3.

Hilbert spaces

In this section we examine a special type of Banach spaces. We start with some
algebraic preliminaries.
Definition. Let K be either R or C, and let Let X and Y be vector spaces
over K. A map φ : X × Y → K is said to be K-sesquilinear, if
• for every x ∈ X, then map φx : Y 3 y 7−→ φ(x, y) ∈ K is linear;
• for every y ∈ Y, then map φy : X 3 y 7−→ φ(x, y) ∈ K is conjugate linear,
i.e. the map X 3 x 7−→ φy (x) ∈ K is linear.
In the case K = R the above properties are equivalent to the fact that φ is bilinear.
Remark 3.1. Let X be a vector space over C, and let φ : X × X → C be a
C-sesquilinear map. Then φ is completely determined by the map
Qφ : X 3 x 7−→ φ(x, x) ∈ C.
This can be see by computing, for k ∈ {0, 1, 2, 3} the quantity

Qφ (x + ik y) = φ(x + ik y, x + ik y) = φ(x, x) + φ(x, ik y) + φ(ik y, x) + φ(ik y, ik y) =

= φ(x, x) + ik φ(x, y) + i−k φ(y, x) + φ(y, y),
which then gives
3
1  X −k
i Qφ (x + ik y) , ∀ x, y ∈ X.

(1) φ(x, y) =
4
k=0

The map Qφ : X → C is called the quadratic form determined by φ. The identity (1)
is referred to as the Polarization Identity. Remark that the Polarization Identity
can also gives
3
1 X k
i Qφ (x + ik y) , ∀ x, y ∈ X.

φ(y, x) =
4
k=0

In particular, we see that the following conditions are equivalent

(i) Qφ is real-valued, i.e. Qφ (x) ∈ R, ∀ x ∈ X;
(ii) φ is sesqui-symmetric, i.e.

φ(y, x) = φ(x, y), ∀ x, y ∈ X.

Definition. Let K be one of the fields R or C, and let X be a vector space

over K. A K-sesquilinear map φ : X × X → K is said to be positive definite, if
• φ is sesqui-symmetric;
• φ(x, x) ≥ 0, ∀ x ∈ X.
As observed before, in the complex case, the second condition actually forces the
first one.

With this terminology, we have the following useful technical result.

79
80 CHAPTER II: ELEMENTS OF FUNCTIONAL ANALYSIS

Proposition 3.1 (Cauchy-Bunyakowski-Schwartz Inequality). Let K be a K-

vector space, and let φ : X × X → K be a positive definite K-sesquilinear map.
Then
(2) |φ(x, y)|2 ≤ φ(x, x) · φ(y, y), ∀ x, y ∈ X.
Moreover, if equality holds then either x = y = 0, or there exists some α ∈ K with
φ(x + αy, x + αy) = 0.
Proof. Fix x, y ∈ X, and choose a number λ ∈ K, with |λ| = 1, such that
|φ(x, y)| = λφ(x, y) = φ(x, λy).
Define the map F : K → K by
F (ζ) = φ(ζλy + x, ζλy + x), ∀ ζ ∈ K.
A simple computation gives
F (ζ) = ζλζ̄ λ̄φ(y, y) + +ζλφ(x, y)ζ̄ λ̄φ(y, x) + φ(x, x) =
= |ζ|2 |λ|2 φ(y, y) + ζλφ(x, y) + ζ̄λφ(x, y) + φ(x, x) =
= |ζ|2 φ(y, y) + ζ|φ(x, y)| + ζ̄|φ(x, y)| + φ(x, x), ∀ z ∈ R.
In particular, when we restrict F to R, it becomes a quadratic function:
F (t) = at2 + bt + c, ∀ t ∈ R,
where a = φ(y, y) ≥ 0, b = 2|φ(x, y)|, c = φ(x, x). Notice that we have
F (t) ≥ 0, ∀ t ∈ R.
This forces either a = b = 0, in which case the inequality (2) is trivial, or a > 0,
and b2 − 4ac ≤ 0, which reads
4|φ(x, y)|2 − 4φ(x, x)φ(y, y) ≤ 0,
and again (2) follows.
Let us examine now when equality holds in (2). We assume that either x 6= 0 or
y 6= 0. In fact we can also assume that φ(x, x) > 0 and φ(y, y) > 0, so the fact that
we have equality in (2) is equivalent to b2 − 4ac = 0, which in terms of quadratic
equations says that the equation
F (t) = at2 + bt + c = 0
has a (unique) solution t0 . If we take α = t0 λ, then the equality F (t0 ) = 0 is
precisely φ(αy + x, αy + x) = 0. 

Definition. An inner product on X is a positive definite K-sesquilinear map

X × X 3 (ξ, η) 7−→ ξ  η ∈ K,


which is non-degenerate, in the sense that

• if ξ ∈ X satisfies ξ  ξ = 0, then ξ = 0.


Proposition 3.2. Let ·  · be an inner product on the K-vector space X.



