SMSTC (2023/2024)

Pure Analysis: Functional Analysis

www.smstc.ac.uk

Contents

4 Weak Topologies 4–1

4.1 A crash course in General Topology for analysts . . . . . . . . . . . . . . . . . . . . . . . 4–1
4.2 Topologies from seminorms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4–5
4.3 Weak topologies and the Banach–Alaoglu Theorem . . . . . . . . . . . . . . . . . . . . . . 4–9
4.4 Probability measures on compact metric spaces . . . . . . . . . . . . . . . . . . . . . . . . 4–12
4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4–13

(i)
 SMSTC (2023/2024)
Pure Analysis: Functional Analysis
Lecture 4: Weak Topologies
Notes by: Tony Carbery, University of Edinburgh
Instructor: Justin Forlanoa

www.smstc.ac.uk

Contents
4.1 A crash course in General Topology for analysts . . . . . . . . . . . . . . . 4–1
4.2 Topologies from seminorms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4–5
4.3 Weak topologies and the Banach–Alaoglu Theorem . . . . . . . . . . . . . . 4–9
4.4 Probability measures on compact metric spaces . . . . . . . . . . . . . . . . 4–12
4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4–13

So far, we have considered the metric space structure of a normed space, so that analytic notions such
as continuity refer to the metric derived from the norm. For example, compactness of the unit ball of a
finite-dimensional normed space provides a powerful tool, but in an infinite-dimensional normed space the
unit ball will never be norm-compact. However, if we are flexible about our notion of topology, equally
powerful substitute results can be obtained. If we weaken the topology so that there are fewer open sets,
we obtain more compact sets, and the Banach–Alaoglu theorem below shows that if we do this weakening
suitably, the unit ball will indeed be compact. We shall also aim to exploit convexity of normed spaces
in a more systematic way..
Thanks to Alastair Gillespie and Sandra Pott who originally developed these notes

4.1 A crash course in General Topology for analysts

In this section, we review briefly the basic ideas of topology that we shall need. More details can be
found in any standard textbook on topology.
Firstly, a topology on a set X is a collection τ of subsets of X with the following properties.

• ∅ ∈ τ and X ∈ τ.
S
• If {Gλ }λ∈Λ is an arbitrary collection of members of τ , then Gλ ∈ τ.
λ∈Λ

N
T
• If {Gk }1≤k≤N is a finite collection of members of τ , then Gk ∈ τ.
k=1

A topological space is a pair (X, τ ), where τ is a topology on X. Given a topological space (X, τ ), we
call the elements of τ the open subsets of X. In practice, we often just talk about a topological space X,
with the understanding that there is a distinguished collection of subsets of X constituting a topology
on X. Given a topological space X, a subset A of X is said to be closed if X\A is open. It is clear from
the definition of a topology that ∅ and X are closed, as are arbitrary intersections and finite unions of
closed sets.
A base for τ is a subset τ0 of τ such that every element of τ is a union of sets in τ0 , whilst τ00 ⊆ τ is a
subbase for τ if the collection of finite intersections elements of τ00 is a base.
Many of the concepts and results in the theory of metric spaces can be expressed or proved using only
the notion of an open (alternatively a closed) subset. For our purposes, we shall use the following ideas.
a [email protected]

4–1
SMST C: Pure Analysis: Functional Analysis 4–2

Closure and interior Given a subset A of a topological space X, the closure A or cl A is defined as

{x ∈ X : G ∩ A 6= ∅ for every open set G such that x ∈ G}

whilst the interior, int A or A◦ , is defined as

{x ∈ X : x ∈ G ⊆ A for some open subset G}.

It is easily seen that the closure of A equals the intersection of all the closed sets containing A, and so
is the smallest closed set containing A, and that the interior of A equals the union of all the open sets
contained in A, and so is the largest open set contained in A.

Continuity Let X and Y be two topological spaces and let f : X → Y . Then f is, by definition,
continuous at a point a ∈ X if, for each open subset G of Y containing f (a), there is an open subset
H of X containing a such that f (H) ⊆ G. An open set containing a point in a topological space is
often called a neighbourhood of the point; thus f is continuous at a if the inverse image under f of every
neighbourhood of f (a) is a neighbourhood of a. As with metric spaces, f is continuous on X (i.e. f is
continuous at every point in X) if and only if f −1 (G) is open in X for every open subset G of Y . (The
proof is essentially the same as in the metric space situation, except that we consider neighbourhoods of
a point rather than open balls centred on the point.)

A neighbourhood base and a neighbourhood subbase of x ∈ X are defined in the obvious way.

Let τ1 and τ2 be two topologies on X. We say that τ1 is weaker than τ2 or that τ2 is stronger than τ1
(often written as τ1 ≤ τ2 or τ2 ≥ τ1 ) if τ1 ⊆ τ2 . Note that, if σ1 and σ2 are two topologies on Y , and
if f : (X, τ1 ) → (Y, σ1 ) is continuous, then f , thought of as a function from (X, τ2 ) to (Y, σ2 ), is also
continuous provided τ2 is stronger than τ1 and σ2 is weaker than σ1 .

If τ1 and τ2 are topologies on X, then τ1 ∩ τ2 is also a topologyTon X, and more generally if τα is a

topology on X for all α ∈ A for some indexing set A, then so is α∈A τα . So if P is a property that a
topology on X may or may not satisfy, and if there is at least one topology on X satisfying P, then the
intersection of all topologies on X satisfying P is the weakest topology on X satisfying P.

Hausdorff spaces If d is a metric on a set X and x, y are distinct points in X, then d(x, y) > 0 and
hence there are two disjoint open balls, one centred on x and the other centred on y. A topological space
X is said to be Hausdorff if it has a similar property: given distinct points x, y ∈ X, there are open sets
G, H with x ∈ G, y ∈ H and G ∩ H = ∅. In a Hausdorff topological space every singleton set {x} is
closed.

Convergence and nets In metric spaces, everything of a topological nature can be done with sequences
as convergence properties of sequences determine the topology. More precisely if d1 and d2 are two metrics
on a set X, then the following are equivalent.

(i) A subset G of X is open in (X, d1 ) if and only if it is open in (X, d2 ).

(ii) A sequence (xn )∞
n=1 converges to x in (X, d1 ) if and only it converges to x in (X, d2 ).

If we replace metrics by topologies, the two statements are no longer equivalent in general; (i) implies
(ii) but (ii) need not imply (i). However there is a generalised notion of a net, which we briefly discuss
below. Nets have the flavour of sequences and often enable us to use sequence-style arguments in general
topological spaces. See [4] for a complete account of nets and generalised convergence.

