X
∗∗∗ : z| ≡ 0}
X
∗∗∗ ⊥ X . Let z ∈ N . Then we have Q(z) = (Q(z) − z) + z ∈ X ⊥ ℓ1 N
L
and hence, X = X
and hence, ∥Q(z)∥ = ∥Q(z) − z∥ + ∥z∥. Since ∥Q∥ ≤ 1, we see that ∥Q(z) − z∥ = 0 and thus,
z = Q(z) ∈ X ∗ . Therefore, we have N ⊆ X ∗ , so N = X ∗ . The proof is complete. □

n
Proposition 12.7. The c0 space is an M -ideal of ℓ∞ .
Proof. We first notice that for h ∈ (ℓ∞ )∗ and ξ ∈ ℓ∞ , then Re(h)(ξ) := Re(h(ξ)) can be viewed as
a R-linear functional on ℓ∞ and ∥h∥ = ∥Re(h)∥.
Using Proposition 12.6, it suffices to show that for g ∈ c∗0 = ℓ1 and f ∈ c⊥
0 , we have ∥g + f ∥(ℓ∞ )∗ =
⊥ ∗
∥g∥(ℓ∞ )∗ + ∥f ∥(ℓ∞ )∗ , where c0 := {f ∈ (ℓ∞ ) : f (c0 ) ≡ 0}. Let ε > 0. By considering the polar
 35

decomposition, then there are elements ξ and ξ ′ in (ℓ∞ )1 of norm-one such that
∞
X
∥f ∥ − ε < f (ξ) and ∥g∥ − ε < g(ξ ′ ) = Re(g)(ξ ′ ) = Re(ξ ′ (n)g(n)).
n=1

Since g ∈ c∗0 = ℓ1 , there is N such that |g(n)| < ε. Now let ξ ′′ be an element in ℓ∞ given by
P
n>N
(
ξ ′ (n) if n ≤ N
ξ ′′ (n) :=
ξ(n) if n > N .

Then ∥ξ ′′ ∥∞ ≤ 1 and ξ ′′ −P
ξ ∈ c00 . Hence we have f (ξ) = f (ξ ′′ ) because f (c0 ) ≡ 0.
On the other hand, since n>N |g(n)| < ε, we have
N
X X N
X
′ ′
∥g∥ − ε < Re(g)(ξ ) ≤ Re(ξ (n)g(n)) + |g(n)| < Re(ξ ′ (n)g(n)) + ε.
n=1 n>N n=1

Thus, we have
∞
X
′′
Re(g)(ξ ) = Re(ξ ′′ (n)g(n))
n=1
XN X
≥ Re(ξ ′ (n)g(n)) − | ξ(n)g(n)|
n=1 n>N
≥ ∥g∥ − 2ε − ε.
Therefore, we have
∥g∥ + ∥f ∥ = ∥Re(g)∥ + ∥f ||
≤ Re(g)(ξ ′′ ) + f (ξ ′′ ) + 4ε
= Re(g + f )(ξ ′′ ) + 4ε
≤ ∥Re(g + f )∥ + 4ε
= ∥g + f ∥ + 4ε.
for all ε > 0. The proof is complete.
□

13. Appendix: Basic sequences

Throughout this section, X always denotes a Banach space.
An infinite sequence (xn ) in X is called a basic sequence if for each element x in X0 := [x1 , x2 , · · · ],
the closed linear span of {x1 , x2 , ...}, then there is a unique sequence of scalars (an ) such that
x= ∞
P
i=1 i xi , that is (xn ) is a Schauder basis for its closed linear span [(xi )].
a
Put ψi the corresponding i-th coordinate Pfunction, i.e., Pψni (x) := ai and Qn : X0 → En := [x1 · · · xn ]
the n-th canonical projection, i.e., Qn ( ∞ i=1 ai x i ) := i=1 ai xi .

Theorem 13.1. Using the notations as above, for each element x ∈ X0 , put
q(x) := sup{∥Qn (x)∥ : n = 1, 2...}.
Then
(i) q is a Banach equivalent norm on X0 .
(ii) Each coordinate projection Qn and coordinate function ψn are bounded in the original norm-
topology.
36 CHI-WAI LEUNG

Proof. Since x = limn Qn x for all x ∈ X0 , ∥x∥ = limn ∥Qn x∥. This implies that the sequence
(∥Qn x)∥) is bounded and hence, q is well defined, in addition, q(·) ≥ ∥ · ∥ on X0 . From this,
together with the Open Mapping Theorem, all assertions follows if we show that q is a Banach
norm on X0 .
Let (xn ) be a Cauchy sequence in X0 with respect to the norm q. Clearly, (xn ) is also a Cauchy
sequence in the ∥ · ∥-topology because q(·) ≥ ∥ · ∥. Let x = limn xn be the limit in X0 in the
∥ · ∥-topology. We are going to show that x is also the limit of (xn ) with respect to the q-topology.
We first note that yk := limn Qk xn exists in X0 for all k = 1, 2, ... by the definition of the norm q.
Claim 1: ∥ · ∥-limk yk = x.
Let ε > 0. Then by the definition of the norm q, there is a positive integer N1 such that ∥Qk xN −
Qk xm ∥ < ε and ∥xN − xm ∥ < ε for all m, N ≥ N1 and for all k = 1, 2.... This gives
∥x − Qk xm ∥ ≤ ∥x − xN1 ∥ + ∥xN1 − Qk xN1 ∥ + ∥Qk xN1 − Qk xm ∥ < 2ε + ∥xN1 − Qk xN1 ∥
for all m ≥ N1 and for all positive integers k. Thus, if we take m → ∞, then we have
∥x − yk ∥ ≤ 2ε + ∥xN1 − Qk xN1 ∥ → 2ε + 0 as k → ∞.
Claim 2: Qk x = yk for all k = 1, 2....
Fix a positive integer k1 . Note that Qk1 yk = yk1 for all k ≥ k1 . Indeed, since Ek and Ek1 are of
finite dimension, the restrictions Qk1 |Ek and Qk |Ek1 both are continuous. This implies that
Qk1 yk = Qk1 (lim Qk xn ) = lim Qk1 Qk (xn ) = lim Qk Qk1 (xn ) = Qk (lim Qk1 xn ) = Qk (yk1 ) = yk1
n n n n

sequence of scalars (βn ) so that yk = ki=1 βi xi for all k = 1, 2...

P
for all k ≥ k1 . Henece, there is a P
∞
P∞ hand, if we let x = i=1 αi xi , then by Claim 1 we have limk (yk − Qk x) = 0 and thus
On the other
we have i=1 (βi − αi )xi = 0. Therefore, we have βi = αi for all i = 1, 2.... The Claim 2 follows.
It remains to show that limn q(xn − x) = 0.
Let η > 0. Then there is a positive integer N so that ∥Qk xn − Qk xm ∥ < η for all m, n ≥ N and for
all positive integers k. Taking m → ∞, Claim 2 gives
∥Qk xn − Qk x∥ = ∥Qk xn − yk ∥ ≤ η
for all n ≥ N and for all positive integers k. Thus, we have q(xn − x) ≤ η for all n ≥ N . The proof
is complete. □

Proposition 13.2. Using the notation as above, a sequence (xi ) is a basic sequence in X if and
only if there is K > 0 such that
X X
ai xi ≤ K ai xi ,
1≤i≤n 1≤i≤m

whenever n ≤ m and (ai ) ∈ c00 .

 37

14. Geometry of Hilbert space I

From now on, all vectors spaces are over the complex field. Recall that an inner product on a
vector space V is a function (·, ·) : V × V → C which satisfies the following conditions.

(i) (x, x) ≥ 0 for all x ∈ V and (x, x) = 0 if and only if x = 0.

(ii) (x, y) = (y, x) for all x, y ∈ V .
(iii) (αx + βy, z) = α(x, z) + β(y, z) for all x, y, z ∈ V and α, β ∈ C.
Consequently, for each x ∈ V , the map y ∈ V 7→ (x, y) ∈ C is conjugate linear by the conditions
(ii) and (iii), i.e., (x, αy + βz) = ᾱ(x, y) + β̄(x, z) for all y, z ∈ V and α, β ∈ C.
In addition, the inner product (·, ·) will give a norm on V which is defined by
p
∥x∥ := (x, x)
for x ∈ V .

We first recall the following useful properties of an inner product space which can be found in the
standard text books of linear algebras.

Proposition 14.1. Let V be an inner product space. For all x, y ∈ V , we always have:
(i): (Cauchy-Schwarz inequality): |(x, y)| ≤ ∥x∥∥y∥ Consequently, the inner product on
V × V is jointly continuous.
(ii): (Parallelogram law): ∥x + y∥2 + ∥x − y∥2 = 2∥x∥2 + 2∥y∥2 .
Furthermore, a norm ∥ · ∥ on a vector space X is induced by an inner product if and only if
it satisfies the Parallelogram law. In this case such inner product is given by the following:
1 1
Re(x, y) = (∥x + y∥2 − ∥x − y∥2 ) and Im(x, y) = (∥x + iy∥2 − ∥x − iy∥2 )
4 4
for all x, y ∈ X.
(iii) Gram-Schmidt process Let {x1 , x2 , ...} be a sequence of linearly independent vectors in
an inner product space V . Put e1 := x1 /∥x1 ∥. Define en inductively on n by
xn − nk=1 (xk , ek )ek
P
en+1 := .
∥xn − nk=1 (xk , ek )ek ∥
P

Then {en : n = 1, 2, ..} forms an orthonormal system in V Moreover, the linear span of
x1 , ..., xn is equal to the linear span of e1 , ..., en for all n = 1, 2....

Clearly, due to the Cauchy-Schwarz inequality, we have the following.

Lemma 14.2. If X is an inner product space, then the inner product (·, ·) : X × X → C is jointly
continuous, that is limn (xn , yn ) = (x, y) whenever (xn ) and (yn ) are convergent sequences in X
with limn xn = x and limn yn = y.
The following is one of the most important classes in mathematics.

Definition 14.3. A Hilbert space is a Banach space whose norm is given by an inner product.

Example 14.4. (1) CN is a Hilbert space under the usual inner product given by (w, z) :=
PN N
k=1 w(k)z(k) for w, z ∈ C .
(2) It follows from Proposition 14.1 immediately that ℓ2 is a Hilbert space and ℓp is not a Hilbert
space for all p ∈ [1, ∞] \ {2}.
38 CHI-WAI LEUNG

In the rest of this section, X always denotes a complex Hilbert space with an inner product (·, ·).
Recall that two vectors x and y in an inner product space V are said to be orthogonal if (x, y) = 0.

Proposition 14.5. (Bessel′ s inequality) : Let {e1 , ..., eN } be an orthonormal set in an inner
product space V , i.e., (ei , ej ) = 1 if i = j, otherwise is equal to 0. Then for any x ∈ V , we have
N
X
|(x, ei )|2 ≤ ∥x∥2 .
i=1

Proof. It can be obtained by the following equality immediately

N
X N
X
2 2
∥x − (x, ei )ei ∥ = ∥x∥ − |(x, ei )|2 .
i=1 i=1
□

Corollary 14.6. Let (ei )i∈I be an orthonormal set in a Hilbert space X. Then for any element
x ∈ X, the set
I(x) := {i ∈ I : (ei , x) ̸= 0}
is countable.
Proof. Note that for each x ∈ V , we have
∞
[
{i ∈ I : (ei , x) ̸= 0} = {i ∈ I : |(ei , x)| ≥ 1/n}.
n=1
Then the Bessel’s inequality implies that the set {i ∈ I : |(ei , x)| ≥ 1/n} must be finite for each
n ≥ 1. Thus the result follows. □

Proposition 14.7. Let P (en ) be a sequence of orthonormal vectors in a Hilbert space X. Then for
any x ∈ V , the series ∞ n=1 (x, en )en is convergent.
Moreover, if (eσ(n) ) is a rearrangement of (en ), i.e., σ : {1, 2...} −→ {1, 2, ..} is a bijection. Then
we have
X∞ ∞
X
(x, en )en = (x, eσ(n) )eσ(n) .
n=1 n=1

of the series ∞
P
Proof. Since X is a Hilbert space, the convergence
Pp n=1 (x, en )en follows from the
Bessel’s inequality. In fact, if we put sp := n=1 (x, en )en , then we have
X
∥sp+k − sp ∥2 = |(x, en )|2 .
p+1≤n≤p+k
P∞ P∞
Now put y = n=1 (x, en )en and z = n=1 (x, eσ(n) )eσ(n) . Note that we have
N
X N
X
(y, y − z) = lim( (x, en )en , (x, en )en − z)
N
n=1 n=1
N
X N
X ∞
X
2
= lim |(x, en )| − lim (x, en ) (x, eσ(j) )(en , eσ(j) )
N N
n=1 n=1 j=1
∞
X N
X
2
= |(x, en )| − lim (x, en )(x, en ) (∵ for each n, there is a unique j such that n = σ(j))
N
n=1 n=1
= 0.
Similarly, we have (z, y − z) = 0. The result follows. □
 39

A family of an orthonormal vectors, say B, in X is said to be complete if it is maximal with

respect to the set inclusion order, i.e., if C is another family of orthonormal vectors with B ⊆ C,
then B = C.
A complete orthonormal subset of X is also called an orthonormal basis of X.