(i) If we define, for every ξ ∈ X the quantity kξk = (ξ|ξ), then the map
p

X 3 ξ 7−→ kξk ∈ [0, ∞) defines a norm on X.

 §3. Hilbert spaces 81

(ii) For every ξ, η ∈ X, one has the Cauchy-Bunyakovski-Schwartz inequality:

 

(3)  ξ  η  ≤ kξk · kηk.

Moreover, if equality holds then ξ and η are proportional, in the sense that
either ξ = 0, or η = 0, or ξ = λη, for some λ ∈ K.
Proof. First of all, remark that by the non-degeneracy of the inner product,
one has the implication
(4) kξk = 0 ⇒ ξ = 0.
The inequality (3) is immediate from Proposition 3.1. Moreover, if wehave equality,


then by Proposition 3.1, we also know that either ξ = 0, or ξ + αη  ξ + αη = 0,
for some α ∈ K. This is equivalent to kξ + αηk2 = 0, so by (4), we must have
ξ = −αη.
Finally, in order to check that k . k is a norm, we must prove
(a) kλξk = |λ| · kξk, ∀ λ ∈ K, ξ ∈ X;
(b) kξ + ηk ≤ kξk + kηk, ∀ ξ, η ∈ X.
Property (a) is trivial, by seqsquilinearity:
kλξk2 = λξ  λξ = λ̄λ λξ  λξ = |λ|2 · kξk2 .



Finally, for ξ, η ∈ X, we have

kξ + ηk2 = ξ + η  ξ + η = ξ  ξ + η  η + ξ  η + η  ξ =






= kξk2 + kηk2 + ξ  η + ξ  η = kξk2 + kηk2 + 2Re ξ  η .





We now use the C-B-S inequality (3), which allows us to continue the above estimate
to get
kξ + ηk2 = kξk2 + kηk2 + 2Re ξ η ≤ kξk2 + kηk2 + 2 ξ η  ≤

 


2
≤ kξk2 + kηk2 + 2kξk · kηk = kξk + kηk ,
so we immediately get kξ + ηk ≤ kξk + kηk. 

Exercise 1. Use the above notations, and assume we have two vectors ξ, η 6= 0,
such that kξ +ηk = kξk+kηk. Prove that there exists some λ > 0 such that ξ = λη.
Lemma 3.1. Let X be a K-vector space, equipped with an inner product.
(ii) [Parallelogram Law] kξ + ηk2 + kξ − ηk2 = 2kξk2 + 2kηk2 .
(i) [Polarization Identities]
(a) If K = R, then

1
ξ  η = kξ + ηk2 − kξ − ηk2 , ∀ ξ, η ∈ X.

4
(b) If K = R, then
3

1X
ξ η =
 i−k kξ + ik ηk2 , ∀ ξ, η ∈ X.
4
k=0

Proof. (i). This is obvious, by the computations from the proof of Proposition
3.2, which give
kξ ± ηk2 = kξk2 + kηk2 ± 2Re ξ  η .


82 CHAPTER II: ELEMENTS OF FUNCTIONAL ANALYSIS

(ii).(a). In the real case, the above identity gives

kξ ± ηk2 = kξk2 + kηk2 ± 2 ξ  η ,


so we immediately get
kξ + ηk2 − kξ − ηk2 = 4 ξ  η .


(b). This case is a particular instance of the Polarization Identity discussed in

Remark 3.1. 

Corollary

3.1. Let X be a K-vector space equipped with an inner product
·  · . Then the map
X × X 3 (ξ, η) 7−→ ξ  η ∈ K


is continuous, with respect to the product topologies.

Proof. Immediate from the polarization identities. 


3.2. Let X and

Corollary
Y be two K-vector spaces equipped with inner
products ·  · X and ·  · Y . If T : X → K is an isometric linear map,
then
T ξ  T η Y = ξ  η X , ∀ ξ, η ∈ X.



Proof. Immediate from the polarization identities. 

Exercise 2. Let X be a normed K-vector space. Assume the norm

satisfies the
Parallelogram Law. Prove that there exists an inner product ·  · on X, such
that q 

kξk = ξ  ξ , ∀ ξ ∈ X.

Hint: Define the inner product by the Polarization Identity, and then prove that it is indeed an
inner product.

Proposition 3.3. Let X be a K-vector space, equipped with an inner product

·  · X . Let Z be the completion of X with respect to the norm defined by the

inner product. Then Z carries a unique inner product ·  · Z , so that




the norm
on Z is defined by ·  · Z . Moreover, this inner product extends ·  · X , in the

sense that
hξi  hηi Z = ξ  η X , ∀ ξ, η ∈ X.



