SMSTC (2023/24)

Pure Analysis: Functional Analysis

www.smstc.ac.uk

Contents

1 Normed Linear Spaces and Inner Product Spaces 1–1

1.1 Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–1
1.2 Inner products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–2
1.3 Geometry of Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–3
1.4 Orthonormal bases in Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–5
1.5 Spaces defined by families of norms or seminorms . . . . . . . . . . . . . . . . . . . . . . . 1–7
1.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–9

(i)
 SMSTC (2023/24)
Pure Analysis: Functional Analysis
Lecture 1: Normed Linear Spaces and Inner Product Spaces
Notes by: Tony Carbery, University of Edinburgh
Instructor: Justin Forlanoa

www.smstc.ac.uk

Contents
1.1 Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–1
1.2 Inner products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–2
1.3 Geometry of Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–3
1.4 Orthonormal bases in Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . 1–5
1.5 Spaces defined by families of norms or seminorms . . . . . . . . . . . . . . . 1–7
1.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1–9

1.1 Norms
Let X be a real or complex vector space. We say that a function

k·k:X →R

is a norm if it satisfies the following conditions:

• kxk ≥ 0 for all x ∈ X, and kxk = 0 if and only if x = 0

• kλxk = |λ| kxk for all vectors x and all scalars λ
• kx + yk ≤ kxk + kyk for all x, y ∈ X

This last condition is called the triangle inequality. The pair (X, k · k) is called a normed linear space
or normed space for short.

The most familiar examples of norms on vector spaces are when X = Rn or Cn and

kxkp := {|x1 |p + |x2 |p + · · · + |xn |p }1/p

for 1 ≤ p < ∞. When p = ∞ we let kxk∞ = max1≤j≤n |xj |; this again defines a norm. When 0 < p < 1,
kxkp does not define a norm.

More interesting are the infinite dimensional spaces `p (1 ≤ p < ∞) consisting of those sequences (xj )
such that  1/p
X
kxkp :=  |xj |p  < ∞;
j

once again, k · kp does define a norm on `p . When p = ∞ we make the usual modification: `∞ consists of
all bounded sequences and the norm is kxk∞ := supj |xj |. Note that we are not specifying the indexing
set that j belongs to: it could be N, Z, Q, Zk or indeed any countable set.
a [email protected]

1–1
SMST C: Pure Analysis: Functional Analysis 1–2

There are also the “continuous” versions of these spaces. Let (X, M, µ) be a measure space and let, as
in Measure & Integration Lecture 6, Lp be the space of measurable (extended-)real valued functions f
such that Z 1/p
kf kp := |f |p dµ <∞
X

(with the corresponding modification for L∞ – see M & I Lecture 7). Once again, these define norms
on Lp upon making the accommodation that f = 0 in Lp if and only if f (x) = 0 a.e. dµ. In fact, these
spaces generalise the discrete `p spaces, simply by taking counting measure on a countable indexing set.
There are also easy modifications when we consider complex-valued functions rather than real-valued
functions.
Underlying the triangle inequality for Lp is the Hölder inequality (M & I Theorem 6.2) and ultimately
Young’s inequality (M & I Lemma 6.1).
There is another, different, generalisation when p = ∞. Let X be a compact Hausdorff space (or if
you prefer, simply a compact metric space) and let C(X ) be the space of continuous complex-valued
functions defined on X . By elementary analysis (compactness of X ), each such function is bounded and
achieves its bounds. So for f ∈ C(X ), kf k∞ := maxt∈X |f (t)| = supt∈X |f (t)| is finite, and defines a norm
(the uniform norm) on C(X ). This is not to be confused with the L∞ norm if X happens also to be a
measure space. If X is only locally compact rather than compact – for example if X is R or Z – then the
natural space to consider is the space C0 (X ) of continuous functions tending to zero at infinity. Thus,
for example, C0 (Z) = c0 (Z) is just the space of two-tailed sequences (aj ) such that aj → 0 as j → ±∞.
A further example, closely related to L1 , is the space M(X ) of complex Borel measures
R on a compact
metric space X . Here, the norm of a measure µ is given by its total variation kµk = X d|µ|(x). See M
& I Lemma 4.1 and Definition 4.2.
The virtue of normed spaces is that one is automatically able to talk about topological and indeed metric
notions:

Lemma 1.1 Let (X, k · k) be a normed space. Let d(x, y) = kx − yk. Then d defines a metric on X, so
that (X, d) is a metric space.