Definition 4.1 A directed set is a non-empty set I with a direction ≤ which is reflexive and transitive
and satisfies the direction property: for all i, j ∈ I, there exists k ∈ I with i ≤ k and j ≤ k. A net in
a topological space X is a function I → X, whose domain is a directed set. [We prefer to write (xi )i∈I
for a net rather than x(i), as it better suggests the analogy with sequences.] We say that a net (xi )i∈I
converges to x ∈ X if, given any open set U containing x, there exists i ∈ I such that xj ∈ U for j ≥ i,
i.e. the net eventually lies in every open neighbourhood of x.
SMST C: Pure Analysis: Functional Analysis 4–3

The natural numbers N with the usual ordering is a directed set. In this way every sequence is a net and
the notion of convergence of nets extends that of sequences. Any non-empty interval in the real line is
also a directed set. Another standard example of a directed set arises from neighbourhood bases. Given
a topological space X and x ∈ X, equip the set Nx = {U ⊆ X : U open, x ∈ U } with the direction
of reverse inclusion, i.e. U ≤ V iff U ⊇ V (this is a direction, as given U, V ∈ Nx , the intersection
U ∩ V ∈ Nx is contained in both U and V ). Given U ∈ Nx , choose xU ∈ U . Then (xU )U ∈Nx is a net
converging to x.
With these notions we have the following properties, which are left as exercises.

• A subset A in a topological space is closed iff A contains the limits of all its convergent nets.
• A function f : X → Y between topological spaces is continuous at x, if and only if, given any net
(xi )i∈I in X converging to x, the net (f (xi ))i∈I converges to f (x).
• A topological space is Hausdorff if and only if every net has at most one limit.

Subspaces and Products Given a subset A of a topological space (X, τ ), we can define a topology
on A by
τA = {A ∩ G : G ∈ τ }.
We then refer to (A, τA ) as a subspace of (X, τ ), and call τA the subspace topology on A. Note that τA
is the weakest topology on A that ensures that the injection A ,→ (X, τ ) is continuous.
Q
In a somewhat similar way, suppose that X = Xλ , where each Xλ is a topological space. There is
λ∈Λ
a weakest topology τ on X, called the product topology, that ensures that each ‘coordinate’ projection
X → Xλ0 , x = (xλ )λ∈Λ 7→ xλ0 is continuous. A subset G of X is open in the product topology if and
only if, for each x = {xλ } ∈ G, there are open subsets Gλ of Xλ with Gλ = Xλ for all but (at most) a
finite number of λ0 s in the index set Λ such that
Y
x = {xλ } ∈ Gλ ⊆ G.
λ∈Λ

It is important to note that Λ could be finite, countably infinite or uncountable here. We shall need to
consider in Section 4.3 the case when Λ is uncountable.
When Λ is finite or countably infinite and each Xλ is a metric space, the product topology is metrisable.
More specifically, take Λ = N (the finite case is even simpler) and Q∞ suppose that (Xn , dn ) is a metric space
for n ∈ N. It can be shown that the product topology on X = n=1 Xn is given by the metric ρ defined
by
∞
X dn (xn , yn )
ρ(x, y) = 2−n
n=1
1 + dn (xn , yn )
for x = (xn )∞ ∞
n=1 , y = (yn )n=1 ∈ X.

Compactness As with metric spaces, a subset A of a topological space X is said to be compact if

every covering of A by open subsets of X has a finite subcovering. In particular, X itself is compact if,
whenever X equals the union of a family of open subsets, there is a finite subfamily whose union also
equals X. Note that, given a subset A of X, A is compact when considered as a subset of X if and only
the subspace (A, τA ) is compact (this is immediate from the definition of τA ).
Many of the familiar results concerning compactness in metric spaces extend in a straightforward way to
topological spaces. In particular, we note the following:

• If X and Y are topological spaces and f : X → Y is continuous, then f (A) is compact in Y for
each compact subset A of X.
• Let X be a topological space. Then X is compact if and only if it has the following property, called
the finite intersection property (FIP):
T
if (Fλ )λ∈Λ is a family of closed subsets of X such that Fλ 6= ∅ for every non-empty finite subset
λ∈Λ0
Λ0 of Λ, then
T
Fλ 6= ∅.
λ∈Λ
SMST C: Pure Analysis: Functional Analysis 4–4

• A closed subset of a compact space is compact, and a compact subset of a Hausdorff space is closed.

A useful consequence of the last item is the following simple fact (compare with the Open Mapping
Theorem).

Proposition 4.1 Let τ1 and τ2 be two topologies on X with τ1 ≤ τ2 . Suppose that (X, τ1 ) is Hausdorff
and (X, τ2 ) is compact. Then τ1 = τ2 .

Proof Let G ⊆ X be τ2 −open. Then X\G is τ2 −closed and hence τ1 −compact, since τ1 ≤ τ2 . Hence
X\G is τ1 −closed since τ1 is Hausdorff, and so G is τ1 −open. Thus τ1 = τ2 . 

The fundamental result concerning compactness that we shall need is deeper than any of the above. This
is Tychonoff’s theorem.
Q
Theorem 4.2 (Tychonoff ’s Theorem) Let Xλ be compact for λ ∈ Λ. Then X = Xλ is also
λ∈Λ
compact with respect to the product topology.

When the index set Λ is finite, this is straightforward to prove – the key step is when we are considering
the product of two compact spaces; induction then extends to an arbitrary finite product. However, if
Λ is infinite (even countably infinite) things are not so easy and the proof requires the axiom of choice.
(Indeed, Tychonoff’s theorem is equivalent to the axiom of choice.) We follow Chernoff’s proof [2], which
illustrates the theory of general convergence. Alternative approaches can be found in [4], [3], [7].
First we set out the net version of sequential compactness for metric spaces. Given a net (xi )i∈I in
a topological space X, we say that x ∈ X is a cluster point of the net (xi )i∈I if for each open set U
containing x and i ∈ I, there exists j ∈ I with j ≥ i, such that xj ∈ U . It is easy to see that a x is a
cluster point of a sequence (xn )∞
n=1 iff x is the limit of a subsequence (xnk ) of (xn ). We define the notion
of subnets, so the analogous result holds in general. A map φ : J → I between directed sets is cofinal if,
given any i ∈ I, there exists j0 ∈ J such that j ≥ j0 =⇒ φ(j) ≥ i. A subnet of the net (xi )i∈I is a net
of the form (xφ(j) )j∈J for some cofinal map φ : J → I. The following proposition relates cluster points
and subnets. It contains a typical method of encoding an iterated limit in a net.

Proposition 4.3 Let (xi )i∈I be a net in a topological space. Then x ∈ X is a cluster point of (xi )i∈I iff
it is the limit of a subnet (xj )j∈J of (xi )i∈I .

Proof If x is the limit of the subnet (xφ(j) )j∈J for some cofinal map φ : J → I, then for any open
U ⊂ X containing x, there exists j0 ∈ J with xφ(j) ∈ U for j ≥ j0 . Given i0 ∈ I, there exists j1 ∈ J such
that j ≥ j1 =⇒ φ(j) ≥ i0 . Take j ∈ J with j ≥ j0 and j ≥ j1 . Then φ(j) ≥ i0 and xφ(j) ∈ U , so x is a
cluster point of (xi )i∈I .
Conversely, suppose x is a cluster point of (xi )i∈I . Write Nx = {U ⊂ X : U open, x ∈ U } and consider
J = {(i, U ) ∈ I × Nx : i ∈ I, xi ∈ U } directed by (i1 , U1 ) ≤ (i2 , U2 ) iff i1 ≤ i2 and U1 ⊇ U2 . Note
that this is a direction: given (i1 , U1 ), (i2 , U2 ) ∈ J fix i0 with i0 ≥ i1 , i0 ≥ i2 and find i3 ≥ i0 such that
xi3 ∈ U1 ∩ U2 (as x is a cluster point). Thus (i3 , U1 ∩ U2 ) ≥ (i1 , U1 ) and (i3 , U1 ∩ U2 ) ≥ (i2 , U2 ). The map
φ(i, U ) = i defines a cofinal map from J → I and (xφ(j) )j∈J is a subnet of (xi )i∈I converging to x. 