Proof. It is obvious that the norm on Z satisfies the Parallelogram Law. We

then apply Exercise 2. 
Definitions. Let K be one of the fields R or C. A Hilbert space over K is a
K-vector space, equipped with an inner product, which is complete with respect to
the norm defined by the inner product. Some textbooks use the term Euclidean for
real Hilbert spaces, and reserve the term Hilbert only for the complex case.
Examples 3.1. For J a non-empty set, the space `2K (J) is a Hilbert space. We
know that this is a Banach space. The inner product defining the norm is

X
αβ = α(j)β(j), ∀ α, β ∈ `2K (J).
j∈J
 §3. Hilbert spaces 83

(The fact that function αβ : J → K belongs to `1K (J) is discussed in Section 1,

Exercise 5.)
More generally, a Banach space whose norm satisfies the Parallelogram Law is
a Hilbert space.

Definitions. Let X be a K-vector space, equipped with 
an inner product

·  · . Two vectors ξ, η ∈ X are said to be orthogonal, if ξ  η = 0. In this case

we write ξ ⊥ η. Given a set M ⊂ X, and a vector ξ ∈ X, we write ξ ⊥ M, if

ξ ⊥ η, ∀ η ∈ M.
Finally, two subsets M, N ⊂ X are said to be orthogonal, and we write M ⊥ N, if
ξ ⊥ η, ∀ ξ ∈ M, η ∈ N.
Notation. Let X be a vector space equipped with an inner product. For a
subset M ⊂ X, we define the set
M⊥ = {ξ ∈ X : ξ ⊥ M .

Remarks 3.2. Let X be a K-vector space equipped with an inner product.

A. The relation ⊥ is symmetric.
B. If ξ, η ∈ X satisfy ξ ⊥ η, then one has the Pythagorean Theorem:
kξ + ηk2 = kξk2 + kηk2 .


This is a consequence of the equality kξ + ηk2 = kξk2 + kηk2 + 2Re ξ  η .
C. If M ⊂ X is an arbitrary subset, then M⊥ is a closed linear subspace of
X. This follows from the linearity of the inner product in the second variable, and
from the continuity.
D. For sets M ⊂ N ⊂ X, one has
M⊥ ⊃ N ⊥ .
E. For any set M ⊂ X, one has

⊥
M⊥ = Span M ,
where Span M denotes the norm closure of the linear span of M. The inclusion

⊥
M⊥ ⊃ Span M
is trivial, since we have M ⊂ Span M. Conversely, if ξ ∈ M⊥ , then M ⊂ {ξ}⊥ . But
since {ξ}⊥ is a closed linear subspace, this gives
Span M ⊂ {ξ}⊥ ,

⊥
i.e. ξ ∈ Span M .
The following result gives a very interesting property of Hilbert spaces.
Proposition 3.4. Let H be a Hilbert space, let C ⊂ H be a non-empty closed
convex set. For every ξ ∈ H, there exists a unique vector ξ 0 ∈ C, such that
kξ − ξ 0 k = dist(ξ, C).
Proof. Denote dist(ξ, C) simply by d. By definition, we have
δ = inf kξ − ηk.
η∈C

Choose a sequence (ηn )n≥1 ⊂ C, such that limn→∞ kξ − ηn k = δ.

84 CHAPTER II: ELEMENTS OF FUNCTIONAL ANALYSIS

Claim: One has the inequality

kηm − ηn k2 ≤ 2kξ − ηm k2 + 2kξ − ηn k2 − 4δ 2 , ∀ m, n ≥ 1.
Use the Parallelogram Law
(5) 2kξ − ηm k2 + 2kξ − ηn k2 = k2ξ − ηm − ηn k2 + kηm − ηn k2 .
We notice that, since 21 (ηm + ηn ) ∈ C, we have
kξ − 12 (ηm + ηn )k ≥ δ,
so we get
k2ξ − ηm − ηn k2 = 4kξ − 12 (ηm + ηn )k2 ≥ 4δ 2 ,
so if we go back to (5) we get
2kξ − ηm k2 + 2kξ − ηn k2 = k2ξ − ηm − ηn k2 + kηm − ηn k2 ≥ 4δ 2 + kηm − ηn k2 ,
and the Claim follows.
Having proven the Claim, we now notice that, since limn→∞ kξ − ηn k = δ, we
immediately get the fact that the sequence (ηn )n≥1 is Cauchy. Since H is complete,
it follows that the sequence is convergent to some point ξ 0 . Since C is closed, it
follows that ξ 0 ∈ C. So far we have
kξ − ξ 0 k = lim kξ − ηn k = δ = dist(ξ, C),
n→∞
thus proving the existence.
Let us prove now the uniqueness. Assume ξ 00 ∈ C is another point such that
kξ − ξ 00 k = δ. Using the Parallelogram Law, we have
4δ 2 = 2kξ − ξ 0 k2 + kξ − ξ 00 k2 = k2ξ − ξ 0 − ξ 00 k2 + kξ 0 − ξ 00 k2 .
If ξ 0 6= ξ 00 , then we will have
4δ 2 > k2ξ − ξ 0 − ξ 00 k2 = 4kξ − 21 (ξ 0 + ξ 00 )k2 ,
so we have a new vector η = 12 (ξ 0 + ξ 00 ) ∈ C, such that
kξ − ηk < δ,
thus contracting the definition of δ. 
Definition. Let H be a Hilbert space, and let X ⊂ H be a closed linear
subspace. For every ξ ∈ H, using the above result, we let PX ξ ∈ X denote the
unique vector in X with the property
kξ − PX ξk = dist(ξ, X).
This way we have constructed a map PX : H → H, which is called the orthogonal
projection ont X.
The properties of the orthogonal projection are summarized in the following
result.
Proposition 3.5. Let H be a Hilbert space, and let X ⊂ H be a closed linear
subspace.
(i) For vectors ξ ∈ H and ζ ∈ X one has the equivalence
ζ = PX ξ ⇐⇒ (ξ − ζ) ⊥ X.