The proof is immediate. Notice that the metric is translation-invariant, which means that d(x+z, y+z) =
d(x, y) for all x, y and z ∈ X. Thus the family of all open sets containing z ∈ X coincides with the translate
by z of the family of all open sets containing 0.
As we all know, the best metric spaces on which to do analysis are the complete metric spaces – i.e.
those in which every Cauchy sequence is convergent. All of undergraduate analysis is based upon the
fact that R with the usual metric is complete – i.e. upon the Cauchy criterion for R. This in turn is
equivalent to the supremum axiom, or the assertion that bounded monotonic sequences are convergent.
The completeness of R can be jacked up to give completeness for many other metric spaces – including all
of the normed spaces we have discussed above. It is instructive to pin down exactly where completeness
of R is used to establish the completeness of any of these spaces, such as Lp . A complete normed space
is called a Banach space.

1.2 Inner products

The case p = 2 of the Lp spaces is special, because here the norm arises from an inner product, and so
the geometry of L2 closely resembles that of a finite-dimensional euclidean space.
If X is a complex vector space, an inner product on X is a sesquilinear map

h·, ·i : X × X → C

satisfying

• hx, yi = hy, xi for all x, y ∈ X

• hx, xi ≥ 0 for all x ∈ X, and hx, xi = 0 if and only if x = 0.
SMST C: Pure Analysis: Functional Analysis 1–3

(“Sesquilinear” means linear in the first variable and conjugate-linear in the second. In the case of a real
vector space, things are easier and the map is required to be bilinear, symmetric and positive definite.)

If X = L2 we can define an inner product on X by

Z
hf, gi = f g.

Note that this is well-defined because of the Cauchy–Schwarz inequality (M & I Theorem 6.2, case p = 2),
and it is easy to check the required properties for an inner product.
p
We now claim that any inner product space can be regarded as a normed space by defining kxk = hx, xi.
The only difficulty is verification of the triangle inequality. As in the case of L2 , we proceed by first
establishing the abstract Cauchy–Schwarz inequality:

Lemma 1.2 Let X be an inner product space. Then, for all x, y ∈ X,

|hx, yi| ≤ kxkkyk.

Proof If x or y is zero there is nothing to prove. By homogeneity we may assume that kxk = kyk = 1.
Consider
0 ≤ hx − y, x − yi = hx, xi − 2<hx, yi + hy, yi = 2 − 2<hx, yi.
So
<hx, yi ≤ 1.
But the same is true replacing x by αx where α is any complex number of modulus 1. Choosing α to
make hαx, yi = |hx, yi| we are done. 

We give another, more geometric proof, below. With the Cauchy–Schwarz inequality proved, the triangle
inequality is now immediate:

Lemma 1.3 Let X be an inner product space. Then, for all x, y ∈ X,

kx + yk ≤ kxk + kyk.

Proof Squaring,
2
kx + yk2 = hx + y, x + yi = kxk2 + 2<hx, yi + kyk2 ≤ kxk2 + 2kxkkyk + kyk2 = (kxk + kyk) .

Now all of our previous discussion about normed spaces applies also to inner product spaces. In particular,
if the metric induced in turn by the inner product and norm is complete, we call the inner product space
a Hilbert space.