Proposition 14.8. Let {ei }i∈I be a family of orthonormal vectors in X. Then the followings are
equivalent:
(i): {ei }i∈I is complete;
(ii): if (x, ei ) = 0 for all i ∈ I, then
P x = 0;
(iii): for any x ∈ X, we have x = i∈I (x, ei )ei ;
(iv): for any x ∈ X, we have ∥x∥2 = i∈I |(x, ei )|2 .
P

Note : there are only countable many (x, ei ) ̸= 0 by Corollary 14.6, so the sums in (iii) and (iv)
are convergent by Proposition 14.7. In this case, the expression of each element x ∈ X in Part (iii)
is unique.

Proposition 14.9. Let X be a Hilbert space. Then

(i) : X processes an orthonormal basis.
(ii) : If {ei }i∈I and {fj }j∈J both are the orthonormal bases for X, then I and J have the same
cardinality. In this case, the cardinality |I| of I is called the orthonormal dimension of X.
Proof. Part (i) follows from Zorn’s Lemma.
For part (ii), if the cardinality |I| is finite, then the assertion is clear since |I| = dim X (vector
space dimension) in this case.
Now assume that |I| is infinite, for each ei , put Jei := {j ∈ J : (ei , fj ) ̸= 0}. Note that since {ei }i∈I
is maximal, Proposition 14.8 implies that we have
[
J= Jei .
i∈I
Note that Jei is countable for each ei by using Proposition 14.6. On the other hand, we have
|N| ≤ |I| because |I| is infinite and thus |N × I| = |I|. Then we have
X X
|J| ≤ |Jei | = |N| = |N × I| = |I|.
i∈I i∈I
Similarly, we also have |I| ≤ |J|. □

Remark 14.10. Recall that a vector space dimension of X is defined by the cardinality of a maximal
linearly independent set in X.
Note that if X is finite dimensional, then the orthonormal dimension is the same as the vector
space dimension.
In addition, the vector space dimension is larger than the orthornormal dimension in general since
every orthogonal set must be linearly independent.

Two Hilbert spaces X and Y are said to be isomorphic if there is linear isomorphism U from X
onto Y such that (U x, U x′ ) = (x, x′ ) for all x, x′ ∈ X. In this case U is called a unitary operator.

Theorem 14.11. Two Hilbert spaces are isomorphic if and only if they have the same orthonornmal
dimension.
Proof. The converse part (⇐) is clear.
Now for the (⇒) part, let X and Y be isomorphic Hilbert spaces. Let U : X −→ Y be a unitary.
Note that if {ei }i∈I is an orthonormal basis of X, then {U ei }i∈I is also an orthonormal basis of Y .
Thus the necessary part follows immediately from Proposition 14.9. □
40 CHI-WAI LEUNG

Corollary 14.12. Every separable Hilbert space is isomorphic to ℓ2 or Cn for some n.

Proof. Let X be a separable Hilbert space.
If dim X < ∞, then it is clear that X is isomorphic to Cn for n = dim X.
Now suppose that dim X = ∞ and its orthonormal dimension is larger √ than |N|, i.e., X has an
orthonormal basis {fi }i∈I with |I| > |N|. Note that since ∥fi − fj ∥ = 2 for all i, j ∈ I with i ̸= j.
This implies that B(fi , 1/4) ∩ B(fj , 1/4) = ∅ for i ̸= j.
On the other hand, if we let D be a countable dense subset of X, then B(fi , 1/4) ∩ D ̸= ∅ for all
i ∈ I. Thusfor each i ∈ I, we can pick up an element xi ∈ D ∩ B(fi , 1/4). Therefore, one can define
an injection from I into D. It is absurd to the countability of D. □

Example 14.13. The followings are important classes of Hilbert spaces.

(i) C n
Pn is a n-dimensional Hilbert space. In this case , nthe inner product is given by (z, w) :=
k=1 zk w k for z = (z1 , ..., zn ) and (w1 , ..., wn ) in C .
The natural basis {e1 , ..., en } forms an orthonormal basis for Cn .
(ii) ℓ2 is a separable Hilbert space of infinite dimension whose inner product is given by (x, y) :=
P∞ 2
n=1 x(n)y(n) for x, y ∈ ℓ .
If we put en (n) = 1 and en (k) = 0 for k ̸= n, then {en } is an orthonormal basis for ℓ2 .
(iii) Let T := {z ∈ C : |z| = 1}. For each f ∈ C(T) (the space of all complex-valued continuous
functions defined on T), the integral of f is defined by
Z Z 2π Z 2π Z 2π
1 1 i
f (z)dz := it
f (e )dt = Ref (e )dt +
it
Imf (eit )dt.
T 2π 0 2π 0 2π 0

An inner product on C(T) is given by

Z
(f, g) := f (z)g(z)dz
T

for each f, g ∈ C(T). We write ∥ · ∥2 for the norm induced by this inner product.
The Hilbert space L2 (T) is defined by the completion of C(T) under the norm ∥ · ∥2 .
Now for each n ∈ Z, put fn (z) = z n . We claim that {fn : n = 0, ±1, ±2, ....} is an orthonor-
mal basis for L2 (T).
In fact, by using the Euler Formula: eiθ = cos θ + i sin θ for θ ∈ R, we see that the family
{fn : n ∈ Z} is orthonormal.
It remains to show that the family {fn } is maximal. By Proposition 14.8, it needs to show
that if (g, fn ) = 0 for all n ∈ Z, then g = 0 in L2 (T). for showing this, we have to make
use the known fact that every element in L2 (T) can be approximated by the polynomial
functions of z and z̄ on T in ∥ · ∥2 -norm due to the the Stone-Weierstrass Theorem:

For a compact metric space E, suppose that a complex subalgebra A of C(E) satisfies the
conditions: (i): the conjugate f¯ ∈ A whenever f ∈ A, (ii): for every pair z, z ′ ∈ E, there is
f ∈ A such that f (z) ̸= f (z ′ ) and (iii): A contains the constant one function. Then A is
dense in C(E) with respect to the sup-norm.

Thus, the algebra of all polynomials functions of z and z̄ on T is dense in C(T). From
this we can find a sequence of polynomials (pn (z, z̄)) such that ∥g − pn ∥2 → 0 as n → 0.
Since (g, fn ) = 0 for all n, we see that (g, pn ) = 0 for all n. Therefore, we have
∥g∥22 = lim(g, pn ) = 0.
n

The proof is complete.

 41

Remark 14.14. In view of Example 14.13(iii) above, L2 (T) can be identified as the space
R 2π
L2 [0, 2π] := {f : [0, 2π] → C : f is a Lebesgue measure function so that 0 |f (x)|2 dx exists}.
1
R 2π
In this case, the inner product in L2 [0, 2π] is given by (f, g) := 2π 0 f (t)g(t)dt. Now for each
x ∈ [0, 2π] and f ∈ L2 [0, 2π], put ek (x) := eikx and
N N Z 2π
X 1 X
(14.1) SN (f, x) := (f, ek )ek (x) = ( f (t)e−ik dt)eikx .
2π 0
k=−N k=−N
∞
X
Then the series (f, ek )ek (x) (need not be convergent) is the usual notation of the Fourier series
−∞
of f at x.
N
X
In the previous result, we have seen that the series lim ∥f − (f, ek )ek ∥L2 → 0
N →∞
k=−N
It is naturally raised the question: for f ∈ C[0, 2π], do we have f (x) = lim SN (f, x) for all
N →∞
x ∈ [0, 2π], that is, f can be written as its Fourier series? The answer is negative. Indeed, we have
the following stronger result.
Theorem 14.15. For each element x0 ∈ [0, 2π], there is a continuous function f on [0, 2π] such
∞
X
that its Fourier series (f, ek )eikx0 at x0 is divergent.
k=−∞

Proof. We give an outline argument in here. The details of the proof is referred to the Katznelson’s
classic book [9].
We keep the notation as above. We fix a point x0 ∈ [0, 2π]. For f ∈ C[0, 2π], Using the Eq 14.1
above, we have
N Z 2π
1 X
φN (f ) := SN (f, x0 ) = ( f (t)e−ik dt)eikx0
2π 0
k=−N
Z 2π N
(14.2) 1 X
= f (t)( eik(x0 −t) )dt
2π 0 k=−N
Z 2π
1
= f (t)DN (x0 − t)dt,
2π 0
N
X
where DN (s) := eiks for s ∈ R. Then φN ∈ C[0, 2π]∗ . It is a fact that we have (see [9,
k=−N
Chapter 2. §2])
Z 2π
1
∥φN ∥ = |DN (x0 − t)|dt → ∞ as N → ∞.
2π 0
From this we have sup{∥φN ∥ : N = 1, 2....} = ∞. This implies that the following set is dense in
C[0, 2π].
D := {f ∈ C[0, 2π] : sup |φN (f )| = ∞}.
N
To see this, suppose not. Then there is an element f0 ∈ C[0, 2π] and r > 0 such that supN |φN (f )| <
∞ for all f ∈ B(f0 , r). By considering B(0, 1) = 1r (B(f0 , r) − f0 ), then we have supN |φN (g)| < ∞
for all g ∈ B(0, 1) and so, supN |φN (g)| < ∞ for all g ∈ C[0, 2π]. The Uniform Boundedness
Theorem implies that supN ∥φN ∥ < ∞ that leads to a contradiction.
42 CHI-WAI LEUNG

∞
X
On the other hand, it is clear that if the series (f, ek )eikx0 is convergent, then supN |SN (f, x0 )| <
k=−∞
∞. Therefore, the Fourier series of each element in D is divergent at x0 as desired. We finish the
proof. □

The above proof is an elegant application of functional analysis.

Students should be made to think, to doubt, to communicate, to question, to learn from their
mistakes, and most importantly have fun in their learning.
∼ Richard Feynman ∼ one of the greatest physicists in 20th century.

15. Geometry of Hilbert space II

In this section, let X always denote a complex Hilbert space.
Proposition 15.1. If D is a closed convex subset of X, then there is a unique element z ∈ D such
that
∥z∥ = inf{∥x∥ : x ∈ D}.
Consequently, for any element u ∈ X, there is a unique element w ∈ D such that
∥u − w∥ = d(u, D) := inf{∥u − x∥ : x ∈ D}.
Proof. We first claim the existence of such z.
Let d := inf{∥x∥ : x ∈ D}. Then there is a sequence (xn ) in D such that ∥xn ∥ → d. Note that (xn )
is a Cauchy sequence. In fact, the Parallelogram Law implies that
xm − xn 2 1 1 xm + xn 2 1 1
∥ ∥ = ∥xm ∥2 + ∥xn ∥2 − ∥ ∥ ≤ ∥xm ∥2 + ∥xn ∥2 − d2 −→ 0
2 2 2 2 2 2
as m, n → ∞, where the last inequality holds because D is convex and hence 12 (xm + xn ) ∈ D. Let
z := limn xn . Then ∥z∥ = d and z ∈ D because D is closed.
For the uniqueness, let z, z ′ ∈ D such that ∥z∥ = ∥z ′ ∥ = d. Thanks to the Parallelogram Law
again, we have
z − z′ 2 1 1 z + z′ 2 1 1
∥ ∥ = ∥z∥2 + ∥z ′ ∥2 − ∥ ∥ ≤ ∥z∥2 + ∥z ′ ∥2 − d2 = 0.
2 2 2 2 2 2
Therefore z = z .′

The last assertion follows by considering the closed convex set u−D := {u−x : x ∈ D} immediately.
□

Remark 15.2. Using the notation given as in Proposition 15.1, we have a well defined function
r : X → X given by x ∈ X 7→ r(x) ∈ D such that ∥x − r(x)∥ = dist(x, D). Clearly, we have
r(x) = x whenever x ∈ D. Moreover, we have the following assertion which are shown in [5].