(ii) PX X = IdX .

(iii) The map PX : H → X is linear, continuous. If X 6= {0}, then kPX k = 1.
 §3. Hilbert spaces 85

(iv) Ran PX = X and Ker PX = X⊥ .

Proof. (i). “⇒.” Assume ζ = PX ξ. Fix an arbitrary vector η ∈ X r {0}, and
choose a number λ ∈ K, with |λ| = 1, such that

 

λ ξ − ζ  η =  ξ − ζ  η .
In particular, we have
 
 

 ξ − ζ  η  = Re ξ − ζ  λη .
Define the map F : R → R by
F (t) = kξ − ζ − tληk2 − kξ − ζk2 .
By the definition of ζ = PX ξ, we have
(6) F (t) > 0, ∀ t ∈ R r {0}.
2


Notice that F (t) = at + bt, ∀ t ∈ R, where a = λη  λη = kηk2 , and b =

 

2Re ξ − ζ λη = 2 ξ − ζ η . Of course, the property
   

at2 + bt > 0, ∀ t ∈ R r {0}



forces b = 0, so we indeed get ξ − ζ  η = 0.
“⇐.” Assume (ξ − ζ) ⊥ X. For any η ∈ X, we have (ξ − ζ) ⊥ (ζ − η), so using
the Pythagorean Theorem, we get
kξ − ηk2 = kξ − ζk2 + kζ − ηk2 ,
which forces
kξ − ηk ≥ kξ − ζk, ∀ η ∈ X.
This proves that
kξ − ζk = dist(ξ, X),
i.e. ζ = PX ξ.
(ii). This property is pretty clear. If ξ ∈ X, then 0 = ξ − ξ is orthogonal to X,
so by (i) we get ξ = PX ξ.
(iii). We prove the linearity of PX . Start with vectors ξ1 , ξ − 2 ∈ H and a scalar
λ ∈ K. Take ζ1 = PX ξ1 and ζ2 = PX ξ2 . Consider the vector ζ = λζ1 + ζ2 . For any
η ∈ X, we have



λξ1 + ξ2 − ζ  η = (λξ1 − λζ1 ) + (ξ2 − ζ2 )  η =





= λξ1 − λζ1  η + ξ2 − ζ2  η = λ ξ1 − ζ1  η + ξ2 − ζ2  η = 0.
By (i) we have (ξ1 − ζ1 ) ⊥ X and (ξ1 − ζ1 ) ⊥ X, so the above computation proves
that
(λξ1 + ξ2 − ζ) ⊥ X,
so using (i) we get
PX (λξ1 + ξ2 ) = ζ = λζ1 + ζ2 = λPX ξ1 + PX ξ2 ,
so PX is indeed linear.
To prove the continuity, we start with an arbitrary vector ξ ∈ H and we use
the fact that (ξ − PX ξ) ⊥ PX ξ. By the Pythagorean Theorem we then have
kξk2 = k(ξ − PX ξ) + PX ξk2 = kξ − PX ξk2 + kPX ξk2 ≥ kPX ξk2 .
In other words, we have
kPX ξk ≤ kξk, ∀ ξ ∈ H,
86 CHAPTER II: ELEMENTS OF FUNCTIONAL ANALYSIS

so PX is indeed continuous, and we have kPX k ≤ 1. Using (ii) we immediately get

that, when X 6= {0}, we have kPX k = 1.
(iv). The equality Ran PX = X is trivial by the construction of PX and by (ii).
If ξ ∈ Ker PX , then by (i), we have ξ ∈ X⊥ . Conversely, if ξ ⊥ X, then ζ = 0
satisfies the condition in (i), i.e. PX ξ = 0. 

Corollary 3.3. If H is a Hilbert space, and X ⊂ H is a closed linear subspace,

then
X + X⊥ = H and X ∩ X⊥ .
In other words the map
(7) X × X⊥ 3 (η, ζ) 7−→ η + ζ ∈ H
is a linear isomorphism.
Proof. If ξ ∈ H then PX ξ ∈ X, and ξ − PX ξ ∈ X⊥ , and then the equality
ξ = PX ξ + (ξ − PX ξ)
proves that ξ ∈ X + X⊥ . The equality X ∩ X⊥ = {0} is trivial, since for ζ ∈ X ∩ X⊥ ,
we must have ζ ⊥ ζ, which forces ζ = 0. 

Exercise 3. Let H be a Hilbert space.