Lemma 4.4 Let A be a subset of a topological space X. The following conditions are equivalent:

(1) A is compact;
(2) every net (ai )i∈I in A has a cluster point in A;
(3) every net in A has a subnet converging to a point in A.

Proof See the Exercises. 

We now prove Tychonoff’s theorem.
Q
Proof Let (Xλ )λ∈Λ be a family of non-empty compact topological Q spacesQand write X = λ∈Λ Xλ .
Given a subset Λ0 ⊂ Λ, we have the canonical projection pΛ0 : λ∈Λ Xλ → λ∈Λ0 Xλ .
Fix a net (xi )i∈I in X. A partial cluster point of (xi )i∈I is a pair (Λ0 , (yλ )λ∈Λ0 ) for some subset Λ0 ⊂ Λ
such that (yλ )λ∈Λ0 is a cluster point of (pΛ0 (xi ))i∈I . Let P be the set of all partial cluster points. This is
SMST C: Pure Analysis: Functional Analysis 4–5

non-empty (either take Λ0 = ∅, or alternatively take Λ0 to have one element say λ0 , when a partial cluster
can be found as Xλ0 is compact). Define a partial ordering on P by (Λ0 , (yλ )λ∈Λ0 ) ≤ (Λ1 , (zλ )z∈Λ1 ) iff
Λ0 ⊆ Λ1 and zλ = yλ for all λ ∈ Λ0 (i.e. the partial cluster (zλ )λ∈Λ1 extends (yλ )λ∈Λ0 ). Given a totally
(s) S (s)
ordered subset C = {(Λs , (yλ )λ∈Λs ) : s ∈ S} of P, define Λ0 = s∈S Λs and for λ ∈ Λ0 , set yλ = yλ
for some s with λ ∈ Λs (this does not depend on the choice of s as CQis totally ordered). To see that
(Λ0 , (yλ )Λ0 ) is a partial cluster point, take a basic neighbourhood U = λ∈Λ0 Uλ of (yλ )Λ0 , so that only
(s)
finitely many Uλ 6= Xλ . Then there is some s ∈ S such that Uλ = Xλ for λ ∈ / Λs . As (Λs , (yλ )Λs ) is a
partial cluster point, we see that (pΛ0 (xi ))I lies in U frequently, so that (Λ0 , (yλ )Λ0 ) is a partial cluster
point. Thus by Zorn’s lemma, P contains a maximal element, say (Λ0 , (yλ )λ∈Λ0 ).
We complete the proof by showing that Λ0 = Λ so that (yλ )λ∈Λ0 is a cluster point of (xi )i∈I , showing
compactness of X. Suppose, by way of a contradiction that Λ0 6= Λ and fix µ ∈ Λ \ Λ0 . Since (yλ )λ∈Λ0 is
a cluster point of (pΛ0 (xi ))i∈I , there exists a subnet (xj )j∈J of (xi )i∈I such that (yλ )λ∈Λ0 is the limit of
(pΛ0 (xj ))j∈J . By compactness of Xµ the net (pµ (xj ))j∈J has a cluster point, yµ ∈ Xµ . Set Λ1 = Λ0 ∪{µ}.
It is easy to check that (Λ1 , (yλ )λ∈Λ1 ) is a partial cluster point of (xi ), contrary to the maximality of
(Λ0 , (yλ )λ∈Λ0 ). Thus Λ0 = Λ, as required. 

4.2 Topologies from seminorms

Here, and in the rest of both this and the next lecture, we shall consider both real and complex vector
spaces, and much of what we discuss will apply in both situations. It is convenient, therefore, to write F
to denote either R or C as the underlying scalar field.

Suppose then that X is a vector space over F. Recall from Section 4 of Lecture 2 that a seminorm
on X is a function p : X → R such that, for all x, y ∈ X and all α ∈ F

• p(x) ≥ 0
• p(αx) = |α|p(x)
• p(x + y) ≤ p(x) + p(y).

For us the most important examples of seminorms will be those of the form x 7→ |f (x)| for f in the dual
of a normed space X.

We say that a family P of seminorms on X is separating or total (the terminology here is not com-
pletely standard) if p(x) = 0 for all p ∈ P implies that x = 0.

A family P of seminorms on X gives rise to a topology τP on X as follows. For a ∈ X, p ∈ P and

ε > 0, let
N (a ; p, ε) = {x ∈ X : p(x − a) < ε}.
Define a subset G of X to be τP –open if, given a ∈ G, there exist p1 , p2 , . . . , pn in P and ε1 , ε2 , . . . , εn > 0
such that
\n
N (a ; pj , εj ) ⊆ G.
j=1

It is easily checked that τP is indeed a topology on X in which the sets N (a; p, ε) are open. Also, this
topology is Hausdorff if P is a separating family. Note also that it is translation-invariant in the sense
that a subset G of X is τP –open if and only if G + a is τP –open for all a ∈ X. It is also invariant under
rescaling in the sense that, given α > 0, G is open if and only if αG is open.

Definition 4.2 A locally convex space is a vector space X with a topology on X defined by a separating
family of seminorms.

Examples

1. Consider the space C(Ω) of continuous complex valued functions on an open subset of Rn . For
each compact subset K of Ω, let pK (f ) = supω∈K |f (ω)|. Then {pK : K is a compact subset of Ω}
SMST C: Pure Analysis: Functional Analysis 4–6

is a separating family of seminorms on C(Ω). The corresponding topology is often referred as the
compact open topology on C(Ω).
There are variants of this example. For instance, Ω could be a more general topological space or we
could take Ω to be a domain in C and replace C(Ω) by the space H(Ω) of functions analytic on Ω.
2. Let S be a set, and let `∞ (S) be the space of all bounded complex-valued functions on S. There
is, of course, a natural norm, kf k∞ = sups∈S |f (s)| with respect to which `∞ (S) is a Banach space.
However, we can define a separating family of seminorms on `∞ (S) given by P = {ps : s ∈ S}, where
ps (f ) = |f (s)|. The corresponding topology τ , which is weaker than the k · k∞ −topology, is called
the topologyQ of pointwise convergence. If we identify `∞ (S) via f ↔ (f (s))s∈S with a subset S̃ of
the product s∈S C, then the topology of pointwise convergence corresponds to the restriction to
S̃ of the product topology on s∈S C. Under this identification, the unit ball of `∞ (S) corresponds
Q
Q
to s∈S Ds , where each Ds is the closed unit disc in C. Applying Tychonoff’s Theorem (Theorem
4.2), it follows that the unit ball of `∞ (S) is compact in the topology of pointwise convergence. In
contrast, the unit ball is not compact in the norm topology. Indeed, the unit ball in an infinite-
dimensional normed space is never compact.
3. Let X be a normed space with dual space X ∗ . For each f ∈ X ∗ , define a seminorm pf on X by
pf (x) = |f (x)|. The Hahn-Banach theorem in the guise of Corollary 2.4 of Lecture 2 shows that the
family PX ∗ = {pf : f ∈ X ∗ } is separating. The corresponding topology on X is usually denoted by
σ(X, X ∗ ) and is called the weak topology on X. It is the weakest topology on X such that all the
maps x 7→ f (x) for f ∈ X ∗ are continuous.