Proposition 15.3. Using the notation as in Remark15.2, the map r : X → X is a contraction,

hence, the map r is a Lipschitz retraction of D in X.
Proof. We first claim that we have Re(x − r(z), r(z) − z) ≥ 0 for all x ∈ D and z ∈ X. In fact, let
z ∈ X and x ∈ D. Then by the definition, for all t ∈ [0, 1] we have
∥r(z) − z∥2 ≤ ∥z − tx − (1 − t)r(z)∥2
= ∥z − r(z) − t(x − r(z))∥2
= ∥z − r(z)∥2 + t2 ∥x − r(z)∥2 + 2tRe(x − r(z), r(z) − z).
 43

This gives t2 ∥x − r(z)∥2 + 2tRe(x − r(z), r(z) − z) ≥ 0 for all 0 ≤ t ≤ 1. This implies that Re(x −
r(z), r(z)−z) ≥ 0 for all x ∈ D and z ∈ X. From this, for a, b ∈ X we have Re(r(b)−r(a), r(a)−a) ≥
0 and Re(r(a)−r(b), r(b)−b) ≥ 0, so we have Re(r(b)−r(a), r(a)−a)+Re(r(b)−r(a), b−r(b)) ≥ 0.
Thus, we have
∥r(b) − r(a)∥2 = Re(r(b) − r(a), r(b) − r(a))
≤ Re(r(b) − r(a), b − a)
≤ |(r(b) − r(a), b − a)|
≤ ∥r(b) − r(a)∥∥b − a∥.
The proof is complete. □

Proposition 15.4. Suppose that M is a closed subspace. Let u ∈ X and w ∈ M . Then the
followings are equivalent:
(i): ∥u − w∥ = d(u, M );
(ii): u − w ⊥ M , i.e., (u − w, x) = 0 for all x ∈ M .
Consequently, for each element u ∈ X, there is a unique element w ∈ M such that u − w ⊥ M .
Proof. Let d := d(u, M ).
For proving (i) ⇒ (ii), fix an element x ∈ M . Then for any t > 0, note that since w + tx ∈ M , we
have
d2 ≤ ∥u − w − tx∥2 = ∥u − w∥2 + ∥tx∥2 − 2Re(u − w, tx) = d2 + ∥tx∥2 − 2Re(u − w, tx).
This implies that
(15.1) 2Re(u − w, x) ≤ t∥x∥2
for all t > 0 and for all x ∈ M . Thus by considering −x in Eq.15.1, we obtain
2|Re(u − w, x)| ≤ t∥x∥2 .
for all t > 0. This implies that Re(u − w, x) = 0 for all x ∈ M . Similarly, putting ±ix into Eq.15.1,
we have Im(u − w, x) = 0. Thus(ii) follows.
For (ii) ⇒ (i), we need to show that ∥u − w∥2 ≤ ∥u − x∥2 for all x ∈ M . Note that since u − w ⊥ M
and w ∈ M , we have u − w ⊥ w − x for all x ∈ M . This gives
∥u − x∥2 = ∥(u − w) + (w − x)∥2 = ∥u − w∥2 + ∥w − x∥2 ≥ ∥u − w∥2 .
Part (i) follows.
The last statement is obtained immediately by Proposition 15.1. □

Theorem 15.5. Let M be a closed subspace. Put

M ⊥ := {x ∈ X : x ⊥ M }.
Then M ⊥ is a closed subspace and we have X = M ⊕ M ⊥ . Consequently, for x ∈ X if x = u ⊕ v
for u ∈ M and v ∈ M ⊥ , then dist(x, M ) = ∥x − u∥.
In this case, M ⊥ is called the orthogonal complement of M .
Proof. Clearly, M ⊥ is a closed subspace and M ∩ M ⊥ = (0). We need to show X = M + M ⊥ .
Let u ∈ X. Then by Proposition 15.4, we can find an element w ∈ M such that u − w ⊥ M . Thus
u − w ∈ M ⊥ and u = w + (u − w).
The last assertion follows immediately from Proposition 15.4. The proof is complete. □

Corollary 15.6. Let M be a closed subspace of X. Then M ⊊ X if and only if there is a non-zero
element z ∈ X such that z ⊥ M .
Proof. It is clear from Theorem 15.5. □
44 CHI-WAI LEUNG

Corollary 15.7. If M is a closed subspace of X, then M ⊥⊥ = M .

Proof. Clearly, we have M ⊆ M ⊥⊥ by the definition of M ⊥⊥ . Then M can be viewed as a closed
subspace of the Hilbert space M ⊥⊥ . Thus, if M ⊊ M ⊥⊥ , then there exists a non-zero element
z ∈ M ⊥⊥ so that z ⊥ M by Corollary 15.6 and hence, z ∈ M ⊥ . This implies that z ⊥ z and hence,
z = 0 which leads to a contradiction. □

Theorem 15.8. Riesz Representation Theorem : For each f ∈ X ∗ , then there is a unique
element vf ∈ X such that
f (x) = (x, vf )
for all x ∈ X and we have ∥f ∥ = ∥vf ∥.
P
Furthermore, if (ei )i∈I is an orthonormal basis of X, then vf = i f (ei )ei .
Proof. We first prove the uniqueness of vf . If z ∈ X also satisfies the condition: f (x) = (x, z) for
all x ∈ X. This implies that (x, z − vf ) = 0 for all x ∈ X. Thusz − vf = 0.
Now for proving the existence of vf , it suffices to show the case ∥f ∥ = 1. Then ker f is a closed
proper subspace. Then by the orthogonal decomposition again, we have
X = ker f ⊕ (ker f )⊥ .
Since f ̸= 0, we have (ker f )⊥ is linear isomorphic to C. Note that the restriction of f on (ker f )⊥ is
of norm one. Hence there is an element vf ∈ (ker f )⊥ with ∥vf ∥ = 1 such that f (vf ) = ∥f |(ker f )⊥ ∥ =
1 and (ker f )⊥ = Cvf . Thusfor each element x ∈ X, we have x = z + αvf for some z ∈ ker f and
α ∈ C. Then f (x) = αf (vf ) = α = (x, vf ) for all x ∈ X.X
Concerning about the last assertion, if we put vf = αi ei , then f (ej ) = (ej , vf ) = αj for all
i∈I
j ∈ I. □

2 ∗
Example R 15.9. Consider the Hilbert space H := L (T) (see Example 14.13). Define φ ∈ H by
R element g ∈ H, there is an element h ∈ ker φ
φ(f ) := T f (z)dz. Using Proposition 15.4, for each
such that ∥g −h∥ = dist(g, ker φ). Then h = g −( gdz)1 where 1 denotes the constant-one function
on T. In fact, consider the orthogonal decomposition H = ker φ ⊕ (ker φ)⊥ . Note that φ(g) = (g, 1)
for all g ∈R H. Thus, for each g ∈ H, we have g = h ⊕ α1. From this, we see that α = (g, 1). Thus,
h = g − ( gdz)1.

Corollary 15.10. Using the notations as in Theorem 15.8, define the map
(15.2) Φ : f ∈ X ∗ 7→ vf ∈ X, i.e., f (y) = (x, Φ(f ))
for all y ∈ X and f ∈ X ∗ .
Moreover, if we define (f, g)X ∗ := (vg , vf )X for f, g ∈ X ∗ , then (X ∗ , (·, ·)X ∗ ) becomes a Hilbert
space, and Φ is an anti-unitary operator from X ∗ onto X, i.e., Φ satisfies the conditions:
Φ(αf + βg) = αΦ(f ) + βΦ(g) and (Φf, Φg)X = (g, f )X ∗
for all f, g ∈ X ∗ and α, β ∈ C.
Furthermore, if we define J : x ∈ X 7→ fx ∈ X ∗ , where fx (y) := (y, x), then J is the inverse of Φ,
and hence, J is an isometric conjugate linear isomorphism.
Proof. The result follows immediately from the observation that vf +g = vf + vg and vαf = αvf for
all f ∈ X ∗ and α ∈ C.
The last assertion is clearly obtained by the Eq.15.2 above. □

Corollary 15.11. Every Hilbert space is reflexive.

 45

Proof. Using the notations as in the Riesz Representation Theorem 15.8, let X be a Hilbert space.
and Q : X → X ∗∗ the canonical isometry. Let ψ ∈ X ∗∗ . To apply the Riesz Theorem on the dual
space X ∗ , there exists an element x∗0 ∈ X ∗ such that
ψ(f ) = (f, x∗0 )X ∗
for all f ∈ X ∗ . By using Corollary 15.10, there is an element x0 ∈ X such that x0 = vx∗0 and thus,
we have
ψ(f ) = (f, x∗0 )X ∗ = (x0 , vf )X = f (x0 )
for all f ∈ X ∗ . Therefore, ψ = Q(x0 ) and so, X is reflexive.
The proof is complete. □

Theorem 15.12. Every bounded sequence in a Hilbert space has a weakly convergent subsequence.
Proof. Let (xn ) be a bounded sequence in a Hilbert space X and M be the closed subspace of X
spanned by {xm : m = 1, 2...}. Then M is a separable Hilbert space.
Method I : Define a map by jM : x ∈ M 7→ jM (x) := (·, x) ∈ M ∗ . Then (jM (xn )) is a bounded
sequence in M ∗ . By Banach’s result, Proposition 6.9, (jM (xn )) has a w∗ -convergent subsequence
w∗
(jM (xnk )). Put jM (xnk ) −−→ f ∈ M ∗ , i.e., jM (xnk )(z) → f (z) for all z ∈ M . The Riesz Rep-
resentation will assure that there is a unique element m ∈ M such that jM (m) = f . Thuswe
have (z, xnk ) → (z, m) for all z ∈ M . In particular, if we consider the orthogonal decomposition
X = M ⊕ M ⊥ , then (x, xnk ) → (x, m) for all x ∈ X and thus (xnk , x) → (m, x) for all x ∈ X. Then
xnk → m weakly in X by using the Riesz Representation Theorem again.
Method II : We first note that since M is a separable Hilbert space, the second dual M ∗∗ is also
separable by the reflexivity of M . Thus, the dual space M ∗ is separable (see Proposition4.11).
Let Q : M −→ M ∗∗ be the natural canonical mapping. To apply the Banach’s result Propo-
sition 6.9 for X ∗ , then Q(xn ) has a w∗ -convergent subsequence, says Q(xnk ). This gives an
element m ∈ M such that Q(m) = w∗ -limk Q(xnk ) because M is reflexive. Thus, we have
f (xnk ) = Q(xnk )(f ) → Q(m)(f ) = f (m) for all f ∈ M ∗ . Using the same argument as in Method I
again, xnk weakly converges to m. □

Remark 15.13. Recall the well known James’s Theorem that a Banach space X is reflexive if
and only if every bounded sequence in X has a weakly convergent subsequence. (see Appendix 7
for the proof of a separable space case). Hence, Theorem 15.12 can be obtained by the James’s
Theorem directly. However, the Riesz Representation Theorem gives a simple proof for the Hilbert
spaces case.

16. Operators on a Hilbert space

Throughout this section, all spaces are complex Hilbert spaces. Let B(X, Y ) denote the space
of all bounded linear operators from X into Y . If X = Y , we write B(X) for B(X, X).
Let T ∈ B(X, Y ). We make use the following simple observation later.

(16.1) (T x, y) = 0 for all x ∈ X; y ∈ Y if and only if T = 0.

Therefore, the elements in B(X, Y ) are uniquely determined by the Eq.16.1, i.e., T = S in B(X, Y )
if and only if (T x, y) = (Sx, y) for all x ∈ X and y ∈ Y .

Remark 16.1. For Hilbert spaces H1 and H2 , we consider their direct sum H := H1 ⊕ H2 . If we
define the inner product on H by
(x1 ⊕ x2 , y1 ⊕ y2 ) := (x1 , y1 )H1 + (x2 , y2 )H2
46 CHI-WAI LEUNG

for x1 ⊕ x2 and y1 ⊕ y2 in H, then H becomes a Hilbert space. Now for each T ∈ B(H1 , H2 ), we
can define an element T̃ ∈ B(H) by T̃ (x1 ⊕ x2 ) := 0 ⊕ T x1 . Therefore, the space B(H1 , H2 ) can be
viewed as a closed subspace of B(H). Thus, we can consider the case of H1 = H2 for studying the
space B(H1 , H2 ).

Proposition 16.2. Let T : X → X be a linear operator. Then we have

(i): T = 0 if and only if (T x, x) = 0 for all x ∈ X. Consequently, for T, S ∈ B(X), T = S if
and only if (T x, x) = (Sx, x) for all x ∈ X.
(ii): T is bounded if and only if sup{|(T x, y)| : x, y ∈ X with ∥x∥ = ∥y∥ = 1} is finite. In this
case, we have ∥T ∥ = sup{|(T x, y)| : x, y ∈ X with ∥x∥ = ∥y∥ = 1}.
Proof. Clearly, the necessary part holds in Part (i). We want to show the sufficient part in Part
(i). We assume that (T x, x) = 0 for all x ∈ X. Then we have
0 = (T (x + iy), x + iy) = (T x, x) + i(T y, x) − i(T x, y) + (T iy, iy) = i(T y, x) − i(T x, y).
Thus, we have (T y, x) − (T x, y) = 0 for all x, y ∈ X. In particular, if we replace y by iy in the
equation, then we get i(T y, x) − i(T x, y) = 0 and hence we have (T y, x) + (T x, y) = 0. Therefore
we have (T x, y) = 0.
For showing part (ii), let α = sup{|(T x, y)| : x, y ∈ X with ∥x∥ = ∥y∥ = 1}. It suffices to show
∥T ∥ = α. Clearly, we have ∥T ∥ ≥ α. We need to show ∥T ∥ ≤ α.
In fact, let x ∈
X with ∥x∥ = 1. If T x ̸= 0, then we take y = T x/∥T x∥. Thus, we have
∥T x∥ = | T x, y | ≤ α, and so ∥T ∥ ≤ α. The proof is complete. □

Proposition 16.3. Let T ∈ B(X). Then there is a unique element T ∗ in B(X) such that
(16.2) (T x, y) = (x, T ∗ y)
In this case, T ∗ is called the adjoint operator of T .
Proof. First, we show the uniqueness. Suppose that there are S1 , S2 in B(X) which satisfy the
Eq.16.2. Then (x, S1 y) = (x, S2 y) for all x, y ∈ X. Eq.16.1 implies that S1 = S2 .
Finally, we prove the existence. Note that if we fix an element y ∈ X, define the map fy (x) :=
(T x, y) for all x ∈ X. Then fy ∈ X ∗ . By applying the Riesz Representation Theorem, there is a
unique element y ∗ ∈ X such that (T x, y) = (x, y ∗ ) for all x ∈ X and ∥fy ∥ = ∥y ∗ ∥. In addition, we
have
|fy (x)| = |(T x, y)| ≤ ∥T ∥∥x∥∥y∥
for all x, y ∈ X and thus ∥fy ∥ ≤ ∥T ∥∥y∥. If we put T ∗ (y) := y ∗ , then T ∗ satisfies the Eq.16.2.
Moreover, we have ∥T ∗ y∥ = ∥y ∗ ∥ = ∥fy ∥ ≤ ∥T ∥∥y∥ for all y ∈ X. Thus, we have T ∗ ∈ B(X) and
∥T ∗ ∥ ≤ ∥T ∥. Hence the operator T ∗ is as desired. □

Remark 16.4. Let S, T : X → X be linear operators (without assuming to be bounded). If they

satisfy the Eq.16.2 above, i.e.,
(T x, y) = (x, Sy)
for all x, y ∈ X. Using the Closed Graph Theorem, we can show that S and T both are automatically
bounded.
In fact, let (xn ) be a sequence in X such that lim xn = x and lim Sxn = y for some x, y ∈ X. Now
for any z ∈ X, we have
(z, Sx) = (T z, x) = lim(T z, xn ) = lim(z, Sxn ) = (z, y).
Thus Sx = y and hence S is bounded by the Closed Graph Theorem.
Similarly, we can also see that T is bounded.
 47

Remark 16.5. Let T ∈ B(X). Let T t : X ∗ → X ∗ be the transpose of T which is defined by

T t (f ) := f ◦ T ∈ X ∗ for f ∈ X ∗ (see Proposition 4.13). Then we have the following commutative
diagram (Check!)
T∗
X −−−−→ X
 

JX y
J
y X
Tt
X ∗ −−−−→ X ∗
where JX : X → X ∗ is the anti-unitary given by the Riesz Representation Theorem (see Corollary
15.10).