(i) Prove that, for any closed subspace X ⊂ H, one has the equality
PX⊥ = I − PX .
(ii) Prove that two closed subspaces X, Y ⊂ H, the following are equivalent:
– X ⊥ Y;
– PX PY = 0;
– PY PX = 0.
(iii) Prove that two closed subspaces X, Y ⊂ H, the following are equivalent:
– X ⊂ Y;
– PX PY = PX ;
– PY PX = PX .
(iv) Let X, Y ⊂ H are closed subspaces, such that X ⊥ Y, then
– X + Y is c closed linear subspace of H;
– PX+Y = PX + PY .
Corollary 3.4. Let H be a Hilbert space, and let X ⊂ H be a linear (not
necessarily closed) subspace. Then on has the equality

⊥
X = X⊥ .

⊥ ⊥
Proof. Denote the closed subspace X⊥ by Z. Since X⊥ = X , by the
previous exercise we have
PZ = I − PX⊥ = I − PX⊥ = I − (I − PX ) = PX ,
which forces
Z = Ran PZ = Ran PX = X.

 §3. Hilbert spaces 87

Theorem 3.1 (Riesz’ Representation Theorem). Let H be a Hilbert space over

K, and let φ : H → K be a linear continuous map. Then there exists a unique
vector ξ ∈ H, such that
φ(η) = ξ  η , ∀ η ∈ H.


Moreover one has kξk = kφk.

Proof. First we show the existence. If φ = 0, we simply take ξ = 0. Assume
φ 6= 0. Define the subspace X = Ker φ. Notice that X is closed. Using the linear
isomorphism (7) we see that the composition
quotient map
X⊥ ,→ H −−−−−−−−→ H/X
is a linear isomorphism. Since
H/X = H/Ker φ ' Ran φ = K,
it follows that dim(X⊥ ) = 1. In other words, there exists ξ0 ∈ X⊥ , ξ0 6= 0, such
that
X⊥ = Kξ.
Start now with some arbitrary vector η ∈ H. On the one hand, using the equality
Kξ0 + X = H, there exists λ ∈ K and ζ ∈ X, such that
η = λξ0 + ζ,
and since ζ ∈ X = Ker φ, we get
φ(η) = φ(λξ0 ) = λφ(ξ0 ).
On the other hand, we have
ξ0  η = ξ0  λξ0 + ξ0  ζ = λkξ0 k2 ,




so if we define ξ = φ(ξ0 )kξ0 k−2 we will have

ξ  η = φ(ξ0 )kξ0 k−2 ξ0 | η = φ(ξ0 )kξ0 k−2 ξ0  η = λφ(ξ0 ) = φ(η).




To prove uniqueness, assume ξ 0 ∈ H is another vector with

φ(η) = ξ 0  η , ∀ η ∈ H.


In particular, we have
kξ − ξ 0 k2 = ξ − ξ 0 | ξ − ξ 0 = ξ | ξ − ξ 0 − ξ 0 | ξ − ξ 0 = φ(ξ − ξ 0 ) − φ(ξ − ξ 0 ) = 0,

which forces ξ = ξ 0 .
Finally, to prove the norm equality, we first observe that when ξ = 0, the
equality is trivial. If ξ 6= 0, then on the one hand, using C-B-S inequality we have
|φ(η)| =  ξ  η  ≤ kξk · kηk, ∀ η ∈ H,
 


so we immediately get kφk ≤ kξk. If we take the vector ζ = kξk−1 ξ, then kζk = 1,
and
φ(ζ) = ξ  kξk−1 ξ = kξk,


so we also have kφk ≥ kξk. 

In the remainder of this section we discuss a Hilbert space notion of linear

independence. This should be thought as a “rigid” linear independence.
88 CHAPTER II: ELEMENTS OF FUNCTIONAL ANALYSIS

Definition. Let X be a K-vector space, equipped with an inner product. A

set F ⊂ X is said to be orthogonal, if 0 6∈ F, and
ξ ⊥ η, ∀ ξ, η ∈ F, with ξ 6= η.
A set F ⊂ X is said to be orthonormal, if it is orthogonal, but it also satisfies:
kξk = 1, ∀ ξ ∈ F.
Remark that, if one starts with an orthogonal set F ⊂ X, then the set
F(1) = kξk−1 ξ : ξ ∈ F


is orthonormal.
Proposition 3.6. Let X be a K-vector space equipped with an inner product.
Any orthogonal set F ⊂ X is linearly independent.
Proof. Indeed, if one starts with a vanishing linear combination
λ1 ξ1 + · · · + λn ξn = 0,
with λ1 , . . . , λn ∈ K, ξ1 , . . . , ξn ∈ X, such that ξk 6= ξ` , for all k, ` ∈ {1, . . . , n} with
k 6= `, then for each k ∈ {1, . . . , n} we clearly have
λk kξk k2 = ξk  λ − 1ξ1 + · · · + λn ξn = 0,


and since ξk 6= 0, we get λk = 0. 