1.3 Geometry of Hilbert spaces

As mentioned above, the geometry of inner product spaces is analogous to classical euclidean geometry.
In particular there is a notion of orthogonality between two vectors (that is, when hx, yi = 0) and the
notion of the angle between vectors (implicit in the Cauchy–Schwarz inequality). We have

• Pythagoras’ theorem: If x and y are orthogonal then kx + yk2 = kxk2 + kyk2

• the parallelogram law: kx + yk2 + kx − yk2 = 2kxk2 + 2kyk2 for all x and y
• the polarisation identity:
1
kx + yk2 − kx − yk2 + ikx + iyk2 − ikx − iyk2


hx, yi =
4
(and in the case of a real inner product space the last two terms on the right-hand side are omitted).
SMST C: Pure Analysis: Functional Analysis 1–4

Given two vectors x and y 6= 0 in an inner product space, it is natural to try to write x as a “component”
parallel to y plus a “component” perpendicular to y. Thus we write x = λy + (x − λy), with the scalar
λ to be chosen appropriately. While λy is parallel to y no matter what λ is, we have that (x − λy) is
perpendicular to y if and only if hx − λy, yi = hx, yi − λkyk2 = 0, i.e. if and only if λ = hx,yi
kyk2 . Thus we
have
hx, yi hx, yi
x= y + (x − y) (1.1)
kyk2 kyk2
where the two summands on the right are orthogonal. The first term on the right, hx,yi
kyk2 y, is called the
projection of x in the direction determined by y. (When y is a unit vector, this reduces to the familiar
hx, yiy.)
hx,yi
Letting λ0 = kyk2 , (1.1) implies that for arbitrary λ,

x − λy = (x − λ0 y) − (λ − λ0 )y (1.2)

where once again the two summands on the right-hand side are orthogonal. So by Pythagoras’ theorem,

kx − λyk2 = kx − λ0 yk2 + |λ − λ0 |2 kyk2 . (1.3)

This equation contains a lot of information:

• Setting λ = 0, we see that kxk2 ≥ |λ0 |2 kyk2 , which upon rearrangement gives

|hx, yi|2 ≤ kxk2 kyk2 ,

thus yielding another proof of the Cauchy–Schwarz inequality

• The function λ 7→ kx − λyk achieves it minimum when λ = λ0 ; that is, the closest point to the
vector x in the linear span of y is precisely the projection of x in the direction determined by y
• the function λ 7→ kx − λyk achieves it minimum when λ = 0 if and only if x and y are orthogonal.

The fact that a one-dimensional subspace Y of an inner product space X contains a unique closest point
to any x ∈ X \ Y can be generalised to closed subspaces of Hilbert spaces. This will be the first time we
use the completeness property of a Hilbert space.

Theorem 1.1 Let H be a Hilbert space and C a nonempty closed convex subset of H. Let x ∈ H \ C.
Then there is a unique y ∈ C such that

kx − yk = inf kx − zk := d(x, C).

z∈C

Proof Without loss of generality x = 0. Let yn ∈ C be such that kyn k → d(0, C) := d. Then, by the
parallelogram law, and by convexity of C,
ym + yn 2
kym − yn k2 = 2kym k2 + 2kyn k2 − 4 k k ≤ 2kym k2 + 2kyn k2 − 4d2
2
which tends to zero as m, n → ∞. So (yn ) is a Cauchy sequence and thus converges to some y ∈ C with
minimal distance from 0. If z were another such, we would have
y+z 2
ky − zk2 = 2kyk2 + 2kzk2 − 4 k k ≤ 2d2 + 2d2 − 4d2 = 0.
2


We now wish to proceed towards defining the projection of x ∈ H onto any closed subspace of H. We
first need a definition.

Definition 1.1 Let H be a Hilbert space, and let A ⊆ H be any nonempty subset of H. We define

A⊥ = {x ∈ H : hx, ai = 0 for all a ∈ A}.

SMST C: Pure Analysis: Functional Analysis 1–5

Lemma 1.4 Let A ⊆ H. Then A⊥ is a closed linear subspace of H.

Proof That it is a subspace is clear. If xn ∈ A⊥ , with xn → x ∈ H, then, by the Cauchy–Schwarz

inequality,
|hx, ai − hxn , ai| = |hx − xn , ai| ≤ kx − xn kkak
which tends to zero as n → ∞. Since hxn , ai = 0 for all a ∈ A we conclude that hx, ai = 0 for all a ∈ A
too. Thus x ∈ A⊥ . 
If X is a subspace of H it is clear that every vector in X is perpendicular to every vector in X ⊥ . In such
a case we call X ⊥ the orthogonal complement of X. We now show that the direct sum of a closed
subspace X and its orthogonal complement is the whole of H:

Theorem 1.2 (Orthogonal Decomposition Theorem) Let H be a Hilbert space and let X be a closed
linear subspace of H. Then
H = X ⊕ X ⊥;
that is, for every h ∈ H, there exists a unique x ∈ X and x⊥ ∈ X ⊥ such that h = x + x⊥ .