We can interchange the roles of X and X ∗ here, defining for x ∈ X a seminorm qx on X ∗ by

qx (f ) = |f (x)|. Again we have a separating family QX = {qx : x ∈ X} of seminorms on X ∗ .
(That QX is separating follows by definition, without recourse to the Hahn-Banach theorem.) The
corresponding topology on X ∗ is usually denoted by σ(X ∗ , X) and is called the weak* topology on
X ∗ . It is the weakest topology on X ∗ such that all the maps f 7→ f (x) for x ∈ X are continuous.
The weak and weak* topologies on X and X ∗ are the most important topologies after those defined
by the respective norms. We shall study them in more detail in the next section.
4. Suppose now that X and Y are normed spaces. There are two natural topologies on the space
B(X, Y ) of bounded linear operators from X to Y defined in a similar way to the weak topology
σ(X, X ∗ ) on X. More precisely, for x ∈ X and f ∈ Y ∗ , define seminorms px and px,f on B(X, Y )
by px (T ) = kT xk and px,f (T ) = |f (T x)|. Then the families of seminorms PX = {px : x ∈ X}
and PX,X ∗ = {px,f : x ∈ X and f ∈ Y ∗ } are separating and define topologies on B(X, Y ). The
former is called the strong operator topology (sot) and the latter the weak operator topology (wot).
The weak operator topology is weaker than the strong operator topology which, in turn, is weaker
than the norm topology on B(X, Y ). These topologies are particularly important when we work
with (algebras of) operators on Hilbert spaces. Here the Riesz Theorem (Theorem 2.1 of Lecture 2)
shows that the weak operator topology on B(H) can be specified by the seminorms T 7→ |hT x, yi|
for x, y ∈ H.

Corresponding to the boundedness criterion for the continuity of linear maps between normed spaces
(Proposition 2.1 of Lecture 2), we have the following.

Proposition 4.5 Let P and Q be families of seminorms on vector spaces X and Y respectively, and let
T : X → Y be a linear mapping. Then T is continuous with respect to τP on X and τQ on Y if and only
if, for every q ∈ Q, there exist a finite set {p1 , . . . , pn } in P and a positive constant α such that
n
X
q(T x) ≤ α pj (x) for all x ∈ X. (4.1)
j=1

Proof Firstly, suppose that T is continuous and fix q ∈ Q. The inverse image Tnunder T of the open set
N (0 ; q, 1) containing 0 in Y is an open set containing 0 in X and so contains j=1 N (0 ; pj , εj ) for some
p1 , . . . , pn ∈ P and some positive ε1 , . . . , εn . Let ε = min{ε1 , . . . , εn } > 0. Then
n
X n
\
pj (x) < ε ⇒ x ∈ N (0 ; pj , εj ) ⇒ q(T x) < 1.
j=1 j=1
SMST C: Pure Analysis: Functional Analysis 4–7

Pn Pn
The positive homogeneity of j=1 pj and q now gives q(T x) ≤ ε−1 j=1 pj (x) for all x ∈ X.
Conversely, suppose that, for each q ∈ Q, there exist p1 , . . . , pn in P and α > 0 such that (4.1) holds.
Fix q and corresponding p1 , . . . , pn , α. Then,
Tn given ε > 0, q(T x) < ε whenever pj (x) < ε(nα)−1 for
−1
j = 1, . . . , n. Thus T maps the open set j=1 N (0 ; pj , ε(nα) ) into N (0 ; q, ε). As these sets form a
subbasis for the neighbourhoods of 0 ∈ Y , it follows that T is continuous at 0 and hence on all of X by
the translation invariance of τP and τQ , together with the linearity of T . 

Corollary 4.6 A linear functional f on X is τP −continuous if and only if there exist p1 , . . . , pn in P

and α > 0 such that
n
X
|f (x)| ≤ α pj (x) f or all x ∈ X. (4.2)
j=1

Proof Note that there is a single seminorm q defining the topology on the underlying scalar field
F = R or C, namely q(λ) = |λ|. 
In Lecture 2, Section 2.4, the Hahn-Banach extension theorem, both in its real and in its complex form,
was discussed and various consequences derived concerning continuous linear functionals on a normed
space. There are important analogues in our present context which we now record. Most of the work
has already been done, as the fundamental extension results (Corollaries 2.1 and 2.2) involve seminorms.
Since there is no concept of the norm of a continuous linear functional on a vector space with a topology
derived from a family of seminorms, more precise results involving norms (in particular, Corollaries 2.3
and 2.4) have to be modified.

Theorem 4.7 Let P be a family of seminorms on a real or complex vector space X and let Y be a linear
subspace of X. Suppose that f is a linear functional defined on Y that is τP −continuous. Then there is
a τP −continuous linear function f˜ defined on X such that f˜(y) = f (y) for all y ∈ Y .

Proof Firstly, when we say that f is τP −continuous all we are referring to is the topology on Y derived
from the restrictions to Y of the seminorms in P or, equivalently, to the topology on Y as a subspace
of (X, τP ). Thus, applying Corollary 4.6, there are seminorms p1 , . . . , pn in P and a positive constant α
such that
Xn
|f (y)| ≤ α pj (y) for all y ∈ Y.
j=1
Pn
Since α j=1 pj is a seminorm on X, we can apply Corollary 2.1 (for real scalars) or Corollary 2.2 (for
complex scalars) to deduce that there exists a linear functional f˜ on X such f˜(y) = f (y) for all y ∈ Y
and
Xn
|f˜(x)| ≤ α pj (x) for all x ∈ X.
j=1

Corollary 4.6 now tells us that f˜ is τP −continuous on X as required. 

We saw in Lecture 2 as a consequence of the Hahn-Banach theorem that, if Y is linear subspace of a

normed linear space X and if d(x, Y ) > 0, then there exists a continuous linear functional f on X with
kf k = 1 such that the restriction of f to Y equals 0 and f (x) = d(x, Y ). In particular, if x 6= 0, then
there exists a continuous linear functional f on X with kf k = 1 and f (x) = kxk. (See Corollaries 2.3
and 2.4.)
We are now working in the context of a vector space with a topology not defined by a norm but by
a family of seminorms, and so cannot expect to get quite such precise results. Nevertheless, a good
deal survives. Indeed, we prove a much stronger separation result involving closed convex sets rather
than closed subspaces. We shall need this again in the next lecture when we discuss the Krein-Milman
theorem. The proof is another application of the Hahn-Banach Theorem, but the setting is a bit more
complicated.