Lemma 3.2. Let X be a K-vector space equipped with an inner product, and let
F ⊂ X be an orthogonal set. Then there exists a maximal (with respect to inclusion)
orthogonal set G ⊂ X with F ⊂ G.
Proof. Consider the sets
A = G : G orthogonal subset of X ,


AF = G ∈ A : G ⊃ F ,


ordered with the inclusion. We are going to apply Zorn’s Lemma to AF . Let
T ⊂ AF be a subcollection, which is totally ordered, i.e. for any G1 , G2 ∈ T one has
G1 ⊂ G2 or G1 ⊃ G2 . Define the set
[
M= G.
G∈T

Since G ⊂ X r {0}, for all G ∈ T, it is clear that M ⊂ X r {0}. If ξ1 , ξ2 ∈ M

are vectors with ξ1 6= ξ2 , then we can find G1 , G2 ∈ T with ξ1 ∈ G1 and ξ2 ∈ G2 .
Using the fact that T is totally ordered, it follows that there is k ∈ {1, 2} such that
ξ1 , ξ2 ∈ Gk , so we indeed get ξ1 ⊥ ξ2 . It is now clear that M ∈ AF , and M ⊃ G, for
all G ∈ T. In other words, we have shown that every totally ordered subset of AF
has an upper bound, in AF . By Zorn’s Lemma, AF has a maximal element. Finally,
it is clear that any maximal element for AF is also a maximal element in A. 
Remark 3.3. Using the notations from the proof above, given an orthonormal
set M ⊂ X, the following are equivalent:
(i) M is maximal in A;
(ii) M is maximal in
A(1) = G : G orthonormal subset of X .

 §3. Hilbert spaces 89

The implication (i) ⇒ (ii) is trivial. Conversely, if M is maximal in A(1) , we use

the Lemma to find a maximal N ∈ A with N ⊃ M. But then N(1) is orthonromal,
and N(1) ⊃ M, which by the maximality of M in A(1) will force N(1) = M. Since N
is linearly independent, the relations
N(1) = M ⊂ N,
will force N = N(1) = M.
Comment. In linear algebra we know that a linearly independent set is max-
imal, if and only if it spans the whole space. In the case of orthogonal sets, this
statement has a version described by the following result.
Theorem 3.2. Let H be a Hilbert space, and let F be an orthogonal set in H.
The following are equivalent:
(i) F is maximal among all orthogonal subsets of H;
(ii) Span F is dense in H in the norm topology.
Proof. (i) ⇒ (ii). Assume F is maximal. We are going to show that Span F is
dense in H, by contradiction. Denote the closure Span F simply by X, and assume
X ( H. Since

⊥
X = X⊥ ,
we see that, the strict inclusion X ( H forces X⊥ 6= {0}. But now if we take a
non-zero vector ξ ∈ X⊥ , we immediately see that the set F ∪ {ξ} is still orthogonal,
thus contradicting the maximality of F.
(ii) ⇒ (i). Assume Span F is dense in H, and let us prove that F is maximal.
We do this by contardiction. If F is not maximal, then there exists ξ ∈ H r F, such
that F ∪ {ξ} is still orthogonal. This would force ξ ⊥ F, so we will also have
ξ ⊥ Span F.
But since Span F is dense in H, this will give ξ ⊥ H. In particular we have ξ ⊥ ξ,
which would force ξ = 0, thus contradicting the fact that F ∪ {ξ} is orthigonal.
(Recall that all elements of an orthigonal set are non-zero.) 
Definition. Let H be a Hilbert space An orthonormal set B ⊂ H, which is
maximal among all orthogonal (or orthonormal) subsets of H, is called an orthonor-
mal basis for H.
By Lemma ??, we know that given any orthonormal set F ⊂ H, there exists an
orthonormal basis B ⊃ F.
By the above result, an orthonormal set B ⊂ H is an orthonormal basis for H,
if and only if Span B is dense in H.
Example 3.2. Let J be a non-empty set. Consider the Hilbert space `2K (I).
For every j ∈ J, let δj : J → K be the function defined by

1 if k = j
δj (k) =
0 if k 6= j
If we consider the set B = {δj : j ∈ J}, then
Span B = finK (J),
which is dense in `2K (J). The above result then says that B is an orthonormal basis
for `2K (J).
90 CHAPTER II: ELEMENTS OF FUNCTIONAL ANALYSIS

Lemma 3.3. Let B be an orthonormal basis for the Hilbert space H, and let
F ( B be an arbitrary non-empty subset.
(i) F is an orthonormal basis for the Hilbert space Span F.

⊥
(ii) Span F = Span(B r F).
Proof. (i). This is clear, since F is orthonormal and has dense span.
(ii). Denote for simplicity Span F = X and Span(B r F) = Y. Since
ξ ⊥ η, ∀ ξ ∈ F, η ∈ B r F,
it is pretty obvious that X ⊥ Y. Since X + Y clearly contains Span B, it follows that
X + Y is dense in H. We know howver that X + Y is closed, so we have in fact the
equality
X + Y = H.
This will then give
I = PH = PX + PY ,
so we get
PY = I − PX = PX⊥ ,
so
X⊥ = Ran PX⊥ = Ran PY = Y.