(Note that we have already seen the special case of this when X is one-dimensional.)

Proof If h ∈ X there is nothing to prove. If h ∈ / X, let x be the closest point in X to h (using Theorem
1.1). Let x⊥ = h − x. Then for all λ ∈ C, all z ∈ X, kx⊥ k ≤ kx⊥ + λzk. Thus, by the third consequence
of (1.3), hx⊥ , zi = 0 for all z ∈ X, and so x⊥ ∈ X ⊥ . The uniqueness assertion is easy. 
The x and x⊥ obtained in this theorem are called the orthogonal projections of h ∈ H onto X and X ⊥
respectively. Thus, if X is a closed subspace of a Hilbert space H, we have the canonical orthogonal
projection maps
ΠX : H → H, ΠX (h) = x
and
ΠX ⊥ : H → H, ΠX ⊥ (h) = x⊥
of H onto X and X ⊥ respectively.

1.4 Orthonormal bases in Hilbert spaces

A subset {eα : α ∈ A} of an inner product space H is called an orthonormal set if keα k = 1 for all
α ∈ A and if heα , eβ i = 0 for all α 6= β. A familiar example of an orthonormal set in `2 (Z) is given by
{ej : j ∈ Z} where ej = (. . . , 0, 0, 1, 0, 0 . . . ) with the 1 in the j’th place. Another example in the complex
space L2 (T) is given by ej (t) = e2πijt for j ∈ Z. Orthonormal sets play the same role in an inner product
space as do the standard unit basis vectors in Rn or Cn .

If {eα : α ∈ A} is an orthonormal set in H, and x ∈ H, then {hx, eα i} are called the Fourier coefficients
of x with respect to {eα : α ∈ A}. We also denote hx, eα i by x
b(α).

Note that if Pan orthonormal set

n P{e
n
1 , . . . , en } is given and if we choose x ∈
/ span {e1 , e2 , . . . en }, then
en+1 := (x− j=1 x b(j)ej )/kx− j=1 x b(j)ej k is a unit vector perpendicular to e1 , . . . en . This observation
is the basis of the Gram-Schmidt procedure for iteratively constructing orthonormal sets in an inner
product space.

Proposition 1.1 Let {e1 , . . . , en } be a finite orthonormal set in an inner product space H. Then, for
all x ∈ H, for all aj ∈ C,
n
X n
X n
X
kx − aj ej k2 = kx − b(j)ej k2 +
x b(j)|2 .
|aj − x (1.4)
j=1 j=1 j=1

Equation (1.4) is analogous to (1.3), is proved similarly, and has similarly rich consequences.
SMST C: Pure Analysis: Functional Analysis 1–6

Proof Multiplying out the left-hand side gives

n
X n
X n
X
kx − aj ej k2 = kxk2 + |aj |2 − 2< b(j)aj .
x (1.5)
j=1 j=1 j=1

Applying this with aj = x

b(j) gives
n
X n
X n
X n
X
kx − b(j)ej k2 = kxk2 +
x x(j)|2 − 2
|b x(j)|2 = kxk2 −
|b x(j)|2 .
|b (1.6)
j=1 j=1 j=1 j=1

Subtracting (1.6) from (1.5), we get

n
X n
X n
X n
X n
X n
X
2 2 2 2
kx − aj ej k − kx − b(j)ej k =
x |aj | − 2< b(j)aj +
x |b
x(j)| = b(j)|2 .
|aj − x
j=1 j=1 j=1 j=1 j=1 j=1


Pn
The proposition can also be observed by noticing that if we set P x = j=1 x b(j)ej and if we let y ∈
lin{ej }, then x − P x and P x − y are orthogonal. We may then apply Pythagoras’ theorem.