Proposition 16.6. Let T, S ∈ B(X). Then we have

(i): T ∗ ∈ B(X) and ∥T ∗ ∥ = ∥T ∥.
(ii): The map T ∈ B(X) 7→ T ∗ ∈ B(X) is an isometric conjugate anti-isomorphism, i.e.,

(αT + βS)∗ = αT ∗ + βS ∗ for all α, β ∈ C; and (T S)∗ = S ∗ T ∗ .

(iii): ∥T ∗ T ∥ = ∥T ∥2 .
Proof. For Part (i), in the proof of Proposition 16.3, we have shown that ∥T ∗ ∥ ≤ ∥T ∥. In addition,
the reverse inequality follows clearly from T ∗∗ = T .
The Part (ii) follows from the adjoint operators which are uniquely determined by the Eq.16.2
above.
For Part (iii), we always have ∥T ∗ T ∥ ≤ ∥T ∗ ∥∥T ∥ = ∥T ∥2 . For the reverse inequality, let x ∈ BX .
Then
∥T x∥2 = (T x, T x) = (T ∗ T x, x) ≤ ∥T ∗ T x∥∥x∥ ≤ ∥T ∗ T ∥.
Therefore, we have ∥T ∥2 ≤ ∥T ∗ T ∥. □

Example 16.7. If X = Cn and D = (aij )n×n an n × n matrix, then D∗ = (aji )n×n . In fact, note
that
aji = (Dei , ej ) = (ei , D∗ ej ) = (D∗ ej , ei ).
Thusif we put D∗ = (dij )n×n , then dij = (D∗ ej , ei ) = aji .

∞
P∞ X
Example 16.8. Let ℓ2 (N) := {x : N → C : 2
i=0 |x(i)| < ∞}, and put (x, y) := x(i)y(i).
i=0
Define the operator D ∈ B(ℓ2 (N)) (called the unilateral shift) by
Dx(i) = x(i − 1)
for i ∈ N, where we set x(−1) := 0, i.e., D(x(0), x(1), ...) = (0, x(0), x(1), ....).
Then D is an isometry and the adjoint operator D∗ is given by
D∗ x(i) := x(i + 1)
for i = 0, 1, .., i.e., D∗ (x(0), x(1), ...) = (x(1), x(2), ....).
Indeed we can directly check that
∞
X ∞
X
(Dx, y) = x(i − 1)y(i) = x(j)y(j + 1) = (x, D∗ y).
i=0 j=0

Note that D∗ is NOT an isometry.

48 CHI-WAI LEUNG

Example 16.9. Let ℓ∞ (N) = {x : N → C : supi≥0 |x(i)| < ∞} and ∥x∥∞ := supi≥0 |x(i)|. For
each x ∈ ℓ∞ , define Mx ∈ B(ℓ2 (N)) by
Mx (ξ) := x · ξ
for ξ ∈ ℓ2 (N),
where (x · ξ)(i) := x(i)ξ(i); i ∈ N.
Then ∥Mx ∥ = ∥x∥∞ and Mx∗ = Mx , where x(i) := x(i).

Definition 16.10. Let T ∈ B(X) and let I be the identity operator on X. T is said to be
(i) : selfadjoint if T ∗ = T ;
(ii) : normal if T ∗ T = T T ∗ ;
(iii) : unitary if T ∗ T = T T ∗ = I.

Proposition 16.11. We have

(i) : Let T : X −→ X be a linear operator. T is a bounded linear selfadjoint operator if and
only if we have
(16.3) (T x, y) = (x, T y) for all x, y ∈ X.
(ii) : T is normal if and only if ∥T x∥ = ∥T ∗ x∥ for all x ∈ X.
Proof. The necessary part of Part (i) is clear.
Now suppose that the Eq.16.3 holds, it needs to show that T is bounded. Indeed, it follows
immediately from Remark16.4.
For Part (ii), note that by Proposition 16.2, T is normal if and only if (T ∗ T x, x) = (T T ∗ x, x).
Thus, Part (ii) follows from
∥T x∥2 = (T x, T x) = (T ∗ T x, x) = (T T ∗ x, x) = (T ∗ x, T ∗ x) = ∥T ∗ x∥2
for all x ∈ X. □

Remark 16.12. In Proposition 16.11(i), if the domain of T is replaced P∞ by dense domain, then
the conclusion does not hold. For example, let D := {x ∈ ℓ2 : n=1 |nx(n)|2 < ∞} and let

T (x)(n) := nx(n) for x ∈ D. Then D is a dense domain because the canonical basis (en ) ⊆ D. It
is noted that T is unbounded on D, but (T x, y) = (x, T y) for all x, y ∈ D.

Proposition 16.13. Let T ∈ B(H). We have the following assertions.

(i) : T is selfadjoint if and only if (T x, x) ∈ R for all x ∈ H.
(ii) : If T is selfadjoint, then ∥T ∥ = sup{|(T x, x)| : x ∈ H with ∥x∥ = 1}.
Proof. Part (i) follows immediately from Proposition16.2.
For Part (ii), if we let a = sup{|(T x, x)| : x ∈ H with ∥x∥ = 1}, then we have a ≤ ∥T ∥. We want
to show the reverse inequality. T is selfadjoint, and so we can directly check that
(T (x + y), x + y) − (T (x − y), x − y) = 4Re(T x, y)
for all x, y ∈ H. Thus if x, y ∈ H with ∥x∥ = ∥y∥ = 1 and (T x, y) ∈ R, then by using the
Parallelogram Law, we have
a a
(16.4) |(T x, y)| ≤ (∥x + y∥2 + ∥x − y∥2 ) = (∥x∥2 + ∥y∥2 ) = a.
4 2
Now for x, y ∈ H with ∥x∥ = ∥y∥ = 1, by considering the polar form of (T x, y) = reiθ , the Eq.16.4
gives
|(T x, y)| = |(T x, eiθ y)| ≤ a.
∥T ∥ = sup |(T x, y)|, and so we have ∥T ∥ ≤ a. The proof is complete. □
∥x∥=∥y∥=1
 49

Proposition 16.14. Let T ∈ B(X). Then we have

ker T = (imT ∗ )⊥ and (ker T )⊥ = imT ∗
where imT denotes the image of T .
Proof. The first equality follows clearly from x ∈ ker T if and only if 0 = (T x, z) = (x, T ∗ z) for all
z ∈ X.
⊥
On the other hand, it is clear that we have M ⊥ = M for any subspace M of X. This, together
with the first equality and Corollary15.7, gives immediately the second equality. □

Proposition 16.15. Let X be a Hilbert space. Let M and N be the closed subspaces of X such
that
X =M ⊕N . . . . . . . . . . . . (∗)
Let Q : X → X be the projection along the decomposition (∗) with im Q = M (note that Q is
bounded by Proposition 12.1). Then N = M ⊥ (and hence (∗) is the orthogonal decomposition of X
with respect to M ) if and only if Q satisfies the conditions: Q2 = Q and Q∗ = Q. In this case, Q
is called the orthogonal projection (or projection for simply) with respect to M .
Proof. Now if N = M ⊥ , then for y, y ′ ∈ M and z, z ′ ∈ N , we have
(Q(y + z), y ′ + z ′ ) = (y, y ′ ) = (y + z, Q(y ′ + z ′ )).
ThusQ∗ = Q.
The converse of the last statement follows immediately from Proposition 16.14 because ker Q = N
and imQ = M .
The proof is complete. □

Proposition 16.16. When X is a Hilbert space, we put M the set of all closed subspaces of X and
P the set of all orthogonal projections on X. Now for each M ∈ M, let PM be the corresponding
projection with respect to the orthogonal decomposition X = M ⊕ M ⊥ . Then there is an one-one
correspondence between M and P which is defined by
M ∈ M 7→ PM ∈ P.
Furthermore, if M, N ∈ M, then we have
(i) : M ⊆ N if and only if PM PN = PN PM = PM .
(ii) : M ⊥N if and only if PM PN = PN PM = 0.
Proof. Using Proposition 16.15, we note that PM ∈ P.
Indeed the inverse of the correspondence is given by the following. If we let Q ∈ P and M = Q(X),
then M is closed. In addition, clearly we have X = Q(X) ⊕ (I − Q)X with M ⊥ = (I − Q)X. Hence
M is the corresponding closed subspace of X, i.e., M ∈ M and PM = Q.
For the final assertion, Part (i) and (ii) follow immediately from the orthogonal decompositions
X = M ⊕ M ⊥ = N ⊕ N ⊥ and together with the fact that M ⊆ N if and only if N ⊥ ⊆ M ⊥ . □

17. Spectral Theory I

Definition 17.1. Let E be a normed space and let T ∈ B(E). The spectrum of T , denoted by
σ(T ), is defined by
σ(T ) := {λ ∈ C : T − λI is not invertible in B(E)}.
50 CHI-WAI LEUNG

Remark 17.2. More precisely, for a normed space E, an operator T ∈ B(E) is said to be invertible
in B(E) if T is an linear isomorphism and the inverse T −1 is also bounded. However, if E is
complete, the Open Mapping Theorem assures that the inverse T −1 is bounded automatically. Thus
if E is a Banach space and T ∈ B(E), then λ ∈ / σ(T ) if and only if T − λ := T − λI is an linear
isomorphism. Thus, λ lies in the spectrum σ(T ) if and only if T − λ is either not one-one or not
surjective.
In particular, if there is a non-zero element v ∈ X such that T v = λv, then λ ∈ σ(T ) and λ is
called an eigenvalue of T with eigenvector v.
In addition, we write σp (T ) for the set of all eigenvalue of T and call σp (T ) the point spectrum.

Example 17.3. Let E = Cn and T = (aij )n×n ∈ Mn (C). Then λ ∈ σ(T ) if and only if λ is an
eigenvalue of T and thus σ(T ) = σp (T ).

Example 17.4. Let E = (c00 (N), ∥ · ∥∞ ) (note that c00 (N) is not a Banach space). Define the
map T : c00 (N) → c00 (N) by
x(k)
T x(k) :=
k+1
for x ∈ c00 (N) and i ∈ N.
Then T is bounded, in fact, ∥T x∥∞ ≤ ∥x∥∞ for all x ∈ c00 (N).
On the other hand, we note that if λ ∈ C and x ∈ c00 (N), then
1
(T − λ)x(k) = ( − λ)x(k).
k+1
From this we see that σp (T ) = {1, 21 , 31 , ...}. In addition, if λ ∈ / {1, 12 , 13 , ...}, then T − λ is an linear
isomorphism and its inverse is given by
1
(T − λ)−1 x(k) = ( − λ)−1 x(k).
k+1
Thus, (T − λ)−1 is unbounded if λ = 0,so 0 ∈ σ(T ).
Besides, if λ ∈/ {0, 1, 21 , 31 , ...}, then (T − λ)−1 is bounded. In fact, if λ = a + ib ̸= 0, for a, b ∈ R,
1
then η := min | − a|2 + |b|2 > 0 because λ ∈ / {1, 12 , 13 , ...}. This gives
k 1+k
1
∥(T − λ)−1 ∥ = sup |( − λ)−1 | < η −1 < ∞.
k∈N k + 1

We can now conclude that σ(T ) = {1, 21 , 13 , ...} ∪ {0}.