Theorem 3.3. Let H be a Hilbert space, and let B be an orthonormal basis

for H, labelled1 as B = {ξj : j ∈ J}. For every vector η ∈ H, let αη : J → K be
the map defined by


η̂(j) = ξj  η , ∀ j ∈ J.
(i) For every η ∈ H, the map η̂ belongs to `2K (J).
(ii) The map
T : H 3 η 7−→ η̂ ∈ `2K (J)
is an isometric linear isomorphism.
Proof. (i). Fix for the moment η ∈ H. One way to prove that η̂ belongs to
`2K (J) is by showing that
X
|η̂(j)|2 : F ∈ Pfin (J) < ∞.

sup
j∈F

For any non-empty finite subset F ⊂ J, we define the subspace

HF = Span{ξj : j ∈ F },
and define the vector X 
ηF = ξj  η ) · ξj .
j∈F

Claim: For every finite set F ⊂ J, one has the equality

ηF = PHF η.

1 This notation implicitly assumes that ξ 6= ξ , for all j, k ∈ J with j 6= k.

j k
 §3. Hilbert spaces 91

It suffices to prove that

(η − ηF ) ⊥ HF .
But this
is obvious, since if we start with some k ∈ F , then using the fact that
ξk  ξj = 0, for all j ∈ F r {k}, together with the equality kξk k = 1, we get


X 





ξk  η − ηF = ξk  η − ξj  η · ξk  ξj = ξk  η − ξk  η · ξk  ξk = 0.
j∈F

Having proven the Claim, let us observe that, since the terms in the sum that
defines ηF are all orthogonal, we get
ξj  η · ξj 2 =  ξj  η 2 · kξj k2 =
X X X
kηF k2 = |η̂(j)|2 .




j∈F j∈F j∈F

Combining this computation with the Claim, we now have

X
|η̂(j)|2 = kηF k2 = kPHF ηk2 ≤ kηk2 ,
j∈F

which proves that

X
|η̂(i)|2 : F ∈ Pfin (J) < kηk2 .

sup
j∈F

(ii). The linearlity of T is obvious. The above inequality actually proves that
kT ηk ≤ kηk, ∀ η ∈ H.
We now prove that in fact T is isometric. Since T is linear and continuous, it
suffices to prove that T Span B is isometric. Start with some vector η ∈ Span B,


P there exists some finite set F ⊂ J, and scalars (λk )k∈F ⊂ K,

which means that
such that η = k∈F λk ξk . Remark that


X
λk if k ∈ F
ξj | η = λk ξ j | ξ j =
0 if k ∈ F
k∈F

so the element η̂ = T η ∈ `2K (J) is defined by


λk if k ∈ F
η̂(k) =
0 if k ∈ F
This gives
X 
X X
kηk2 = λj λ k ξ j  ξ k = |λk |2 = |η̂(k)|2 = kη̂k2 ,
j,k∈F k∈F k∈F

so we indeed get
kηk = kT ηk, ∀ η ∈ Span B.
Let us prove that T is surjective. Notice that, the above computation, applied to
singleton sets F = {k}, k ∈ J, proves that
T ξk = δk , ∀ k ∈ J.
In particular, we have
Ran T ⊃ T Span B = Span T (B) =

= Span{T ξk : k ∈ J} = Span{δk : k ∈ J} = finK (J),

which proves that Ran T is dense in `2K (J). We know however that T is isometric,
so Ran T ⊂ `2K (J) is closed. This forces Ran T = `2K (J). 
92 CHAPTER II: ELEMENTS OF FUNCTIONAL ANALYSIS

Corollary 3.5 (Parseval Identity). Let H be a Hilbert space, and let B =

{ξj : j ∈ J} be an orthonormal basis for H. One has:

X 

ζ  ξj · ξj  η , ∀ ζ, η ∈ H.


ζ η =
j∈I



Proof. If we define α(j) = ξj  ζ and ξj  η , ∀ j ∈ I, then by construction
we have α = T ζ and β = T η. Using the fact that T is isometric, the right hand
side of the above equality is the equal to
X 



α(j)β(j) = α  β = T ζ  T η = ζ  η .
j∈J


Notation. Let H be a Hilbert space, let B = {ξj : j ∈ J} be an orthonormal
basis for H, and let T : H → `2K (J) be the isometric linear isomorphism defined in
the previous theorem. Given an element α ∈ `2K (J), we denote the vector T −1 α ∈ H
by X
α(j)ξj .
j∈J

The summation notation is justified by the following fact.