Corollary 1.1 Let {e1 , . . . , en } be a finite orthonormal set in an inner product space H, and let x ∈ H.
Pn
(i) The closest point to x in the linear span of {e1 , . . . , en } is j=1 x b(j)ej .
Pn
(ii) x is in the linear span of {e1 , . . . , en } if and only if x = j=1 x b(j)ej .
n n n
(iii) kxk2 = kx − j=1 x b(j)ej k2 + j=1 |b x(j)|2 = d(x, lin{ej })2 + j=1 |b x(j)|2 .
P P P

Proof Part (i) is obvious from (1.4) as is part (ii). Part (iii) follows by taking aj = 0 for all j. 

Corollary 1.2 (Bessel’s inequality) Let {eα : α ∈ A} be an orthonormal set in an inner product space
H. Then, for all x ∈ H, X
|hx, eα i|2 ≤ kxk2 .
α∈A
P P
Proof Recall that an uncountable sum of nonnegative terms α∈A aα is defined to be supA0 ⊆A α∈A0 aα
where the “sup” is taken over all finite subsets A0 of A. So it suffices to prove that for any finite or-
thonormal set {e1 , . . . , en } we have
Xn
x(j)|2 ≤ kxk2 ;
|b
j=1

this is an immediate consequence of (1.4) upon taking all the aj to be zero. 

Suppose now that {e1 , e2 , . . . } is a countable orthornormal set in a Hilbert

Pn space H. Let En = lin{e1 , . . . , en },
E be the closed linear span of {e1 , e2 , . . . } and, for x ∈ H, let xn = j=1 x b(j)ej . By Corollary 1.1, part
(iii),
n
X n
X ∞
X
kxk2 = kx − xn k2 + x(j)|2 = d(x, En )2 +
|b x(j)|2 = d(x, E)2 +
|b x(j)|2
|b (1.7)
j=1 j=1 j=1

by Pletting n → ∞. Now, by Bessel’s inequality, (xn ) is a Cauchy sequence in H which therefore converges
∞
to j=1 x b(j)ej ∈ E; moreover we have that x ∈ E if and only if d(x, E) = 0 if and only if d(x, En ) → 0
if and only if (using (1.7))
X∞
x= x
b(j)ej
j=1
if and only if
∞
X
kxk2 = x(j)|2 .
|b (1.8)
j=1

Equation (1.8) is called Parseval’s formula.

This discussion leads to the following definition:
SMST C: Pure Analysis: Functional Analysis 1–7

Definition 1.2 A countable orthonormal set {e1 , e2 . . . } is called a complete orthonormal set or an
orthonormal basis for a Hilbert space H if the closed linear span of {e1 , e2 . . . } is all of H.

Note that the use of the word “complete” here has nothing at all to do with the notion of completeness
of a metric space; it is a historical accident that the same word is used to describe two different things.
With the above discussion we have:

Theorem 1.3 (Riesz–Fischer Theorem) Let H be a Hilbert space and let {e1 , e2 . . . } be a countable
orthonormal set in H. Then the following are equivalent:
(i) {e1 , e2 . . . } is an orthonormal basis for H.
(ii) For all x ∈ H,
∞
X
x= x
b(j)ej .
j=1

(iii) For all x ∈ H, Parseval’s formula

∞
X
kxk2 = x(j)|2
|b
j=1

holds.
(iv) x
b(j) = 0 for all j implies x = 0.

Proof We have already seen the equivalence of the first three statements. If (iii) or (ii) holds, and if
b(j) = 0 for all j, then clearly x = 0, so (iv) holds. If (i) fails to hold, there will be a nonzero x ∈ H which
x
is perpendicular to the closed linear span of {e1 , e2 . . . }, by the Orthogonal Decomposition Theorem 1.2,
showing that (iv) fails too. 

Theorem 1.4 Let H be a Hilbert space with an orthonormal basis {e1 , e2 . . . }. Then the map

x 7→ (b
x(j))j

defines an isometric isomorphism of H onto `2 .

Proof That the map preserves the norm structure is the Parseval identity, (and that it also preserves
the inner product structurePis an exercise below). The map is clearly linear, and is surjective
P∞ since if
n
(an ) ∈ `2 , the partial sums j=1 aj ej form a Cauchy sequence in H, and so converge to j=1 aj ej whose
Fourier coefficients are easily verified to be (aj ). 