Theorem 4.8 Let P be a family of seminorms on a vector space X, let K be a non-empty τP −closed
convex subset of X and let x ∈ X\K. Then there exists a τP −continuous linear functional f on X and
c ∈ R such that
Rf (y) < c < Rf (x)
SMST C: Pure Analysis: Functional Analysis 4–8

for all y ∈ K.

Proof Firstly, suppose that the underlying field F = R. By translating if necessary, we may assume that
0 ∈ K. Since x ∈ / K, there exist a τP −continuous seminorm p and ε > 0 such that N (x ; p, ε) ∩ K = ∅.
This requires a moment’s thought; p can be taken to be an appropriate finite sum of seminorms in
P. Let U = N (0 ; p, ε/2). Then (K + U ) ∩ (x + U ) = ∅. (For if y + u1 = x + u2 for some y ∈ K and
u1 , u2 ∈ U , we would have p(y−x) = p(u2 −u1 ) ≤ p(u1 )+p(u2 ) < ε and hence y ∈ K ∩N (x ; p, ε)∩K = ∅.)

Given an arbitrary z ∈ X, there exists some λ > 0 such that p(λz) < ε/2, i.e. z ∈ λ−1 U ⊆ λ−1 (K + U )
(recall that we are assuming 0 ∈ K). We can thus define g : X → R by

g(z) = inf{µ > 0 : z ∈ µ(K + U )}.

We show that g is sublinear. Firstly, it is clear that g(αz) = αg(z) for z ∈ X and α > 0. Notice that
K + U is convex (since both K and U are), from which it follows that g is subadditive. To see this,
suppose that z1 ∈ µ1 (K + U ) and z2 ∈ µ2 (K + U ). Then

z1 + z2 ∈ µ1 (K + U ) + µ2 (K + U )
 
µ1 µ2
= (µ1 + µ2 ) (K + U ) + (K + U )
µ1 + µ2 µ1 + µ2
⊆ (µ1 + µ2 )(K + U ) ,

the final inclusion by convexity. Taking infima over µ1 and µ2 separately gives g(z1 + z2 ) ≤ g(z1 ) + g(z2 ).
Thus g is a sublinear functional on X.

We claim that g(y) ≤ 1 for all y ∈ K and that g(x) > 1. To see this, note first that, since 0 ∈ K ⊆ K + U
and K + U is convex,
λ(K + U ) ⊆ K + U whenever 0 ≤ λ ≤ 1. (4.3)
It follows from (4.3) that g(z) ≤ 1 for z ∈ K + U and hence, in particular, for y ∈ K. Further, (4.3)
implies that λ(K + U ) ∩ (x + U ) = ∅ for 0 ≤ λ ≤ 1 and so g(z) ≥ 1 for all z ∈ x + U . Now there
exists η with 0 < η < 1 such that p(−ηx) = ηp(x) < ε/2, i.e. −ηx ∈ U . Hence (1 − η)x ∈ x + U
and so, again by (4.3), (1 − η)x ∈ / λ(K + U ) for 0 ≤ λ ≤ 1. Hence g((1 − η)x) ≥ 1 and thus
g(x) = (1 − η)−1 g((1 − η)x) ≥ (1 − η)−1 > 1.

We have now set the scene to apply the real version of the Hahn-Banach Theorem (Lecture 2, Theorem
2.3). On the one-dimensional subspace M spanned by x, define a linear functional by f (λx) = λg(x).
Then f (z) ≤ g(z) for all z ∈ M since f (λx) = g(λx) for λ ≥ 0 and f (λx) < 0 ≤ g(λx) for λ < 0. Now
apply the real Hahn-Banach Theorem to extend f to all of X so that f (z) ≤ g(z) for all z ∈ X. (Rather
than introduce a new symbol, denote the extension by f as well.) Then f (z) ≤ g(z) ≤ 1 for all z ∈ K + U
and hence for all y ∈ K, whilst f (x) = g(x) > 1.

It only remains to show that f is τP −continuous. Since U ⊆ K + U , f (z) ≤ g(z) ≤ 1 for z ∈ U .

But U = −U and so also −f (z) = f (−z) ≤ 1 for z ∈ U . Hence |f (z)| ≤ 1 for z ∈ U . Thus p(z) < ε/2
implies that |f (z)| ≤ 1, so |f (z)| ≤ (2/ε)p(z) for z ∈ X. Proposition 4.5 now gives the required continuity
and the proof is complete in the real case.
The case when X is a complex vector space is an easy consequence of the real case. Considering X as a
vector space over R, we obtain a τP −continuous real linear functional f0 on X and c ∈ R such that

f0 (y) < c < f0 (x)

for all y ∈ K. Now define f : X → C by f (z) = f0 (z) − if0 (iz). Then f is also τP −continuous and has
the required properties.


Corollary 4.9 Let P be a family of seminorms on a vector space X.

(i) Suppose that P is a separating family and let x ∈ X with x 6= 0. Then there exists a τP −continuous
linear functional on X such that f (x) 6= 0.
SMST C: Pure Analysis: Functional Analysis 4–9

(ii) Let Y be a τP −closed linear subspace of X and let x ∈ X\Y . Then there exists a τP −continuous
linear functional on X such that f (y) = 0 for all y ∈ Y and f (x) 6= 0.

Proof Firstly, when P is separating, the subspace {0} is τP −closed and so (i) follows from (ii). For (ii),
take K = Y in Theorem 4.8 and note that, if Rf (y) < c for all y ∈ Y , then λRf (y) = Rf (λy) < c for
all λ ∈ R and all y ∈ Y . Hence Rf (x) > c > 0 and Rf (y) = 0 for all y ∈ Y . When the underlying field
F = R, this completes the proof. When F = C, note that If (y) = −Rf (iy) to conclude that f = 0 on Y .


4.3 Weak topologies and the Banach–Alaoglu Theorem

We now apply the results of the previous section to the weak and weak* topologies on, respectively, a
normed linear space X and its dual space X ∗ . Recall from Example 3 early in the previous section that the
weak topology σ(X, X ∗ ) on X and the weak* topology on X ∗ are the topologies derived respectively from
the families of seminorms {pf : f ∈ X ∗ } on X and {qx : x ∈ X} on X ∗ given by pf (x) = qx (f ) = |f (x)|.
These definitions have a certain symmetry about them and, in order to discuss them within a single
framework, we consider a sightly more general situation. Suppose that E is a vector space over F = R or
C. A set F of linear functionals on E is said to be total if, for each non-zero x ∈ E, there exists f ∈ F
such that f (x) 6= 0. Such a set gives rise to a separating family PF = {pf : f ∈ F } of seminorms on
E, where pf (x) = |f (x)| and the resulting topology is denoted by σ(E, F ). The first thing to observe is
that we might as well take F to be a linear subspace of the (algebraic) dual of E. To see this, we use the
following simple result.