Proposition 17.5. Let E be a Banach space and T ∈ B(E). Then

(i) : I − T is invertible in B(E) whenever ∥T ∥ < 1.
(ii) : If |λ| > ∥T ∥, then λ ∈
/ σ(T ).
(iii) : σ(T ) is a compact subset of C.
(iv) : If we let GL(E) the set of all invertible elements in B(E), then GL(E) is an open subset
of B(E) with respect to the ∥ · ∥-topology. Moreover, the map T ∈ GL(E) 7→ T −1 ∈ GL(E)
is continuous in the norm-topology.
Proof. Note that since B(E) is complete, Part (i) follows immediately from the following equality.
(I − T )(I + T + T 2 + · · · · · · + T N −1 ) = I − T N
for all N ∈ N.
For Part (ii), if |λ| > ∥T ∥, then by Part (i) , we see that I − λ1 T is invertible and so is λI − T .
This implies λ ∈/ σ(T ).
For Part (iii), since σ(T ) is bounded by Part (ii), we need to show that σ(T ) is closed.
 51

Let c ∈ C \ σ(T ). We need to find r > 0 such that µ ∈ / σ(T ) as |µ − c| < r. Note that since T − c
is invertible, then for µ ∈ C, we have T − µ = (T − c) − (µ − c) = (T − c)(I − (µ − c)(T − c)−1 ).
Therefore, if ∥(µ − c)(T − c)−1 )∥ < 1, then T − µ is invertible by Part (i). Thus, if we take
1
0<r< , then r is as desired, i.e., B(c, r) ⊆ C \ σ(T ). Hence σ(T ) is closed.
∥(T − c)−1 ∥
For the last assertion, let T ∈ GL(E). Note that for any S ∈ B(E), we have S = S − T + T =
T (1 − T −1 (T − S)). Thus, if 1 − T −1 (T − S) is invertible, then so is S. Using Part (i), if
∥T − S∥ < 1/∥T −1 ∥, then 1 − T −1 (T − S) is invertible. Therefore we have B(T, ∥T 1−1 ∥ ) ⊆ GL(E).
Finally, we show the inverse map is continuous. It suffices to show that if (Tn ) is Pa sequence in
GL(E) so that Tn → I, then Tn−1 → 1. Note that if ∥Tn − 1∥ < 1/2, then Tn−1 = ∞ k
k=0 (1 − Tn ) ,
hence, we may assume that (Tn−1 ) is uniformly bounded by 2. Therefore,
∥Tn−1 − 1∥ ≤ ∥Tn−1 ∥∥Tn − 1∥ ≤ 2∥Tn − 1∥.
The proof is complete. □
Corollary 17.6. If U is a unitary operator on a Hilbert space X, then σ(U ) ⊆ {λ ∈ C : |λ| = 1}.
Proof. Since ∥U ∥ = 1, we have σ(U ) ⊆ {λ ∈ C : |λ| ≤ 1} by Proposition 17.5(ii).
Now if |λ| < 1, then ∥λU ∗ ∥ < 1. By using Proposition 17.5 again, we have I − λU ∗ is invertible.
This implies that U − λ = U (I − λU ∗ ) is invertible and thus λ ∈
/ σ(U ). □

Example 17.7. Let E = ℓ2 (N) and let D ∈ B(E) be the right unilateral shift operator as in
Example16.8. Recall that Dx(k) := x(k − 1) for k ∈ N and x(−1) := 0. Then σp (D) = ∅ and
σ(D) = {λ ∈ C : |λ| ≤ 1}.
We first claim that σp (D) = ∅.
Suppose that λ ∈ C and x ∈ ℓ2 (N) satisfy the equation Dx = λx. Then by the definition of D, we
have
x(k − 1) = λx(k) · · · · · · · · · (∗)
for all k ∈ N.
If λ ̸= 0, then we have x(k) = λ−1 xk−1 for all k ∈ N. Since x(−1) = 0, this forces x(k) = 0 for all
i, i.e., x = 0 in ℓ2 (N).
On the other hand if λ = 0, the Eq.(∗) gives x(k − 1) = 0 for all k and so x = 0 again.
Therefore σp (D) = ∅.
Finally, we are going to show σ(D) = {λ ∈ C : |λ| ≤ 1}.
Note that since D is an isometry, ∥D∥ = 1. Proposition 17.5 tells us that
σ(D) ⊆ {λ ∈ C : |λ| ≤ 1}.
Note that since σp (D) is empty, it suffices to show that D − µ is not surjective for all µ ∈ C with
|µ| ≤ 1.
Now suppose that there is λ ∈ C with |λ| ≤ 1 such that D − λ is surjective.
We consider the case where |λ| = 1 first.
Let e1 = (1, 0, 0, ...) ∈ ℓ2 (N). Then by the assumption, there is x ∈ ℓ2 (N) such that (D − λ)x = e1
and thus Dx = λx + e1 . This implies that
x(k − 1) = Dx(k) = λx(k) + e1 (k)
for all k ∈ N. From this we have x(0) = −λ−1 and x(k) = −λ−k x(0) for all k ≥ 1 because e1 (0) = 1
and e1 (k) = 0 for all k ≥ 1. Moreover, since |λ| = 1, it turns out that |x(0)| = |x(k)| for all k ≥ 1.
As x ∈ ℓ2 (N), this forces x = 0. However, it is absurd because Dx = λx + e1 .
Now we consider the case where |λ| < 1.
By Proposition 16.14, we have
⊥
im(D − λ) = ker(D − λ)∗ = ker(D∗ − λ).
52 CHI-WAI LEUNG

Thus if D − λ is surjective, we have ker(D∗ − λ) = (0) and hence λ ∈ / σp (D∗ ).

∗
Note that the adjoint D of D is given by the left shift operator, i.e.,
D∗ x(k) = x(k + 1) · · · · · · · · · (∗∗)
for all k ∈ N.
Now when D∗ x = µx for some µ ∈ C and x ∈ ℓ2 (N), by using Eq.(∗∗), which is equivalent to saying
that
x(k + 1) = µx(k)
k
for all k ∈ N. Therefore, if |λ| = |λ| < 1 and we set x(0) = 1 and x(k + 1) = λ x(0) for all k ≥ 1,
then x ∈ ℓ2 (N) and D∗ x = λx. Hence λ ∈ σp (D∗ ) which leads to a contradiction. The proof is
complete.

18. Spectral Theory II

Throughout this section, let H be a complex Hilbert space.

Lemma 18.1. Let T ∈ B(H) be a normal operator (recall that T ∗ T = T T ∗ ). Then T is invertible
in B(H) if and only if there is c > 0 such that ∥T x∥ ≥ c∥x∥ for all x ∈ H.
Proof. The necessary part is obvious.
Now we want to show the converse. We first show the case where T is selfadjoint. Clearly, T
is injective from the assumption. By the Open Mapping Theorem, we need to show that T is
surjective.
⊥
In fact since ker T = imT ∗ and T = T ∗ , we see that the image of T is dense in H.
Now if y ∈ H, then there is a sequence (xn ) in H such that T xn → y. Thus, (T xn ) is a Cauchy
sequence. From this and the assumption give us that (xn ) is also a Cauchy sequence. If xn converges
to x ∈ H, then y = T x. Therefore the assertion is true when T is selfadjoint.
Now if T is normal, then we have ∥T ∗ x∥ = ∥T x∥ ≥ c∥x∥ for all x ∈ H by Proposition 16.11(ii).
Therefore, we have ∥T ∗ T x∥ ≥ c∥T x∥ ≥ c2 ∥x∥. Hence T ∗ T still satisfies the assumption. Note that
T ∗ T is selfadjoint. Therefore, we can apply the previous case to know that T ∗ T is invertible. This
implies that T is also invertible because T ∗ T = T T ∗ .
The proof is complete. □

Definition 18.2. Let T ∈ B(H). We say that T is positive, denoted by T ≥ 0, if (T x, x) ≥ 0 for

all x ∈ H. For a pair of selfadjoint operators S and T , we say that S ≤ T if T − S ≥ 0.

Remark 18.3. Clearly, a positive operator is selfadjoint by Proposition 16.13.

In particular, all projections are positive.

Proposition 18.4. If T is an invertible operator in B(H), then the inverse T −1 of T belongs to

the closed ∗-subalgebra of B(H) generated by T and I.
Proof. Put S := T ∗ T . Then S is invertible in B(H). Now we may assume that ∥S∥ ≤ 1. Lemma
18.1 gives c > 0 such that (x, x) ≥ (S 2 x, x) ≥ c(x, x) for all x ∈ H. We choose a positive integer N
such that N c ≥ 1. Then we have
1 1 1
(x, x) ≥ (x, x) ≥ (S 2 x, x) ≥ (x, x)
Nc Nc N
for all x ∈ H. Thus, we have
S2 1
0≤I− ≤ I − I < I.
Nc N
 53

S2
If we let R := I − Nc , then (I − R)−1 exists in B(H) and hence we have
∞
S 2 −1 X S2 n
( ) = (I − R)−1 = (I − ) .
Nc Nc
n=0
Then the result follows from
∞
1 X (T ∗ T )2 n ∗ ∗
T −1 = (I − ) T TT .
Nc Nc
n=0
□

Proposition 18.5. Let T ∈ B(H). We have

(i) : If T ≥ 0, then T + I is invertible.
(ii) : If T is self-adjoint, then σ(T ) ⊆ R. In particular, if T ≥ 0, we have σ(T ) ⊆ [0, ∞).
Proof. For Part (i), we assume that T ≥ 0. This implies that
∥(I + T )x∥2 = ∥x∥2 + ∥T x∥2 + 2(T x, x) ≥ ∥x∥2
for all x ∈ H. Thus, the invertibility of I + T follows from Lemma 18.1.
For Part (ii), we first claim that T +i is invertible. Indeed, it follows immediately from (T +i)∗ (T +
i) = T 2 + I and Part (i).
Now if λ = a + ib where a, b ∈ R with b ̸= 0, then T − λ = −b( −1 b (T − a) + i) is invertible because
−1
b (T − a) is selfadjoint. Thus, σ(T ) ⊆ R.
Finally we want to show σ(T ) ⊆ [0, ∞) when T ≥ 0. Note that since σ(T ) ⊆ R, it suffices to show
that T − c is invertible if c < 0. Indeed, if c < 0, then we see that T − c = −c(I + ( −1 c T )) is
−1
invertible by the previous assertion because c T ≥ 0.
The proof is complete. □

Remark 18.6. In Proposition 18.5, we have shown that if T is selfadjoint, then σ(T ) ⊆ R. How-
ever, the converse does not hold. For example, consider H = C2 and
 
0 1
T = .
0 0

Example 18.7. Notice that the multiplication defines an isometry M : x ∈ ℓ∞ 7→ M (x) ∈ B(ℓ2 )
by M (x)(ξ)(n) := x(n)ξ(n); n = 1, 2... for ξ ∈ ℓ2 . Then M (x̄) = M (x)∗ for x ∈ ℓ∞ , and so, M (x)
is self-adjoint if and only if x is a R-sequence. Now let x ∈ ℓ∞ be a R-sequence. For simply for
each element x ∈ ℓ∞ , we also write x for M (x) as an element in B(ℓ2 ).
Now we claim that if x ∈ ℓ∞ is self-adjoint, then λ ∈ σ(x) if and only if inf |x(n) − λ| = 0.
n
Consequently, σp (x) = {xn : n = 1, 2...} and σ(x) = {x(n) : n = 1, 2....}.
In fact, for showing (⇐), let λ ∈ R such that inf |x(n) − λ| = 0. If x − λ is invertible in B(ℓ2 ), then
n
by Lemma 18.1, there is c > 0 such that ∥(x − λ)ξ∥ ≥ c for all ξ ∈ ℓ2 of norm one. In particular,
for each n = 1, 2..., we have |x(n) − λ| = ∥(x − λ)(en )∥ ≥ c > 0. It leads to a contradiction.
For showing (⇒), let λ ∈ R such that c := inf |x(n) − λ| > 0. Then x(n) ̸= λ for all n = 1, 2...
n
This implies that x − λ is injective. On the other hand, for any η ∈ ℓ2 , if (x(n) − λ)ξ(n) = η(n)
η(n)
for all n, then we have ξ(n) = x(n)−λ and so, |ξ(n)| ≤ |η(n)|
c . This gives ξ ∈ ℓ2 . Therefore, x − λ is
surjective and thus, x − λ is invertible. Hence, λ ∈/ σ(x).
From this, the last assertion follows because λ ∈ σ(x) if and only if λ = xn for some n or there is
a subsequence (xnk ) of (xn ) that converges to λ.
54 CHI-WAI LEUNG

Theorem 18.8. Let T ∈ B(H) be a selfadjoint operator. Put

M (T ) := sup (T x, x) and m(T ) = inf (T x, x).
∥x∥=1 ∥x∥=1

For convenience, we also write M = M (T ) and m = m(T ) if there is no confusion.