In `2 the standard basis vectors give an orthonormal basis; more interesting is the following:

Proposition 1.2 In the complex Hilbert space L2 (T), ej (t) := e2πijt for j ∈ Z forms a complete or-
thonormal set.

Proof We have already seen from last semester’s lecture on Fourier Analysis, Theorem 1.3 and Corollary
1.4, that the second and third conditions of Theorem 1.3 are satisfied.


1.5 Spaces defined by families of norms or seminorms

In analysis we frequently need to consider vector spaces of functions which do not arise naturally as
normed spaces, but instead whose metric or topological properties are defined by a family of norms or
even seminorms. A seminorm satisfies the same properties as a norm except that we do not assume
that kxk = 0 imples x = 0. For example, the space Lp of genuine measurable functions whose Lp norm
is finite is a seminormed space, and measurable functions which are zero almost everywhere have “norm”
SMST C: Pure Analysis: Functional Analysis 1–8

zero; recall that we turn this into the genuine normed space Lp by identifying functions that are equal
almost everywhere. See M & I Lecture 6. However, this example is not typical for what follows. A more
pertinent example for us is the seminorm kf k = supt∈[0,1] |f 0 (t)| defined on the space C 1 ([0, 1]). Another,
which will feature strongly in Lecture 4, is the map x 7→ |f (x)| where f : X → R or C is a continuous
linear map (see Lecture 2).
Let us consider another example.
Example 1. Let X be the vector space of continuous complex-valued functions defined on R, X = {f :
R → C : f is continuous}. Then, for n ∈ N, we can define a seminorm kf kn := supt∈[−n,n] |f (t)|. From
this we define a “semimetric” ρn (f, g) = kf −gkn as usual, (the term semimetric here meaning that ρn has
all the properties of a metric save that ρn (f, g) = 0 does not imply that f = g). Note that an equivalent
semimetric is given by dn (f, g) = ρn (f, g)/(1 + ρn (f, g)). (Equivalent here means that ρn (fj , f ) → 0 as
j → ∞ if and only if dn (fj , f ) → 0 as j → ∞.) Finally, let
∞
X
d(f, g) = 2−n dn (f, g).
n=1

Then (X, d) is a genuine metric space, and we claim that fj converges to f in this metric space if and
only if for every compact set K, fj converges uniformly on K to f .
To see this, suppose first that d(fj , f ) → 0. Then for all n we have dn (fj , f ) → 0, and hence ρn (fj , f ) → 0
as j → ∞. So, for all n, fj converges to f uniformly on P∞ [−n, n]. Conversely, suppose that for every n,
ρn (fj , f ) → 0 as j → ∞. Let  > 0. Choose N so that j=N +1 2−n < /2. So there is a J such that for
j ≥ J and 1 ≤ n ≤ N we have ρn (fj , f ) < /2N . Hence for j ≥ J,
∞
X N
X ∞
X
d(fj , f ) = 2−n dn (fj , f ) ≤ ρn (fj , f ) + 2−n < N /2N + /2 = ,
n=1 n=1 n=N +1

as required.
The construction of Example 1 works whenever we have a countable family of seminorms k·kn on a vector
space X, leading to a metric d on X such that d(xj , x) → 0 if and only if for each n, kxj − xkn → 0.
In order to show that d(x, y) = 0 only for x = y we need the additional property that for every nonzero
x ∈ X there is some n with kxkn 6= 0. This property is described by saying that the family of seminorms
is sufficient or separating for X, (and makes sense whether or not the family of seminorms is countable).
Example 2. Let D = Cc∞ be the vector space consisting of the complex-valued C ∞ functions defined on
R which have compact support. This space is called the space of test functions. [It is not immediately
2
obvious that D is nonempty; one sees this by suitably modifying the function e−1/t . Once one has this,
it is not hard to see that the space D is in fact quite large: indeed it is dense in Lp (R) for 1 ≤ p < ∞.
For all this see M & I Lecture 6.]
For each n ∈ N and each i = 0, 1, 2, . . . , we define a seminorm k · kn,i on D by

kf kn,i = sup |f (i) (t)|.

t∈[−n,n]

This family is obviously sufficient for D. So we can put a metric d on D such that fj converges to f in
(i)
this metric if and only if for every compact set K and for every i, fj converges uniformly to f (i) on K.