Proposition 4.10 Let F be a total set of linear functionals on a vector space E. Then σ(E, F ) =
σ(E, span F ).

Proof Since F ⊆ span F , σ(E, F ) is weaker than σ(E, span F ) (there are fewer seminorms defining the
former). On the other hand, if f = α1 f1 + · · · + αn fn with each fj ∈ F , then

|f (x)| ≤ |α1 f1 (x)| + · · · + |αn fn (x)|

for all x ∈ E and so f is σ(E, F ) continuous and each set {x : |f (x)| < ε} is open in σ(E, F ). 

Henceforth we shall assume that F is a total linear subspace of the algebraic dual of E. In this situation
we can identify E with a linear subspace of the algebraic dual F via the the correspondence x ↔ ϕx ,
where ϕx (f ) = f (x) for f ∈ F (this corresponds to the injection ι of a normed space X into X ∗∗ discussed
in Lecture 2, Section 2.3) and it is customary to do so when convenient. This is sometimes described by
saying that the pair (E, F ) are in duality and the notation (x, f ) used to denote both value of f ∈ F at
x ∈ E and the value of x ∈ E (identified with a subspace of the algebraic dual of F ) at f ∈ F . This puts
E and F on an equal footing.
We first identify F with the space of σ(E, F )− continuous linear functionals on E. For this we need the
following simple lemma.

Lemma 4.11 Let {f1 , . . . , fn } be a finite set of linear functionals on E and suppose that ϕ is a linear
functional on E such that
\n
ker fj ⊆ ker ϕ .
j=1

Then ϕ ∈ span {f1 , . . . , fn }.

Proof Define a linear mapping θ : E → Fn by

θ(x) = (f1 (x), . . . , fn (x))

and, on the range θ(E) of θ, define a functional Φ by

Φ(θ(x)) = ϕ(x).
SMST C: Pure Analysis: Functional Analysis 4–10

The assumed conditions on ϕ ensure that Φ is well defined and linear. Extend Φ to a linear functional Φ̃
on all of Fn . This will have the form

Φ̃(u) = α1 u1 + · · · + αn un (u = (u1 , . . . , un ))

for some (α1 , . . . , αn ) ∈ Fn . Hence ϕ = α1 f1 + · · · + αn fn . 

Theorem 4.12 Let (E, F ) be two vector spaces in duality. A linear functional f on E is continuous
with respect to the σ(E, F ) topology if and only if f ∈ F .

Proof From the definition of σ(E, F ) it is clear that each f ∈ F is σ(E, F )−continuous. Conversely,
suppose that ϕ is a σ(E, F )−continuous linear functional on E. By Proposition 4.5 and the fact that
σ(E, F ) is the topology on E determined by the seminorms, there exist f1 , . . . , fn ∈ F and α > 0 such
that
|ϕ(x)| ≤ α(|f1 (x)| + · · · + |fn (x)|)
Tn
for all x ∈ E. Then j=1 ker fj ⊆ ker ϕ and so, by Lemma 4.11, ϕ ∈ F . 

As in the normed space setting, we call the space (obviously a linear space) of σ(E, F )−continuous linear
functionals on E the dual space of (E, σ(E, F )) and denote it by (E, σ(E, F ))∗ . Interchanging the roles
of E and F , and with identification of E as a subspace of the algebraic dual of F , we also have that
the σ(E, F )−continuous linear functionals on F are precisely those in E . Thus we have the satisfying
symmetric results that

(E, σ(E, F ))∗ = F and (F, σ(F, E))∗ = E. (4.4)

Taking E = X and F = X ∗ , where X is a normed linear space, we thus have the following, recalling that
σ(X, X ∗ ) is called the weak topology on X and σ(X ∗ , X) is the weak* topology on X ∗ .

Theorem 4.13 For a normed linear space X,

(i) a linear functional on X is norm continuous if and only if it is weakly continuous;

(ii) a linear functional on X ∗ is weak* continuous if and only if it belongs to X (when we identify X
with a subspace of X ∗∗ ).

Somewhat paradoxically, the trivial implication in part (i) here is the ‘if’ statement, i.e. that weak
continuity implies norm continuity.

Proof The ‘if’ parts of both results follow from the definitions of the weak and weak* topologies. The
‘only if’ parts are immediate consequences of (4.4), taking E = X and F = X ∗ . More concretely for
part (i), if f is norm continuous, Corollary 4.6 with n = 1, p = f and α = 1 shows that f is weakly
continuous. 

It is important to observe that, if X ∗∗ 6= X under the identification of X with a subspace of X ∗∗ , then

there will be norm continuous linear functionals on X ∗ that are not weak* continuous. However, if
X = X ∗ under this identification − when this happens, we say that X is reflexive (see Exercise 2-7),
then X equals the space of weak* continuous linear functionals. This is the case, for instance, when X
is an Lp space with 1 < p < ∞.
Another result with a similar flavour is Mazur’s theorem. Since the weak topology on a normed space
is contained in the norm topology, every weakly closed set is norm closed. Mazur’s theorem (also often
referred to as the Hahn–Banach theorem without further comment in some of the literature) states that
if a set is norm closed and convex, then it is weakly closed.

Theorem 4.14 (Mazur’s theorem) Let K be a convex subset of a normed linear space. Then the
closure of K in the norm topology equals the closure of K in the weak topology.
w
Proof Let K denote the closure of K in the norm topology and K the closure in the weak topology.
w
Since the weak topology is weaker than the norm topology, K ⊆ K . Suppose that x ∈ X\K. Now K
SMST C: Pure Analysis: Functional Analysis 4–11

is convex (this follows from the joint continuity of addition in X with respect the norm topology). So
there exists, by Theorem 4.8 applied to X with the norm topology, an f ∈ X ∗ such that

Rf (x) > sup{Rf (y) : y ∈ K}.

w
Then x ∈
/ K , and we have the desired equality. 

Remarks on weak and weak* convergence In Definition 4.1 we defined what it means for a sequence
(or a net) to converge in a general topological space. For what comes next it will be useful to understand
what this means in the specific examples of the weak topology on a normed space X and the weak*
topology on X ∗ . In these cases they boil down to: xn converges to x in the weak topology (xn converges
weakly to x) iff f (xn ) → f (x) for all f ∈ X ∗ , and fn converges to f in the weak* topology (fn converges
weak* to f ) iff fn (x) → f (x) for all x ∈ X respectively; and similarly for nets. Naturally, if xn converges
to x in norm, it also converges weakly, and if fn converges to f in norm, it also converges weak*. We
leave the verification of all these assertions as an exercise.
In general terms, compactness is a property that has many desirable and useful consequences. The closed
and bounded subsets of a finite-dimensional normed space are always compact (in the norm topology)
but this result fails dramatically in infinite dimensions. The unit ball of an infinite-dimensional normed
space X is never compact in the norm topology. See Exercise 2-9, where an infinite sequence (xn )∞ n=1 of
vectors in the unit ball of X with kxn − xm k ≥ 1/2 for all n 6= m is constructed; such a sequence cannot
have a convergent subsequence. However, we do have compactness in the dual space of normed space
provided we weaken the norm topology, and work instead with its weak* topology. This is the content
of the following important theorem which has numerous applications in analysis, PDE and beyond.