Proposition 3.7. With the above notations, the net (ηF )F ∈Pfin (J) ⊂ H of
partial sums, defined by
X
ηF = α(k)ξk , ∀ F ∈ Pfin (J),
k∈F
P
is convergent in the norm topology to the vector j∈J α(j)ξj .
P
Proof. Define the vector η = j∈I α(j)ξj . By construction we have T η = α.
Likewise, if we define, for each finite set F ⊂ J, the element αF ∈ `2K (J) by

α(k) if k ∈ F
αF (k) =
0 if k ∈ J r F
then T −1 αF = ηF . Using the fact that T is an isometry, we have
kη − ηF k = kT η − T ηF k = kα − αF k,
and the desired property follows from the well-known properties of `2K (J). 
Exercise 4. Let H be a Hilbert space, let F = {ξj : j ∈ J} be an orthonormal
set. Define the closed linear subspace HF = Span F. Prove that the orthogonal
projection PHF is defined by
X
ξj  η ξj , ∀ η ∈ H.


PHF η =
j∈J

Hints: Extend F to an orthonormal basis B. Let B be labelled as {ξi : i ∈ I} for some set
I ⊃ J. First prove that for any η ∈ H, the map β η = T η J belongs to `2K (J). In particular, the
sum X 

ηF = ξj  η ξ j
j∈J
is “legitimate” and defines an element in HF (use the fact that F is an orthonormal basis for HF ).
Finally, prove that (η − ηF ) ⊥ F, using Parseval Identity.
 §3. Hilbert spaces 93

The next result discusses the appropriate notion of dimension for Hilbert spaces.
Theorem 3.4. Let H be a Hilbert space. Then any two orthonormal bases of
H have the same cardinality.
Proof. Fix two orthonormal bases B and B0 . There are two possible cases.
Case I: One of the sets B or B0 is finite.
In this case H is finite dimensional, since the linear span of a finite set is
automatically closed. Since both B and B0 are linearly independent, it follows that
both B and B0 are finite, hence their linear spans are both closed. It follows that
Span B = Span B0 = H,
so B and B0 are in fact linear bases for H, and then we get
Card B = Card B0 = dim H.
Case II: Both B and B0 are infinite.
The key step we need in this case is the following.
Claim 1: There exists a dense subset Z ⊂ H, with
Card Z = Card B0 .
To prove this fact, we define the set
X = SpanQ B0 .
It is clear that
Card X = Card B0 .
Notice that X is dense in SpanR B0 . If we work over K = R, then we are done. If
we work over K = C, we define
Z = X + iX,
and we will still have
Card Z = Card X = Card B0 .
Now we are done, since clearly Z is dense in SpanC B0 .
Choose Z as in Claim 1. For every ξ ∈ B we choose a vector ζξ ∈ Z, such that
√
2−1
kξ − ζξ k ≤ .
2
Claim 2: The map B 3 ξ 7−→ ζξ ∈ Z is injective.
Start with two vectors ξ1 , ξ2 ∈ B, such that ξ1 6= ξ2 . In particular, ξ1 ⊥ ξ2 , so we
also have ξ1 ⊥ (−ξ2 ), and using the Pythogorean Theorem we get
kξ1 − ξ2 k2 = kξ2 k2 + k − ξ2 k2 = 2,
which gives √
kξ1 − ξ2 k = 2.
Using the triangle inequality, we now have
√ √
2 = kξ1 − ξ2 k ≤ kξ1 − ζξ1 k + kξ2 − ζξ2 k + kζξ1 − ζξ2 k ≤ 2 − 1 + kζξ1 − ζξ2 k.
This gives
kζξ1 − ζξ2 k ≥ 1,
which forces ζξ1 6= ζξ2 .
94 CHAPTER II: ELEMENTS OF FUNCTIONAL ANALYSIS

Using Claim 2, we have constructed an injective map B → Z. In particular,

using Claim 1 and the cardinal arithmetic rules, we get
Card B ≤ Card Z = Card B0 .
By symmetry we also have
Card B0 ≤ Card B,
and then using the Cantor-Bernstein Theorem, we finally get
Card B = Card B0 .

Corollary 3.6 (of the proof). A Hilbert space is separable, in the norm topol-
ogy, if and only of it has an orthonormal basis which is at most countable.
Proof. Use Claims 1 and 2 from the proof of the Theorem. 
Definition. Let H be a Hilbert space, and let B be an orthonormal basis for
H. By the above theorem, the cardinal number Card B does not depend on the
choice of B. This number is called the hilbertian (or orthogonal ) dimension of H,
and is denoted by h-dim H.
Corollary 3.7. For two Hilbert spaces H and H0 , the following are equivalent:
(i) h-dim H = h-dim H0 ;
(ii) There exists an isometric linear isomorphism U : H → H0 .
Proof. (i) ⇒ (ii). Choose a set I with h-dim H = h-dim H0 = Card I.
Apply Theorem ?? to produce isometric linear isomorphisms T : H → `2K (I) and
T 0 : H0 → `2K (I). Then define U = T 0−1 ◦ T .
(ii) ⇒ (i). Assume one has an isometric linear isomorphism U : H → H0 .
Choose an orthonormal basis B for H. Then U (B) is clearly and orthonormal basis
for H0 , and since U : B → U (B) is bijective, we get
h-dim H = Card B = Card U (B) = h-dim H0 .