Then we have
(i) : ∥T ∥ = max{|m|, |M |}.
(ii) : {m, M } ⊆ σ(T ).
(iii) : σ(T ) ⊆ [m, M ].
Proof. Note that m and M are well defined because (T x, x) is real for all x ∈ H by Proposition 16.13
(ii). To see this part (i), since (T x, x) ≤ ∥T ∥ for all x ∈ H with ∥x∥ = 1, we have M (T ) ≤ ∥T ∥.
Hence we also have −m(T ) = M (−T ) ≤ ∥ − T ∥ = ∥T ∥ and so −∥T ∥ ≤ m(T ). Then we have
−∥T ∥ ≤ m(T ) ≤ M (T ) ≤ ∥T ∥. This gives max(|m(T )|, |M (T )|) ≤ ∥T ∥. Now if we consider the
case |m| ≤ |M | first, then we have −|M | ≤ −|m| ≤ m ≤ (T x, x) ≤ M ≤ |M | for all x ∈ H with
∥x∥ = 1. Thus, by using Lemma 16.13 (ii) again, we have ∥T ∥ ≤ |M | = max(|m|, |M |). We will
complete the proof of Part (i) if we consider (−T ) for the case |M (T )| ≤ |m(T )|.
For Part (ii), we first claim that M ∈ σ(T ) if T ≥ 0. Note that 0 ≤ m ≤ M = ∥T ∥ in this
case by Lemma 16.13. Then there is a sequence (xn ) in H with ∥xn ∥ = 1 for all n such that
(T xn , xn ) → M = ∥T ∥. Then we have
∥(T − M )xn ∥2 = ∥T xn ∥2 + M 2 ∥xn ∥2 − 2M (T xn , xn ) ≤ ∥T ∥2 + M 2 − 2M (T xn , xn ) → 0.
Hence, by Lemma 18.1 we have shown that T − M is not invertible and hence M ∈ σ(T ) if T ≥ 0.
Now for any selfadjoint operator T if we consider T − m, then T − m ≥ 0. Thus we have M − m =
M (T − m) ∈ σ(T − m) by the previous case. Clearly, we have σ(T − c) = σ(T ) − c for all c ∈ C.
Therefore we have M ∈ σ(T ) for any self-adjoint operator.
We claim that that m(T ) ∈ σ(T ). Note that M (−T ) = −m(T ). Thus, we have −m(T ) ∈ σ(−T ).
It is clear that σ(−T ) = −σ(T ). Then m(T ) ∈ σ(T ).
Finally, we want to show σ(T ) ⊆ [m, M ].
Indeed, since T − m ≥ 0, then by Proposition 18.5, we have σ(T ) − m = σ(T − m) ⊆ [0, ∞). This
gives σ(T ) ⊆ [m, ∞).
On the other hand, we consider M − T ≥ 0. Then we get M − σ(T ) = σ(M − T ) ⊆ [0, ∞). This
implies that σ(T ) ⊆ (−∞, M ]. The proof is complete. □

19. Appendix: σ(T ) ̸= ∅

Let X be a complex Banach space. In this appendix, we will show that the spectrum σ(T ) is
non-empty for any T ∈ B(X).
First we recall some basic result in Complex Analysis. Students can refer to any standard text
book of Complex Analysis, see for example [1].
A function g : C → C is called an entire function if g is differentiable on C, i.e., the following limit
exists for all c ∈ C
g(z) − g(c)
g ′ (c) := lim .
z→c z−c
The following result is one of important properties of entire functions (see [1, p.122]).

Theorem 19.1. Liouville’s Theorem Every bounded entire function is a constant function.

Theorem 19.2. Using the notion as before, let T ∈ B(X). Then the spectrum σ(T ) ̸= ∅.
Proof. Assume that σ(T ) = ∅. Fix f ∈ B(X)∗ , define the map g(z) := f ((z − T )−1 ) is defined
for all z ∈ C. Note that g is continuous on C by considering the composition λ ∈ C 7→ λ − T 7→
 55

(λ − T )−1 ∈ B(X) and using Proposition 17.5 (iv). Moreover, we have limz→∞ |g(z)| = 0. Thus, g
is a bounded function on C. On the other hand, if we fix a point c ∈ C, then we see that
g(z) − g(c)
lim = −f ((c − T )−2 ).
z→c z−c
Therefore, g is a bounded entire function. By the Liouville’s Theorem, f ((z − T )−1 ) is a constant
function on C. Then the Hahn-Banach Theorem implies that the function z ∈ C 7→ (z − T )−1 ∈
B(X) is constant on C. It leads to a contradiction. □

20. Appendix: Existence of the square root of a positive operator

This section is based on the note of the course Functional Analysis taught by my teacher Dr.
Chow Hing Lun in 1984-85 when I was an undergraduate student in the CUHK.
Throughout this section, let H be a complex Hilbert space and let T be a positive bounded operator
on H. The aim of this section is to show that there is a unique positive operator S (called the
square root of T ) on H such that S 2 = T . The main feature of the proof here is without using the
functional calculus.

Proposition 20.1. Let S, T ∈ B(H) such that ST = T S. If S, T both are positive operators, then
so is ST .
S
Proof. If S = 0, then the assertion is clear. Now we assume that S ̸= 0. Put S1 := ∥S∥ . Set
Sn+1 := Sn − Sn2
for n = 1, 2, ....
Claim 1: 0 ≤ Sn ≤ I for all n = 1, .... The assertion will be obtained by induction on n.
Notice that as n = 1, clearly we have 0 ≤ S1 ≤ I. Suppose that the Claim 1 is true for n,
i.e., 0 ≤ Sn ≤ I and thus, we have 0 ≤ I − Sn ≤ I. This implies that for all x ∈ H we
have (Sn2 (I − Sn )x, x) = ((I − Sn )Sn x, Sn x) ≥ 0. This gives Sn2 (I − Sn ) ≥ 0. Similarly, we have
Sn (I − Sn )2 ≥ 0. Hence, we have 0 ≤ Sn2 (I − Sn ) + Sn (I − Sn )2 = Sn − Sn2 = Sn+1 . On the other
hand, we have 0 ≤ (I − Sn ) + Sn2 = I − Sn+1 because Sn2 ≥ 0 and I − Sn ≥ 0. Therefore Claim 1
follows from the induction.
The proof will be complete if we show that (ST x, x) ≥ 0 for all x ∈ H.
In fact, notice that we have
S1 = S12 + S2 = S12 + S22 + S3 = · · · = S12 + · · · + Sn2 + Sn+1 .
This implies that
S12 + · · · + Sn2 = S1 − Sn+1 ≤ S1
for all n = 1, 2.. because Sn+1 ≥ 0. Thus, we have
Xn Xn n
X
2
∥Sk x∥ = (Sk x, Sk x) = (Sk2 x, x) ≤ (S1 x, x)
k=1 k=1 k=1
P∞
for all n. This gives k=1 ∥Sk x∥2 < ∞ and so, Sn x → 0. This implies that
n
X
( Sk2 )x = S1 (x) − Sn+1 (x) → 0
k=1
P∞ 2
for all x ∈ H and so we have k=1 Sk (x) = S1 (x) for all x ∈ H. Finally, we complete the proof by
the following
∞
X ∞
X
2
(ST x, x) = ∥S∥(T S1 x, x) = ∥S∥ (T Sk x, x) = ∥S∥ (T Sk x, Sk x) ≥ 0
k=1 k=1
for all x ∈ H. □
56 CHI-WAI LEUNG

Proposition 20.2. Let Tn , n = 1, 2, .... and K be the bounded linear selfadjoint operators on H.
Suppose that
(1) T1 ≤ T2 ≤ · · · ≤ K.
(2) Tn Tm = Tn Tm and KTn = Tn K for all m, n = 1, 2.....
Then there is a bounded selfadjoint operator T on H with T ≤ K such that lim Tn x = T x for all
x ∈ H.
Proof. Now let Sn := K − Tn for n = 1, 2, ... Then 0 ≤ Sn for all n = 1, 2, .... By using Proposition
2 − S S = (S − S )S ≥ 0 and hence, S 2 ≥ S S for n ≥ m. Similarly, we
20.1, we see that Sm n m m n m m n m
also have Sn Sm ≥ Sn2 for n ≥ m. Therefore, we have
2
(20.1) Sm ≥ Sn Sm ≥ Sn2
for n ≥ m. Thus, ((Sm 2 x, x))∞ 2
m=1 is a decreasing sequence of non-negative numbers and so lim(Sn x, x)
exists for all x ∈ H. Moreover since Sn and Sm commutes to each other, Eq 20.1 gives
∥Sm x − Sn x∥2 = ((Sm − Sn )2 x, x)
2 2
= (Sm x, x) − 2(Sm Sn x, x) + (Sm x, x)
2
≤ (Sm x, x) − (Sn2 x, x) → 0
for n ≥ m and for all x ∈ H. This implies that (Sn x) is a Cauchy sequence and hence, lim Sn x
exists for all x ∈ H. This implies that T (x) := lim Tn (x) = K − lim Sn x exists for all x ∈ H. The
Uniform Boundedness Theorem tells us that T ∈ B(H). In addition T is selfadjoint because each
Tn is selfadjoint. The proof is complete. □

We now come to the main result in this section.

Theorem 20.3. If T is a bounded positive operator on H, then there is a unique positive operator
S such that S 2 = T . In this case, we call S the square root of T .
Proof. We show the existence first.
T
Clearly, we may assume that T ̸= 0 and T ≤ I by considering the operator ∥T ∥ . Put S0 = 0 and
1 2
Sn = Sn−1 + (T − Sn−1 )
2
for n = 1, 2, ..... Then Sn is a polynomial of T and so, all Sn ’s are selfadjoint operators and commute
to each other. Notice that since 0 < T ≤ I and by the definition of Sn , we have
1 2 1 1
I − Sn = I − Sn−1 − (T − Sn−1 ) = (I − Sn−1 )2 + (I − T ) ≥ 0.
2 2 2
Thus Sn ≤ I for all n = 0, 1, 2.... On the other hand, we have
1 1 1
(20.2) Sn+1 − Sn = Sn + (T − Sn2 ) − Sn−1 − (T − Sn−1 2
) = (Sn − Sn−1 )(I − (Sn + Sn−1 ))
2 2 2
1
for all n = 0, 1, 2.... Since Sn ≤ I, I − 2 (Sn + Sn−1 ) ≥ 0. Using Proposition 20.1 and the Eq 20.2,
we can apply induction on n to see that 0 = S0 ≤ · · · ≤ Sn ≤ Sn+1 ≤ · · · ≤ I for all n = 0, 1, 2....
Proposition 20.2 tells us that Sx := lim Sn x exists for all x ∈ H and S ∈ B(H). In addition S is
positive because Sn ≥ 0 for all n = 0, 1, 2.... Also, since Sn x = Sn−1 x + 21 (T − Sn−1
2 )x for all x ∈ H,
2
by taking n → ∞, we see that T x = S x for all x. Thus the operator S is as desired.
Finally, we show the uniqueness.
Now let R be another positive bounded operator on H such that R2 = T . Notice that RT = R3 =
T R. This implies that RS = SR because S is the ∥ · ∥-limit of the polynomials of T by the above
construction of S. Now we take any x ∈ H and put y := (S − R)x. Then we have
0 ≤ (Sy, y) + (Ry, y) = ((S + R)(S − R)x, y) = ((S 2 − R2 )x, y) = 0.
 57

This implies that (Sy, y) = (Ry, y) = 0 because both are non-negative numbers. On the other
hand, since S ≤ 0, by above there is another positive operator W such that W 2 = S, and so we
have 0 = (Sy, y) = (W y, W y) that gives Sy = 0. Similarly, we also have Ry = 0. Finally, we have
∥(S − R)x∥2 = ((S − R)2 x, x) = ((S − R)y, x) = 0.
Thus, S = R as desired. The proof is complete. □

21. Compact operators on a Hilbert space

Throughout this section, let H be a complex Hilbert space.

Definition 21.1. A linear operator T : H → H is said to be compact if for every bounded sequence
(xn ) in H, (T (xn )) has a norm convergent subsequence.
Write K(H) for the set of all compact operators on H and K(H)sa for the set of all compact
selfadjoint operators.

Remark 21.2. Let U be the closed unit ball of H. Clearly, T is compact if and only if the norm
closure T (U ) is a compact subset of H. Thus if T is compact, then T is bounded automatically
because every compact set is bounded. In particular, if T has finite rank, that is dim imT < ∞,
then T must be compact because every closed and bounded subset of a finite dimensional normed
space is compact. In addition, clearly we have the following result.

Proposition 21.3. The identity operator I : H → H is compact if and only if dim H < ∞.

Example 21.4. Let H = ℓ2 ({1, 2...}). Define T x(k) := x(k)

k for k = 1, 2.... Then T is compact.
2
In fact, if we let (xn ) be a bounded sequence in ℓ , then by the diagonal argument, we can find
a subsequence ym := T xm of T xn such that lim ym (k) = y(k) exists for all k = 1, 2... Let
m→∞
L := supn ∥xn ∥22 . Since |ym (k)|2 ≤ kL2 for all m, k, we have y ∈ ℓ2 . Now let ε > 0. Then one can
find a positive integer N such that k≥N 4L/k 2 < ε. Thus we have
P

X X 4L
|ym (k) − y(k)|2 < <ε
k2
k≥N k≥N

for all m. On the other hand, since lim ym (k) = y(k) for all k, we can choose a positive integer
m→∞
M such that
N
X −1
|ym (k) − y(k)|2 < ε
k=1
for all m ≥ M . Finally, we have ∥ym − y∥22 < 2ε for all m ≥ M .