Theorem 4.15 (Banach–Alaoglu) Let X be a normed space. Then the closed unit ball BX ∗ of the
dual X ∗ of X is compact in the weak* topology.

Note that the Banach–Alaoglu theorem does not state that the closed unit sphere of X ∗ is compact in
the weak* topology.
Proof We shall appeal to Tychonoff’s Theorem (Theorem 4.2) to prove this.
Q
For x ∈ X, consider the closed disc Dx = {λ ∈ F : |λ| ≤ kxk} in F, and then the product Ω = Dx ,
x∈X
with the product topology inherited from the standard topology on F = R or C. The product topology
is Hausdorff and, since each Dx is compact, is compact by Tychonoff’s Theorem.
Now define a mapping θ : BX ∗ → Ω given by

θf = (f (x))x∈X .

This mapping is clearly injective. Moreover, the subspace topology on θ(BX ∗ ) inherited from the product
topology on Ω is precisely the weakest topology making all the projections θf 7→ f (x) for x ∈ X continu-
ous; that is, it is the image of the weak* topology on BX ∗ under θ. Thus θ is a homeomorphism (that is,
θ and θ−1 are continuous) from BX ∗ with the weak* topology to its image with the product topology on Ω.

Observe that θ(BX ∗ ) equals the intersection over all x, y ∈ X and all λ ∈ F of the sets

Fx,y,λ = {(α(z))z∈X ∈ Ω : α(x + y) = α(x) + α(y) and α(λx) = λα(x)}.

Each set Fx,y,λ is closed in Ω since its definition only involves a finite number of coordinates. Thus
θ(BX ∗ ) is a closed subset of the compact Hausdorff space Ω, and hence is compact. Continuity of θ−1
now implies that BX ∗ is weak* compact. 

Consider now the second dual X ∗∗ of a normed space X. We know from Lecture 2, Section 2.3 that X
embeds in X ∗∗ isometrically via the mapping ι given by ι(x)(f ) = f (x) for x ∈ X and f ∈ X ∗ . (See
Corollary 2.7.) As mentioned above, when this embedding maps X onto X ∗∗ we say that X is reflexive
(see Exercises 2-7 and 2-21). Since dual spaces are always complete (by Lecture 2, Proposition 2.3), a
reflexive space is automatically a Banach space. The most important examples of reflexive spaces are
Hilbert spaces and Lp spaces when 1 < p < ∞, the former from the identification of the dual space of a
Hilbert space with itself via the F. Riesz Duality Theorem (Lecture 2, Theorem 2.1) and the latter from
SMST C: Pure Analysis: Functional Analysis 4–12

the identification of the dual of Lp with Lq , where p−1 + q −1 = 1 as discussed at the start of Lecture 2,
Section 2.2.
An immediate consequence of the Banach–Alaoglu Theorem is the following.

Theorem 4.16 The unit ball of a reflexive Banach space is weakly compact. In particular, the unit ball
of a Hilbert space is weakly compact.

Proof Identifying X with X ∗∗ via the map ι, we see that the topologies σ(X, X ∗ ) and σ(X ∗∗ , X ∗ ) on
X coincide − both are generated by the seminorms x → |f (x)| (f ∈ X ∗ ). Now apply Theorem 4.15 to
X ∗∗ = (X ∗ )∗ . 

Another helpful property that a topology τ may have is metrisability − that is, there is a metric d that
gives the same topology as τ . The following is often useful.

Theorem 4.17 Let X be a separable normed space. Then the weak* topology on the unit ball of X ∗ is
metrisable.

Proof Let (xn )∞ n=1 be a countable (norm) dense sequence in X (this is the definition of separability)
and, for f, g ∈ X ∗ , let
∞
X |f (xn ) − g(xn )|
d(f, g) = 2−n .
n=1
1 + |f (xn ) − g(xn )|
Then d is a metric on X ∗ . The main point here is to note that d(f, g) = 0 ⇒ f = g since {xn } is dense;
the other properties of a metric are easily checked. Let τ denote the topology on X ∗ given by d. We
have, for all N ∈ N,
N
X
d(f, g) ≤ 2−n |f (xn ) − g(xn )| + 2−N . (4.5)
n=1

It follows from (4.5) that, for a given ε > 0, d(f, g) < ε if N is large enough to have 2−N < ε/2 and then,
with that N fixed, |f (xn ) − g(xn )| < ε/(2N ) for 1 ≤ n ≤ N . Hence τ ≤ σ(X ∗ , X). Since τ is Hausdorff
and BX ∗ is σ(X ∗ , X) compact, it follows that these topologies coincide on BX ∗ by Proposition 4.1. 

Corollary 4.1 Let X be a separable normed space, and suppose fn ∈ X ∗ satisfies kfn k ≤ 1 for all n ∈ N.
Then there is a subsequence (fnk ) of (fn ) and an f ∈ X ∗ satisfying kf k ≤ 1 such that for all x ∈ X we
have
fnk (x) → f (x).

Proof By Theorem 4.17 the weak* topology on X ∗ is metrisable, so that the weak* compactness of the
unit ball of X ∗ given by the Banach–Alaoglu theorem is equivalent to weak* sequential compactness. We
thus have the existence of a subsequence (fnk ) of (fn ) which converges weak* to some f ∈ X ∗ . And this
means precisely (see the remark after Theorem 4.14) that fnk (x) → f (x) for all x ∈ X. 

4.4 Probability measures on compact metric spaces

We illustrate the ideas of the previous section with a concrete example which typifies how they are applied
in analysis.
Let K be a compact metric space and let X = C(K), the space of complex valued continuous functions
on K, equipped with norm kf k∞ = sup{|f (x)| : x ∈ K}. Then X is a Banach space, and the dual space
of X can be Ridentified with the space M(K) of complex Borel measures on K with the total variation
norm kµk = K d|µ|(x) (see Measure & Integration Theorem 7.4 for the case when we consider the space
of real valued continuous functions on K; the complex version can be deduced from the real version).
R
We thus have a natural weak* topology on M(K) given by the family of seminorms µ 7→ | X f (x)dµ(x)|,
with f ranging over C(K). The following lemma will allow us to apply the remark following Theorem
4.17.

Lemma 4.1 With the above notation, X is a separable Banach space.

SMST C: Pure Analysis: Functional Analysis 4–13

Proof The metric space K is separable since it is compact. Let {xn } be a countable dense sequence
of points in X . For n ∈ N and rational r > 0, there is a function fn,r ∈ X such that 0 ≤ fn,r ≤ 1,
fn,r (x) = 1 if d(x, xn ) ≤ r and fn,r (x) = 0 if d(x, xn ) ≥ 2r. Such a function can easily be written down:
assuming there are points x with d(x, xn ) ≥ 2r, let

d(x, Kn,r )
fn,r (x) = ,
d(x, Kn,r ) + d(x, Ln,r )

where Kn,r = {x : d(x, xn ) ≥ 2r} and Ln,r is the closed ball with centre x and radius r; otherwise just
take f = 1. It is then straightforward to check by approximating an arbitrary f ∈ X that the set of finite
linear combinations of the fn,r ’s with coefficients in a dense countable subset of C is a dense countable
subset of X. (An alternative proof can be given by showing that, if µ ∈ M(X ) satisfies
Z
fn,r (x)dµ(x) = 0
K

for all n and r, then µ = 0; now apply Corollary 2.8 from Lecture 2.) 