In the setting of this discussion, the fact that xj converges to x in the metric space if and only if
kxj − xkn → 0 for each n means that all the seminorms k · kn are continuous. In particular, for every
finite subset S of N, and for every choice of rα > 0, the sets ∩α∈S {x ∈ X : kxkα < rα } are open sets in
the metric space. In fact any open set containing 0 will itself contain such a set. This collection of “basic
open neighbourhoods of 0” thus determines completely the collection of open sets for the metric space
via translation.
In Lecture 4 we shall need to examine the situation when on a vector space X we are given a family of
seminorms k · kα , with α belonging to some uncountable indexing set A. We shall assume that the family
of seminorms is sufficient for X, i.e that for every non-zero x ∈ X, there exists some α with kxkα 6= 0.
SMST C: Pure Analysis: Functional Analysis 1–9

We cannot now proceed by defining a metric on X as above, but we want nevertheless to be able to put
a “topology” on X which has the property that all of the seminorms k · kα are continuous. To do this,
we proceed directly to describe the “open sets” of the topology. With the discussion above in mind, it is
very natural to define a basic open neighbourhood of 0 to be a set of the form ∩α∈A0 {x ∈ X : kxkα < rα }
for some finite subset A0 of A and some choice of rα > 0. A basic open neighbourhood of ξ ∈ X is defined
to be a translate by ξ of a basic open neighbourhood of 0. A set U is defined to be open if for each ξ ∈ U
there is a basic open neighbourhood of ξ which is contained in U .

In this setting, the notion of convergence is a bit more tricky, and we have to consider nets (x ) rather
than sequences (xn ). It will turn out that a set F is closed in this setting if and only if whenever (x ) is
a net with x ∈ F for all  and kx − xkα → 0 for all α ∈ A, then automatically x ∈ F .

For a much more systematic discussion of these matters, see the forthcoming Lecture 4 on Weak Topolo-
gies.

1.6 Exercises
Exercises 6, 8, 9, 10, 13, 15 and 16 will be supported in tutorials.

1–1. Let K be a bounded open convex set in Rn with 0 as an interior point, and which satisfies x ∈ K
if and only if −x ∈ K. Find a norm on Rn for which K is the open unit ball. Conversley, let L be
the open unit ball for some norm on Rn . Show that L is bounded, convex, 0 is in the interior of L
and x ∈ L if and only if −x ∈ L.
1–2. (i) Let X be a vector space, and let k · k and k| · k| be norms on X. Show that for 1 ≤ p ≤ ∞,
1/p
x 7→ (kxkp + k|xk|p ) defines a norm on X.
1/p
(ii) (Harder.) What about x 7→ (kxkp − k|xk|p ) ? Can you write down any (nontrivial) case

1/4
where this does define a norm? (Hint: Consider 2kxk42 − kxk44 .
1–3. Let 1 ≤ p < ∞ and consider the space of continuous complex-valued functions defined on [0, 1] with
the Lp norm. Is this space a Banach space?
1–4. Show that there is equality in the Cauchy–Schwarz inequality with equality (for nonzero x, y) if
and only if x and y are parallel.
1–5. Show that on the space C 1 ([0, 1]) of functions whose first derivatives exist and are continuous on
R1
[0, 1], hf, gi = 0 (f g + f 0 g 0 ) defines an inner product.
1–6. Suppose that X is a normed space in which the parallelogram law holds. Show that the norm arises
from an inner product.
1–7. Let X be a finite-dimensional vector space and suppose that k · k and k| · k| are norms on X. Show
that they are equivalent, i.e. there exist constants A and B so that

kxk ≤ Ak|xk| and k|xk| ≤ Bkxk

for all x ∈ X. (In fact, and this is harder, if k · k2 is the standard euclidean norm on Rn , and if k · k
is any other norm, then there exists an invertible n × n real matrix A such that

kxk ≤ kAxk2 ≤ n1/2 kxk

for all x ∈ Rn . This is part of the John ellipsoid theorem.)