Theorem 21.5. Let T ∈ B(H). Then T is compact if and only if T maps every weakly convergent
sequence in H to a norm convergent sequence.
Proof. We first assume that T ∈ K(H). Let (xn ) be a weakly convergent sequence in H. Since H is
reflexive, (xn ) is bounded by the Uniform Boundedness Theorem. Thus we can find a subsequence
(xj ) of (xn ) such that (T xj ) is norm convergent. Let y := limj T xj . We claim that y = limn T xn .
Suppose that y ̸= limn T xn . Then by the compactness of T again, we can find a subsequence (xi )
of (xn ) such that T xi converges to y ′ with y ̸= y ′ . Thus there is z ∈ H such that (y, z) ̸= (y ′ , z).
58 CHI-WAI LEUNG

On the other hand, if we let x be the weakly limit of (xn ), then (xn , w) → (x, w) for all w ∈ H.
Thus we have
(y, z) = lim(T xj , z) = lim(xj , T ∗ (z)) = (x, T ∗ z) = (T x, z).
j j
Similarly, we also have (y ′ , z) = (T x, z) and hence (y, z) = (y ′ , z) that contradicts to the choice of
z.
For the converse, let (xn ) be a bounded sequence. Then by Theorem 15.12, (xn ) has a weakly
convergent subsequence. Thus T (xn ) has a norm convergent subsequence by the assumption. Thus
T is compact. □

Proposition 21.6. Let S, T ∈ K(H). Then we have

(i) : αS + βT ∈ K(H) for all α, β ∈ C;
(ii) : T Q and QT ∈ K(H) for all Q in B(H);
(iii) : T ∗ ∈ K(H).
Moreover K(H) is normed closed in B(H), and hence K(H) is a closed ∗-ideal of B(H).
Proof. (i) and (ii) are clear.
For property (iii), let (xn ) be a bounded sequence. Then (T ∗ xn ) is also bounded. Thus T T ∗ xn has
a convergent subsequence T T ∗ xnk by the compactness of T . Note that we have
∥T ∗ xnk − T ∗ xnl ∥2 = (T T ∗ (xnk − xnl ), xnk − xnl )
for all nk , nl . This implies that (T ∗ xnk ) is a Cauchy sequence and thus is convergent.
Finally we want to show that K(H) is closed. Let (Tm ) be a sequence in K(H) such that Tm → T in
norm. Let (xn ) be a bounded sequence in H. Then by the diagonal argument there is a subsequence
(xnk ) of (xn ) such that lim Tm xnk exists for all m. Now let ε > 0. Since limm Tm = T , there is a
k
positive integer N such that ∥T − TN ∥ < ε. On the other hand, there is a positive integer K such
that ∥TN xnk − TN xnk′ ∥ < ε for all k, k ′ ≥ K. Thus we can now have
∥T xnk − T xnk′ ∥ ≤ ∥T xnk − TN xnk ∥ + ∥TN xnk − TN xnk′ ∥ + ∥TN xnk′ − T xnk′ ∥ ≤ (2L + 1)ε
for all k, k ′ ≥ K where L := supn ∥xn ∥. Thus limk T xnk exists. We can now conclude that
T ∈ K(H). □

Example 21.7. Let k(z, w) ∈ C(T × T). Define an operator T : L2 (T) → L2 (T) by
Z
T ξ(z) := k(z, w)ξ(w)dw
T
for z ∈ T and ξ ∈ L2 (T). Then T is a compact operator.
Proof. Clearly, we have ∥T ∥ ≤ ∥k∥∞ . On the other hand, Stone-Weiestrass Theorem tells us the
polynomials of (z, z̄; w, w̄) are ∥ · ∥∞ -dense in C(T × T). Therefore, by using Proposition 21.6, it
suffices to show for the case k(z, w) = N i j
P
i,j=1 aij (z, z̄)w w̄ where aij (z, z̄) is a polynomial of (z, z̄)
of degree N . From this, we have
XN Z
T ξ(z) = aij (z, z̄) wi w̄j ξ(w)dw
i,j=1 T

for ξ ∈ L2 (T). Thus, T (ξ) ∈ span{z i z̄ j : 0 ≤ i, j ≤ N } which is of finite dimension for all ξ ∈ L2 (T).
This implies that T has finite dimensional range and thus, T is compact. The proof is complete. □

Corollary 21.8. Let T ∈ K(H). If dim H = ∞, then 0 ∈ σ(T ).

/ σ(T ). Then T −1 exists in B(H). Proposition 21.6 gives I = T T −1 ∈
Proof. Suppose that 0 ∈
K(H). This implies dim H < ∞. □
 59

Proposition 21.9. Let T ∈ K(H) and let c ∈ C with c ̸= 0. Then T − c has a closed range.
1 1
Proof. Note that T ∈ K(H). Thus if we consider T −I, we may assume that c = 1. Let S = T −I.
c c
Let (xn ) be a sequence in H such that Sxn → x ∈ H in norm. By considering the orthogonal
decomposition H = ker S ⊕ (ker S)⊥ , we write xn = yn ⊕ zn for yn ∈ ker S and zn ∈ (ker S)⊥ . We
first claim that (zn ) is bounded. Suppose that (zn ) is unbounded. By considering a subsequence
zn
of (zn ), we may assume that we may assume that ∥zn ∥ → ∞. Put vn := ∈ (ker S)⊥ .
∥zn ∥
Since Szn = Sxn → x, we have Svn → 0. On the other hand, since T is compact, and (vn ) is
bounded, by passing a subsequence of (vn ), we may also assume that T vn → w. Since S = T − I,
vn = T vn − Svn → w − 0 = w ∈ (ker S)⊥ . In addition from this we have Svn → Sw. On the other
hand, we have Sw = limn Svn = limn T vn − limn vn = w − w = 0. Thus w ∈ ker S ∩ (ker S)⊥ . It
follows that w = 0. However, since vn → w and ∥vn ∥ = 1 for all n. It leads to a contradiction.
Thus (zn ) is bounded.
Finally we are going to show that x ∈ imS. Now since (zn ) is bounded, (T zn ) has a convergent
subsequence (T znk ). Let limk T znk = z. Then we have
znk = Sznk − T znk = Sxnk − T znk → x − z.
It follows that x = limk Sxnk = limk Sznk = S(x − z) ∈ imS. The proof is complete. □

Remark 21.10. Proposition 21.9 dose not hold as c = 0. Example 21.4 gives a counter example
about this. In fact, we can directly check that the vector y := (1, 21 , 13 , .....) ∈ im(T ) \ im(T ), where
T is the compact operator defined in Example 21.4.

Theorem 21.11. Fredholm Alternative Theorem : Let T ∈ K(H)sa and let 0 ̸= λ ∈ C. Then
T − λ is injective if and only if T − λ is surjective.
Proof. Since T is selfadjoint, σ(T ) ⊆ R. Thus if λ ∈ C \ R, then T − λ is invertible. Thus the result
holds automatically.
Now consider the case λ ∈ R \ {0}.
Then T − λ is also selfadjoint. From this and Proposition 16.14, we have ker(T − λ) = (im(T − λ))⊥
and (ker(T − λ))⊥ = im(T − λ).
Thus the proof is complete immediately by using Proposition 21.9. □

Corollary 21.12. Let T ∈ K(H)sa . Then we have σ(T ) \ {0} = σp (T ) \ {0}. Consequently if
the values m(T ) and M (T ) which are defined in Theorem 18.8 are non-zero, then both are the
eigenvalues of T and ∥T ∥ = max |λ|.
λ∈σp (T )

Proof. It follows immediately from the Fredholm Alternative Theorem. This, together with Theo-
rem 18.8, implies the last assertion. □

Example 21.13. Let T ∈ B(ℓ2 ) be defined as in Example 21.4. We have shown that T ∈ K(ℓ2 ) and
T is selfadjoint. Then by Corollary 21.12 and Corollary 21.8, we see that σ(T ) = {0, 1, 12 , 13 , .....}.

Lemma 21.14. Let T ∈ K(H)sa and let Eλ := {x ∈ H : T x = λx} for λ ∈ σ(T ) \ {0}, that is the
eigenspace of T corresponding to λ. Then dim Eλ < ∞.
Proof. It is because the restriction T |Eλ : Eλ → Eλ is also a compact operator on Eλ , then
dim Eλ < ∞ for all λ ∈ σ(T ) \ {0} = σp (T ) \ {0}. □
60 CHI-WAI LEUNG

Theorem 21.15. Let T ∈ K(H)sa . Suppose that dim H = ∞. Then σ(T ) = {λk : k = 1, ..., N } ∪
{0}, where 1 ≤ N ≤ ∞ and (λn ) is a sequence of non-zero real numbers with |λ1 | ≥ |λ2 | ≥ · · · and
λi ̸= λj for i ̸= j. Moreover, if (λn ) is an infinite sequence, then |λn | ↓ 0.
Proof. Note that since dim H = ∞, 0 ∈ σ(T ). In addition we have ∥T ∥ = max(|M (T )|, |m(T )|)
and σ(T ) \ {0} = σp (T ) \ {0}. Thus by Corollary 21.12, there is |λ1 | = max |λ| = ∥T ∥. Since
λ∈σp (T )
dim Eλ1 < ∞, then Eλ⊥1 ̸= 0. By considering the restriction of ⊥
T2 := T |Eλ1 , if T2 ̸= 0, then there is
0 ̸= |λ2 | = maxλ∈σp (T2 ) |λ| = ∥T2 ∥. Note that λ2 ∈ σp (T ) and |λ2 | ≤ |λ1 | because ∥T2 ∥ ≤ ∥T ∥. To
repeat the same step, if TN +1 = 0 for some N , then 0 ∈ σp (T ). Otherwise, we can get an infinite
sequence (λn ) such that (|λn |) is decreasing.
Now we claim that if (λn ) is an infinite sequence, then limn |λn | = 0.
Otherwise, there is η > 0 such that |λn | ≥ η for all n. If we let vn ∈ Eλn with ∥vn ∥ = 1 for all
n. Note that since dim H = ∞ and dim Eλ < ∞, for any λ ∈ σp (T ) \ {0}, there are infinite many
λn ’s. Then wn := |λ1n | vn is a bounded sequence and ∥T wn − T wm ∥2 = ∥vn − vm ∥2 = 2 for m ̸= n.
This is a contradiction since T is compact. Thus limn |λn | = 0.
Finally we need to check σ(T ) = {λ1 , λ2 , ...} ∪ {0}.
In fact, let µ ∈ σp (T ). Since |λ1 | = ∥T ∥ ≥ |µ|, |λm+1 | < |µ| ≤ |λm |. Note that we have Eα ⊥Eβ
for α and β in σp (T ) with α ̸= β. Then by the construction of λn ’s, we have µ = λm . For
example, if |λ2 | < |µ| ≤ |λ1 | and µ ̸= λ1 , then Eµ ⊥Eλ1 . Hence, we have Eµ ⊆ (Eλ1 )⊥ . Then
by the construction of λ2 , that is |λ2 | = ∥T2 ∥ ≥ |µ| which leads to a contradiction. Thus, if
|λ2 | < |µ| ≤ |λ1 |, then µ = λ1 . The proof is complete. □

Theorem 21.16. Spectral Decomposition Theorem: Let T ∈ K(H)sa and let (λn )N n=1 , (1 ≤
N ≤ ∞), be a sequence of eigenvalues of T given as in Theorem 21.15. For each λ ∈ σp (T ) \ {0},
put d(λ) := dim Eλ < ∞. Let {eλ,i : i = 1, ..., d(λ)} be an orthonormal basis for Eλ . Then we have
the following orthogonal decomposition:
N
M
(21.1) H = ker T ⊕ Eλn
n=1
∞
M
In here as N = ∞, the space Eλn means that the closed linear spanned by the set {xk : xk ∈
n=1
Eλk }.
Moreover B := {eλ,i : λ ∈ σp (T ) \ {0}; i = 1, .., d(λ)} forms an orthonormal basis of T (H), and we
have
N d(λ
X X n)

(21.2) Tx = λn (x, eλn ,i )eλn ,i

n=1 i=1
for all x ∈ H.
∞
X
In addition, if N = ∞, then the series λn Pn norm converges to T , where Pn is the orthogonal
n=1
d(λn )
X
projection from H onto Eλn , that is, Pn (x) := (x, eλn ,i )eλn ,i , for x ∈ H.
i=1
N
M
Proof. Put E = Eλn . Clearly, we have ker T ⊆ E ⊥ . On the other hand, if the restriction
n=1
T0 := T |E ⊥ =
̸ 0, then there exists an non-zero element µ ∈ σp (T0 ) ⊆ σp (T ) because T0 ∈ K(E ⊥ ).
It is absurd because µ ̸= λi for all i. Thus T |E ⊥ = 0 and hence E ⊥ ⊆ ker T . Therefore, we have
 61

the decomposition (21.1). Moreover, from this we see that the family B forms an orthonormal basis
of (ker T )⊥ . On the other hand, we have (ker T )⊥ = imT ∗ = imT . Therefore, B is an orthonormal
basis for T (H) and Equation 21.2 follows.
For the last assertion,Pit needs to show that the series ∞
P
n=1 λn Pn converges to T in norm. Note
that if we put Sm := m n=1 λ n P n , then by the decomposition (21.1), lim Sm x = T x for all x ∈ H.
m→∞
Thus it suffices to show that (Sm )∞
m=1 is a Cauchy sequence in B(H). In fact we have

∥λm+1 Pm+1 + · · · · · · + λm+p Pm+p ∥ = |λm+1 |

for all m, p ∈ N because Eλn ⊥Eλm for m ̸= n and |λn | is decreasing. This gives that (Sn ) is a
Cauchy sequence since |λn | ↓ 0 as N = ∞. The proof is complete. □

Corollary 21.17. T ∈ K(H) if and only if T can be approximated by finite rank operators.
Proof. The sufficient condition follows immediately from Proposition 21.6.
Conversely, for a general compact operator T , we can consider the decomposition:
1 1
T = (T + T ∗ ) + i( (T − T ∗ )).
2 2i
Note that Re(T ) := 12 (T + T ∗ ) (call the real part of T ) and Im(T ) := 2i
1
(T − T ∗ ) (call the imaginary
part of T ) both are the self-adjoint compact operators. From this, we see that the T can be
approximated by finite ranks operators by using Theorem 21.16. □

22. Unbounded operators

Throughout this section, let H be a complex Hilbert space. An operator T on H means that
T is a linear operator defined in a vector subspace of T (it is not necessarily bounded). We write
D(T ) for the domain of T . We say that T is densely defined if the domain D(T ) is dense in H. An
operators S is said to be an extension of T if D(T ) ⊆ D(S) and Sx = T x for all x ∈ D(T ), denoted
it by T ⊂ S.
In addition, if T and S are operators on H, then we naturally define the domains of the following
operations.
(i) D(T + S) := D(T ) ∩ D(S).
(ii) D(S ◦ T ) := {x ∈ D(T ) : T x ∈ D(S)}.