Thus the weak* topology on the unit ball of M(K) is metrisable by Theorem 4.17.
A Borel measure µ on K is a probability measure if µ is positive and µ(M) = 1. Such a measure is
necessarily bounded, with kµk = µ(M) = 1. Thus the set P(K) of probability measures on K forms a
subset of the unit ball of M(K). Since
\ Z Z
P(K) = {µ : f (x)dµ(x) ≥ 0 and 1dµ(x) = 1},
f ≥0 K K

we see that P(K) is a weak* closed subset of the unit ball of M(K). Using Corollary 4.1 we thus have
the following.

Theorem 4.18 Let (µn )∞ n=1 be a sequence of probability measures on K. Then there is a subsequence
(µnk ) and a probability measure µ on K such that
Z Z
f (x)dµnk (x) → f (x)dµ(x)
K K

for every continuous function f on K.

4.5 Exercises
Q
4–1. For n ∈ N, let (Xn , dn ) be a compact metric space and let X = Xn .
n∈N

(a) Verify that the product topology on X is given by the metric

∞
X dn (xn , yn )
ρ(x, y) = 2−n ,
n=1
1 + dn (xn , yn )

where x = {xn }, y = {yn } ∈ X.

(b) Let {x(m) }m∈N be a sequence in X, with x(m) = {xmn }n∈N . Use the sequential compactness
of X1 to obtain a convergent subsequence {x1,n }n∈N1 of x(1) . Now obtain a convergent sub-
sequence {x2,n }n∈N2 of {x2,n }n∈N1 . Carry on in this way to obtain convergent subsequences
{xm,n }n∈Nm of x(m) for m ∈ N with each Nm+1 a subsequence of Nm .
(c) From the construction in (b), obtain a subsequence of x(m) that converges in (X, ρ). Thus
(X, ρ) is sequentially compact and, being a metric space, is also compact.
This gives a proof of Tychonoff’s theorem for the case of a countable product of compact metric
spaces.
4–2. Prove Lemma 4.4 as follows.
 SMST C: Pure Analysis: Functional Analysis 4–14

(a) Let (xi )i∈I be a net in a topological space X. For each i ∈ I, write Ai = {xj : j ∈ I, j ≥ i}.
Show that x ∈ X is a cluster point of (xi )i∈I iff x lies in the closure of Ai for each i ∈ I.
(b) Suppose X is compact and (xi )i∈I is a net in X. Using the notation above, apply the finite
intersection property to the closures of the Ai to show that (xi )i∈I has a cluster point.
(c) Now let X be a topological space such that every T net has a cluster point. Given a family
(Fλ )λ∈Λ of closed subsets with
T the property that λ∈Λ0 Fλ 6= ∅ for all finite subsets T
Λ0 ⊂ Λ,
define a suitable net on I = { λ∈Λ0 Fλ 6= ∅ : Λ0 is a finite subset of Λ} to show that λ∈Λ Fλ
is non-empty.

4–3. Let X and Y be normed spaces over F (= R or C) and let ϕ : B(X, Y ) → F be a linear functional.
Prove that the following statements are equivalent.
(a) There exist x1 , . . . , xn in X and f1 , . . . , fn in Y ∗ such that
n
X
ϕ(T ) = fj (T xj )
j=1

for T ∈ B(X, Y ).
(b) ϕ is continuous with respect to the weak operator topology on B(X, Y ).
(c) ϕ is continuous with respect to the strong operator topology on B(X, Y ).
[For (c) ⇒ (a), you may find it helpful to consider a linear functional defined initially on a subspace
of the product of n copies of Y of the form {(T x1 , . . . , T xn ) : T ∈ B(X, Y )}, where x1 , . . . , xn ∈ X
come from the strong continuity of ϕ.]
4–4. Let X and Y be normed spaces and let T : X → Y be a linear mapping. Prove that T is continuous
with respect to the norm topologies on X and Y if and only if it is continuous with respect to the
weak topologies on X and Y .

4–5. (a) Let xn → x weakly in a normed space X. Prove that there is a sequence of (finite) convex
combinations of the xn ’s converging in norm to x.
(b) Illustrate the result of (a) when X is the sequence space `p (1 < p < ∞) with the usual norm
and xn is the element of X with 1 in the nth place and 0 elsewhere. What happens when
p = 1?

4–6. (a) Let xn → x weakly in a normed space. Prove that kxk ≤ lim inf n→∞ kxn k.
(b) Prove the corresponding result for a sequence {fn } converging with respect to the weak*
topology to f in the dual of a normed space, and for a sequence {Tn } of linear mappings in
B(X, Y ) converging either in the strong or in the weak operator topology to T ∈ B(X, Y ), X
and Y being normed spaces.
(c) Let xn → x weakly in a Hilbert space and suppose that kxn k → kxk. Prove that kxn −xk → 0.
4–7. Let {en }n∈N be an orthonormal basis of a Hilbert space H and, for n, m ∈ N with 1 ≤ m < n, let
xm,n = em + men . Let A = {xm,n : 1 ≤ m < n}. Prove that 0 belongs to the weak closure of A
but that no sequence in A converges weakly to 0. This shows that the weak topology on H is not
metrisable (though the weak topology on the unit ball of H is).

4–8. Let X be a reflexive Banach space, and let T ∈ B(X). Prove that T (BX ), the image under T of
the closed unit ball BX of X, is closed in the norm topology. (Note the result of Ex. 14-3.)
4–9. Let X be a reflexive Banach space. Prove that the unit ball of B(X) is compact in the weak operator
topology.

4–10. Let X be a Banach space and consider X as embedded in X ∗∗ via the natural embedding. Prove
that the unit ball of X is σ(X ∗∗ , X ∗ ) dense in the unit ball of X ∗∗ .
4–11. Let X be a Banach space and suppose that the unit ball of X is compact in the weak topology
σ(X, X ∗ ). Prove that X is reflexive.
SMST C: Pure Analysis: Functional Analysis 4–15

References
[1] J. B. Conway, A Course in Functional Analysis, Springer-Verlag, 1985.
[2] P. Chernoff, A Simple proof of Tychonoff’s theorem via nets, Amer. Math. Monthly, 932-934, 1992.
[3] N. Dunford and J. T. Schwartz, Linear Operators, Vol. 1, Wiley-Interscience, 1957.

[4] J. Kelley, General Topology, van Nostrand, 1955.

[5] H. L. Royden, Real Analysis, Macmillan, 1968.
[6] W. Rudin, Functional Analysis, Tata McGraw Hill, 1974.
[7] G. F. Simmons, Introduction to Topology and Modern Analysis, McGraw Hill, 1963.