1–8. Show that Theorem 1.1 fails in a normed space setting. (In finite dimensions, there may be many
closest points in a closed convex set to a given point outside that set. In infinite dimensions, there
may be no such closest point, even if the closed convex set is a linear subspace.)

1–9. Let X be a normed space and Y a proper closed linear subspace. Show that for all  > 0 there
exists an x ∈ X with kxk = 1 and d(x, Y ) ≥ 1 − .
 SMST C: Pure Analysis: Functional Analysis 1–10

1–10. Use Theorem 1.1 to show that if C is a closed convex subset of a real Hilbert space H, and if
x ∈ H \ C, then there is a closed hyperplane K such that C lies in one open half-space determined
by K while x lies in the other open half-space. (A closed hyperplane is some translate of a closed
subspace whose complementary subspace is one-dimensional.)
1–11. Show that if X is a closed subspace of a real Hilbert space H, then the projection map ΠX is linear,
satisfies Π2X = ΠX and hΠX u, vi = hu, ΠX vi for all u, v ∈ H. Show that kΠX uk ≤ kuk for all u ∈ H
and that kΠX uk = kuk if and only if u ∈ X. What are im ΠX and ker ΠX ?
1–12. Let {eα : α ∈ A} be an orthonormal set in an inner product space H. Show that for all x ∈ H,
x
b(α) = 0 for all but countably many α. (Hint: for each  > 0 consider the cardinality of the set
{α ∈ A : |b
x(α)| > }.)
1–13. Prove Selberg’s inequality: let yj , 1 ≤ j ≤ N be arbitrary nonzero elements of an inner product
space H and let x ∈ H. Then
N
X |hx, yj i|2
PN ≤ kxk2 .
j=1 k=1 |hyj , yk i|

1–14. Show that

Pif e1 , e2 . . . is an orthonormal basis for a Hilbert space H, then we have Parseval’s identity
∞
hx, yi = j=1 xb(j)b y (j).

1–15. Let S be a linear subspace of L2 ([0, 1]) such that there exists a K satisfying kf k∞ ≤ Kkf k2 for all
f ∈ S. Show that the dimension of S is at most K 2 .
1–16. Quotient Spaces. Let X be a normed space and Y a proper closed linear subspace. Let X/Y be
the usual quotient space {x + Y : x ∈ X} of equivalence classes of X under the equivalence relation
x ∼ y iff x − y ∈ Y .
(i) Show that X/Y is a vector space with a suitable notion of addition and scalar multiplication.
(ii) Show that |x + Y k = inf y∈Y kx + yk defines a norm on X/Y .
(iii) Show that if X is complete, so is X/Y .
(iv) Let π : X → X/Y be the natural map. Show that kπk = 1.
1–17. Haar wavelets. For each n = 0, 1, 2 . . . , let

In = {[j2−n , (j + 1)2−n ) : j = 0, 1, . . . , 2n − 1}.

These are half-open intervals of length 2−n that partition the unit interval [0, 1). Let I = n In . We
S
call I the collection of dyadic intervals in [0, 1). For I ∈ I, write I = [aI , bI ), mI = (aI + bI )/2,
and for x ∈ [0, 1), define 
 (bI − aI )1/2 x ∈ [aI , mI )
hI (x) = −(bI − aI )1/2 x ∈ [mI , bI )
0 otherwise


The functions hI are called Haar wavelets. Show that the collection {hI : I ∈ I} is an orthonormal
basis for L2 ([0, 1)). (Hint: First show that all continuous functions are in the closed linear span of
{hI : I ∈ I}, and then use the fact that uniformly continuous functions are dense in L2 ([0, 1)).)

References
[1] G. B. Folland, Real Analysis, Wiley Interscience, 1999.

[2] H. L. Royden, Real Analysis, Macmillan, 1968.

[3] W. Rudin, Real and Complex Analysis, McGraw Hill, 1973.
[4] W. Rudin, Functional Analysis, Tata McGraw Hill, 1974.
[5] N. Young, An Introduction to Hilbert Space, CUP, 1988.