Definition 22.1. Let T be a densely defined operator on H. Put

D(T ∗ ) := {y ∈ H : there is z ∈ H such that (T x, y) = (x, z) for all x ∈ D(T )}.
Clearly, D(T ∗ ) is a vector subspace of H. In addition, since T is densely defined, for each element
y ∈ D(T ∗ ), there is a unique element in H, denoted it by T ∗ y, satisfying
(T x, y) = (x, T ∗ y)
for all x ∈ D(T ). We call T ∗ the adjoint operator of T .
We call an operator T symmetric (resp. self-adjoint) if T ⊂ T ∗ (resp. T = T ∗ ).
Note that T is symmetric if and only if we have
(T x, y) = (x, T y)
for all x, y ∈ D(T ).

Proposition 22.2. Let S, T be the operators on H. Assume that T , S and ST are densely defined.
Then T ∗ S ∗ ⊂ (ST )∗ .
62 CHI-WAI LEUNG

Proof. We first claim that T ∗ S ∗ ⊂ (ST )∗ . Let x ∈ D(ST ) and y ∈ D(T ∗ S ∗ ). Then S ∗ y is defined
and S ∗ y ∈ D(T ∗ ). Since x ∈ D(ST ) we have x ∈ D(T ) and T x ∈ D(S). Thus we have
(ST x, y) = (T x, S ∗ y) = (x, T ∗ S ∗ y).
This implies that y ∈ D(ST )∗ and (ST )∗ (y) = T ∗ S ∗ y and hence T ∗ S ∗ ⊂ (ST )∗ . □

Example 22.3. First we recall that a function f : [a, b] → C is called an indefinite integral if there
is an element φ ∈ L1 [a, b] such that
Z x
f (x) = f (a) + φ(t)dt
a

for all x ∈ [a, b], where dt is the Lebesgue measure on [a, b]. In this case we have f ′ (x) = φ(x)
almost everywhere in (a, b).
Let
D := {f : [a, b] → C : f is an indefinite integral with f (a) = f (b) and f ′ ∈ L2 [a, b]}.
Note that D is dense subspace of L2 [a, b]. Define an operator T with D(T ) = D by
T f := if ′ .
for f ∈ D. We claim that T is self-adjoint. The proof is divided by several steps.
Claim 1: T ⊂ T ∗ .
In fact, let f, g ∈ D. Then we have
Z b
(T f, g) = if ′ (t)g(t)dt
a
Z b
= ig(t)df (t)
a
(22.1) Z b
= if (t)g(t)|ba −i f (t)g ′ (t)dt
a
Z b
= f (t)ig ′ (t)dt = (f, T g).
a

Therefore, the Claim 1 follows. Next we want to show D(T ∗ ) ⊆ D(T ).

Let g ∈ D(T ∗ ∗ 2 1 2 1
R x ). Put φ := T g ∈ L [a, b]. Note that φ ∈ L [a, b] because L [a, b] ⊆ L [a, b]. Thus,
Φ(x) := a φ(t)dt for x ∈ [a, b] is an indefinite integral of φ.
Claim 2: There is a constant c so that g(t) + iΦ(t) = c for all t ∈ [a, b]. Note that for any f ∈ D,
we have
(T f, g) = (f, T ∗ g)
Z b
= f (t)φ(t)dt
a
Z b
= f (t)dΦ(t)
a
Z b
= f (b)Φ(b) − Φ(t)f ′ (t)dt
a
= Φ(b) − (T f, iΦ).
From this if we take f ≡ 1 ∈ D in above, then Φ(b) = 0. Therefore, we have
(T f, g) = −(T f, iΦ)
 63

for all f ∈ D. This implies that (g + iΦ)⊥im(T ). If we let 1 ∈ L2 [a, b] be the function of constant
one in [a, b], then we have
Z b
(T f, 1) = if ′ (t)dt = i(f (b) − f (a)) = 0
a
for all f ∈ D, hence C1⊥im(T ). On the other hand, note that for any ξ ∈ L2 [a, b] if we put
Rb Rb Rx
ξ1 = ξ − a ξ(t)dt ∈ L2 [a, b], then a ξ1 (t)dt = 0. Let h(x) := i a ξ1 (t)dt. Then h ∈ D and
T h = ξ1 . Therefore, we have L2 [a, b] = C1+im(T ) and hence we have the orthogonal decomposition
L2 [a, b] = C1 ⊕ im(T ). In particular, (im(T ))⊥ = C1. This implies that g + iΦ = c for some
constant c. Then g ′ = −iΦ′ = −iφ ∈ L2 [a, b], so g is an indefinite integral because g ′ ∈ L1 [a, b].
Moreover, we see that g(b) = g(a) = c because Φ(b) = Φ(a) = 0. We can now conclude that g ∈ D.
The proof is complete.

Example 22.4. Using the notation as in Example 22.3, we let

D1 := {f ∈ D : f (a) = f (b) = 0}.
Then D1 is dense subspace of L2 [a, b]. Define T1 : D1 → L2 [a, b] by
T1 f = if ′
for f ∈ D1 . Then T1 is symmetric but it is not self-adjoint.
By using the similar calculation as in Eq 22.1 in Example 22.3 above, we see that T1 ⊂ T1∗ . Let
D2 := {f : [a, b] → C : f is an indefinite integral and f ′ ∈ L2 [a, b] }. Then D2 ⊆ D(T1∗ ). In fact,
let f ∈ D1 and g ∈ D2 , using the same argument as in Eq 22.1 again, we have
Z b Z b
(T1 f, g) = if (t)g(t)|ba − i f (t)g ′ (t)dt = f (t)ig ′ (t)dt = (f, T2 g)
a a
because f (a) = f (b) = 0, where T2 (g) := ig ′ for g ∈ D2 . Therefore D(T1 ) ⊊ D(T1∗ ) since D(T1 ) =
D1 ⊊ D2 . The proof is complete.

Definition 22.5. An operator T on H is said to be closed if its graph of T , denoted it by G(T ) :=

{(x, T x) ∈ H × H : x ∈ D(T )}, is closed in H × H. More precisely, if (xn ) is a sequence in D(T )
such that xn → x and T xn → y, then x ∈ D(T ) and T x = y.

Define an operator V : H × H → H × H by V (x, y) = (−y, x) for (x, y) ∈ H × H. Then

(V (x, y), V (x′ , y ′ )) = ((x, y), (x′ , y ′ )) for all (x, y) and (x′ , y ′ ) in H × H and hence, the operator
preserves the orthogonality on H × H.

Proposition 22.6. Using the notation as above, let T be a densely operator on H. Then G(T ∗ ) =
(V (G(T )))⊥ . Consequently, the adjoint operator T ∗ is closed. In particular, if T is self-adjoint,
then T is closed.
Proof. Note that for x ∈ D(T ∗ ) and y ∈ D(T ), we have ((x, T ∗ x), V (y, T y)) = 0 Therefore, we
have G(T ∗ ) ⊆ (V (G(T )))⊥ . On the other hand, if (u, v)⊥(−T y, y) for all y ∈ D(T ). Then we
have (v, y) = (u, T y) and hence, u ∈ D(T ∗ ) and T ∗ u = v. Therefore, (u, v) ∈ G(T ∗ ). The proof is
complete. □

Proposition 22.7. Let T be a symmetric operator on H. Then the following statements are
equivalent.
(i) T is self-adjoint.
(ii) T is closed and ker(T ∗ ± i) = {0}.
(iii) im(T ± i) = H.
64 CHI-WAI LEUNG

Proof. For (i) ⇒ (ii), assume that T is self-adjoint. Then by Proposition 22.6, T is closed. Next
we show ker(T ∗ − i) = {0}. Let y ∈ D(T ∗ ) such that T ∗ y = iy. Since D(T ) = D(T ∗ ), we have
i(y, y) = (T y, y) = (y, T ∗ y) = −i(y, y), so y = 0. Similarly, we have ker(T ∗ + i) = {0}.
For (ii) ⇒ (iii), we first claim that im(T + i) is dense in H. Let z⊥im(T + i). Then z⊥(T + i)x for
all x ∈ D(T ), and thus we have (T x, z) = (x, −iz). This implies that z ∈ D(T ∗ ) and T ∗ z = −iz.
Thus, z ∈ ker(T ∗ + i), so z = 0. Therefore, it suffices to show that im(T + i) is closed. Let (xn ) be
a sequence in D(T ) such that lim(T − i)xn = y. Since T is symmetric, we have
∥T (xm − xn ) + i(xm − xn )∥2 = ∥T (xm − xn )∥2 + ∥(xm − xn )∥2
for all m, n. From this we see that u := lim xn and v := lim T xn both exist. T is closed by the
assumption, so u ∈ D(T ) and T u = v. Therefore, we have
y = lim(T xn + ixn ) = v + iu = (T + i)u ∈ im(T + i).
Hence im(T + i) = H. Similarly, we have im(T − i) = H.
For the last implication (iii) ⇒ (i), since T ⊂ T ∗ , we need to show that D(T ∗ ) ⊆ D(T ). Let
u ∈ D(T ∗ ). Since im(T − i) = H, there is an element v ∈ D(T ) such that
(T − i)v = (T ∗ − i)u.
Since T ⊂ T ∗ , we have (T − i)v = (T ∗ − i)v, thus, v − u ∈ ker(T ∗ − i). Then for any z ∈ D(T ), we
have
((T + i)z, v − u) = (z, (T + i)∗ (v − u)) = (z, (T ∗ − i)(v − u)) = 0.
im(T + i) = H by assumption, so u = v ∈ D(T ). The proof is complete. □
Proposition 22.8. Let T be a symmetric operator on H. Then there is the smallest closed exten-
sion of T , denoted it by T . We call T the closure of T . In addition, G(T ) = G(T ) and T = T ∗∗ .
Proof. Let D(T ) := {x ∈ H : (x, y) ∈ G(T ) for some y ∈ H}. We first note for each element
x ∈ D(T ), there is a unique element y ∈ H so that (x, y) ∈ G(T ). In fact, if (x, y) ∈ G(T ), there is
a sequence (xn ) in D(T ) such that lim xn = x and lim T xn = y. Note that for any u ∈ D(T ), since
T is symmetric, we have
(T u, x) = lim(T u, xn ) = lim(u, T xn ) = (u, y).
Therefore, y is uniquely determined by x because D(T ) is dense in H. Hence, we can define T x = y
for x ∈ D(T ). Clearly, we have G(T ) = G(T ) by the construction of T , and hence T is closed.
Moreover, we can directly show that T is the smallest closed extension of T .
For the last assertion, since T ⊂ T ∗ , T ∗ is densely defined, so T ∗∗ := (T ∗ )∗ is defined. Since
V 2 = −I and V is an isometry and an orthogonal preserver, by using Proposition 22.6, we have
G(T ∗∗ ) = [V G(T ∗ )]⊥
= V [G(T ∗ )⊥ ]
= V [V (G(T ))]
= V 2 (G(T ))
= G(T ).
Thus, T = T ∗∗ . □

References
[1] L.V. Ahlfors, Complex analysis, 3rd edition, Mc Graw Hill, (1979).
[2] S. Banach, Theory of linear operators, English translation by F. Jallett, North-Holland, (1987).
[3] A. Bowers and N.J. Kalton, An introductory course in functional analysis, Springer, (2014).
[4] N.L. Carothers, A short course on Banach space theory, Cambridge University Press (2005).
 65

[5] W. Cheney, Ward and A. Goldstein, Proximity maps for convex sets. Proc. Amer. Math. Soc. 10 (1959), 448–450.
[6] J. Diestel, Sequences and series in Banach spaces, Springer-Verlag, (1984).
[7] P.R. Halmos, Naive set theory, Dover publications, (2017).
[8] R.C. James, A non-refexivity of Banach space isometric with its second conjugate space, Proc. N.A.S. USA, 37
(1951), 174-177.
[9] Y. Katznelson, An introduction to Harmonic analysis, 3rd edition, Cambridge Univ Press, (2004).
[10] R.E. Megginson, An introduction to Banach space theory, GTM vol. 183, Springer, (1998).

(Chi-Wai Leung) Department of Mathematics, The Chinese University of Hong Kong, Shatin, Hong
Kong
Email address: cwleung@math.cuhk.edu.hk